chore(main): release 0.4.0 (#625 )

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
fix(sandbox): scrub dynamic credential env vars (#682 )
2026-11-08 09:49:34 -10:00 · 2026-07-17 13:25:31 +08:00 · 2026-07-17 13:04:12 +08:00 · 2026-07-17 13:03:32 +08:00 · 2026-07-17 13:03:14 +08:00 · 2026-07-17 12:49:38 +08:00
164 changed files with 14953 additions and 421 deletions
@@ -1 +1 @@
-{".":"0.3.0"}
+{".":"0.4.0"}
@@ -6,6 +6,60 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 aims to follow [Semantic Versioning](https://semver.org/spec/v2.0.0.html) once the first release is
 tagged. Until then, source builds report the version `dev`.

+## [0.4.0](https://github.com/Gitlawb/zero/compare/v0.3.0...v0.4.0) (2026-07-17)
+
+
+### Features
+
+* **aimlapi:** AI/ML API provider with guided onboarding (top-up + key issuance) ([#655](https://github.com/Gitlawb/zero/issues/655)) ([6b9c2f0](https://github.com/Gitlawb/zero/commit/6b9c2f0c083e02bcd3aad68fb7af0501a9e7cd61))
+* **cli:** show ZERO wordmark on --version ([#673](https://github.com/Gitlawb/zero/issues/673)) ([9acb411](https://github.com/Gitlawb/zero/commit/9acb4113cc3337b3f361a16278e9cf11ca105e34))
+* **cli:** wire MCP serve WorkspaceRoot and --add-dir scope ([#694](https://github.com/Gitlawb/zero/issues/694)) ([75a78e7](https://github.com/Gitlawb/zero/commit/75a78e715bf23154f69c8c79cb58eb4b535b2a2a))
+* **npm:** ship the native binary as platform optionalDependencies ([#626](https://github.com/Gitlawb/zero/issues/626)) ([5e1405d](https://github.com/Gitlawb/zero/commit/5e1405d0b7abff5b3ccb3cfdb66d64d6d3322922))
+* **perf:** emit prompt-prefix hash fingerprint per turn ([#704](https://github.com/Gitlawb/zero/issues/704)) ([1c5c6e7](https://github.com/Gitlawb/zero/commit/1c5c6e78a8a0e228bdf53d7be90934fbce9d98c3))
+* **providers:** add AI/ML API preset (rebased onto main) ([#621](https://github.com/Gitlawb/zero/issues/621)) ([d66a9dd](https://github.com/Gitlawb/zero/commit/d66a9dda69c32aa59f4bea903cfefe00d4b7adef))
+* **providers:** refresh MiniMax model coverage ([#665](https://github.com/Gitlawb/zero/issues/665)) ([fa3052a](https://github.com/Gitlawb/zero/commit/fa3052a1422a4ad30a3a6295829564f42dee31a8))
+* **skills:** discover shared ~/.agents/skills with multi-root skill loading ([#696](https://github.com/Gitlawb/zero/issues/696)) ([7d57999](https://github.com/Gitlawb/zero/commit/7d579996b43741a2e57e3e38fcdd8484fcfc34e9))
+* **tui:** Ctrl+X leader chords and emacs menu navigation ([#699](https://github.com/Gitlawb/zero/issues/699)) ([7f669f4](https://github.com/Gitlawb/zero/commit/7f669f455021be51319ee3b8298cd48a17f745c7))
+* **tui:** press up to edit queued messages ([#656](https://github.com/Gitlawb/zero/issues/656)) ([4c986d3](https://github.com/Gitlawb/zero/commit/4c986d327b5ed26338d295c23bea5839681db8d0))
+
+
+### Bug Fixes
+
+* **acp:** make truncateHint rune-safe ([#614](https://github.com/Gitlawb/zero/issues/614)) ([ddc4927](https://github.com/Gitlawb/zero/commit/ddc4927aac5544bf4dd2c46615aeda5a81c96576))
+* **agent,tui:** resolve git branch detection when starting Zero in subdirectories ([#613](https://github.com/Gitlawb/zero/issues/613)) ([0184581](https://github.com/Gitlawb/zero/commit/0184581ec234ea414e0a47bc33e8b0f4ddfb497b))
+* **agent:** raise default and deep-mode turn budgets ([#650](https://github.com/Gitlawb/zero/issues/650)) ([635c93a](https://github.com/Gitlawb/zero/commit/635c93af51ebc20f3e0917917e55a79edfe27c35))
+* **cli:** prevent consuming positional arguments as flag values ([#619](https://github.com/Gitlawb/zero/issues/619)) ([5b4f48d](https://github.com/Gitlawb/zero/commit/5b4f48d2dcb66402c13bde0c3cfe9c9371da19fb))
+* **config:** enforce MCP trust boundary so project config cannot override user disable ([#609](https://github.com/Gitlawb/zero/issues/609)) ([4d8c31c](https://github.com/Gitlawb/zero/commit/4d8c31cf16a3080344e3be7039408fcae71c075d)), closes [#512](https://github.com/Gitlawb/zero/issues/512)
+* **config:** surface unknown/typo'd config fields instead of silently dropping them ([#645](https://github.com/Gitlawb/zero/issues/645)) ([893b7b4](https://github.com/Gitlawb/zero/commit/893b7b424cc203a2fcf92327a4e25c84286a90e0))
+* **cron:** prevent cron job Mutate from clobbering concurrent updates ([#630](https://github.com/Gitlawb/zero/issues/630)) ([e4bd703](https://github.com/Gitlawb/zero/commit/e4bd703cfb28dab2dfa3c2ddba46237e1bb2e164))
+* **daemon:** handle os.ErrPermission as collision during O_EXCL lock creation ([#616](https://github.com/Gitlawb/zero/issues/616)) ([8ea5384](https://github.com/Gitlawb/zero/commit/8ea53841a5d54c37a153779ae86bea010659433c))
+* **exec:** stop false INCOMPLETE downgrades on conversational final messages ([#608](https://github.com/Gitlawb/zero/issues/608)) ([b6117af](https://github.com/Gitlawb/zero/commit/b6117af86d6bc87a4ee66910e99d76cb16b03fed))
+* harden MCP credential boundaries ([#597](https://github.com/Gitlawb/zero/issues/597)) ([fdddb05](https://github.com/Gitlawb/zero/commit/fdddb05ba84b1600ae6c3a20028bf83afe474c44))
+* **hooks:** fail closed on launch failures for beforeTool hooks ([#629](https://github.com/Gitlawb/zero/issues/629)) ([dc06fe7](https://github.com/Gitlawb/zero/commit/dc06fe72caf45f72d2cba1e8a835c0f5b405c1e8))
+* **hooks:** run sessionEnd hooks after Esc/Ctrl+C interrupts ([#606](https://github.com/Gitlawb/zero/issues/606)) ([824ecdb](https://github.com/Gitlawb/zero/commit/824ecdbcf9c467c35ef4e2666770fdadcb5bf402))
+* **keyring:** pass generic password via stdin on macOS ([#574](https://github.com/Gitlawb/zero/issues/574)) ([91ea6de](https://github.com/Gitlawb/zero/commit/91ea6ded7503538834a84d090f78670a363c62d3))
+* **lock:** prevent POSIX lock file overwrite and leak on Windows/Unix ([#628](https://github.com/Gitlawb/zero/issues/628)) ([da41c3a](https://github.com/Gitlawb/zero/commit/da41c3a75b782d6e0836fe13346321e40a90fbb4))
+* **openai:** omit prompt_cache_key for openai-compatible providers ([#636](https://github.com/Gitlawb/zero/issues/636)) ([1af5882](https://github.com/Gitlawb/zero/commit/1af58828eb3c22567599c000736c913a290959d2)), closes [#624](https://github.com/Gitlawb/zero/issues/624)
+* **plugins:** resolve relative executable paths against plugin root ([#627](https://github.com/Gitlawb/zero/issues/627)) ([2efe6d5](https://github.com/Gitlawb/zero/commit/2efe6d539e29374b3ef39c2290bdea81f33a228b))
+* **sandbox:** remove windowsWriteRestricted flag to fix DenyRead bypass ([#612](https://github.com/Gitlawb/zero/issues/612)) ([3d96ac7](https://github.com/Gitlawb/zero/commit/3d96ac7e55c760a97f28c0e6ceaf1ec3b4ab717a))
+* **sandbox:** scrub dynamic credential env vars ([#682](https://github.com/Gitlawb/zero/issues/682)) ([9043bae](https://github.com/Gitlawb/zero/commit/9043baedcff7776c7373645b57563cac06b31847))
+* **sandbox:** scrub sensitive credentials from sandbox environment ([#660](https://github.com/Gitlawb/zero/issues/660)) ([6fc1220](https://github.com/Gitlawb/zero/commit/6fc1220f6ac66fb3ae67b637cbbed7068d2213c0))
+* **sandbox:** unblock git fetch/commit/add under the write-restricted sandbox ([#654](https://github.com/Gitlawb/zero/issues/654)) ([5c4815a](https://github.com/Gitlawb/zero/commit/5c4815a66ed07d9cf90b825adfd936d3ac07639d))
+* **sandbox:** use WRITE_RESTRICTED token when no DenyRead paths are configured ([#658](https://github.com/Gitlawb/zero/issues/658)) ([a5d2e32](https://github.com/Gitlawb/zero/commit/a5d2e327c8681671aa8a9e5378801215b747edcf))
+* **securefile,credstore:** call Sync on temp file before close and rename ([#631](https://github.com/Gitlawb/zero/issues/631)) ([212734a](https://github.com/Gitlawb/zero/commit/212734adf3b5f982e22385161205aa1afe4634fe))
+* **securefile:** reclaim stale lock files to prevent permanent DOS ([#615](https://github.com/Gitlawb/zero/issues/615)) ([8536cc8](https://github.com/Gitlawb/zero/commit/8536cc87f7a885f9e436d6ef28f5f325201623dc))
+* **swarm:** wait for job.Runs directly in scheduler skip test ([#667](https://github.com/Gitlawb/zero/issues/667)) ([1bb6b57](https://github.com/Gitlawb/zero/commit/1bb6b5745af90321d1e657a12a1976cded5dd1bd))
+* **tools:** classify silent wrapped Windows command failures as sandbox denials ([#659](https://github.com/Gitlawb/zero/issues/659)) ([8bd9742](https://github.com/Gitlawb/zero/commit/8bd9742fa95c41b93ea4e718628aed0ff3ae9dd0))
+* **tools:** preserve SysProcAttr during PTY fallback ([#618](https://github.com/Gitlawb/zero/issues/618)) ([f78b36c](https://github.com/Gitlawb/zero/commit/f78b36c770daa4577a9f99265b18a354454e36eb))
+* **tui:** resolve pending askUser callbacks to prevent runner hangs ([#620](https://github.com/Gitlawb/zero/issues/620)) ([aa73a76](https://github.com/Gitlawb/zero/commit/aa73a76f1bd1b6fe97bac2fbff2d61b7474139f2))
+* **tui:** stop the composer cursor blinking while typing or unfocused ([#672](https://github.com/Gitlawb/zero/issues/672)) ([2b42cd5](https://github.com/Gitlawb/zero/commit/2b42cd567b96d6f7e0818594e53ff31cce1e42e9))
+* **windows:** resolve absolute path for taskkill to prevent hijacking ([#617](https://github.com/Gitlawb/zero/issues/617)) ([2db00ee](https://github.com/Gitlawb/zero/commit/2db00ee3d57db97e0fbe23cb8628e8bcb47f6f09))
+
+
+### Performance Improvements
+
+* **tools:** add explicit effect metadata for safe concurrency ([#705](https://github.com/Gitlawb/zero/issues/705)) ([8ef8576](https://github.com/Gitlawb/zero/commit/8ef8576df7d0775a5b815bc0776a794cadb75c34))
+
 ## [0.3.0](https://github.com/Gitlawb/zero/compare/v0.2.0...v0.3.0) (2026-07-09)


@@ -127,14 +181,19 @@ tagged. Until then, source builds report the version `dev`.
 ## [Unreleased]

 ### Added
+- Shared multi-agent skills discovery: when present, `~/.agents/skills` is searched after the primary
+  Zero skills dir (and before plugin skill roots). `zero skills list` / `info` and the runtime `skill`
+  tool share one multi-root discovery path; install/remove/lock still target only the Zero skills directory.
 - `SECURITY.md` with a private vulnerability-reporting path, `CODE_OF_CONDUCT.md`, this changelog, and
  GitHub issue/PR templates.
 - Interactive `/theme` picker: bare `/theme` opens a popup that live-previews each palette as you move
  and applies on select (Esc reverts).
- Ten built-in color themes alongside the `dark`/`light` built-ins — `dracula`, `nord`, `gruvbox`,
-  `tokyo-night`, `catppuccin`, `one-dark`, `solarized-dark`, `rose-pine`, `everforest`, and
-  `solarized-light` — selectable via `/theme <name>`, `--theme <name>`, or `ZERO_THEME`. Every palette
-  is contrast-audited to WCAG AA. The built-in light theme was reworked for legibility.
+- Twelve built-in color themes alongside the `dark`/`light` built-ins — `dracula`, `nord`, `gruvbox`,
+  `tokyo-night`, `catppuccin`, `one-dark`, `solarized-dark`, `rose-pine`, `everforest`,
+  `solarized-light`, `dune`, and `neon` — selectable via `/theme <name>`, `--theme <name>`, or
+  `ZERO_THEME`. Every palette is contrast-audited to WCAG AA, and the new presets are additionally
+  audited after xterm-256 downsampling; see [docs/THEMES.md](docs/THEMES.md). The built-in light
+  theme was reworked for legibility.
 - `--theme <name>` flag for the TUI, accepting `auto` or any registered theme (previously only the
  `ZERO_THEME` env var existed).
 - "Accessibility / Appearance" section in the README documenting `NO_COLOR`, `ZERO_THEME`, `/theme`,
@@ -1,7 +1,7 @@
 # Zero build/test/lint targets. AGENTS.md says "Build with `make`" and "Run `make
 # lint` before opening a PR" — these targets back those instructions.
 .DEFAULT_GOAL := build
-.PHONY: build build-all test test-race vet fmt fmt-check lint tidy clean help
+.PHONY: build build-all test test-race vet fmt fmt-check lint tidy clean baseline help

 # Build the main CLI binary into ./zero.
 build:
@@ -40,5 +40,19 @@ clean:
 	rm -f zero
 	go clean ./...

+# Run the per-turn benchmark harness over the checked-in baseline manifest and
+# write the JSON result to internal/perfbench/reports/baseline.json. Requires a
+# built `zero` binary and a model; set ZERO_BENCH_MODEL (required) and
+# ZERO_BENCH_BINARY (defaults to ./zero) to configure the run. The report is
+# machine-specific and regenerated, not hand-edited.
+baseline: build
+	@if [ -z "$(ZERO_BENCH_MODEL)" ]; then echo "Set ZERO_BENCH_MODEL (and optionally ZERO_BENCH_BINARY) before running 'make baseline'"; exit 2; fi
+	@ZERO_BIN="$${ZERO_BENCH_BINARY:-./zero}"; \
+	go run ./cmd/zero-perf-bench turn \
+		--suite internal/perfbench/manifests/baseline.json \
+		--model $(ZERO_BENCH_MODEL) \
+		--binary "$$ZERO_BIN" \
+		--output internal/perfbench/reports/baseline.json
+
 help:
-	@echo "Targets: build (default), build-all, test, test-quick, vet, fmt, fmt-check, lint, tidy, clean"
+	@echo "Targets: build (default), build-all, test, test-quick, vet, fmt, fmt-check, lint, tidy, clean, baseline"
@@ -129,11 +129,18 @@ the key in the wizard:
 export OPENAI_API_KEY=sk-...
 export ANTHROPIC_API_KEY=...
 export GEMINI_API_KEY=...
+export AIMLAPI_API_KEY=...
 export LONGCAT_API_KEY=...
 export MINIMAX_API_KEY=...
 export MINIMAXI_API_KEY=...
 ```

+To configure AI/ML API directly, run:
+
+```bash
+zero providers setup aimlapi --set-active
+```
+
 To configure Meituan LongCat (LongCat-2.0) directly, run:

 ```bash
@@ -184,6 +191,8 @@ Useful controls:
 |---|---|
 | `Enter` | send the prompt |
 | `/` | open slash-command suggestions |
+| `Ctrl+X` then letter | common slash commands (e.g. `m` → `/model`; `Ctrl+X` `?` for full list) |
+| `Ctrl+P` / `Ctrl+N` | previous / next item in menus (arrows still work) |
 | `Shift+Tab` | cycle permission mode |
 | `Ctrl+B` | show/hide the sidebar |
 | `Ctrl+C` | cancel or exit |
@@ -332,7 +341,7 @@ manifest.
 | Control | Effect |
 |---|---|
 | `NO_COLOR=<anything>` | disables color output |
-| `ZERO_THEME=<name>` | selects the startup theme (`auto`, `dark`, `light`, or a color theme like `dracula`, `nord`, `gruvbox`, `tokyo-night`, `catppuccin`, `one-dark`, `solarized-dark`, `rose-pine`, `everforest`, `solarized-light`) |
+| `ZERO_THEME=<name>` | selects the startup theme (`auto`, `dark`, `light`, or a color theme like `dracula`, `nord`, `gruvbox`, `tokyo-night`, `catppuccin`, `one-dark`, `solarized-dark`, `rose-pine`, `everforest`, `neon`, `solarized-light`, `dune`) |
 | `--theme <name>` | selects the TUI theme from the CLI (same names) |
 | `/theme` | opens the theme picker inside the TUI (live preview; `/theme <name>` switches directly) |
 | `ZERO_NO_FADE=1` | disables streaming fade animation |
@@ -379,6 +388,7 @@ go run ./cmd/zero-release build --goos windows --goarch amd64 --output dist/zero

 - [Install](docs/INSTALL.md)
 - [Update flow](docs/UPDATE.md)
+- [Themes](docs/THEMES.md)
 - [Stream-JSON protocol](docs/STREAM_JSON_PROTOCOL.md)
 - [Specialists](docs/SPECIALISTS.md)
 - [GitHub Action](docs/GITHUB_ACTION.md)
@@ -234,7 +234,7 @@ zero update           检查更新版本
 | 控制 | 效果 |
 |---|---|
 | `NO_COLOR=<任意值>` | 禁用颜色输出 |
-| `ZERO_THEME=<名称>` | 选择启动主题（`auto`、`dark`、`light`，或颜色主题如 `dracula`、`nord`、`gruvbox`、`tokyo-night`、`catppuccin`、`one-dark`、`solarized-dark`、`rose-pine`、`everforest`、`solarized-light`） |
+| `ZERO_THEME=<名称>` | 选择启动主题（`auto`、`dark`、`light`，或颜色主题如 `dracula`、`nord`、`gruvbox`、`tokyo-night`、`catppuccin`、`one-dark`、`solarized-dark`、`rose-pine`、`everforest`、`neon`、`solarized-light`、`dune`） |
 | `--theme <名称>` | 从 CLI 选择 TUI 主题（相同名称） |
 | `/theme` | 在 TUI 中打开主题选择器（实时预览；`/theme <名称>` 直接切换） |
 | `ZERO_NO_FADE=1` | 禁用流式淡入动画 |
@@ -261,6 +261,7 @@ go run ./cmd/zero-release build --goos windows --goarch amd64 --output dist/zero

 - [安装](docs/INSTALL.md)
 - [更新流程](docs/UPDATE.md)
+- [主题](docs/THEMES.md)
 - [Stream-JSON 协议](docs/STREAM_JSON_PROTOCOL.md)
 - [专家](docs/SPECIALISTS.md)
 - [GitHub Action](docs/GITHUB_ACTION.md)
@@ -39,6 +39,9 @@ func run(args []string, getenv func(string) string, stdout io.Writer, stderr io.
 	if len(args) > 0 && args[0] == "tasks" {
 		return runTasksCommand(args[1:], getenv, stdout, stderr)
 	}
+	if len(args) > 0 && args[0] == "turn" {
+		return runTurnCommand(args[1:], getenv, stdout, stderr)
+	}
 	options, err := parseArgs(args, getenv)
 	if err != nil {
 		_, _ = fmt.Fprintln(stderr, err.Error())
@@ -206,6 +209,7 @@ func helpText() string {
 	return strings.Join([]string{
 		"Usage: zero-perf-bench [options]",
 		"       zero-perf-bench tasks [options]   (Terminal-Bench-style task harness; see `tasks --help`)",
+		"       zero-perf-bench turn  [options]   (per-turn tracing benchmark; see `turn --help`)",
 		"",
 		"Options:",
 		"  --iterations <n>             Measured samples to collect (default: 5)",
@@ -0,0 +1,238 @@
+package main
+
+import (
+	"bytes"
+	"context"
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+	"strings"
+
+	"github.com/Gitlawb/zero/internal/perfbench"
+)
+
+// turnOptions configures the `zero-perf-bench turn` subcommand: the per-turn
+// benchmark harness that runs ZERO headlessly with --trace, parses each turn's
+// NDJSON trace, and records per-span latency plus the top controllable latency
+// sources — the Phase 0 baseline's "do not proceed until" criterion.
+type turnOptions struct {
+	SuitePath   string
+	Model       string
+	Mode        string
+	SelfCorrect bool
+	Binary      string
+	Iterations  int
+	Version     string
+	Commit      string
+	Output      string
+	JSON        bool
+	DryRun      bool
+	Help        bool
+}
+
+func runTurnCommand(args []string, getenv func(string) string, stdout io.Writer, stderr io.Writer) int {
+	options, err := parseTurnArgs(args, getenv)
+	if err != nil {
+		_, _ = fmt.Fprintln(stderr, err.Error())
+		return 2
+	}
+	if options.Help {
+		_, _ = fmt.Fprint(stdout, turnHelpText())
+		return 0
+	}
+
+	set, err := perfbench.LoadTaskSet(options.SuitePath)
+	if err != nil {
+		_, _ = fmt.Fprintln(stderr, "[zero] Turn benchmark failed: "+err.Error())
+		return 1
+	}
+
+	// The dry-run path records a zero-iteration run without a binary, so the
+	// manifest loads and the report path is exercised in CI without a model.
+	if options.DryRun {
+		_, _ = fmt.Fprintln(stdout, "[zero] turn benchmark: dry run (no agent invoked)")
+		return 0
+	}
+
+	binary, err := perfbench.ResolveBinary(options.Binary)
+	if err != nil {
+		_, _ = fmt.Fprintln(stderr, "[zero] Turn benchmark failed: "+err.Error())
+		return 2
+	}
+
+	result, err := perfbench.RunTurnBench(context.Background(), set, perfbench.TurnBenchConfig{
+		Model:       options.Model,
+		Mode:        options.Mode,
+		SelfCorrect: options.SelfCorrect,
+		Version:     options.Version,
+		Commit:      options.Commit,
+		Iterations:  options.Iterations,
+		Runner:      perfbench.NewTurnExecRunner(binary),
+	})
+	if err != nil {
+		_, _ = fmt.Fprintln(stderr, "[zero] Turn benchmark failed: "+err.Error())
+		return 1
+	}
+
+	if options.Output != "" {
+		if err := writeTurnReport(options.Output, result); err != nil {
+			_, _ = fmt.Fprintln(stderr, "[zero] Turn benchmark failed: "+err.Error())
+			return 1
+		}
+	}
+	if options.JSON {
+		if err := perfbench.WriteTurnBenchJSON(stdout, result); err != nil {
+			_, _ = fmt.Fprintln(stderr, "[zero] Turn benchmark failed: "+err.Error())
+			return 1
+		}
+		return 0
+	}
+	_, _ = fmt.Fprintln(stdout, perfbench.FormatTurnBenchSummary(result))
+	return 0
+}
+
+func parseTurnArgs(args []string, getenv func(string) string) (turnOptions, error) {
+	options := turnOptions{
+		Iterations: 1,
+		Version:    strings.TrimSpace(getenv("ZERO_BENCH_VERSION")),
+		Commit:     strings.TrimSpace(getenv("ZERO_BENCH_COMMIT")),
+	}
+	for index := 0; index < len(args); index++ {
+		arg := args[index]
+		flag, inlineValue := splitFlagValue(arg)
+		switch flag {
+		case "--suite":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.SuitePath = value
+			index = next
+		case "--model":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.Model = value
+			index = next
+		case "--mode":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.Mode = value
+			index = next
+		case "--binary":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.Binary = value
+			index = next
+		case "--iterations":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			parsed, err := parsePositiveInteger(flag, value)
+			if err != nil {
+				return options, err
+			}
+			options.Iterations = parsed
+			index = next
+		case "--version":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.Version = value
+			index = next
+		case "--commit":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.Commit = value
+			index = next
+		case "--output":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.Output = value
+			index = next
+		case "--self-correct":
+			if strings.Contains(arg, "=") {
+				return options, fmt.Errorf("%s does not accept a value", flag)
+			}
+			options.SelfCorrect = true
+		case "--json":
+			if strings.Contains(arg, "=") {
+				return options, fmt.Errorf("%s does not accept a value", flag)
+			}
+			options.JSON = true
+		case "--dry-run":
+			if strings.Contains(arg, "=") {
+				return options, fmt.Errorf("%s does not accept a value", flag)
+			}
+			options.DryRun = true
+		case "-h", "--help":
+			if strings.Contains(arg, "=") {
+				return options, fmt.Errorf("%s does not accept a value", flag)
+			}
+			options.Help = true
+		default:
+			return options, fmt.Errorf("unknown option: %s", arg)
+		}
+	}
+	if options.Help {
+		return options, nil
+	}
+	if strings.TrimSpace(options.SuitePath) == "" {
+		return options, fmt.Errorf("--suite is required")
+	}
+	if strings.TrimSpace(options.Model) == "" && !options.DryRun {
+		return options, fmt.Errorf("--model is required (or pass --dry-run)")
+	}
+	return options, nil
+}
+
+func writeTurnReport(path string, result perfbench.TurnBenchResult) error {
+	if dir := filepath.Dir(path); dir != "" {
+		if err := os.MkdirAll(dir, 0o755); err != nil {
+			return err
+		}
+	}
+	var buffer bytes.Buffer
+	if err := perfbench.WriteTurnBenchJSON(&buffer, result); err != nil {
+		return err
+	}
+	return os.WriteFile(path, buffer.Bytes(), 0o644)
+}
+
+func turnHelpText() string {
+	return strings.Join([]string{
+		"Usage: zero-perf-bench turn [options]",
+		"",
+		"Runs ZERO headlessly with --trace against a per-turn benchmark task set and",
+		"records per-span latency plus the top controllable latency sources (the",
+		"Phase 0 baseline's \"do not proceed until\" criterion). Each task is a fresh",
+		"`zero exec` process, so iterations are cold-start samples; a warm path needs",
+		"an in-process runner (future work).",
+		"",
+		"Options:",
+		"  --suite <path>      Task set JSON file (required)",
+		"  --model <model>     Model to run (required unless --dry-run)",
+		"  --mode <name>       Exec mode preset to apply",
+		"  --self-correct      Enable the post-edit verify-and-correct loop",
+		"  --binary <path>     Path to the `zero` binary (default: zero on PATH / repo root)",
+		"  --iterations <n>    Times to run each task (default: 1)",
+		"  --version <v>       Record the ZERO version (default: $ZERO_BENCH_VERSION)",
+		"  --commit <sha>      Record the ZERO commit (default: $ZERO_BENCH_COMMIT)",
+		"  --output <path>     Write the JSON result to path",
+		"  --json              Print only the JSON result",
+		"  --dry-run           Load the manifest and exit without invoking the agent",
+		"  -h, --help          Show this help",
+	}, "\n") + "\n"
+}
@@ -102,14 +102,24 @@ The full format spec (frontmatter fields, tool scopes, prompt conventions) is in

 Skills are markdown instruction packs the agent can pull in on demand. Each skill is a directory containing a `SKILL.md`. Standalone project skill directories are not supported in this version (shared project-wide skills must go in `AGENTS.md` or as a hook). However, project plugins (section 6) may bundle skills, which are merged into the active run.

-Discovery root: `$ZERO_SKILLS_DIR` → `$XDG_DATA_HOME/zero/skills` → `~/.local/share/zero/skills/`. A missing directory is fine — Zero just reports "no skills".
+Discovery roots (earlier wins on name collisions):
+
+1. **Primary Zero dir** — `$ZERO_SKILLS_DIR` if set, else `$XDG_DATA_HOME/zero/skills`, else `~/.local/share/zero/skills/`
+2. **Shared multi-agent dir** — `~/.agents/skills/` when present (read-only discovery; never an install target)
+3. **Plugin skill roots** — skills bundled by active plugins (section 6)
+
+A missing directory is fine — Zero just omits it. Management commands (`zero skills add` / `remove` / `lock`) always write to the primary Zero dir only; `list` / `info` search primary + `~/.agents/skills`.

 ```text
-~/.local/share/zero/skills/
+~/.local/share/zero/skills/          # primary (install/remove/lock target)
  run-benchmarks/
    SKILL.md
  write-changelog/
    SKILL.md
+
+~/.agents/skills/                    # optional shared multi-agent root
+  shared-review/
+    SKILL.md
 ```

 `SKILL.md` format:
@@ -126,9 +136,9 @@ description: Run the project's benchmark suite and summarize the deltas.
 3. Report any regression > 5% with the function name and the previous value.
 ```

-Only `name` and `description` are recognized in the frontmatter today. The `name` defaults to the directory name. Within a single skills root, duplicate names are resolved by lexicographic directory order. During an agent run, Zero loads the default skills directory before plugin skill roots; earlier roots win name collisions silently. Plugin-declared skills (section 6) are merged into the active agent run at plugin activation time, so bundled skills appear in the available skills list and can be loaded with the `skill` tool.
+Only `name` and `description` are recognized in the frontmatter today. The `name` defaults to the directory name. Within a single skills root, duplicate names are resolved by lexicographic directory order. Across roots, Zero loads the primary directory first, then `~/.agents/skills`, then plugin skill roots; earlier roots win name collisions. Plugin-declared skills (section 6) are merged into the active agent run at plugin activation time, so bundled skills appear in the available skills list and can be loaded with the `skill` tool.

-The `skill` core tool lets the agent load any discovered skill by name.
+The `skill` tool lets the agent load any discovered skill by name (primary, agents, or plugin).

 ## 4. Hooks

@@ -240,10 +250,13 @@ OAuth server names ending in `.<32 hex chars>` are reserved for token storage.

 ```bash
 zero serve --mcp
+zero serve --mcp -C /path/to/workspace
+zero serve --mcp --add-dir /path/to/extra
 ```

 The server speaks MCP over stdio. Configure it from the receiving side as a `stdio` server whose command is `zero serve --mcp`.

+`-C/--cwd` sets the workspace root exposed as MCP resources and used by core tools. `--add-dir` (repeatable) widens that resource/tool scope beyond the workspace without granting the sandbox temp roots used by interactive runs.
 ## 6. Plugins

 A plugin is a self-contained directory that bundles tools, hooks, and skills for one capability. Plugins live at:
@@ -733,9 +733,10 @@ flowchart TD
 - **Skills** are instruction packs the model can load on demand.
 - **Specialists** are sub-agents callable through the `Task` tool.
 - **MCP servers** contribute external tools.
- **Plugins** can add tools, hooks, and skill roots. Bootstrap merges plugin skill
-  roots into the runtime skill tool and prompt-visible skill list; the base
-  `internal/skills` scanner still scans one root at a time.
+- **Plugins** can add tools, hooks, and skill roots. Bootstrap always registers a
+  multi-root skill tool: primary Zero skills dir, optional `~/.agents/skills`,
+  then plugin skill roots (earlier wins). `internal/skills` owns that merge via
+  `LoadFromRoots` / `DiscoveryRoots`; single-root `Load` remains for install/write.
 - **Hooks** can observe or block tool lifecycle events.

 ## End-to-End Data Flow
@@ -0,0 +1,39 @@
+# TUI Themes
+
+Zero's TUI ships a set of built-in color themes. Pick one with `--theme <name>`,
+the `ZERO_THEME` environment variable, or the `/theme <name>` command while
+running. `auto` (the default) follows the terminal's detected background.
+
+Run `/theme` with no argument to open a picker: move through the list to
+live-preview each theme, press Enter to apply the highlighted one, or Esc to
+cancel and restore the previously active theme. Run `/theme list` to print
+the active theme and the registered names without opening the picker.
+
+## Dune (`dune`)
+
+A warm sand-and-cream palette: sand/cream surface, charcoal ink, and a soft
+amber accent.
+
+## Neon (`neon`)
+
+A neon-on-black palette: pitch-black surface, bright green ink, and a cyan accent.
+
+## Adding a theme
+
+Every theme is a `palette` (a table of color hex tokens) plus one entry in
+`themeRegistry`, both in `internal/tui/theme_palettes.go`. `buildTheme` in
+`internal/tui/theme.go` turns a `palette` into the resolved `lipgloss.Style`
+set every renderer reads from the active `zeroTheme`. Adding a theme means
+adding a new `palette{...}` literal, a `themeRegistry` entry, and test
+coverage for the new palette (see below).
+
+Registry-wide tests in `internal/tui/theme_select_test.go` assert the basic
+WCAG AA text tokens, the gray-ramp order, the diff word-span pairs, and the
+selected-row band for every entry. The rendered-surface invariants beyond
+those (permission surfaces, selected-row secondary text, diff gutters, and
+the xterm-256 downsampling checks) are asserted per palette, not against the
+whole registry: `TestExtendedThemeContrastInvariants` and
+`TestExtendedThemeANSI256Contrast` enumerate the palettes they cover. A new
+theme must be added to those tests (or given equivalent palette-specific
+assertions), or CI can stay green while its permission, selected-row, and
+diff surfaces ship unreadable.
@@ -3,9 +3,9 @@ module github.com/Gitlawb/zero
 go 1.26.5

 require (
-	charm.land/bubbles/v2 v2.1.0
-	charm.land/bubbletea/v2 v2.0.7
-	charm.land/lipgloss/v2 v2.0.4
+	charm.land/bubbles/v2 v2.1.1
+	charm.land/bubbletea/v2 v2.0.8
+	charm.land/lipgloss/v2 v2.0.5
 	github.com/alecthomas/chroma/v2 v2.27.0
 	github.com/atotto/clipboard v0.1.4
 	github.com/aymanbagabas/go-udiff v0.4.1
@@ -14,18 +14,18 @@ require (
 	github.com/charmbracelet/x/term v0.2.2
 	github.com/coder/websocket v1.8.15
 	github.com/ledongthuc/pdf v0.0.0-20250511090121-5959a4027728
-	golang.org/x/sys v0.46.0
+	golang.org/x/sys v0.47.0
 	mvdan.cc/sh/v3 v3.13.1
 )

 require (
-	github.com/charmbracelet/ultraviolet v0.0.0-20260615092913-2399af76d5b1 // indirect
+	github.com/charmbracelet/ultraviolet v0.0.0-20260703014108-f5a850f9c2b7 // indirect
 	github.com/charmbracelet/x/exp/golden v0.0.0-20260615092313-b57e5e6d29bb // indirect
 	github.com/charmbracelet/x/termios v0.1.1 // indirect
 	github.com/charmbracelet/x/windows v0.2.2 // indirect
 	github.com/clipperhouse/displaywidth v0.11.0 // indirect
 	github.com/clipperhouse/uax29/v2 v2.7.0 // indirect
-	github.com/dlclark/regexp2/v2 v2.2.1 // indirect
+	github.com/dlclark/regexp2/v2 v2.5.0 // indirect
 	github.com/go-quicktest/qt v1.102.0 // indirect
 	github.com/lucasb-eyer/go-colorful v1.4.0 // indirect
 	github.com/mattn/go-runewidth v0.0.24 // indirect
@@ -34,5 +34,5 @@ require (
 	github.com/rogpeppe/go-internal v1.15.0 // indirect
 	github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e // indirect
 	golang.org/x/exp v0.0.0-20260611194520-c48552f49976 // indirect
-	golang.org/x/sync v0.21.0 // indirect
+	golang.org/x/sync v0.22.0 // indirect
 )
@@ -1,9 +1,9 @@
-charm.land/bubbles/v2 v2.1.0 h1:YSnNh5cPYlYjPxRrzs5VEn3vwhtEn3jVGRBT3M7/I0g=
-charm.land/bubbles/v2 v2.1.0/go.mod h1:l97h4hym2hvWBVfmJDtrEHHCtkIKeTEb3TTJ4ZOB3wY=
-charm.land/bubbletea/v2 v2.0.7 h1:7qw2tTAVar7m7klOPBYfTB0mniv/RuexsYwMRNxSeL0=
-charm.land/bubbletea/v2 v2.0.7/go.mod h1:DGW2q8gvzHnOpMpZTORs0aySVHCox5C+2Svk0fci1qs=
-charm.land/lipgloss/v2 v2.0.4 h1:lcPeVtcp23SNra7lHy8iYE4UC2aIipVQ47sbGyyxR5Q=
-charm.land/lipgloss/v2 v2.0.4/go.mod h1:0653x8epbZSzdDfO/XPS1a/uYPOBeSsCssOpJOqDzik=
+charm.land/bubbles/v2 v2.1.1 h1:7r55WzBxpo/R3z98hGmY7KKPd3ET6vsf0Fb9sDHOV60=
+charm.land/bubbles/v2 v2.1.1/go.mod h1:GE6M31gaWZVXzGw73OeuTTgy4lX+OtkH0E5ymnNsHxo=
+charm.land/bubbletea/v2 v2.0.8 h1:SxTJMhCAI3lbPmy4SgX5LWZ24AdINr4I6UEqzZvYJuY=
+charm.land/bubbletea/v2 v2.0.8/go.mod h1:2SkdgoTXluXJHOUwAoRlRXF/28vklb1rFl6GcgV1/ss=
+charm.land/lipgloss/v2 v2.0.5 h1:kbNxgeeUOYv5J0YdpxFjfvf3dFvqH8Aci4zB6xqFtrY=
+charm.land/lipgloss/v2 v2.0.5/go.mod h1:9oqhxt4yxIMe6q5A4kHr44DremZk7J9UNh74GlWa5nc=
 github.com/alecthomas/assert/v2 v2.11.0 h1:2Q9r3ki8+JYXvGsDyBXwH3LcJ+WK5D0gc5E8vS6K3D0=
 github.com/alecthomas/assert/v2 v2.11.0/go.mod h1:Bze95FyfUr7x34QZrjL+XP+0qgp/zg8yS+TtBj1WA3k=
 github.com/alecthomas/chroma/v2 v2.27.0 h1:FodwmyOBgJULFYmDqibcp9pvfDLWdtPRh9v/r5BXYZs=
@@ -16,8 +16,8 @@ github.com/aymanbagabas/go-udiff v0.4.1 h1:OEIrQ8maEeDBXQDoGCbbTTXYJMYRCRO1fnodZ
 github.com/aymanbagabas/go-udiff v0.4.1/go.mod h1:0L9PGwj20lrtmEMeyw4WKJ/TMyDtvAoK9bf2u/mNo3w=
 github.com/charmbracelet/colorprofile v0.4.3 h1:QPa1IWkYI+AOB+fE+mg/5/4HRMZcaXex9t5KX76i20Q=
 github.com/charmbracelet/colorprofile v0.4.3/go.mod h1:/zT4BhpD5aGFpqQQqw7a+VtHCzu+zrQtt1zhMt9mR4Q=
-github.com/charmbracelet/ultraviolet v0.0.0-20260615092913-2399af76d5b1 h1:4+r3uOJ69ueRBt4okgEfWZeXs3BD36HcDBmOIAUlETk=
-github.com/charmbracelet/ultraviolet v0.0.0-20260615092913-2399af76d5b1/go.mod h1:f/jRa757WUmaOZrbPspXymbg/GnbF+rwe4OLsG7aXYo=
+github.com/charmbracelet/ultraviolet v0.0.0-20260703014108-f5a850f9c2b7 h1:3FmWoGNWK4STvqg0O0Aeav2T7rodWJAPeF0QpH+8gFw=
+github.com/charmbracelet/ultraviolet v0.0.0-20260703014108-f5a850f9c2b7/go.mod h1:f/jRa757WUmaOZrbPspXymbg/GnbF+rwe4OLsG7aXYo=
 github.com/charmbracelet/x/ansi v0.11.7 h1:kzv1kJvjg2S3r9KHo8hDdHFQLEqn4RBCb39dAYC84jI=
 github.com/charmbracelet/x/ansi v0.11.7/go.mod h1:9qGpnAVYz+8ACONkZBUWPtL7lulP9No6p1epAihUZwQ=
 github.com/charmbracelet/x/exp/golden v0.0.0-20260615092313-b57e5e6d29bb h1:m1Uub9uM7meZ8ofVyeRfK7M/knCi3sF2gKn9kldVW0I=
@@ -34,8 +34,8 @@ github.com/clipperhouse/uax29/v2 v2.7.0 h1:+gs4oBZ2gPfVrKPthwbMzWZDaAFPGYK72F0NJ
 github.com/clipperhouse/uax29/v2 v2.7.0/go.mod h1:EFJ2TJMRUaplDxHKj1qAEhCtQPW2tJSwu5BF98AuoVM=
 github.com/coder/websocket v1.8.15 h1:6B2JPeOGlpff2Uz6vOEH1Vzpi0iUz20A+lPVhPHtNUA=
 github.com/coder/websocket v1.8.15/go.mod h1:NX3SzP+inril6yawo5CQXx8+fk145lPDC6pumgx0mVg=
-github.com/dlclark/regexp2/v2 v2.2.1 h1:mf4KkFUj0gJuarK8P+LgiS+Lit7m9N1yAwEfPbee7R0=
-github.com/dlclark/regexp2/v2 v2.2.1/go.mod h1:avUrQvPaLz2DrFNHJF0taWAFFX2C1GMSSoeiqFjcBmU=
+github.com/dlclark/regexp2/v2 v2.5.0 h1:liNiWIPCvCh5HBcYfsjd+P16AG79fwd6T1Toy2gOtEA=
+github.com/dlclark/regexp2/v2 v2.5.0/go.mod h1:avUrQvPaLz2DrFNHJF0taWAFFX2C1GMSSoeiqFjcBmU=
 github.com/go-quicktest/qt v1.102.0 h1:HSQxCeh5YZH3EL3W39ixjtyaEhcWSXQHtHnMBzSs474=
 github.com/go-quicktest/qt v1.102.0/go.mod h1:p4lGIVX+8Wa6ZPNDvqcxq36XpUDLh42FLetFU7odllI=
 github.com/google/go-cmp v0.7.0 h1:wk8382ETsv4JYUZwIsn6YpYiWiBsYLSJiTsyBybVuN8=
@@ -62,9 +62,9 @@ github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e h1:JVG44RsyaB9T2KIHavM
 github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e/go.mod h1:RbqR21r5mrJuqunuUZ/Dhy/avygyECGrLceyNeo4LiM=
 golang.org/x/exp v0.0.0-20260611194520-c48552f49976 h1:X8Hz2ImujgbmetVuW+w2YkyZChE3cBpZi2P158rTG9M=
 golang.org/x/exp v0.0.0-20260611194520-c48552f49976/go.mod h1:vnf4pv9iKZXY58sQE1L86zmNWJ4159e1RkcWiLCkeEY=
-golang.org/x/sync v0.21.0 h1:HLII4xRRTtCRkxYp4HNFF0Js/Og6q2i++KXbg0gHCwM=
-golang.org/x/sync v0.21.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
-golang.org/x/sys v0.46.0 h1:noSf2Fq6F8DBgS+LysIkx7rIExoNHJsxOAtPp4rthXw=
-golang.org/x/sys v0.46.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
+golang.org/x/sync v0.22.0 h1:SZjpbeLmrCk4xhRSZFNZW5gFUeCeFgjekvI/+gfScek=
+golang.org/x/sync v0.22.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
+golang.org/x/sys v0.47.0 h1:o7XGOvZQCADBQQ4Y7VNq2dRWQR7JmOUW8Kxx4ZsNgWs=
+golang.org/x/sys v0.47.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
 mvdan.cc/sh/v3 v3.13.1 h1:DP3TfgZhDkT7lerUdnp6PTGKyxxzz6T+cOlY/xEvfWk=
 mvdan.cc/sh/v3 v3.13.1/go.mod h1:lXJ8SexMvEVcHCoDvAGLZgFJ9Wsm2sulmoNEXGhYZD0=
@@ -6,6 +6,7 @@ import (
 	"errors"
 	"strings"

+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )

@@ -430,6 +431,12 @@ func (state *compactionState) maybeCompact(
 		state.lowWaterMark = size
 		return messages
 	}
+	// Only count a compaction when it actually shrank the history, so the
+	// compaction counter reflects real context reductions rather than paid
+	// no-ops that left the token budget untouched.
+	if r := trace.FromContext(ctx); r != nil {
+		r.Counter(trace.CounterCompactionCount, 1)
+	}
 	state.lowWaterMark = newSize
 	return compacted
 }
@@ -479,7 +486,11 @@ func (state *compactionState) recover(
 	// one-shot budget now so a provider that keeps returning context-limit errors
 	// after a successful compaction can't loop forever. Store the low-water mark in
 	// the SAME combined (messages + tool-defs) domain maybeCompact uses, so the
-	// proactive shrink-guard compares like with like.
+	// proactive shrink-guard compares like with like. Count it now that the shrink
+	// is confirmed, so the counter mirrors maybeCompact's real-reduction policy.
+	if r := trace.FromContext(ctx); r != nil {
+		r.Counter(trace.CounterCompactionCount, 1)
+	}
 	state.reactiveAttempted = true
 	state.lowWaterMark = state.calibratedTokens(estimateTokens(result) + estimateToolDefTokens(tools))
 	return result, true, nil
@@ -15,6 +15,7 @@ import (
 	"github.com/Gitlawb/zero/internal/sandbox"
 	"github.com/Gitlawb/zero/internal/streamjson"
 	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )

@@ -114,6 +115,26 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 		return Result{}, errors.New("agent provider is required")
 	}

+	// Tracing is opt-in. When a recorder is wired, thread it into ctx so the
+	// providerio seam and reconnect helper can reach it via trace.FromContext,
+	// mark the run's start, and wrap OnUsage so token counters accumulate for
+	// free alongside the existing per-request plumbing. nil recorder leaves the
+	// loop byte-identical: FromContext returns nil, the no-op stamps cost
+	// nothing, and the OnUsage wrapper is not installed.
+	if options.Trace != nil {
+		ctx = trace.WithContext(ctx, options.Trace)
+		options.Trace.Start()
+		originalOnUsage := options.OnUsage
+		options.OnUsage = func(usage Usage) {
+			if originalOnUsage != nil {
+				originalOnUsage(usage)
+			}
+			options.Trace.Counter(trace.CounterInputTokens, int64(usage.InputTokens))
+			options.Trace.Counter(trace.CounterCachedInputTokens, int64(usage.CachedInputTokens))
+			options.Trace.Counter(trace.CounterOutputTokens, int64(usage.OutputTokens))
+		}
+	}
+
 	maxTurns := options.MaxTurns
 	if maxTurns <= 0 {
 		maxTurns = 12
@@ -164,7 +185,11 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 	result = Result{Messages: copyMessages(messages)}
 	dispatchSessionStart(ctx, options)
 	defer func() {
-		dispatchSessionEnd(ctx, options, result, err)
+		// Use a context detached from ctx's cancellation: on Esc/Ctrl+C the run
+		// context is already canceled by the time this defer runs, and a canceled
+		// parent would make Hooks.Dispatch's context.WithTimeout derive an
+		// already-expired context, so the sessionEnd hook would never actually run.
+		dispatchSessionEnd(context.WithoutCancel(ctx), options, result, err)
 	}()
 	for turn := 0; turn < maxTurns; turn++ {
 		result.Turns = turn + 1
@@ -184,12 +209,40 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 		// the tool-definition tokens (they ride on every request) in its estimate.
 		// partitionTools depends only on registry/permissions/options/loaded, not on
 		// the messages, so computing it before compaction is safe.
+		// Build the per-turn tool list first so proactive compaction can include
+		// the tool-definition tokens (they ride on every request) in its estimate.
+		// partitionTools depends only on registry/permissions/options/loaded, not on
+		// the messages, so computing it before compaction is safe.
+		toolPartitionSpan := options.Trace.Span(trace.SpanToolPartition)
 		exposed, _ := partitionToolsCached(registry, permissionMode, options, loaded, toolDefCache)
+		toolPartitionSpan.End()
+
+		// Prompt-prefix fingerprint: hash the seven cacheable sub-components
+		// of this turn's request (system prompt sections, project context,
+		// skills, partitioned tool list, tool schemas) and emit one trace
+		// event. Stable across turns => cacheable. Drift in any sub-hash
+		// names the sub-component that broke the cache for this turn. The
+		// computation is pure and cheap (seven SHA-256s of small strings)
+		// and is a no-op when tracing is off.
+		if options.Trace != nil {
+			fp := ComputePrefixFingerprint(options, exposed)
+			options.Trace.EmitPrefixHash(trace.PrefixHash{
+				BaseInstructionsHash:   fp.BaseInstructionsHash,
+				ConfirmationPolicyHash: fp.ConfirmationPolicyHash,
+				ProjectContextHash:     fp.ProjectContextHash,
+				SkillsHash:             fp.SkillsHash,
+				ToolsHash:              fp.ToolsHash,
+				SchemaHash:             fp.SchemaHash,
+				CompletePrefixHash:     fp.CompletePrefixHash,
+			})
+		}

 		// PROACTIVE compaction: if the history is approaching the model's
 		// context window, summarize the oldest middle before building the
 		// request. A no-op when ContextWindow == 0 (compaction disabled).
+		compactionSpan := options.Trace.Span(trace.SpanCompaction)
 		messages = compactor.maybeCompact(ctx, provider, messages, exposed)
+		compactionSpan.End()
 		request := zeroruntime.CompletionRequest{
 			Messages:        copyMessages(messages),
 			Tools:           exposed,
@@ -255,11 +308,39 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 		// conversation-state duplication.
 		forwardedVisibleText := false
 		forwardingOpts := zeroruntime.CollectOptions{OnUsage: options.OnUsage}
-		if options.OnText != nil {
-			forwardingOpts.OnText = func(s string) { forwardedVisibleText = true; options.OnText(s) }
+		// Install text/reasoning forwarding handlers whenever EITHER a user
+		// callback OR a trace recorder is set. A headless traced run (e.g. `zero
+		// exec --trace`) sets Trace but no OnText/OnReasoning; without these
+		// handlers the stream is still collected, but FirstTokenAt would never
+		// stamp and the trace would lose its TTFT signal. The trace recorder's
+		// stamp methods are nil-safe, but we guard on `trace != nil` so a run with
+		// a user callback but no recorder still works. forwardedVisibleText stays
+		// tied to the USER callback only, preserving the stall-retry semantics
+		// (a trace-only handler forwards nothing the user would see duplicated).
+		rec := options.Trace
+		onText := options.OnText
+		if onText != nil || rec != nil {
+			forwardingOpts.OnText = func(s string) {
+				if rec != nil {
+					rec.StampFirstVisibleEvent()
+					rec.StampFirstToken()
+				}
+				if onText != nil {
+					forwardedVisibleText = true
+					onText(s)
+				}
+			}
 		}
-		if options.OnReasoning != nil {
-			forwardingOpts.OnReasoning = func(s string) { options.OnReasoning(s) }
+		onReasoning := options.OnReasoning
+		if onReasoning != nil || rec != nil {
+			forwardingOpts.OnReasoning = func(s string) {
+				if rec != nil {
+					rec.StampFirstToken()
+				}
+				if onReasoning != nil {
+					onReasoning(s)
+				}
+			}
 		}
 		if options.OnToolCallStart != nil {
 			forwardingOpts.OnToolCallStart = func(id, name string) { options.OnToolCallStart(id, name) }
@@ -297,14 +378,18 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 				if retryStreamErr != nil {
 					return collected, retryStreamErr
 				}
+				genSpan := options.Trace.Span(trace.SpanGeneration)
 				collected = zeroruntime.CollectStreamWithOptions(ctx, retryStream, zeroruntime.CollectOptions{
 					OnUsage: options.OnUsage,
 				})
+				genSpan.End()
 			}
 			return collected, nil
 		}

+		generationSpan := options.Trace.Span(trace.SpanGeneration)
 		collected := zeroruntime.CollectStreamWithOptions(ctx, stream, forwardingOpts)
+		generationSpan.End()
 		if collected.Error != "" {
 			updated, stop := recoverStreamError(collected)
 			collected = updated
@@ -357,7 +442,9 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 				result.Messages = copyMessages(messages)
 				return result, retryErr
 			}
+			stallGenSpan := options.Trace.Span(trace.SpanGeneration)
 			collected = zeroruntime.CollectStreamWithOptions(ctx, retryStream, forwardingOpts)
+			stallGenSpan.End()
 		}
 		if collected.Error != "" {
 			// Route a reissued stream's non-stall error through the SAME recovery as
@@ -459,6 +546,7 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 				if cue || planPending {
 					if continueNudges < maxContinueNudges {
 						continueNudges++
+						options.Trace.Counter(trace.CounterCompletionNudges, 1)
 						reason := "your message ended mid-step"
 						if !cue {
 							reason = "pending plan items remain — finish them, or mark them complete with update_plan if you are done"
@@ -488,6 +576,7 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 				// (no admission, no cue) then finalizes as success on the next turn.
 				if options.SelfCorrect != nil && !acceptanceRequested {
 					acceptanceRequested = true
+					options.Trace.Counter(trace.CounterAcceptanceChecks, 1)
 					messages = append(messages, zeroruntime.Message{
 						Role:    zeroruntime.MessageRoleUser,
 						Content: acceptanceVerificationNudge(),
@@ -548,20 +637,26 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 					runEnd++
 				}
 				if runEnd-index >= 2 {
+					batchSpan := options.Trace.Span(trace.SpanToolExecution)
 					precomputed = executeParallelReadBatch(ctx, registry, collected.ToolCalls, index, runEnd, permissionMode, options)
+					batchSpan.End()
 					precomputedStart, precomputedEnd = index, runEnd
 				}
 			}
 			if options.OnToolCall != nil {
 				options.OnToolCall(call)
 			}
+			options.Trace.StampFirstUsefulAction()
 			var toolResult ToolResult
 			var abortErr error
 			if index >= precomputedStart && index < precomputedEnd {
 				toolResult, abortErr = precomputed[index-precomputedStart].result, precomputed[index-precomputedStart].abortErr
 			} else {
+				toolSpan := options.Trace.Span(trace.SpanToolExecution)
 				toolResult, abortErr = executeToolCall(ctx, registry, call, permissionMode, options)
+				toolSpan.End()
 			}
+			options.Trace.Counter(trace.CounterToolCalls, 1)
 			if options.OnToolResult != nil {
 				options.OnToolResult(toolResult)
 			}
@@ -666,6 +761,7 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 				// context-window sizing, and usage attribution follow the new model.
 				provider = newProvider
 				options.Model = turnRequestedModel
+				options.Trace.Counter(trace.CounterModelSwitches, 1)
 				// KNOWN LIMITATION (deferred): the compactor's context-window budget
 				// is fixed at run start from options.ContextWindow and is NOT updated
 				// here, so a switch to a model with a different window keeps compacting
@@ -764,6 +860,7 @@ func finalAnswerAfterMaxTurns(ctx context.Context, provider Provider, messages [
 	if err != nil {
 		return "", messages, ""
 	}
+	finalGenSpan := options.Trace.Span(trace.SpanGeneration)
 	collected := zeroruntime.CollectStreamWithOptions(ctx, stream, zeroruntime.CollectOptions{
 		OnText:          options.OnText,
 		OnReasoning:     options.OnReasoning,
@@ -771,6 +868,7 @@ func finalAnswerAfterMaxTurns(ctx context.Context, provider Provider, messages [
 		OnToolCallStart: options.OnToolCallStart,
 		OnToolCallDelta: options.OnToolCallDelta,
 	})
+	finalGenSpan.End()
 	if ctx.Err() != nil || collected.Error != "" || strings.TrimSpace(collected.Text) == "" {
 		return "", messages, ""
 	}
@@ -1874,7 +1972,10 @@ func requestPermission(ctx context.Context, request PermissionRequest, options O
 	if options.OnPermissionRequest == nil {
 		return PermissionDecision{Action: PermissionDecisionDeny, Reason: request.Reason}, nil
 	}
-	return options.OnPermissionRequest(ctx, request)
+	permSpan := options.Trace.Span(trace.SpanPermissionWait)
+	decision, err := options.OnPermissionRequest(ctx, request)
+	permSpan.End()
+	return decision, err
 }

 func normalizePermissionDecisionAction(action PermissionDecisionAction) PermissionDecisionAction {
@@ -6,6 +6,7 @@ import (
 	"errors"
 	"fmt"
 	"os"
+	"os/exec"
 	"path/filepath"
 	"runtime"
 	"strings"
@@ -15,6 +16,7 @@ import (
 	"github.com/Gitlawb/zero/internal/sandbox"
 	"github.com/Gitlawb/zero/internal/specmode"
 	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )

@@ -92,6 +94,87 @@ func TestRunDispatchesSessionLifecycleHooks(t *testing.T) {
 	}
 }

+// cancelingProvider simulates an Esc/Ctrl+C abort: it cancels the run's own
+// context mid-stream (as the TUI's interrupt handler would) and returns no
+// events, so Run observes ctx.Err() and returns early with an already-canceled
+// context in scope for the deferred sessionEnd dispatch.
+type cancelingProvider struct {
+	cancel context.CancelFunc
+}
+
+func (provider *cancelingProvider) StreamCompletion(ctx context.Context, request zeroruntime.CompletionRequest) (<-chan zeroruntime.StreamEvent, error) {
+	provider.cancel()
+	ch := make(chan zeroruntime.StreamEvent)
+	close(ch)
+	return ch, nil
+}
+
+// TestRunDispatchesSessionEndHookAfterContextCancellation is a regression test
+// for a bug where an interrupted run (Esc/Ctrl+C) canceled ctx before the
+// deferred sessionEnd dispatch ran, so Hooks.Dispatch's context.WithTimeout(ctx, ...)
+// derived an already-canceled context and the hook command never actually launched.
+func TestRunDispatchesSessionEndHookAfterContextCancellation(t *testing.T) {
+	goBinary, err := exec.LookPath("go")
+	if err != nil {
+		// The test binary runs on the build machine, so its build-time GOROOT
+		// still identifies the toolchain that produced it.
+		goRoot := runtime.GOROOT() //nolint:staticcheck // Safe for this non-portable test binary.
+		goBinary = filepath.Join(goRoot, "bin", "go")
+		if runtime.GOOS == "windows" {
+			goBinary += ".exe"
+		}
+		if _, statErr := os.Stat(goBinary); statErr != nil {
+			t.Skipf("go binary unavailable on PATH or in GOROOT: %v", statErr)
+		}
+	}
+	audit, err := hooks.NewAuditStore(hooks.AuditStoreOptions{AuditPath: filepath.Join(t.TempDir(), "audit.jsonl")})
+	if err != nil {
+		t.Fatalf("NewAuditStore: %v", err)
+	}
+	dispatcher := hooks.NewDispatcher(hooks.DispatcherOptions{
+		Config: hooks.Config{
+			Enabled: true,
+			Hooks: []hooks.Definition{
+				{ID: "zero.session-end", Event: hooks.EventSessionEnd, Command: goBinary, Args: []string{"version"}, Enabled: true},
+			},
+		},
+		Audit: audit,
+	})
+
+	ctx, cancel := context.WithCancel(context.Background())
+	provider := &cancelingProvider{cancel: cancel}
+
+	_, err = Run(ctx, "hi", provider, Options{
+		SessionID:    "session-cancel",
+		Cwd:          t.TempDir(),
+		ProviderName: "test-provider",
+		Model:        "test-model",
+		Hooks:        dispatcher,
+	})
+	if !errors.Is(err, context.Canceled) {
+		t.Fatalf("Run err = %v, want context.Canceled", err)
+	}
+
+	events, err := audit.ReadEvents()
+	if err != nil {
+		t.Fatalf("ReadEvents: %v", err)
+	}
+	var completedStatus hooks.AuditStatus
+	found := false
+	for _, event := range events {
+		if event.Type == "hook_execution_completed" && event.Event == hooks.EventSessionEnd {
+			completedStatus = event.Status
+			found = true
+		}
+	}
+	if !found {
+		t.Fatalf("no sessionEnd hook_execution_completed audit event, events=%#v", events)
+	}
+	if completedStatus != hooks.AuditCompleted {
+		t.Fatalf("sessionEnd hook status = %q, want %q (hook must actually run despite ctx cancellation)", completedStatus, hooks.AuditCompleted)
+	}
+}
+
 type recordingWebSearchTool struct {
 	calls []map[string]any
 }
@@ -3461,3 +3544,73 @@ func TestRunDoesNotFlagCleanToolOutput(t *testing.T) {
 		t.Errorf("clean output should not get a reminder, got %q", captured.Output)
 	}
 }
+
+// TestRunTracingWrapperStampsUsage verifies the per-turn tracing setup in Run:
+// a wired recorder is Started, OnUsage is wrapped so it still forwards to the
+// caller's callback AND stamps token counters, and the run completes.
+func TestRunTracingWrapperStampsUsage(t *testing.T) {
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{{
+		{Type: zeroruntime.StreamEventUsage, Usage: zeroruntime.Usage{InputTokens: 100, CachedInputTokens: 20, OutputTokens: 40}},
+		{Type: zeroruntime.StreamEventText, Content: "done"},
+		{Type: zeroruntime.StreamEventDone},
+	}}}
+	onUsageCalls := 0
+	rec := trace.NewRecorder("tracing-session", "run-1", "test")
+	if _, err := Run(context.Background(), "hi", provider, Options{
+		SessionID:    "tracing-session",
+		Cwd:          t.TempDir(),
+		ProviderName: "test-provider",
+		Model:        "test-model",
+		Trace:        rec,
+		OnUsage:      func(Usage) { onUsageCalls++ },
+	}); err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	tr := rec.Finish()
+	if tr.StartedAt.IsZero() {
+		t.Fatal("tracing wrapper did not Start the recorder")
+	}
+	// FirstTokenAt must stamp even though no OnText/OnReasoning user callback is
+	// set: a headless traced run (e.g. `zero exec --trace`) sets Trace but no UI
+	// callbacks, so the loop installs trace-only forwarding handlers to capture
+	// TTFT. Without them FirstTokenAt stays zero and the trace loses its signal.
+	if tr.FirstTokenAt.IsZero() {
+		t.Fatal("FirstTokenAt not stamped for a traced run with no OnText/OnReasoning callbacks")
+	}
+	if got := tr.Counter(trace.CounterInputTokens); got != 100 {
+		t.Fatalf("input token counter = %d, want 100", got)
+	}
+	if got := tr.Counter(trace.CounterCachedInputTokens); got != 20 {
+		t.Fatalf("cached input token counter = %d, want 20", got)
+	}
+	if got := tr.Counter(trace.CounterOutputTokens); got != 40 {
+		t.Fatalf("output token counter = %d, want 40", got)
+	}
+	if onUsageCalls == 0 {
+		t.Fatal("wrapped OnUsage did not forward to the caller's callback")
+	}
+}
+
+// TestRunNilTraceForwardsUsage verifies a nil recorder leaves the loop
+// byte-identical: OnUsage is not wrapped, so the caller's callback still fires
+// and nothing panics.
+func TestRunNilTraceForwardsUsage(t *testing.T) {
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{{
+		{Type: zeroruntime.StreamEventUsage, Usage: zeroruntime.Usage{InputTokens: 7}},
+		{Type: zeroruntime.StreamEventText, Content: "done"},
+		{Type: zeroruntime.StreamEventDone},
+	}}}
+	onUsageCalls := 0
+	if _, err := Run(context.Background(), "hi", provider, Options{
+		SessionID:    "nil-trace-session",
+		Cwd:          t.TempDir(),
+		ProviderName: "test-provider",
+		Model:        "test-model",
+		OnUsage:      func(Usage) { onUsageCalls++ },
+	}); err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	if onUsageCalls == 0 {
+		t.Fatal("OnUsage not forwarded when Trace is nil")
+	}
+}
@@ -0,0 +1,300 @@
+package agent
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+	"encoding/json"
+	"fmt"
+	"sort"
+	"strings"
+
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// promptSubstrings are the seven cacheable sub-components of the prompt a run
+// sends to a model. Each is the literal content the sub-component would emit
+// into the system prompt or tool list, hashed so a downstream observer can
+// detect drift turn-over-turn. The substrings are produced by
+// buildPromptSubstrings and consumed by ComputePrefixFingerprint.
+type promptSubstrings struct {
+	baseInstructions   string
+	confirmationPolicy string
+	projectContext     string
+	skills             string
+	tools              string
+	schema             string
+}
+
+// buildPromptSubstrings assembles the seven cacheable sub-components of the
+// prompt without joining them. The system prompt sections are produced by the
+// same private builders that buildSystemPrompt uses, so the substrings are
+// byte-identical to what would have appeared in the joined prompt. The tools
+// and schema substrings are derived from the partitioned tool list (caller-
+// provided) so this helper does not need to re-run the partition.
+//
+// This is the seam ComputePrefixFingerprint reads from. It exists as a
+// separate function (rather than returning both prompt and substrings from
+// buildSystemPrompt) so existing callers of buildSystemPrompt are unaffected
+// and the substrings helper is independently testable.
+func buildPromptSubstrings(options Options, exposed []zeroruntime.ToolDefinition) promptSubstrings {
+	core := strings.TrimSpace(options.SystemPrompt)
+	if core == "" {
+		core = strings.TrimSpace(coreSystemPrompt)
+	}
+	if core == "" {
+		core = fallbackSystemPrompt
+	}
+
+	// The confirmation policy is appended unconditionally in buildSystemPrompt;
+	// the substring here is the same constant.
+	policy := strings.TrimSpace(confirmationPolicy)
+
+	// projectContext is the joined output of workspaceContext, which itself
+	// walks the AGENTS.md / ZERO.md / .zero/AGENTS.md chain. Hashing the
+	// joined string captures every guideline file's contribution and the
+	// ordering between them.
+	project := workspaceContext(options.Cwd)
+
+	// skills is the joined output of skillsContext, with each skill on its own
+	// line. The substring is the literal text the model would see.
+	skills := skillsContext(options)
+
+	// tools and schema are derived from the partitioned tool list. The list
+	// passed in is already in stable order (partitionToolsCached is the
+	// canonical stable partitioner) so concatenating in slice order is
+	// sufficient — we sort names again defensively in case a future caller
+	// hands us a list from a different partitioner.
+	toolsSubstr, schemaSubstr := toolSubstrings(exposed)
+
+	return promptSubstrings{
+		baseInstructions:   core,
+		confirmationPolicy: policy,
+		projectContext:     project,
+		skills:             skills,
+		tools:              toolsSubstr,
+		schema:             schemaSubstr,
+	}
+}
+
+// toolSubstrings extracts a stable, name-sorted digest of the partitioned tool
+// list. The tools substring is "name1\nname2\n..." (sorted) so a reordering of
+// the partition produces a different hash. The schema substring is the
+// concatenation of each tool's JSON schema in name order, so a schema change
+// is detected independently of the tool ordering.
+func toolSubstrings(exposed []zeroruntime.ToolDefinition) (toolsSubstr, schemaSubstr string) {
+	if len(exposed) == 0 {
+		return "", ""
+	}
+	names := make([]string, 0, len(exposed))
+	byName := make(map[string]zeroruntime.ToolDefinition, len(exposed))
+	for _, def := range exposed {
+		names = append(names, def.Name)
+		byName[def.Name] = def
+	}
+	sort.Strings(names)
+	var toolsSB, schemaSB strings.Builder
+	for i, name := range names {
+		if i > 0 {
+			toolsSB.WriteByte('\n')
+		}
+		toolsSB.WriteString(name)
+		def := byName[name]
+		// Parameters is rendered to a canonical JSON string per tool. The
+		// schema-render cache in internal/agent (used by partitionToolsCached)
+		// guarantees the same render is byte-identical across turns for the
+		// same tool, so this substring is a stable hash input.
+		schemaSB.WriteString(name)
+		schemaSB.WriteByte('\n')
+		schemaSB.WriteString(canonicalSchemaString(def.Parameters))
+		schemaSB.WriteByte('\n')
+	}
+	return toolsSB.String(), schemaSB.String()
+}
+
+// canonicalSchemaString renders a tool's parameter schema to a stable string.
+// tools.ToolDef.Parameters is a map[string]any in the common case; Go's
+// fmt.Sprintf("%v", m) iterates a map in random order, which would produce
+// a different hash for the same Parameters value across calls and defeat
+// the fingerprint. encoding/json marshals maps with keys sorted
+// alphabetically, so json.Marshal is the primary stable render. The
+// fallback for the rare non-JSON-compatible value (functions, channels,
+// NaN/Inf floats, cyclic references) is a stable key-sorted stringifier
+// that walks the value with sorted keys for maps and a leading
+// "__non_json:" prefix so a future schema change is visible in the trace
+// (a SchemaHash collision from the same value) rather than a silent hash
+// drift.
+//
+// Note: SchemaHash is a stability signal, not a wire-identity signal. The
+// bytes json.Marshal produces for a Go map[string]any may differ from
+// the bytes the provider actually sends on the wire (provider encoders
+// use their own JSON conventions; an Anthropic schema may serialize with
+// different whitespace, key order, or number formatting). A consumer
+// must NOT assume SchemaHash matches the provider's wire schema — only
+// that two turns with the same Go-side Parameters produce the same
+// SchemaHash. This is sufficient for the trace's purpose (drift
+// detection) but not for a content-based equality check.
+func canonicalSchemaString(params map[string]any) string {
+	if len(params) == 0 {
+		return ""
+	}
+	data, err := json.Marshal(params)
+	if err != nil {
+		return "__non_json:" + stableStringify(params)
+	}
+	return string(data)
+}
+
+// stableStringify renders v to a deterministic string. Maps have their keys
+// sorted alphabetically; slices are walked in order; primitives use their
+// natural Go format. The function is recursive but bounded by the size of
+// the input, so a cyclic reference is not reachable here (the json.Marshal
+// fallback is hit when json.Marshal itself returns an error, which for the
+// common case is a non-JSON-compatible value, not a cycle — cycles are
+// rare in tool schemas and acceptable to mis-render in the fallback path
+// since the trace's contract is "the hash is stable for the same input,"
+// not "the fallback is lossless"). Used only when json.Marshal fails.
+func stableStringify(v any) string {
+	var sb strings.Builder
+	writeStable(&sb, v)
+	return sb.String()
+}
+
+func writeStable(sb *strings.Builder, v any) {
+	switch x := v.(type) {
+	case nil:
+		sb.WriteString("null")
+	case bool:
+		if x {
+			sb.WriteString("true")
+		} else {
+			sb.WriteString("false")
+		}
+	case string:
+		sb.WriteByte('"')
+		sb.WriteString(x)
+		sb.WriteByte('"')
+	case float64:
+		// Use %g for compact, deterministic float rendering. NaN and Inf are
+		// not JSON-encodable (which is why we are in the fallback path) and
+		// render as "NaN" / "+Inf" / "-Inf" — distinct, stable strings.
+		fmt.Fprintf(sb, "%g", x)
+	case float32:
+		fmt.Fprintf(sb, "%g", float64(x))
+	case int:
+		fmt.Fprintf(sb, "%d", x)
+	case int64:
+		fmt.Fprintf(sb, "%d", x)
+	case int32:
+		fmt.Fprintf(sb, "%d", x)
+	case uint:
+		fmt.Fprintf(sb, "%d", x)
+	case uint64:
+		fmt.Fprintf(sb, "%d", x)
+	case uint32:
+		fmt.Fprintf(sb, "%d", x)
+	case map[string]any:
+		sb.WriteByte('{')
+		keys := make([]string, 0, len(x))
+		for k := range x {
+			keys = append(keys, k)
+		}
+		sort.Strings(keys)
+		for i, k := range keys {
+			if i > 0 {
+				sb.WriteByte(',')
+			}
+			sb.WriteByte('"')
+			sb.WriteString(k)
+			sb.WriteString(`":`)
+			writeStable(sb, x[k])
+		}
+		sb.WriteByte('}')
+	case []any:
+		sb.WriteByte('[')
+		for i, item := range x {
+			if i > 0 {
+				sb.WriteByte(',')
+			}
+			writeStable(sb, item)
+		}
+		sb.WriteByte(']')
+	default:
+		// Last-resort: include the type so distinct values produce distinct
+		// strings even if their default format collides. Without the type
+		// tag, fmt.Sprintf("%v", x) for two different types could produce
+		// the same bytes (rare, but the type prefix makes it impossible).
+		fmt.Fprintf(sb, "<%T:%v>", x, x)
+	}
+}
+
+// ComputePrefixFingerprint returns a trace.PrefixHash (a 7-field fingerprint
+// of the prompt prefix) for one turn of a run. The seven sub-hashes are
+// independent SHA-256s of the corresponding sub-component; the complete-prefix
+// hash is a SHA-256 of the canonical concatenation of the other six, so any
+// sub-component drift is observable both individually and in aggregate.
+//
+// I/O side effects: buildPromptSubstrings calls workspaceContext, which runs
+// git (gitBranchForPrompt, FindProjectGitRoot) and reads the AGENTS.md /
+// ZERO.md / .zero/AGENTS.md chain plus the repo map. The same workspace-
+// context work is performed separately by buildSystemPrompt a few lines
+// later in the request-build path, so on every traced turn the file and
+// git reads happen twice. The duplication is intentional for this PR (a
+// follow-up will pass substrings through) and is a known perf cost only on
+// the opt-in trace path.
+//
+// A stable CompletePrefixHash across turns means the four captured
+// sub-components (baseInstructions, confirmationPolicy, projectContext,
+// skills) are byte-identical. It does NOT rule out drift in the seven
+// uncaptured sections of buildSystemPrompt (modelPromptAddendum,
+// sessionRuntimeContext, approvedCommandPrefixContext, workspaceSeedContext,
+// userGuidelines, specialistDelegationContext, responseStyleContext). For
+// default Options the four captured substrings are the full prompt; for
+// non-default Options the seven uncaptured sections may contribute, and a
+// consumer correlating CompletePrefix stability with cached_input_tokens
+// must cross-check the model_switches counter (modelPromptAddendum changes
+// on a model switch) to disambiguate "no drift" from "drift in an
+// uncaptured section."
+
+func ComputePrefixFingerprint(options Options, exposed []zeroruntime.ToolDefinition) prefixFingerprint {
+	subs := buildPromptSubstrings(options, exposed)
+	base := sha256hex(subs.baseInstructions)
+	policy := sha256hex(subs.confirmationPolicy)
+	project := sha256hex(subs.projectContext)
+	skills := sha256hex(subs.skills)
+	toolsH := sha256hex(subs.tools)
+	schema := sha256hex(subs.schema)
+	complete := sha256hex(strings.Join([]string{
+		base, policy, project, skills, toolsH, schema,
+	}, "|"))
+	return prefixFingerprint{
+		BaseInstructionsHash:   base,
+		ConfirmationPolicyHash: policy,
+		ProjectContextHash:     project,
+		SkillsHash:             skills,
+		ToolsHash:              toolsH,
+		SchemaHash:             schema,
+		CompletePrefixHash:     complete,
+	}
+}
+
+// prefixFingerprint is the agent-side shape of a prompt-prefix fingerprint. It
+// is converted to a trace.PrefixHash at the loop boundary (see EmitPrefixHash
+// in loop.go) so the trace package does not need to import this one. The field
+// names match the trace.PrefixHash JSON tags 1:1.
+type prefixFingerprint struct {
+	BaseInstructionsHash   string
+	ConfirmationPolicyHash string
+	ProjectContextHash     string
+	SkillsHash             string
+	ToolsHash              string
+	SchemaHash             string
+	CompletePrefixHash     string
+}
+
+// sha256hex returns the hex-encoded SHA-256 of s. Empty input produces the
+// hash of the empty string, which is a constant; callers that want to
+// distinguish "absent" from "empty" should check s before calling.
+func sha256hex(s string) string {
+	sum := sha256.Sum256([]byte(s))
+	return hex.EncodeToString(sum[:])
+}
@@ -0,0 +1,243 @@
+package agent
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// TestComputePrefixFingerprintStableAcrossCalls asserts the headline property
+// the trace is meant to expose: two ComputePrefixFingerprint calls over the
+// same Options and the same exposed tool list produce byte-identical hashes.
+// This is the regression catch for "did anyone introduce non-determinism into
+// the prompt prefix."
+func TestComputePrefixFingerprintStableAcrossCalls(t *testing.T) {
+	opts := Options{Cwd: t.TempDir(), SystemPrompt: "core"}
+	exposed := []zeroruntime.ToolDefinition{
+		{Name: "read_file", Parameters: map[string]any{"type": "object"}},
+		{Name: "grep", Parameters: map[string]any{"type": "object"}},
+	}
+	first := ComputePrefixFingerprint(opts, exposed)
+	second := ComputePrefixFingerprint(opts, exposed)
+	if first != second {
+		t.Fatalf("fingerprint must be stable across calls with the same input:\n  first=%+v\n second=%+v", first, second)
+	}
+	if first.CompletePrefixHash == "" {
+		t.Fatalf("CompletePrefixHash must be non-empty for a non-trivial prompt")
+	}
+}
+
+// TestComputePrefixFingerprintToolsHashIsNameSetSensitive asserts the
+// name-set sensitivity of the ToolsHash. The test name used to read as
+// "reorder changes hash" but the assertion is the opposite: the hash is
+// name-set sensitive, not name-order sensitive (toolSubstrings sorts
+// names before hashing). The partitioner is expected to produce a stable
+// order, so a name-order change is not a hash signal — a name-set change
+// is. The test asserts both: permutation of the same tools produces the
+// same hash, and adding a tool produces a different hash.
+func TestComputePrefixFingerprintToolsHashIsNameSetSensitive(t *testing.T) {
+	opts := Options{Cwd: t.TempDir(), SystemPrompt: "core"}
+	a := []zeroruntime.ToolDefinition{
+		{Name: "read_file", Parameters: map[string]any{"schema": "schema-read"}},
+		{Name: "grep", Parameters: map[string]any{"schema": "schema-grep"}},
+	}
+	b := []zeroruntime.ToolDefinition{
+		{Name: "grep", Parameters: map[string]any{"schema": "schema-grep"}},
+		{Name: "read_file", Parameters: map[string]any{"schema": "schema-read"}},
+	}
+	// toolSubstrings sorts names internally, so the ToolsHash is identical
+	// for a permutation of the same tools. We assert that here to document
+	// the contract: name-order independence, name-set sensitivity.
+	fpA := ComputePrefixFingerprint(opts, a)
+	fpB := ComputePrefixFingerprint(opts, b)
+	if fpA.ToolsHash != fpB.ToolsHash {
+		t.Fatalf("ToolsHash must be name-set sensitive, not name-order sensitive: a=%s b=%s", fpA.ToolsHash, fpB.ToolsHash)
+	}
+	// A different tool set must produce a different ToolsHash.
+	c := append([]zeroruntime.ToolDefinition{}, a...)
+	c = append(c, zeroruntime.ToolDefinition{Name: "bash", Parameters: map[string]any{"schema": "schema-bash"}})
+	fpC := ComputePrefixFingerprint(opts, c)
+	if fpA.ToolsHash == fpC.ToolsHash {
+		t.Fatalf("ToolsHash must change when the tool set changes: a=%s c=%s", fpA.ToolsHash, fpC.ToolsHash)
+	}
+}
+
+// TestComputePrefixFingerprintSchemaChangeChangesSchemaHash asserts that a
+// change to a tool's Parameters schema is visible in the SchemaHash. A schema
+// edit that should be cache-stable (e.g. field reorder, doc tweak) will move
+// the hash and surface in the trace; a schema edit that legitimately
+// invalidates the cache also moves the hash, which is the right behavior.
+func TestComputePrefixFingerprintSchemaChangeChangesSchemaHash(t *testing.T) {
+	opts := Options{Cwd: t.TempDir(), SystemPrompt: "core"}
+	a := []zeroruntime.ToolDefinition{
+		{Name: "read_file", Parameters: map[string]any{"version": "v1"}},
+	}
+	b := []zeroruntime.ToolDefinition{
+		{Name: "read_file", Parameters: map[string]any{"version": "v2"}},
+	}
+	fpA := ComputePrefixFingerprint(opts, a)
+	fpB := ComputePrefixFingerprint(opts, b)
+	if fpA.SchemaHash == fpB.SchemaHash {
+		t.Fatalf("SchemaHash must change when tool Parameters change: a=%s b=%s", fpA.SchemaHash, fpB.SchemaHash)
+	}
+}
+
+// TestComputePrefixFingerprintSystemPromptChangeChangesBaseHash asserts the
+// most common drift path: a system-prompt edit (e.g. an agent.md update, a
+// model addendum change) must move the base-instructions hash, which moves
+// the complete-prefix hash, which is what the trace is for.
+func TestComputePrefixFingerprintSystemPromptChangeChangesBaseHash(t *testing.T) {
+	optsA := Options{Cwd: t.TempDir(), SystemPrompt: "core v1"}
+	optsB := Options{Cwd: t.TempDir(), SystemPrompt: "core v2"}
+	exposed := []zeroruntime.ToolDefinition{{Name: "noop", Parameters: map[string]any{}}}
+	fpA := ComputePrefixFingerprint(optsA, exposed)
+	fpB := ComputePrefixFingerprint(optsB, exposed)
+	if fpA.BaseInstructionsHash == fpB.BaseInstructionsHash {
+		t.Fatalf("BaseInstructionsHash must change when SystemPrompt changes: a=%s b=%s", fpA.BaseInstructionsHash, fpB.BaseInstructionsHash)
+	}
+	if fpA.CompletePrefixHash == fpB.CompletePrefixHash {
+		t.Fatalf("CompletePrefixHash must change when any sub-hash changes: a=%s b=%s", fpA.CompletePrefixHash, fpB.CompletePrefixHash)
+	}
+}
+
+// TestComputePrefixFingerprintCompleteIsAggregateOfSubHashes asserts the
+// canonical-join property: CompletePrefixHash must depend on every sub-hash.
+// This is the regression catch for "did anyone reorder or drop a sub-hash
+// from the canonical join without updating CompletePrefixHash."
+func TestComputePrefixFingerprintCompleteIsAggregateOfSubHashes(t *testing.T) {
+	opts := Options{Cwd: t.TempDir(), SystemPrompt: "core"}
+	exposed := []zeroruntime.ToolDefinition{{Name: "read_file", Parameters: map[string]any{"k": "v"}}}
+	fp := ComputePrefixFingerprint(opts, exposed)
+	// Manually compute what the canonical join should be.
+	expected := sha256hex(strings.Join([]string{
+		fp.BaseInstructionsHash,
+		fp.ConfirmationPolicyHash,
+		fp.ProjectContextHash,
+		fp.SkillsHash,
+		fp.ToolsHash,
+		fp.SchemaHash,
+	}, "|"))
+	if fp.CompletePrefixHash != expected {
+		t.Fatalf("CompletePrefixHash must equal sha256hex of the canonical join of the other six:\n  got:      %s\n  expected: %s", fp.CompletePrefixHash, expected)
+	}
+}
+
+// TestBuildPromptSubstringsDefaultOptions asserts the invariants of the
+// substrings helper for default Options: which substrings are non-empty,
+// which are empty, and that the substring-to-hash round-trip is lossless.
+func TestBuildPromptSubstringsDefaultOptions(t *testing.T) {
+	opts := Options{
+		Cwd:          t.TempDir(),
+		SystemPrompt: "test core",
+	}
+	subs := buildPromptSubstrings(opts, nil)
+	// Invariants the trace depends on (default Options):
+	//   1. baseInstructions substring equals the core system prompt bytes.
+	//   2. confirmationPolicy substring is non-empty (the embedded policy
+	//      is always present, post TrimSpace).
+	//   3. skills, tools, and schema substrings are empty (no skills or
+	//      tools configured for default Options).
+	if subs.baseInstructions != "test core" {
+		t.Fatalf("baseInstructions substring must equal the core system prompt: got %q want %q", subs.baseInstructions, "test core")
+	}
+	if subs.confirmationPolicy == "" {
+		t.Fatalf("confirmationPolicy substring must be non-empty (the embedded policy is always present)")
+	}
+	if subs.skills != "" {
+		t.Fatalf("skills substring must be empty for default Options: got %q", subs.skills)
+	}
+	// Round-trip: the corresponding fingerprint hashes must equal
+	// sha256hex of the substrings, with no truncation or reformatting
+	// between the two layers.
+	fp := ComputePrefixFingerprint(opts, nil)
+	if fp.BaseInstructionsHash != sha256hex(subs.baseInstructions) {
+		t.Fatalf("BaseInstructionsHash in fingerprint must equal sha256hex of the substring")
+	}
+	if fp.ConfirmationPolicyHash != sha256hex(subs.confirmationPolicy) {
+		t.Fatalf("ConfirmationPolicyHash in fingerprint must equal sha256hex of the substring")
+	}
+}
+
+// TestCanonicalSchemaStringStableForMapParameters asserts the headline
+// property the trace needs: the same map[string]any produces the same
+// canonical string across calls, even though Go's fmt.Sprintf("%v", m)
+// iterates a map in random order. json.Marshal sorts map keys
+// alphabetically, which is the fix.
+func TestCanonicalSchemaStringStableForMapParameters(t *testing.T) {
+	params := map[string]any{
+		"description": "read a file",
+		"properties": map[string]any{
+			"path": map[string]any{"type": "string"},
+		},
+		"required": []any{"path"},
+		"type":     "object",
+	}
+	first := canonicalSchemaString(params)
+	second := canonicalSchemaString(params)
+	if first != second {
+		t.Fatalf("canonicalSchemaString must be stable across calls for the same Parameters:\n  first=%s\n second=%s", first, second)
+	}
+	if first == "" {
+		t.Fatal("canonicalSchemaString must produce a non-empty string for a non-empty map")
+	}
+	// Two maps with the same key/value pairs (different declaration
+	// order in source) must produce identical canonical strings. This
+	// is the property that defeats Go's map iteration randomization.
+	reordered := map[string]any{
+		"type":     "object",
+		"required": []any{"path"},
+		"properties": map[string]any{
+			"path": map[string]any{"type": "string"},
+		},
+		"description": "read a file",
+	}
+	if canonicalSchemaString(params) != canonicalSchemaString(reordered) {
+		t.Fatal("canonicalSchemaString must produce identical output for maps with the same key/value pairs in different declaration orders")
+	}
+}
+
+// TestCanonicalSchemaStringFallbackStableForNonJSONValue exercises the
+// non-JSON fallback in canonicalSchemaString: when Parameters contains a
+// value json.Marshal cannot encode (here, a channel), the function must
+// fall back to a key-sorted stringification that is stable across calls
+// and across map-iteration order. The previous fallback used
+// fmt.Sprintf("%v", params), which iterates the map in random order and
+// produced a different hash for the same value across calls — defeating
+// the fingerprint. The current fallback is stable by construction.
+func TestCanonicalSchemaStringFallbackStableForNonJSONValue(t *testing.T) {
+	ch := make(chan int) // channels are not JSON-encodable
+	params := map[string]any{
+		"description": "a tool with a non-JSON value",
+		"channel":     ch,
+		"z_first":     1,
+		"a_second":    2,
+	}
+	first := canonicalSchemaString(params)
+	second := canonicalSchemaString(params)
+	if first == "" {
+		t.Fatal("canonicalSchemaString must produce a non-empty string for the fallback path")
+	}
+	if first != second {
+		t.Fatalf("canonicalSchemaString fallback must be stable across calls:\n  first=%s\n second=%s", first, second)
+	}
+	// A re-ordered map (same key/value pairs, different declaration order)
+	// must produce the same fallback string. This is the property the
+	// key-sorted stringifier provides that fmt.Sprintf("%v", m) does not.
+	reordered := map[string]any{
+		"a_second":    2,
+		"z_first":     1,
+		"channel":     ch,
+		"description": "a tool with a non-JSON value",
+	}
+	if canonicalSchemaString(params) != canonicalSchemaString(reordered) {
+		t.Fatalf("canonicalSchemaString fallback must produce identical output for maps with the same key/value pairs in different declaration orders:\n  first=%s\n second=%s", first, canonicalSchemaString(reordered))
+	}
+	// The fallback must include the "__non_json:" prefix so a consumer
+	// can tell the bytes are not a JSON-marshaled schema (a hash collision
+	// between a json.Marshal result and a stableStringify result is
+	// possible in theory; the prefix makes the source observable).
+	if !strings.HasPrefix(first, "__non_json:") {
+		t.Fatalf("canonicalSchemaString fallback must start with the __non_json: prefix, got: %s", first)
+	}
+}
@@ -8,6 +8,7 @@ import (
 	"time"

 	"github.com/Gitlawb/zero/internal/errhint"
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )

@@ -74,14 +75,21 @@ func stallRetryNoticeFor(options Options) reconnectNotifier {
 // already past its deadline is returned immediately (no retry) — those have
 // their own handling (compaction for context-limit, image-rejection, etc.).
 func streamWithReconnect(ctx context.Context, provider Provider, request zeroruntime.CompletionRequest, notify reconnectNotifier) (<-chan zeroruntime.StreamEvent, error) {
+	recorder := trace.FromContext(ctx)
 	stream, err := provider.StreamCompletion(ctx, request)
 	if err == nil {
+		if recorder != nil {
+			recorder.Counter(trace.CounterModelRequests, 1)
+		}
 		return stream, nil
 	}
 	for attempt := 1; attempt <= maxStreamReconnects; attempt++ {
 		if !shouldReconnect(ctx, err) {
 			return nil, err
 		}
+		if recorder != nil {
+			recorder.Counter(trace.CounterReconnectCount, 1)
+		}
 		if notify != nil {
 			notify(attempt, maxStreamReconnects)
 		}
@@ -90,6 +98,9 @@ func streamWithReconnect(ctx context.Context, provider Provider, request zerorun
 		}
 		stream, err = provider.StreamCompletion(ctx, request)
 		if err == nil {
+			if recorder != nil {
+				recorder.Counter(trace.CounterModelRequests, 1)
+			}
 			return stream, nil
 		}
 	}
@@ -9,6 +9,7 @@ import (

 	"github.com/Gitlawb/zero/internal/lsp"
 	"github.com/Gitlawb/zero/internal/redaction"
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/verify"
 )

@@ -133,7 +134,9 @@ func (sc *SelfCorrector) inspect(ctx context.Context, changedFiles []string) Cor
 	}

 	if sc.cfg.IncludeTests && sc.verifier != nil {
+		verifySpan := trace.FromContext(ctx).Span(trace.SpanVerification)
 		vr, err := sc.verifier.Verify(ctx)
+		verifySpan.End()
 		if err != nil {
 			// "No plan / no tests" is not an error (DetectPlan returns an empty
 			// plan), so a non-nil error means verification could not run at all —
@@ -7,6 +7,7 @@ import (
 	"github.com/Gitlawb/zero/internal/sandbox"
 	"github.com/Gitlawb/zero/internal/streamjson"
 	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )

@@ -279,6 +280,16 @@ type Options struct {
 	// request), so every existing caller is unaffected. A returned error is
 	// non-fatal: the run continues on the current model.
 	ModelSwitcher func(ctx context.Context, modelID string) (Provider, error)
+	// Trace, when set, records per-turn timing for the run: the loop stamps
+	// spans (prompt build, generation, tool execution, permission wait,
+	// compaction, provider connect) and counters (model requests, tool calls,
+	// retries, tokens) into it. nil DISABLES tracing entirely — every stamp is
+	// nil-safe and the loop is byte-identical to an untraced run. The caller
+	// owns the recorder: Run stamps into it but does not Finish or emit it.
+	// A fresh Recorder is required per Run — reusing one across runs merges
+	// their spans, counters, and first-event timestamps, and Finish freezes a
+	// recorder so no further stamps take.
+	Trace *trace.Recorder
 	// SelfCorrect, when set, runs a post-edit verify-and-correct cycle after a
 	// mutating tool call: it verifies the changed files (LSP diagnostics + project
 	// tests) and feeds failures back to the model to fix, bounded by an attempt
@@ -0,0 +1,259 @@
+package aimlapi
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"net/url"
+	"strings"
+	"time"
+)
+
+// PaymentMethod selects how a partner-checkout top-up is paid for.
+type PaymentMethod string
+
+const (
+	PaymentMethodCard   PaymentMethod = "card"
+	PaymentMethodCrypto PaymentMethod = "crypto"
+)
+
+// SessionStatus is the lifecycle state of a partner-checkout session, as reported
+// by the aimlapi.com backend while polling for payment.
+type SessionStatus string
+
+const (
+	SessionStatusPendingAuth    SessionStatus = "pending_auth"
+	SessionStatusPendingPayment SessionStatus = "pending_payment"
+	SessionStatusPaid           SessionStatus = "paid"
+	SessionStatusExchanging     SessionStatus = "exchanging"
+	SessionStatusExchanged      SessionStatus = "exchanged"
+	SessionStatusCancelled      SessionStatus = "cancelled"
+	SessionStatusExpired        SessionStatus = "expired"
+	SessionStatusFailed         SessionStatus = "failed"
+)
+
+// AuthResult is the bearer token minted by a passwordless sign-in / sign-up.
+type AuthResult struct {
+	Token string `json:"token"`
+	Exp   int64  `json:"exp"`
+}
+
+// PartnerCheckoutSession is a partner-attributed top-up session; its SessionToken
+// addresses it on later pay/exchange/poll calls.
+type PartnerCheckoutSession struct {
+	ID             string        `json:"id"`
+	SessionToken   string        `json:"sessionToken"`
+	PartnerID      string        `json:"partnerId"`
+	PartnerName    *string       `json:"partnerName"`
+	UserID         *int64        `json:"userId"`
+	AmountUSDMinor *int64        `json:"amountUsdMinor"`
+	Status         SessionStatus `json:"status"`
+	IssuedKeyID    *string       `json:"issuedKeyId"`
+	ReturnURL      *string       `json:"returnUrl"`
+}
+
+// PaymentSession is the hosted payment-provider checkout; PayURL is the page the
+// user opens to pay.
+type PaymentSession struct {
+	ProviderSessionID string `json:"providerSessionId"`
+	PayURL            string `json:"payUrl"`
+}
+
+// PayResult pairs the hosted checkout with the updated partner-checkout session
+// returned when a top-up payment is initiated.
+type PayResult struct {
+	Checkout        PaymentSession         `json:"checkout"`
+	PartnerCheckout PartnerCheckoutSession `json:"partnerCheckout"`
+}
+
+// ExchangeResult holds the API key minted when a paid session is exchanged.
+type ExchangeResult struct {
+	APIKey   string `json:"apiKey"`
+	APIKeyID string `json:"apiKeyId"`
+}
+
+// TopUpByKeyRequest funds a partner-checkout session with the API key that owns
+// the account (Path A / env key), instead of an email session.
+type TopUpByKeyRequest struct {
+	SessionToken     string
+	AmountUSDMinor   int
+	PaymentSessionID string // idempotency: same id → same checkout, never a second charge
+	AutoTopUp        bool   // enroll the account in automatic top-up at checkout (card, saved off-session)
+	SuccessURL       string
+	CancelURL        string
+}
+
+// TopUpByKeyResult is the hosted checkout returned by TopUpByKey.
+type TopUpByKeyResult struct {
+	Checkout PaymentSession `json:"checkout"`
+}
+
+// APIError is a non-2xx HTTP response from the aimlapi.com API. Status carries the
+// code so callers can branch (e.g. 401 = bad key, 5xx = retryable).
+type APIError struct {
+	Message string
+	Status  int
+	Body    string
+}
+
+// Error renders the method, endpoint, status, and (when present) the response body.
+func (e APIError) Error() string {
+	if strings.TrimSpace(e.Body) != "" {
+		return fmt.Sprintf("%s: HTTP %d: %s", e.Message, e.Status, strings.TrimSpace(e.Body))
+	}
+	return fmt.Sprintf("%s: HTTP %d", e.Message, e.Status)
+}
+
+// Client talks to the aimlapi.com auth, app, and inference APIs for the onboarding
+// and partner-checkout top-up flows.
+type Client struct {
+	endpoints  Endpoints
+	httpClient *http.Client
+}
+
+func NewClient(endpoints Endpoints, httpClient *http.Client) *Client {
+	if httpClient == nil {
+		httpClient = &http.Client{Timeout: 60 * time.Second}
+	}
+	return &Client{endpoints: endpoints, httpClient: httpClient}
+}
+
+// CreateSession opens a partner-checkout session attributed to partnerID; the
+// returned SessionToken addresses it on the pay/exchange/poll calls.
+func (c *Client) CreateSession(ctx context.Context, partnerID string, partnerName string, returnURL string) (PartnerCheckoutSession, error) {
+	body := map[string]any{"partnerId": partnerID}
+	if strings.TrimSpace(partnerName) != "" {
+		body["partnerName"] = strings.TrimSpace(partnerName)
+	}
+	if strings.TrimSpace(returnURL) != "" {
+		body["returnUrl"] = strings.TrimSpace(returnURL)
+	}
+	var result PartnerCheckoutSession
+	err := c.request(ctx, http.MethodPost, strings.TrimRight(c.endpoints.AppBaseURL, "/")+"/v3/partner-checkout/sessions", "", body, &result)
+	return result, err
+}
+
+// GetSession fetches the current state of a partner-checkout session; used to poll
+// for payment completion.
+func (c *Client) GetSession(ctx context.Context, sessionToken string) (PartnerCheckoutSession, error) {
+	var result PartnerCheckoutSession
+	err := c.request(ctx, http.MethodGet, strings.TrimRight(c.endpoints.AppBaseURL, "/")+"/v3/partner-checkout/sessions/"+urlPathEscape(sessionToken), "", nil, &result)
+	return result, err
+}
+
+// Pay initiates payment for a session and returns the hosted checkout URL.
+// autoTopUp enrolls the account in automatic top-up at checkout time.
+func (c *Client) Pay(ctx context.Context, bearer string, sessionToken string, amountUSDMinor int, paymentSessionID string, method PaymentMethod, successURL string, cancelURL string, autoTopUp bool) (PayResult, error) {
+	body := map[string]any{
+		"amountUsdMinor":   amountUSDMinor,
+		"paymentSessionId": paymentSessionID,
+		"method":           method,
+	}
+	if strings.TrimSpace(successURL) != "" {
+		body["successUrl"] = strings.TrimSpace(successURL)
+	}
+	if strings.TrimSpace(cancelURL) != "" {
+		body["cancelUrl"] = strings.TrimSpace(cancelURL)
+	}
+	// Only sent when enabled: enrolls the account in automatic top-up. The backend
+	// honours this field at checkout time.
+	if autoTopUp {
+		body["autoTopUp"] = true
+	}
+	var result PayResult
+	err := c.request(ctx, http.MethodPost, strings.TrimRight(c.endpoints.AppBaseURL, "/")+"/v3/partner-checkout/sessions/"+urlPathEscape(sessionToken)+"/pay", bearer, body, &result)
+	return result, err
+}
+
+// Exchange converts a paid session into a freshly minted API key (new-account flow).
+func (c *Client) Exchange(ctx context.Context, bearer string, sessionToken string) (ExchangeResult, error) {
+	var result ExchangeResult
+	err := c.request(ctx, http.MethodPost, strings.TrimRight(c.endpoints.AppBaseURL, "/")+"/v3/partner-checkout/sessions/"+urlPathEscape(sessionToken)+"/exchange", bearer, nil, &result)
+	return result, err
+}
+
+// TopUpByKey funds a partner-checkout session using the raw API key that owns the
+// account (Path A / env key), returning the hosted checkout. It binds the session
+// to the key's account server-side, so no email session is needed and no key is
+// exchanged. req.PaymentSessionID makes a retry idempotent (same id → same
+// checkout, never a second charge).
+func (c *Client) TopUpByKey(ctx context.Context, apiKey string, req TopUpByKeyRequest) (TopUpByKeyResult, error) {
+	body := map[string]any{
+		"sessionToken":     req.SessionToken,
+		"amountUsdMinor":   req.AmountUSDMinor,
+		"paymentSessionId": req.PaymentSessionID,
+	}
+	if req.AutoTopUp {
+		body["autoTopUp"] = true
+	}
+	if strings.TrimSpace(req.SuccessURL) != "" {
+		body["successUrl"] = strings.TrimSpace(req.SuccessURL)
+	}
+	if strings.TrimSpace(req.CancelURL) != "" {
+		body["cancelUrl"] = strings.TrimSpace(req.CancelURL)
+	}
+	var result TopUpByKeyResult
+	err := c.request(ctx, http.MethodPost, c.billingV2URL("/billing/topup"), apiKey, body, &result)
+	return result, err
+}
+
+// billingV2URL builds a v2 billing API URL from the inference base. The top-up
+// endpoint lives next to the balance endpoint on the API gateway, one version up
+// (".../v1" → ".../v2/billing/..."), so it re-anchors the version while keeping any
+// host/prefix an override supplied.
+func (c *Client) billingV2URL(path string) string {
+	base := strings.TrimRight(strings.TrimSpace(c.endpoints.InferenceBaseURL), "/")
+	base = strings.TrimSuffix(base, "/v1")
+	return strings.TrimRight(base, "/") + "/v2" + path
+}
+
+func (c *Client) request(ctx context.Context, method string, endpoint string, bearer string, body any, out any) error {
+	var reader io.Reader
+	if body != nil {
+		data, err := json.Marshal(body)
+		if err != nil {
+			return err
+		}
+		reader = bytes.NewReader(data)
+	}
+	request, err := http.NewRequestWithContext(ctx, method, endpoint, reader)
+	if err != nil {
+		return err
+	}
+	request.Header.Set("Accept", "application/json")
+	if body != nil {
+		request.Header.Set("Content-Type", "application/json")
+	}
+	if strings.TrimSpace(bearer) != "" {
+		request.Header.Set("Authorization", "Bearer "+strings.TrimSpace(bearer))
+	}
+	response, err := c.httpClient.Do(request)
+	if err != nil {
+		return fmt.Errorf("network request to %s failed: %w", endpoint, err)
+	}
+	defer response.Body.Close()
+	text, readErr := io.ReadAll(io.LimitReader(response.Body, 1<<20))
+	if readErr != nil {
+		return readErr
+	}
+	if response.StatusCode < 200 || response.StatusCode >= 300 {
+		return APIError{Message: method + " " + endpoint, Status: response.StatusCode, Body: string(text)}
+	}
+	if out == nil {
+		return nil
+	}
+	if len(bytes.TrimSpace(text)) == 0 {
+		return APIError{Message: method + " " + endpoint + " returned empty body", Status: response.StatusCode}
+	}
+	if err := json.Unmarshal(text, out); err != nil {
+		return APIError{Message: method + " " + endpoint + " returned non-JSON body", Status: response.StatusCode, Body: string(text)}
+	}
+	return nil
+}
+
+func urlPathEscape(value string) string {
+	return url.PathEscape(value)
+}
@@ -0,0 +1,106 @@
+package aimlapi
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+	"strings"
+)
+
+// Additional client methods backing the interactive TUI onboarding flow. They
+// target the endpoints shipped in the aimlapi backend:
+//   - GET  {inference}/billing/balance            (Path A: validate key + balance)
+//   - PATCH{auth}/v1/auth/account                 (Path B: does the account exist?)
+//   - POST {auth}/v1/auth/sign-in/code(/verify)   (Path B existing: passwordless login)
+//   - POST {auth}/v1/auth/account/passwordless    (Path B new: email-only signup)
+//   - POST {app}/v1/keys                          (mint a key once we hold a session)
+
+// BalanceResult is the wallet balance returned by GetBalance, with the backend's
+// low-balance flag and threshold.
+type BalanceResult struct {
+	Balance             float64 `json:"balance"`
+	LowBalance          bool    `json:"lowBalance"`
+	LowBalanceThreshold float64 `json:"lowBalanceThreshold"`
+}
+
+// CheckResult reports whether an email already has an account, driving the
+// sign-in vs sign-up branch of the onboarding flow.
+type CheckResult struct {
+	// "sign-in" (account exists) or "sign-up" (no account yet).
+	Action   string  `json:"action"`
+	Provider *string `json:"provider"`
+}
+
+// CreatedKey is an API key minted by CreateKey.
+type CreatedKey struct {
+	Key string `json:"key"`
+	ID  string `json:"id"`
+}
+
+func (c *Client) authURL(path string) string {
+	return strings.TrimRight(c.endpoints.AuthBaseURL, "/") + path
+}
+
+func (c *Client) appURL(path string) string {
+	return strings.TrimRight(c.endpoints.AppBaseURL, "/") + path
+}
+
+func (c *Client) inferenceURL(path string) string {
+	return strings.TrimRight(c.endpoints.InferenceBaseURL, "/") + path
+}
+
+// GetBalance doubles as key validation: a 401 means the key is invalid, a 2xx
+// returns the wallet balance. Auth is the raw API key itself.
+func (c *Client) GetBalance(ctx context.Context, apiKey string) (BalanceResult, error) {
+	var result BalanceResult
+	err := c.request(ctx, http.MethodGet, c.inferenceURL("/billing/balance"), apiKey, nil, &result)
+	return result, err
+}
+
+// CheckAccount reports whether an email already has an aimlapi.com account
+// (Action "sign-in") or not ("sign-up").
+func (c *Client) CheckAccount(ctx context.Context, email string) (CheckResult, error) {
+	var result CheckResult
+	err := c.request(ctx, http.MethodPatch, c.authURL("/v1/auth/account"), "", map[string]any{"email": email}, &result)
+	return result, err
+}
+
+// SendSignInCode emails a 6-digit sign-in code to an existing account.
+func (c *Client) SendSignInCode(ctx context.Context, email string) error {
+	return c.request(ctx, http.MethodPost, c.authURL("/v1/auth/sign-in/code"), "", map[string]any{"email": email}, nil)
+}
+
+// VerifySignInCode exchanges an emailed code for a session bearer token.
+func (c *Client) VerifySignInCode(ctx context.Context, email string, code string) (AuthResult, error) {
+	var result AuthResult
+	err := c.request(ctx, http.MethodPost, c.authURL("/v1/auth/sign-in/code/verify"), "", map[string]any{"email": email, "code": code}, &result)
+	if err == nil && strings.TrimSpace(result.Token) == "" {
+		err = fmt.Errorf("aimlapi.com did not return an auth token")
+	}
+	return result, err
+}
+
+// CreatePasswordlessAccount registers a new email-only account and returns its
+// session bearer token.
+func (c *Client) CreatePasswordlessAccount(ctx context.Context, email string) (AuthResult, error) {
+	var result AuthResult
+	err := c.request(ctx, http.MethodPost, c.authURL("/v1/auth/account/passwordless"), "", map[string]any{"email": email}, &result)
+	if err == nil && strings.TrimSpace(result.Token) == "" {
+		err = fmt.Errorf("aimlapi.com did not return an auth token")
+	}
+	return result, err
+}
+
+// CreateKey mints a fresh API key for a session-holding user.
+func (c *Client) CreateKey(ctx context.Context, bearer string, name string) (CreatedKey, error) {
+	body := map[string]any{}
+	if strings.TrimSpace(name) != "" {
+		body["name"] = strings.TrimSpace(name)
+	}
+	var result CreatedKey
+	err := c.request(ctx, http.MethodPost, c.appURL("/v1/keys"), bearer, body, &result)
+	if err == nil && strings.TrimSpace(result.Key) == "" {
+		err = fmt.Errorf("aimlapi.com did not return an API key")
+	}
+	return result, err
+}
@@ -0,0 +1,159 @@
+package aimlapi
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+)
+
+func TestCreateKeySendsAuthenticatedRequestAndDecodesKey(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.Method != http.MethodPost || r.URL.Path != "/v1/keys" {
+			t.Fatalf("request = %s %s, want POST /v1/keys", r.Method, r.URL.Path)
+		}
+		if got := r.Header.Get("Authorization"); got != "Bearer session-token" {
+			t.Fatalf("Authorization = %q", got)
+		}
+		var body map[string]string
+		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
+			t.Fatal(err)
+		}
+		if body["name"] != "Zero CLI" {
+			t.Fatalf("name = %q", body["name"])
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"key":"sk-issued","id":"key-id"}`))
+	}))
+	defer server.Close()
+
+	client := NewClient(Endpoints{AppBaseURL: server.URL}, server.Client())
+	key, err := client.CreateKey(context.Background(), " session-token ", " Zero CLI ")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if key.Key != "sk-issued" || key.ID != "key-id" {
+		t.Fatalf("key = %#v", key)
+	}
+}
+
+func TestCreateKeyRejectsEmptyKeyResponse(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"id":"key-id"}`))
+	}))
+	defer server.Close()
+
+	client := NewClient(Endpoints{AppBaseURL: server.URL}, server.Client())
+	if _, err := client.CreateKey(context.Background(), "session-token", "Zero"); err == nil {
+		t.Fatal("CreateKey accepted a response without an API key")
+	}
+}
+
+func TestAuthMethodsRejectEmptyTokenResponses(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"token":""}`))
+	}))
+	defer server.Close()
+
+	client := NewClient(Endpoints{AuthBaseURL: server.URL}, server.Client())
+	tests := []struct {
+		name string
+		call func() error
+	}{
+		{
+			name: "verify sign-in code",
+			call: func() error {
+				_, err := client.VerifySignInCode(context.Background(), "[email protected]", "123456")
+				return err
+			},
+		},
+		{
+			name: "create passwordless account",
+			call: func() error {
+				_, err := client.CreatePasswordlessAccount(context.Background(), "[email protected]")
+				return err
+			},
+		},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			if err := test.call(); err == nil {
+				t.Fatal("accepted a response without an auth token")
+			}
+		})
+	}
+}
+
+func TestClientReturnsTypedAPIError(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		http.Error(w, "invalid key", http.StatusUnauthorized)
+	}))
+	defer server.Close()
+
+	client := NewClient(Endpoints{InferenceBaseURL: server.URL}, server.Client())
+	_, err := client.GetBalance(context.Background(), "bad-key")
+	var apiErr APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("error = %T %v, want APIError", err, err)
+	}
+	if apiErr.Status != http.StatusUnauthorized || apiErr.Body != "invalid key\n" {
+		t.Fatalf("APIError = %#v", apiErr)
+	}
+}
+
+func TestTypedResponseRejectsEmptySuccessBody(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.WriteHeader(http.StatusNoContent)
+	}))
+	defer server.Close()
+
+	client := NewClient(Endpoints{InferenceBaseURL: server.URL}, server.Client())
+	if _, err := client.GetBalance(context.Background(), "key"); err == nil {
+		t.Fatal("GetBalance accepted an empty successful response")
+	}
+}
+
+func TestBodylessOperationAcceptsEmptySuccessBody(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.WriteHeader(http.StatusNoContent)
+	}))
+	defer server.Close()
+
+	client := NewClient(Endpoints{AuthBaseURL: server.URL}, server.Client())
+	if err := client.SendSignInCode(context.Background(), "[email protected]"); err != nil {
+		t.Fatalf("SendSignInCode() error = %v", err)
+	}
+}
+
+func TestParseAmountUSD(t *testing.T) {
+	tests := []struct {
+		name    string
+		value   string
+		want    int
+		wantErr bool
+	}{
+		{name: "default", want: DefaultAmountUSDMinor},
+		{name: "minimum", value: "20", want: 2000},
+		{name: "cents", value: "25.49", want: 2549},
+		{name: "below minimum", value: "19.99", wantErr: true},
+		{name: "above maximum", value: "10001", wantErr: true},
+		{name: "not a number", value: "twenty", wantErr: true},
+		{name: "nan", value: "NaN", wantErr: true},
+		{name: "infinity", value: "Inf", wantErr: true},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			got, err := ParseAmountUSD(test.value)
+			if (err != nil) != test.wantErr {
+				t.Fatalf("ParseAmountUSD(%q) error = %v", test.value, err)
+			}
+			if got != test.want {
+				t.Fatalf("ParseAmountUSD(%q) = %d, want %d", test.value, got, test.want)
+			}
+		})
+	}
+}
@@ -0,0 +1,113 @@
+package aimlapi
+
+import (
+	"net/url"
+	"os"
+	"strings"
+)
+
+type Endpoints struct {
+	AuthBaseURL      string
+	AppBaseURL       string
+	InferenceBaseURL string
+	PayBaseURL       string
+	// VerificationBaseURL is the aimlapi web-app frontend base. The browser is sent
+	// back here after the partner checkout completes (see BuildPartnerReturnURL), so
+	// it must point at the same environment's front as the one the CLI is talking to
+	// to land the user there. Empty falls back to DefaultReturnURL.
+	VerificationBaseURL string
+}
+
+const (
+	DefaultPartnerID   = "part_62yQoGYDq4Yqnrj2R1iGrDNJ"
+	DefaultPartnerName = "Gitlawb"
+	PartnerHeaderName  = "X-AIMLAPI-Partner-ID"
+	// DefaultReturnURL is the https fallback the browser is sent to after the
+	// checkout / OAuth consent finishes when no frontend base is known. It must be
+	// a real web page the browser can open — NOT a custom scheme like
+	// zero://aimlapi/complete, which fails with "the scheme does not have a
+	// registered handler" because the CLI registers no OS protocol handler. The
+	// CLI learns of success by polling; this URL is purely the browser's landing.
+	DefaultReturnURL = "https://aimlapi.com/app"
+	DefaultModel     = "anthropic/claude-sonnet-5"
+
+	MinAmountUSDMinor     = 2000
+	MaxAmountUSDMinor     = 1000000
+	DefaultAmountUSDMinor = 2500
+)
+
+func ResolveEndpoints() Endpoints {
+	return Endpoints{
+		AuthBaseURL:         envOrDefault("AIMLAPI_AUTH_URL", "https://auth.aimlapi.com"),
+		AppBaseURL:          envOrDefault("AIMLAPI_APP_URL", "https://app.aimlapi.com"),
+		InferenceBaseURL:    envOrDefault("AIMLAPI_INFERENCE_URL", "https://api.aimlapi.com/v1"),
+		PayBaseURL:          envOrDefault("AIMLAPI_PAY_URL", "https://pay.aimlapi.com"),
+		VerificationBaseURL: envOrDefault("AIMLAPI_VERIFICATION_BASE_URL", "https://aimlapi.com/app"),
+	}
+}
+
+// ResolvePartnerID applies the same explicit-option, environment, and default
+// precedence everywhere partner attribution is used. Keeping checkout creation
+// and saved inference profiles on one resolver prevents them from attributing
+// the same user to different partners.
+func ResolvePartnerID(explicit string) string {
+	if value := strings.TrimSpace(explicit); value != "" {
+		return value
+	}
+	return envOrDefault("AIMLAPI_PARTNER_ID", DefaultPartnerID)
+}
+
+// WithResolvedPartnerHeader returns a copy with the effective partner ID set.
+// Header matching is case-insensitive so an existing spelling is replaced rather
+// than duplicated. Callers remain responsible for applying catalog attribution
+// only to the canonical AIMLAPI endpoint.
+func WithResolvedPartnerHeader(headers map[string]string) map[string]string {
+	resolved := make(map[string]string, len(headers)+1)
+	for key, value := range headers {
+		if strings.EqualFold(strings.TrimSpace(key), PartnerHeaderName) {
+			continue
+		}
+		resolved[key] = value
+	}
+	resolved[PartnerHeaderName] = ResolvePartnerID("")
+	return resolved
+}
+
+func BuildPartnerCheckoutReturnURLs(appBaseURL string, sessionToken string) (successURL string, cancelURL string) {
+	base := strings.TrimRight(strings.TrimSpace(appBaseURL), "/")
+	if base == "" {
+		// With no base there is nothing to anchor an absolute URL to. Return empty
+		// so the caller omits the return URLs entirely (client.Pay skips empty ones,
+		// leaving the backend's generic checkout) instead of emitting a broken
+		// relative "/checkout?...".
+		return "", ""
+	}
+	token := url.QueryEscape(sessionToken)
+	return base + "/checkout?checkout=success&partnerCheckout=1&sessionToken=" + token,
+		base + "/checkout?checkout=cancel&partnerCheckout=1&sessionToken=" + token
+}
+
+// BuildPartnerReturnURL returns the https page the browser is sent back to once
+// the partner checkout or OAuth consent completes — the aimlapi web app the user
+// opened the flow from. frontendBaseURL is normally endpoints.VerificationBaseURL
+// (the front that hosts the login/consent page), so the return follows the same
+// environment. AIMLAPI_RETURN_URL overrides it outright; an empty base falls back
+// to DefaultReturnURL. It is deliberately NOT a custom scheme (zero://…): no
+// browser can hand a custom scheme off without an OS-registered handler, which a
+// CLI does not install.
+func BuildPartnerReturnURL(frontendBaseURL string) string {
+	if override := strings.TrimSpace(os.Getenv("AIMLAPI_RETURN_URL")); override != "" {
+		return override
+	}
+	if base := strings.TrimRight(strings.TrimSpace(frontendBaseURL), "/"); base != "" {
+		return base
+	}
+	return DefaultReturnURL
+}
+
+func envOrDefault(name string, fallback string) string {
+	if value := strings.TrimSpace(os.Getenv(name)); value != "" {
+		return value
+	}
+	return fallback
+}
@@ -0,0 +1,55 @@
+package aimlapi
+
+import "testing"
+
+func TestResolveEndpointsDefaults(t *testing.T) {
+	t.Setenv("AIMLAPI_AUTH_URL", "")
+	t.Setenv("AIMLAPI_APP_URL", "")
+	t.Setenv("AIMLAPI_INFERENCE_URL", "")
+	t.Setenv("AIMLAPI_PAY_URL", "")
+
+	endpoints := ResolveEndpoints()
+	if endpoints.AuthBaseURL != "https://auth.aimlapi.com" ||
+		endpoints.AppBaseURL != "https://app.aimlapi.com" ||
+		endpoints.InferenceBaseURL != "https://api.aimlapi.com/v1" ||
+		endpoints.PayBaseURL != "https://pay.aimlapi.com" {
+		t.Fatalf("unexpected endpoints: %#v", endpoints)
+	}
+}
+
+func TestResolvedPartnerHeaderUsesEnvironmentOverrideWithoutDuplicates(t *testing.T) {
+	t.Setenv("AIMLAPI_PARTNER_ID", "part_override")
+
+	headers := WithResolvedPartnerHeader(map[string]string{
+		"x-aimlapi-partner-id":       "part_catalog",
+		"X-AIMLAPI-Integration-Repo": "Gitlawb/zero",
+	})
+
+	if got := headers[PartnerHeaderName]; got != "part_override" {
+		t.Fatalf("partner header = %q, want part_override", got)
+	}
+	if _, ok := headers["x-aimlapi-partner-id"]; ok {
+		t.Fatalf("case-insensitive partner header was duplicated: %#v", headers)
+	}
+	if got := headers["X-AIMLAPI-Integration-Repo"]; got != "Gitlawb/zero" {
+		t.Fatalf("unrelated header changed: %#v", headers)
+	}
+}
+
+func TestBuildPartnerReturnURL(t *testing.T) {
+	t.Setenv("AIMLAPI_RETURN_URL", "")
+	// The frontend base (the web app the flow was opened from) is returned as-is,
+	// trailing slash trimmed — an https page the browser can actually land on.
+	if got := BuildPartnerReturnURL("https://aimlapi.com/app/"); got != "https://aimlapi.com/app" {
+		t.Fatalf("return URL = %q, want the frontend base", got)
+	}
+	// An empty base falls back to the https default, never a custom scheme.
+	if got := BuildPartnerReturnURL(""); got != DefaultReturnURL {
+		t.Fatalf("empty base return URL = %q, want %q", got, DefaultReturnURL)
+	}
+
+	t.Setenv("AIMLAPI_RETURN_URL", "https://example.test/done")
+	if got := BuildPartnerReturnURL("https://aimlapi.com/app"); got != "https://example.test/done" {
+		t.Fatalf("env override return URL = %q", got)
+	}
+}
@@ -0,0 +1,42 @@
+package aimlapi
+
+// UI strings shared by the first-run and provider-wizard onboarding surfaces.
+
+const (
+	// Path A — existing API key. Pick-path prompt + option labels.
+	MsgAPIKeyInputPrompt = "Enter your aimlapi.com key."
+	MsgAPIKeyInvalid     = "API key is invalid. Please make sure you enter a valid aimlapi.com key."
+	MsgPickPathPrompt    = "Do you have aimlapi.com key?"
+	MsgPickPathHaveKey   = "I already have aimlapi.com key"
+	MsgPickPathHaveHint  = "Proceed to paste the key"
+	MsgPickPathNewUser   = "I am a new user"
+	MsgPickPathNewHint   = "One click set up"
+	// Path B — email.
+	MsgEnterEmail           = "Enter your email.\nTo access aimlapi.com dashboard."
+	MsgEmailInvalid         = "Email format is incorrect."
+	MsgAccountActionInvalid = "aimlapi.com returned an unsupported account action. Please try again."
+	// %s = email
+	MsgCodeSent      = "We sent a 6-digit code to %s.\nEnter it below to continue."
+	MsgCodeIncorrect = "Code you've entered is incorrect."
+
+	// Balance check.
+	MsgLowBalance     = "Your aimlapi.com balance is running low.\nIt is recommended to top up your balance."
+	MsgEverythingRuns = "Everything is ready."
+	// Low-balance choices (spec wording).
+	MsgLowBalanceTopUp = "Sure, let's do that"
+	MsgLowBalanceSkip  = "I'll skip topping up the balance for now"
+
+	// Top-up workflow.
+	MsgTopUpPrompt    = "Add credits.\nEnter an amount (min $20)."
+	MsgAmountRequired = "Please enter a top-up amount."
+	// %s = checkout URL
+	MsgTopUpBrowserFallback = "If the browser did not open automatically please use this link to top up your account: %s"
+	MsgTopUpFailed          = "Top up failed. Please try again."
+	MsgTopUpSuccess         = "Top-up successful."
+
+	// Success — key delivered.
+	// %s = amount in dollars (e.g. "25")
+	MsgTopUpAddedFmt = "$%s has been added to your balance."
+	// %s = email
+	MsgSuccessMagicLink = "We've emailed you a magic link to %s. Use it to access your aimlapi.com account and review your usage and balance."
+)
@@ -0,0 +1,26 @@
+package aimlapi
+
+import "testing"
+
+func TestValidEmailRejectsIncompleteDomains(t *testing.T) {
+	tests := []struct {
+		value string
+		valid bool
+	}{
+		{value: "[email protected]", valid: true},
+		{value: "[email protected]", valid: true},
+		{value: "123@123.", valid: false},
+		{value: "123@123", valid: false},
+		{value: "[email protected]", valid: false},
+		{value: "[email protected]", valid: false},
+		{value: "[email protected]", valid: false},
+		{value: "[email protected]", valid: false},
+	}
+	for _, test := range tests {
+		t.Run(test.value, func(t *testing.T) {
+			if got := ValidEmail(test.value); got != test.valid {
+				t.Fatalf("ValidEmail(%q) = %v, want %v", test.value, got, test.valid)
+			}
+		})
+	}
+}
@@ -0,0 +1,335 @@
+package aimlapi
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+	"net/url"
+	"strings"
+)
+
+// StreamTopUpOptions drives a session-based partner-checkout top-up for the
+// interactive (TUI) onboarding. The caller already holds a user session —
+// obtained from an email-code sign-in (Path B existing) or a
+// passwordless new account (Path B sign-up) — so this skips authentication and
+// runs create-session → pay → wait-for-payment, optionally exchanging the paid
+// session into a fresh key (new accounts) versus keeping the caller's key.
+type StreamTopUpOptions struct {
+	SessionToken string // user session bearer (required)
+	AmountUSD    string // dollars; parsed via ParseAmountUSD (min $20)
+	// InferenceBaseURL pins key-bound billing and the returned provider profile to
+	// the endpoint that validated the credential. Empty uses ResolveEndpoints.
+	InferenceBaseURL string
+	Method           PaymentMethod
+	AutoTopUp        bool // enroll the account in automatic top-up at checkout time
+	Exchange         bool // exchange the paid session into a new key (new accounts)
+	Model            string
+	PartnerID        string
+	PartnerName      string
+	NoOpen           bool
+
+	// ResumeSessionToken, when set, resumes that partner-checkout session instead
+	// of opening a new one, so a retry after an ambiguous failure (a lost Pay/
+	// Exchange response) can never double-charge or strand an already-paid session.
+	ResumeSessionToken string
+	// PaymentSessionID is generated once per amount/auto-top-up intent and reused
+	// on retry so an ambiguous Pay response cannot create a second checkout.
+	PaymentSessionID string
+
+	HTTPClient  *http.Client
+	OpenBrowser func(string) error
+	OnStatus    func(Status, string)
+	// OnSession reports the partner-checkout session token the moment it exists (a
+	// fresh CreateSession, or the resumed one), so the caller can retain it and
+	// resume on a later failure. It fires with "" when the session is dead and the
+	// retained token should be dropped (a fresh run must start over).
+	OnSession func(sessionToken string)
+}
+
+// topupPhase is where StreamTopUp enters its create-checkout → poll → exchange
+// pipeline. A fresh run starts at phasePay; a resume skips ahead to whatever the
+// existing session's status makes safe.
+type topupPhase int
+
+const (
+	phasePay          topupPhase = iota // (re-)open the Stripe checkout, then poll + exchange
+	phasePoll                           // checkout already open; poll for payment, never re-pay
+	phaseExchange                       // already paid; exchange only (recovers a lost key)
+	phaseWaitExchange                   // an exchange claim is in flight; poll, never call Exchange twice
+)
+
+// StreamTopUp performs the browser-checkout top-up and reports progress via
+// OnStatus.
+func StreamTopUp(ctx context.Context, options StreamTopUpOptions) (ProvisionedKey, error) {
+	endpoints := ResolveEndpoints()
+	if baseURL := strings.TrimSpace(options.InferenceBaseURL); baseURL != "" {
+		endpoints.InferenceBaseURL = baseURL
+	}
+	client := NewClient(endpoints, options.HTTPClient)
+	partnerID := ResolvePartnerID(options.PartnerID)
+	partnerName := firstNonEmpty(options.PartnerName, DefaultPartnerName)
+	method := options.Method
+	if method != PaymentMethodCrypto {
+		method = PaymentMethodCard
+	}
+	amount, err := ParseAmountUSD(options.AmountUSD)
+	if err != nil {
+		return ProvisionedKey{}, err
+	}
+	if strings.TrimSpace(options.SessionToken) == "" {
+		return ProvisionedKey{}, fmt.Errorf("a session is required to top up")
+	}
+	if strings.TrimSpace(options.PaymentSessionID) == "" {
+		return ProvisionedKey{}, fmt.Errorf("a payment session id is required to top up")
+	}
+
+	status(options.OnStatus, StatusCreatingSession, "")
+	sessionToken, phase, err := resolveTopupSession(ctx, client, options, partnerID, partnerName, endpoints)
+	if err != nil {
+		return ProvisionedKey{}, err
+	}
+	// Retain the live token so an ambiguous failure below can resume this session
+	// rather than open a second checkout on the next attempt.
+	reportSession(options.OnSession, sessionToken)
+
+	if phase <= phasePay {
+		successURL, cancelURL := BuildPartnerCheckoutReturnURLs(endpoints.PayBaseURL, sessionToken)
+		status(options.OnStatus, StatusOpeningCheckout, "")
+		pay, err := client.Pay(ctx, options.SessionToken, sessionToken, amount, options.PaymentSessionID, method, successURL, cancelURL, options.AutoTopUp)
+		if err != nil {
+			return ProvisionedKey{}, err
+		}
+		checkoutURL, err := validatedCheckoutURL(pay.Checkout.PayURL)
+		if err != nil {
+			return ProvisionedKey{}, err
+		}
+		announceCheckout(options.OnStatus, options.OpenBrowser, options.NoOpen, checkoutURL)
+	}
+
+	paidToken := sessionToken
+	if phase <= phasePoll {
+		status(options.OnStatus, StatusWaitingPayment, "")
+		paid, err := pollUntilPaid(ctx, client, sessionToken, options.OnSession)
+		if err != nil {
+			return ProvisionedKey{}, err
+		}
+		paidToken = paid.SessionToken
+		if paid.Status == SessionStatusExchanging {
+			// Another/lost request claimed exchange while we were polling payment.
+			// Carry that fact into the provisioning branch so it follows the claim
+			// instead of issuing a second single-flight Exchange request.
+			phase = phaseWaitExchange
+		}
+	}
+
+	result := ProvisionedKey{
+		BaseURL: endpoints.InferenceBaseURL,
+		Model:   firstNonEmpty(options.Model, DefaultModel),
+	}
+	if options.Exchange {
+		status(options.OnStatus, StatusProvisioningKey, "")
+		if phase == phaseWaitExchange {
+			return ProvisionedKey{}, pollUntilExchangeSettled(ctx, client, sessionToken, options.OnSession)
+		}
+		exchange, err := client.Exchange(ctx, options.SessionToken, paidToken)
+		if err != nil {
+			return ProvisionedKey{}, err
+		}
+		if strings.TrimSpace(exchange.APIKey) == "" {
+			return ProvisionedKey{}, fmt.Errorf("aimlapi.com did not return an API key")
+		}
+		result.APIKey = exchange.APIKey
+		result.APIKeyID = exchange.APIKeyID
+	}
+	return result, nil
+}
+
+// resolveTopupSession returns the partner-checkout session token to drive and the
+// phase to enter. A fresh run opens a new session at phasePay. A resume inspects
+// the existing session and enters the earliest phase that cannot re-charge:
+//   - pending_auth    → phasePay: payment never started, so (re-)Pay is safe.
+//   - pending_payment → phasePoll: a checkout is already open; poll, never re-Pay.
+//   - paid            → phaseExchange: exchange the paid session once.
+//   - exchanging      → phaseWaitExchange: another/lost request owns the atomic
+//     exchange claim; poll it, never issue a second Exchange request.
+//   - exchanged       → terminal: the one-shot raw key is gone; point at dashboard
+//     recovery and keep the token so a re-run repeats that guidance, not a charge.
+//   - cancelled/expired/failed → terminal: drop the dead token so a re-run is fresh.
+func resolveTopupSession(ctx context.Context, client *Client, options StreamTopUpOptions, partnerID, partnerName string, endpoints Endpoints) (string, topupPhase, error) {
+	resume := strings.TrimSpace(options.ResumeSessionToken)
+	if resume == "" {
+		session, err := client.CreateSession(ctx, partnerID, partnerName, BuildPartnerReturnURL(endpoints.VerificationBaseURL))
+		if err != nil {
+			return "", phasePay, err
+		}
+		return session.SessionToken, phasePay, nil
+	}
+	session, err := client.GetSession(ctx, resume)
+	if err != nil {
+		if isTerminalSessionAPIError(err) {
+			reportSession(options.OnSession, "")
+		}
+		return "", phasePay, err
+	}
+	switch session.Status {
+	case SessionStatusPendingAuth:
+		return resume, phasePay, nil
+	case SessionStatusPendingPayment:
+		return resume, phasePoll, nil
+	case SessionStatusPaid:
+		return resume, phaseExchange, nil
+	case SessionStatusExchanging:
+		return resume, phaseWaitExchange, nil
+	case SessionStatusExchanged:
+		return "", phasePay, fmt.Errorf("session was already exchanged; rotate the key from the aimlapi.com dashboard")
+	default:
+		reportSession(options.OnSession, "") // drop the dead token → next run starts fresh
+		return "", phasePay, fmt.Errorf("payment %s; re-run the top-up to try again", session.Status)
+	}
+}
+
+func reportSession(callback func(string), sessionToken string) {
+	if callback != nil {
+		callback(sessionToken)
+	}
+}
+
+// announceCheckout opens the hosted checkout in the browser (unless suppressed)
+// and reports the URL either way, so the TUI can always show the direct link.
+func announceCheckout(onStatus func(Status, string), open func(string) error, noOpen bool, checkoutURL string) {
+	if noOpen || open == nil {
+		status(onStatus, StatusOpeningCheckout, checkoutURL)
+		return
+	}
+	if err := open(checkoutURL); err != nil {
+		status(onStatus, StatusOpeningCheckout, "Open manually: "+checkoutURL)
+		return
+	}
+	status(onStatus, StatusOpeningCheckout, checkoutURL)
+}
+
+func validatedCheckoutURL(value string) (string, error) {
+	checkoutURL := strings.TrimSpace(value)
+	parsed, err := url.Parse(checkoutURL)
+	if err != nil || !parsed.IsAbs() || !strings.EqualFold(parsed.Scheme, "https") || parsed.Hostname() == "" || parsed.User != nil {
+		return "", fmt.Errorf("payment provider did not return a valid HTTPS checkout URL")
+	}
+	return checkoutURL, nil
+}
+
+// StreamTopUpByKeyOptions drives a top-up funded with the API key that already
+// owns the account (Path A pasted key, or AIMLAPI_API_KEY). Unlike StreamTopUp it
+// needs no email session and never exchanges a key — the caller already holds one,
+// so it only funds the wallet: create-session → pay-by-key → wait-for-payment.
+type StreamTopUpByKeyOptions struct {
+	APIKey           string // raw key that owns the account (required)
+	AmountUSD        string // dollars; parsed via ParseAmountUSD (min $20)
+	InferenceBaseURL string // validated key endpoint; empty uses ResolveEndpoints
+	AutoTopUp        bool   // enroll the account in automatic top-up at checkout time
+	PartnerID        string
+	PartnerName      string
+	NoOpen           bool
+
+	// ResumeSessionToken resumes an existing partner-checkout session instead of
+	// opening a new one. PaymentSessionID is the idempotency handle carried across
+	// retries so a re-issued pay returns the same checkout, never a second charge.
+	ResumeSessionToken string
+	PaymentSessionID   string // required; generate once via NewPaymentSessionID, reuse on retry
+
+	HTTPClient  *http.Client
+	OpenBrowser func(string) error
+	OnStatus    func(Status, string)
+	OnSession   func(sessionToken string)
+}
+
+// StreamTopUpByKey funds the account behind APIKey via the hosted checkout and
+// reports progress through OnStatus. It returns an empty ProvisionedKey: the
+// pasted/env key is unchanged, only the balance grows.
+func StreamTopUpByKey(ctx context.Context, options StreamTopUpByKeyOptions) (ProvisionedKey, error) {
+	endpoints := ResolveEndpoints()
+	if baseURL := strings.TrimSpace(options.InferenceBaseURL); baseURL != "" {
+		endpoints.InferenceBaseURL = baseURL
+	}
+	client := NewClient(endpoints, options.HTTPClient)
+	partnerID := ResolvePartnerID(options.PartnerID)
+	partnerName := firstNonEmpty(options.PartnerName, DefaultPartnerName)
+	amount, err := ParseAmountUSD(options.AmountUSD)
+	if err != nil {
+		return ProvisionedKey{}, err
+	}
+	if strings.TrimSpace(options.APIKey) == "" {
+		return ProvisionedKey{}, fmt.Errorf("an API key is required to top up")
+	}
+	if strings.TrimSpace(options.PaymentSessionID) == "" {
+		return ProvisionedKey{}, fmt.Errorf("a payment session id is required to top up")
+	}
+
+	status(options.OnStatus, StatusCreatingSession, "")
+	sessionToken, phase, err := resolveByKeySession(ctx, client, options, partnerID, partnerName, endpoints)
+	if err != nil {
+		return ProvisionedKey{}, err
+	}
+	reportSession(options.OnSession, sessionToken)
+
+	if phase <= phasePay {
+		successURL, cancelURL := BuildPartnerCheckoutReturnURLs(endpoints.PayBaseURL, sessionToken)
+		status(options.OnStatus, StatusOpeningCheckout, "")
+		pay, err := client.TopUpByKey(ctx, options.APIKey, TopUpByKeyRequest{
+			SessionToken:     sessionToken,
+			AmountUSDMinor:   amount,
+			PaymentSessionID: options.PaymentSessionID,
+			AutoTopUp:        options.AutoTopUp,
+			SuccessURL:       successURL,
+			CancelURL:        cancelURL,
+		})
+		if err != nil {
+			return ProvisionedKey{}, err
+		}
+		checkoutURL, err := validatedCheckoutURL(pay.Checkout.PayURL)
+		if err != nil {
+			return ProvisionedKey{}, err
+		}
+		announceCheckout(options.OnStatus, options.OpenBrowser, options.NoOpen, checkoutURL)
+	}
+
+	if phase <= phasePoll {
+		status(options.OnStatus, StatusWaitingPayment, "")
+		if _, err := pollUntilPaid(ctx, client, sessionToken, options.OnSession); err != nil {
+			return ProvisionedKey{}, err
+		}
+	}
+	// No exchange: the account is funded and the caller keeps its existing key.
+	return ProvisionedKey{}, nil
+}
+
+// resolveByKeySession mirrors resolveTopupSession for the by-key flow. Because
+// by-key never exchanges, a paid/exchanging/exchanged session all mean "already
+// funded, done" (phaseExchange skips straight past pay + poll to the empty result).
+func resolveByKeySession(ctx context.Context, client *Client, options StreamTopUpByKeyOptions, partnerID, partnerName string, endpoints Endpoints) (string, topupPhase, error) {
+	resume := strings.TrimSpace(options.ResumeSessionToken)
+	if resume == "" {
+		session, err := client.CreateSession(ctx, partnerID, partnerName, BuildPartnerReturnURL(endpoints.VerificationBaseURL))
+		if err != nil {
+			return "", phasePay, err
+		}
+		return session.SessionToken, phasePay, nil
+	}
+	session, err := client.GetSession(ctx, resume)
+	if err != nil {
+		if isTerminalSessionAPIError(err) {
+			reportSession(options.OnSession, "")
+		}
+		return "", phasePay, err
+	}
+	switch session.Status {
+	case SessionStatusPendingAuth:
+		return resume, phasePay, nil
+	case SessionStatusPendingPayment:
+		return resume, phasePoll, nil
+	case SessionStatusPaid, SessionStatusExchanging, SessionStatusExchanged:
+		return resume, phaseExchange, nil
+	default:
+		reportSession(options.OnSession, "") // drop the dead token → next run starts fresh
+		return "", phasePay, fmt.Errorf("payment %s; re-run the top-up to try again", session.Status)
+	}
+}
@@ -0,0 +1,457 @@
+package aimlapi
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+)
+
+func TestStreamTopUpByKeyFundsWithoutExchange(t *testing.T) {
+	var topupCalls, exchangeCalls int
+	var topupBody map[string]any
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch {
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/partner-checkout/sessions"):
+			_ = json.NewEncoder(w).Encode(PartnerCheckoutSession{SessionToken: "pcs_tok", Status: SessionStatusPendingAuth})
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/v2/billing/topup"):
+			topupCalls++
+			_ = json.NewDecoder(r.Body).Decode(&topupBody)
+			_ = json.NewEncoder(w).Encode(TopUpByKeyResult{Checkout: PaymentSession{PayURL: "https://pay/checkout"}})
+		case r.Method == http.MethodGet && strings.Contains(r.URL.Path, "/partner-checkout/sessions/"):
+			_ = json.NewEncoder(w).Encode(PartnerCheckoutSession{SessionToken: "pcs_tok", Status: SessionStatusPaid})
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/exchange"):
+			exchangeCalls++
+			t.Errorf("by-key must never exchange a key")
+		default:
+			t.Errorf("unexpected request: %s %s", r.Method, r.URL.Path)
+		}
+	}))
+	defer server.Close()
+	t.Setenv("AIMLAPI_APP_URL", server.URL)
+	t.Setenv("AIMLAPI_INFERENCE_URL", server.URL+"/v1")
+
+	var opened string
+	result, err := StreamTopUpByKey(context.Background(), StreamTopUpByKeyOptions{
+		APIKey:           "sk-user",
+		AmountUSD:        "20",
+		AutoTopUp:        true,
+		PaymentSessionID: "pay-1",
+		OpenBrowser:      func(u string) error { opened = u; return nil },
+	})
+	if err != nil {
+		t.Fatalf("StreamTopUpByKey = %v", err)
+	}
+	if result.APIKey != "" {
+		t.Fatalf("by-key must not mint a key, got %q", result.APIKey)
+	}
+	if topupCalls != 1 || exchangeCalls != 0 {
+		t.Fatalf("topup=%d exchange=%d, want 1/0", topupCalls, exchangeCalls)
+	}
+	if opened != "https://pay/checkout" {
+		t.Fatalf("opened checkout = %q", opened)
+	}
+	if topupBody["autoTopUp"] != true || topupBody["paymentSessionId"] != "pay-1" {
+		t.Fatalf("topup body missing autoTopUp/paymentSessionId: %#v", topupBody)
+	}
+}
+
+func TestValidatedCheckoutURLRequiresAbsoluteHTTPS(t *testing.T) {
+	tests := []struct {
+		name    string
+		value   string
+		want    string
+		wantErr bool
+	}{
+		{name: "https", value: " https://pay.example.test/checkout?id=1 ", want: "https://pay.example.test/checkout?id=1"},
+		{name: "empty", wantErr: true},
+		{name: "relative", value: "/checkout", wantErr: true},
+		{name: "http", value: "http://pay.example.test/checkout", wantErr: true},
+		{name: "file", value: "file:///tmp/checkout", wantErr: true},
+		{name: "custom protocol", value: "wallet://checkout/123", wantErr: true},
+		{name: "userinfo", value: "https://[email protected]/checkout", wantErr: true},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			got, err := validatedCheckoutURL(test.value)
+			if (err != nil) != test.wantErr {
+				t.Fatalf("validatedCheckoutURL(%q) error = %v", test.value, err)
+			}
+			if got != test.want {
+				t.Fatalf("validatedCheckoutURL(%q) = %q, want %q", test.value, got, test.want)
+			}
+		})
+	}
+}
+
+func TestStreamTopUpByKeyPinsValidatedInferenceEndpoint(t *testing.T) {
+	validatedCalls := 0
+	validated := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch {
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/partner-checkout/sessions"):
+			_ = json.NewEncoder(w).Encode(PartnerCheckoutSession{SessionToken: "pcs_tok", Status: SessionStatusPendingAuth})
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/v2/billing/topup"):
+			validatedCalls++
+			_ = json.NewEncoder(w).Encode(TopUpByKeyResult{Checkout: PaymentSession{PayURL: "https://pay/checkout"}})
+		case r.Method == http.MethodGet && strings.Contains(r.URL.Path, "/partner-checkout/sessions/"):
+			_ = json.NewEncoder(w).Encode(PartnerCheckoutSession{SessionToken: "pcs_tok", Status: SessionStatusPaid})
+		default:
+			t.Errorf("unexpected validated-endpoint request: %s %s", r.Method, r.URL.Path)
+		}
+	}))
+	defer validated.Close()
+	override := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		t.Errorf("validated key leaked to process override: %s %s", r.Method, r.URL.Path)
+	}))
+	defer override.Close()
+	t.Setenv("AIMLAPI_APP_URL", validated.URL)
+	t.Setenv("AIMLAPI_INFERENCE_URL", override.URL)
+
+	_, err := StreamTopUpByKey(context.Background(), StreamTopUpByKeyOptions{
+		APIKey:           "production-key",
+		AmountUSD:        "20",
+		InferenceBaseURL: validated.URL,
+		PaymentSessionID: "payment-1",
+		OpenBrowser:      func(string) error { return nil },
+	})
+	if err != nil {
+		t.Fatalf("StreamTopUpByKey() error = %v", err)
+	}
+	if validatedCalls != 1 {
+		t.Fatalf("validated endpoint top-up calls = %d, want 1", validatedCalls)
+	}
+}
+
+func TestStreamTopUpEmailRetrySendsStablePaymentSessionID(t *testing.T) {
+	getCalls := 0
+	var paymentSessionID string
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch {
+		case r.Method == http.MethodGet && strings.Contains(r.URL.Path, "/partner-checkout/sessions/"):
+			status := SessionStatusPendingAuth
+			if getCalls > 0 {
+				status = SessionStatusPaid
+			}
+			getCalls++
+			_ = json.NewEncoder(w).Encode(PartnerCheckoutSession{SessionToken: "pcs_retry", Status: status})
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/pay"):
+			var body map[string]any
+			_ = json.NewDecoder(r.Body).Decode(&body)
+			paymentSessionID, _ = body["paymentSessionId"].(string)
+			_ = json.NewEncoder(w).Encode(PayResult{Checkout: PaymentSession{PayURL: "https://pay/checkout"}})
+		default:
+			t.Errorf("unexpected request: %s %s", r.Method, r.URL.Path)
+		}
+	}))
+	defer server.Close()
+	t.Setenv("AIMLAPI_APP_URL", server.URL)
+
+	_, err := StreamTopUp(context.Background(), StreamTopUpOptions{
+		SessionToken:       "bearer",
+		ResumeSessionToken: "pcs_retry",
+		PaymentSessionID:   "payment-stable",
+		AmountUSD:          "20",
+		NoOpen:             true,
+	})
+	if err != nil {
+		t.Fatalf("StreamTopUp() error = %v", err)
+	}
+	if paymentSessionID != "payment-stable" {
+		t.Fatalf("paymentSessionId = %q, want stable retry id", paymentSessionID)
+	}
+}
+
+// resumeServer is a partner-checkout backend fake that fails the test if the
+// re-charging endpoints (create-session / pay) are ever hit, and serves a
+// scripted GetSession status sequence so a resume can be driven deterministically.
+type resumeServer struct {
+	t          *testing.T
+	statuses   []string // GetSession status per call, last value repeats
+	getCalls   int
+	payCalls   int
+	createCall int
+	exchanged  int
+}
+
+func (rs *resumeServer) handler() http.Handler {
+	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch {
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/partner-checkout/sessions"):
+			rs.createCall++
+			rs.t.Errorf("resume must not create a new session")
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/pay"):
+			rs.payCalls++
+			rs.t.Errorf("resume must not re-pay (double charge)")
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/exchange"):
+			rs.exchanged++
+			_ = json.NewEncoder(w).Encode(ExchangeResult{APIKey: "sk-resumed", APIKeyID: "key_1"})
+		case r.Method == http.MethodGet:
+			status := rs.statuses[len(rs.statuses)-1]
+			if rs.getCalls < len(rs.statuses) {
+				status = rs.statuses[rs.getCalls]
+			}
+			rs.getCalls++
+			_ = json.NewEncoder(w).Encode(PartnerCheckoutSession{SessionToken: "pc_tok", Status: SessionStatus(status)})
+		default:
+			rs.t.Errorf("unexpected request: %s %s", r.Method, r.URL.Path)
+		}
+	})
+}
+
+func TestStreamTopUpResumePaidExchangesWithoutRecharge(t *testing.T) {
+	rs := &resumeServer{t: t, statuses: []string{"paid"}}
+	server := httptest.NewServer(rs.handler())
+	defer server.Close()
+	t.Setenv("AIMLAPI_APP_URL", server.URL)
+
+	var seen string
+	result, err := StreamTopUp(context.Background(), StreamTopUpOptions{
+		SessionToken:       "bearer",
+		ResumeSessionToken: "pc_tok",
+		PaymentSessionID:   "payment-1",
+		AmountUSD:          "20",
+		Exchange:           true,
+		NoOpen:             true,
+		OnSession:          func(token string) { seen = token },
+	})
+	if err != nil {
+		t.Fatalf("StreamTopUp resume = %v", err)
+	}
+	if result.APIKey != "sk-resumed" {
+		t.Fatalf("APIKey = %q, want recovered key", result.APIKey)
+	}
+	if rs.exchanged != 1 {
+		t.Fatalf("exchange calls = %d, want 1", rs.exchanged)
+	}
+	if seen != "pc_tok" {
+		t.Fatalf("OnSession token = %q, want retained pc_tok", seen)
+	}
+}
+
+func TestStreamTopUpResumePendingPaymentPollsNeverRepays(t *testing.T) {
+	// First GetSession (resume resolution) sees the still-open checkout; the next
+	// (poll) sees it settle. Pay must never be called — the fake fails the test if
+	// it is.
+	rs := &resumeServer{t: t, statuses: []string{"pending_payment", "paid"}}
+	server := httptest.NewServer(rs.handler())
+	defer server.Close()
+	t.Setenv("AIMLAPI_APP_URL", server.URL)
+
+	result, err := StreamTopUp(context.Background(), StreamTopUpOptions{
+		SessionToken:       "bearer",
+		ResumeSessionToken: "pc_tok",
+		PaymentSessionID:   "payment-1",
+		AmountUSD:          "20",
+		Exchange:           true,
+		NoOpen:             true,
+	})
+	if err != nil {
+		t.Fatalf("StreamTopUp resume = %v", err)
+	}
+	if result.APIKey != "sk-resumed" {
+		t.Fatalf("APIKey = %q, want recovered key", result.APIKey)
+	}
+	if rs.payCalls != 0 {
+		t.Fatalf("pay calls = %d, want 0", rs.payCalls)
+	}
+}
+
+func TestStreamTopUpPollingDeadSessionClearsRetainedToken(t *testing.T) {
+	rs := &resumeServer{t: t, statuses: []string{"pending_payment", "expired"}}
+	server := httptest.NewServer(rs.handler())
+	defer server.Close()
+	t.Setenv("AIMLAPI_APP_URL", server.URL)
+
+	seen := []string{}
+	_, err := StreamTopUp(context.Background(), StreamTopUpOptions{
+		SessionToken:       "bearer",
+		ResumeSessionToken: "pc_tok",
+		PaymentSessionID:   "payment-1",
+		AmountUSD:          "20",
+		Exchange:           true,
+		NoOpen:             true,
+		OnSession:          func(token string) { seen = append(seen, token) },
+	})
+	if err == nil || !strings.Contains(err.Error(), "payment expired") {
+		t.Fatalf("err = %v, want expired payment error", err)
+	}
+	if len(seen) == 0 || seen[len(seen)-1] != "" {
+		t.Fatalf("OnSession events = %#v, want terminal clear", seen)
+	}
+}
+
+func TestPollUntilPaidTerminal4xxClearsRetainedToken(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		http.Error(w, "session rejected", http.StatusNotFound)
+	}))
+	defer server.Close()
+
+	seen := "retained"
+	_, err := pollUntilPaid(context.Background(), NewClient(Endpoints{AppBaseURL: server.URL}, server.Client()), "pc_missing", func(token string) {
+		seen = token
+	})
+	if err == nil {
+		t.Fatal("pollUntilPaid() error = nil, want terminal 4xx")
+	}
+	if seen != "" {
+		t.Fatalf("retained session = %q, want cleared", seen)
+	}
+}
+
+func TestPollUntilExchangeSettledClearsDeadSession(t *testing.T) {
+	for _, test := range []struct {
+		name       string
+		statusCode int
+		status     SessionStatus
+	}{
+		{name: "terminal API rejection", statusCode: http.StatusNotFound},
+		{name: "expired exchange claim", status: SessionStatusExpired},
+	} {
+		t.Run(test.name, func(t *testing.T) {
+			server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+				if test.statusCode != 0 {
+					http.Error(w, "session rejected", test.statusCode)
+					return
+				}
+				_ = json.NewEncoder(w).Encode(PartnerCheckoutSession{SessionToken: "pcs_dead", Status: test.status})
+			}))
+			defer server.Close()
+
+			seen := "retained"
+			err := pollUntilExchangeSettled(context.Background(), NewClient(Endpoints{AppBaseURL: server.URL}, server.Client()), "pcs_dead", func(token string) {
+				seen = token
+			})
+			if err == nil {
+				t.Fatal("pollUntilExchangeSettled() error = nil")
+			}
+			if seen != "" {
+				t.Fatalf("retained session = %q, want cleared", seen)
+			}
+		})
+	}
+}
+
+func TestResumeHelpersClearOnlyRejectedSessions(t *testing.T) {
+	tests := []struct {
+		name      string
+		status    int
+		wantClear bool
+	}{
+		{name: "terminal 4xx", status: http.StatusNotFound, wantClear: true},
+		{name: "retryable 5xx", status: http.StatusServiceUnavailable, wantClear: false},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+				http.Error(w, "session lookup failed", test.status)
+			}))
+			defer server.Close()
+			client := NewClient(Endpoints{AppBaseURL: server.URL}, server.Client())
+
+			for _, flow := range []struct {
+				name    string
+				resolve func(func(string)) error
+			}{
+				{
+					name: "session bearer",
+					resolve: func(onSession func(string)) error {
+						_, _, err := resolveTopupSession(context.Background(), client, StreamTopUpOptions{
+							ResumeSessionToken: "pc_rejected", OnSession: onSession,
+						}, DefaultPartnerID, DefaultPartnerName, Endpoints{})
+						return err
+					},
+				},
+				{
+					name: "api key",
+					resolve: func(onSession func(string)) error {
+						_, _, err := resolveByKeySession(context.Background(), client, StreamTopUpByKeyOptions{
+							ResumeSessionToken: "pc_rejected", OnSession: onSession,
+						}, DefaultPartnerID, DefaultPartnerName, Endpoints{})
+						return err
+					},
+				},
+			} {
+				t.Run(flow.name, func(t *testing.T) {
+					seen := "retained"
+					if err := flow.resolve(func(token string) { seen = token }); err == nil {
+						t.Fatal("resolve error = nil")
+					}
+					if cleared := seen == ""; cleared != test.wantClear {
+						t.Fatalf("session cleared = %v, want %v", cleared, test.wantClear)
+					}
+				})
+			}
+		})
+	}
+}
+
+func TestStreamTopUpResumeExchangedReturnsRecoveryError(t *testing.T) {
+	rs := &resumeServer{t: t, statuses: []string{"exchanged"}}
+	server := httptest.NewServer(rs.handler())
+	defer server.Close()
+	t.Setenv("AIMLAPI_APP_URL", server.URL)
+
+	_, err := StreamTopUp(context.Background(), StreamTopUpOptions{
+		SessionToken:       "bearer",
+		ResumeSessionToken: "pc_tok",
+		PaymentSessionID:   "payment-1",
+		AmountUSD:          "20",
+		Exchange:           true,
+		NoOpen:             true,
+	})
+	if err == nil || !strings.Contains(err.Error(), "already exchanged") {
+		t.Fatalf("err = %v, want an already-exchanged recovery error", err)
+	}
+	if rs.exchanged != 0 {
+		t.Fatalf("exchange calls = %d, want 0 (no second key)", rs.exchanged)
+	}
+}
+
+func TestStreamTopUpResumeExchangingPollsWithoutSecondExchange(t *testing.T) {
+	rs := &resumeServer{t: t, statuses: []string{"exchanging", "exchanged"}}
+	server := httptest.NewServer(rs.handler())
+	defer server.Close()
+	t.Setenv("AIMLAPI_APP_URL", server.URL)
+
+	_, err := StreamTopUp(context.Background(), StreamTopUpOptions{
+		SessionToken:       "bearer",
+		ResumeSessionToken: "pc_tok",
+		PaymentSessionID:   "payment-1",
+		AmountUSD:          "20",
+		Exchange:           true,
+		NoOpen:             true,
+	})
+	if err == nil || !strings.Contains(err.Error(), "already exchanged") {
+		t.Fatalf("err = %v, want dashboard recovery guidance", err)
+	}
+	if rs.exchanged != 0 {
+		t.Fatalf("exchange calls = %d, want 0 while an exchange claim is in flight", rs.exchanged)
+	}
+	if rs.createCall != 0 || rs.payCalls != 0 {
+		t.Fatalf("resume created/paid a new checkout: create=%d pay=%d", rs.createCall, rs.payCalls)
+	}
+}
+
+func TestStreamTopUpPollObservesExchangingWithoutSecondExchange(t *testing.T) {
+	rs := &resumeServer{t: t, statuses: []string{"pending_payment", "exchanging", "exchanged"}}
+	server := httptest.NewServer(rs.handler())
+	defer server.Close()
+	t.Setenv("AIMLAPI_APP_URL", server.URL)
+
+	_, err := StreamTopUp(context.Background(), StreamTopUpOptions{
+		SessionToken:       "bearer",
+		ResumeSessionToken: "pc_tok",
+		PaymentSessionID:   "payment-1",
+		AmountUSD:          "20",
+		Exchange:           true,
+		NoOpen:             true,
+	})
+	if err == nil || !strings.Contains(err.Error(), "already exchanged") {
+		t.Fatalf("err = %v, want dashboard recovery guidance", err)
+	}
+	if rs.exchanged != 0 {
+		t.Fatalf("exchange calls = %d, want 0 after poll observed an existing claim", rs.exchanged)
+	}
+}
@@ -0,0 +1,186 @@
+package aimlapi
+
+import (
+	"context"
+	"crypto/rand"
+	"errors"
+	"fmt"
+	"math"
+	"net/http"
+	"strconv"
+	"strings"
+	"time"
+)
+
+// Status is a progress phase reported by StreamTopUp through its OnStatus callback.
+type Status string
+
+const (
+	StatusCreatingSession Status = "creating-session"
+	StatusOpeningCheckout Status = "opening-checkout"
+	StatusWaitingPayment  Status = "waiting-payment"
+	StatusProvisioningKey Status = "provisioning-key"
+)
+
+// ProvisionedKey is the outcome of a top-up: the (optionally minted) API key plus
+// the inference base URL and model to write into the provider profile.
+type ProvisionedKey struct {
+	APIKey   string
+	APIKeyID string
+	BaseURL  string
+	Model    string
+}
+
+const (
+	pollInterval = 3 * time.Second
+	pollTimeout  = 20 * time.Minute
+)
+
+// ParseAmountUSD parses a dollar string into USD minor units (cents), enforcing the
+// min/max top-up bounds. An empty value yields the default amount.
+func ParseAmountUSD(value string) (int, error) {
+	value = strings.TrimSpace(value)
+	if value == "" {
+		return DefaultAmountUSDMinor, nil
+	}
+	dollars, err := strconv.ParseFloat(value, 64)
+	// NaN/±Inf parse without error and slip past the min/max comparisons below
+	// (every ordered comparison with NaN is false), so reject them explicitly
+	// before the cents conversion turns them into a garbage minor-unit amount.
+	if err != nil || math.IsNaN(dollars) || math.IsInf(dollars, 0) || dollars <= 0 {
+		return 0, fmt.Errorf("invalid amount %q; pass a positive USD amount", value)
+	}
+	minor := int(dollars*100 + 0.5)
+	if minor < MinAmountUSDMinor {
+		return 0, fmt.Errorf("minimum top-up is $%d", MinAmountUSDMinor/100)
+	}
+	if minor > MaxAmountUSDMinor {
+		return 0, fmt.Errorf("maximum top-up is $%d", MaxAmountUSDMinor/100)
+	}
+	return minor, nil
+}
+
+// NewPaymentSessionID returns a random UUIDv4-shaped idempotency id for a top-up.
+// The caller generates it once per top-up intent and reuses it on retry so the
+// backend (and Stripe) coalesce retries onto the same checkout instead of a
+// second charge.
+func NewPaymentSessionID() (string, error) {
+	var b [16]byte
+	if _, err := rand.Read(b[:]); err != nil {
+		return "", err
+	}
+	b[6] = (b[6] & 0x0f) | 0x40 // version 4
+	b[8] = (b[8] & 0x3f) | 0x80 // variant 10
+	return fmt.Sprintf("%x-%x-%x-%x-%x", b[0:4], b[4:6], b[6:8], b[8:10], b[10:16]), nil
+}
+
+func pollUntilPaid(ctx context.Context, client *Client, sessionToken string, onSession func(string)) (PartnerCheckoutSession, error) {
+	deadline := time.Now().Add(pollTimeout)
+	for time.Now().Before(deadline) {
+		session, err := client.GetSession(ctx, sessionToken)
+		if err != nil {
+			// Context cancellation/deadline is terminal (also covers an in-flight
+			// request cancelled mid-poll, whose transport error wraps ctx.Err()).
+			if errors.Is(err, context.Canceled) || errors.Is(err, context.DeadlineExceeded) {
+				return PartnerCheckoutSession{}, err
+			}
+			// A 4xx API response is terminal. Everything else — a 5xx or a transient
+			// transport error (non-APIError, e.g. a dropped connection during the
+			// wait) — is retried until the poll deadline.
+			if isTerminalSessionAPIError(err) {
+				reportSession(onSession, "")
+				return PartnerCheckoutSession{}, err
+			}
+			if err := sleepContext(ctx, pollInterval); err != nil {
+				return PartnerCheckoutSession{}, err
+			}
+			continue
+		}
+		switch session.Status {
+		case SessionStatusPaid, SessionStatusExchanging:
+			return session, nil
+		case SessionStatusExchanged:
+			return PartnerCheckoutSession{}, fmt.Errorf("session was already exchanged; rotate the key from the aimlapi.com dashboard")
+		case SessionStatusCancelled, SessionStatusExpired, SessionStatusFailed:
+			// The checkout cannot recover from these states. Drop the retained token
+			// (and the TUI's paired by-key idempotency ID) now so the very next retry
+			// creates a fresh payment intent.
+			reportSession(onSession, "")
+			return PartnerCheckoutSession{}, fmt.Errorf("payment %s; re-run the top-up to try again", session.Status)
+		default:
+			if err := sleepContext(ctx, pollInterval); err != nil {
+				return PartnerCheckoutSession{}, err
+			}
+		}
+	}
+	return PartnerCheckoutSession{}, fmt.Errorf("timed out waiting for payment; re-run once the payment clears")
+}
+
+func isTerminalSessionAPIError(err error) bool {
+	var apiErr APIError
+	return errors.As(err, &apiErr) && apiErr.Status >= http.StatusBadRequest && apiErr.Status < http.StatusInternalServerError
+}
+
+// pollUntilExchangeSettled follows an exchange claim that was already in flight
+// when a retry resumed the session. Exchange is single-flight and the raw key is
+// returned only to the winning request, so issuing Exchange again is never safe.
+// Once the claim completes, direct the user to the dashboard recovery path.
+func pollUntilExchangeSettled(ctx context.Context, client *Client, sessionToken string, onSession func(string)) error {
+	deadline := time.Now().Add(pollTimeout)
+	for time.Now().Before(deadline) {
+		session, err := client.GetSession(ctx, sessionToken)
+		if err != nil {
+			if errors.Is(err, context.Canceled) || errors.Is(err, context.DeadlineExceeded) {
+				return err
+			}
+			if isTerminalSessionAPIError(err) {
+				reportSession(onSession, "")
+				return err
+			}
+			if err := sleepContext(ctx, pollInterval); err != nil {
+				return err
+			}
+			continue
+		}
+		switch session.Status {
+		case SessionStatusExchanged:
+			return fmt.Errorf("session was already exchanged; rotate the key from the aimlapi.com dashboard")
+		case SessionStatusExchanging:
+			if err := sleepContext(ctx, pollInterval); err != nil {
+				return err
+			}
+		case SessionStatusCancelled, SessionStatusExpired, SessionStatusFailed:
+			reportSession(onSession, "")
+			return fmt.Errorf("key provisioning %s; rotate the key from the aimlapi.com dashboard", session.Status)
+		default:
+			return fmt.Errorf("key provisioning returned to %s; re-run the top-up", session.Status)
+		}
+	}
+	return fmt.Errorf("timed out waiting for key provisioning; retry to check the same session")
+}
+
+func sleepContext(ctx context.Context, delay time.Duration) error {
+	timer := time.NewTimer(delay)
+	defer timer.Stop()
+	select {
+	case <-ctx.Done():
+		return ctx.Err()
+	case <-timer.C:
+		return nil
+	}
+}
+
+func status(callback func(Status, string), value Status, detail string) {
+	if callback != nil {
+		callback(value, detail)
+	}
+}
+
+func firstNonEmpty(values ...string) string {
+	for _, value := range values {
+		if value = strings.TrimSpace(value); value != "" {
+			return value
+		}
+	}
+	return ""
+}
@@ -0,0 +1,38 @@
+package aimlapi
+
+import (
+	"net/mail"
+	"strings"
+)
+
+// ValidEmail performs the lightweight validation used before starting the
+// passwordless onboarding request.
+func ValidEmail(value string) bool {
+	value = strings.TrimSpace(value)
+	address, err := mail.ParseAddress(value)
+	if err != nil || address.Address != value {
+		return false
+	}
+	at := strings.LastIndexByte(value, '@')
+	if at <= 0 || at == len(value)-1 {
+		return false
+	}
+	domain := value[at+1:]
+	if strings.HasPrefix(domain, ".") || strings.HasSuffix(domain, ".") || strings.Contains(domain, "..") {
+		return false
+	}
+	dot := strings.LastIndexByte(domain, '.')
+	if dot <= 0 || dot == len(domain)-1 {
+		return false
+	}
+	tld := domain[dot+1:]
+	if len(tld) < 2 {
+		return false
+	}
+	for _, char := range tld {
+		if char < 'A' || char > 'Z' && char < 'a' || char > 'z' {
+			return false
+		}
+	}
+	return true
+}
@@ -42,6 +42,7 @@ import (
 	"github.com/Gitlawb/zero/internal/worktrees"
 	"github.com/Gitlawb/zero/internal/zerogit"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
+	"github.com/charmbracelet/x/term"
 )

 var version = "dev"
@@ -339,7 +340,17 @@ func runWithDeps(args []string, stdout io.Writer, stderr io.Writer, deps appDeps
 				return 0
 			}
 		}
-		if _, err := fmt.Fprintf(stdout, "zero %s\n", version); err != nil {
+		// Pipes and redirects keep the machine-readable contract exactly as it
+		// was: a single "zero <version>" line. Scripts, command substitutions,
+		// and the NPM wrapper smoke check all parse that record, so the banner
+		// is strictly a TTY affordance.
+		if !stdoutIsTerminal(stdout) {
+			if _, err := fmt.Fprintf(stdout, "zero %s\n", version); err != nil {
+				return 1
+			}
+			return 0
+		}
+		if _, err := fmt.Fprintf(stdout, "%s\n\nzero %s\n", tui.Wordmark(), version); err != nil {
 			return 1
 		}
 		return 0
@@ -755,11 +766,12 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 	}
 	sandboxBackend := deps.selectSandboxBackend(sandbox.BackendOptions{})
 	sandboxEngine := sandbox.NewEngine(sandbox.EngineOptions{
-		WorkspaceRoot: workspaceRoot,
-		Policy:        applyConfiguredSandboxPolicy(sandbox.DefaultPolicy(), resolved.Sandbox),
-		Store:         sandboxStore,
-		Backend:       sandboxBackend,
-		Scope:         scope,
+		WorkspaceRoot:    workspaceRoot,
+		Policy:           applyConfiguredSandboxPolicy(sandbox.DefaultPolicy(), resolved.Sandbox),
+		Store:            sandboxStore,
+		Backend:          sandboxBackend,
+		Scope:            scope,
+		SensitiveEnvKeys: providerSensitiveEnvKeys(resolved),
 	})
 	lastKnownMCPConfig := mcpConfig
 	fileTracker := tools.NewFileTracker()
@@ -1115,6 +1127,17 @@ func closeSpecialistRuntime(stderr io.Writer, runtime *agentToolRuntime) {
 	}
 }

+// stdoutIsTerminal reports whether w is an interactive terminal. Tests pass
+// bytes.Buffer writers and pipes/redirects fail term.IsTerminal, so both take
+// the machine-readable path.
+func stdoutIsTerminal(w io.Writer) bool {
+	f, ok := w.(*os.File)
+	if !ok {
+		return false
+	}
+	return term.IsTerminal(f.Fd())
+}
+
 func writeAppError(stderr io.Writer, message string, exitCode int) int {
 	if _, err := fmt.Fprintf(stderr, "[zero] %s\n", message); err != nil {
 		return 1
@@ -41,6 +41,8 @@ func TestRunPrintsVersion(t *testing.T) {
 	if exitCode != 0 {
 		t.Fatalf("expected exit code 0, got %d", exitCode)
 	}
+	// Non-terminal stdout keeps the machine-readable contract: exactly one
+	// "zero <version>" record, no wordmark banner, no ANSI.
 	if got := stdout.String(); got != "zero dev\n" {
 		t.Fatalf("expected version output, got %q", got)
 	}
@@ -49,6 +51,33 @@ func TestRunPrintsVersion(t *testing.T) {
 	}
 }

+// TestRunVersionRedirectedFile covers the `zero --version > file` path: stdout
+// is a real *os.File but not a terminal, so the single-line contract must hold.
+func TestRunVersionRedirectedFile(t *testing.T) {
+	stdout, err := os.CreateTemp(t.TempDir(), "version-stdout")
+	if err != nil {
+		t.Fatalf("create temp stdout: %v", err)
+	}
+	defer stdout.Close()
+	var stderr bytes.Buffer
+
+	exitCode := Run([]string{"--version"}, stdout, &stderr)
+
+	if exitCode != 0 {
+		t.Fatalf("expected exit code 0, got %d", exitCode)
+	}
+	content, err := os.ReadFile(stdout.Name())
+	if err != nil {
+		t.Fatalf("read temp stdout: %v", err)
+	}
+	if got := string(content); got != "zero dev\n" {
+		t.Fatalf("expected version output, got %q", got)
+	}
+	if stderr.Len() != 0 {
+		t.Fatalf("expected empty stderr, got %q", stderr.String())
+	}
+}
+
 func TestRunPrintsHelp(t *testing.T) {
 	for _, args := range [][]string{
 		{"--help"},
@@ -348,6 +348,83 @@ func TestRunProvidersAddWritesCatalogProfile(t *testing.T) {
 	}
 }

+func TestRunProvidersAddAimlapiWritesDefaultHeaders(t *testing.T) {
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	configPath := filepath.Join(t.TempDir(), "zero", "config.json")
+
+	exitCode := runWithDeps([]string{"providers", "add", "aimlapi"}, &stdout, &stderr, providerSetupDeps(configPath))
+
+	if exitCode != exitSuccess {
+		t.Fatalf("expected exit code %d, got %d: %s", exitSuccess, exitCode, stderr.String())
+	}
+	cfg := readFileConfig(t, configPath)
+	if len(cfg.Providers) != 1 {
+		t.Fatalf("providers = %#v, want one provider", cfg.Providers)
+	}
+	profile := cfg.Providers[0]
+	if profile.Name != "aimlapi" ||
+		profile.CatalogID != "aimlapi" ||
+		profile.BaseURL != "https://api.aimlapi.com/v1" ||
+		profile.Model != "anthropic/claude-sonnet-5" ||
+		profile.APIKeyEnv != "AIMLAPI_API_KEY" {
+		t.Fatalf("unexpected provider profile: %#v", profile)
+	}
+	if profile.CustomHeaders["X-AIMLAPI-Partner-ID"] != "part_62yQoGYDq4Yqnrj2R1iGrDNJ" ||
+		profile.CustomHeaders["X-AIMLAPI-Integration-Repo"] != "Gitlawb/zero" {
+		t.Fatalf("missing aimlapi.com default headers: %#v", profile.CustomHeaders)
+	}
+}
+
+func TestRunProvidersAddAimlapiMixedCaseHeaderOverride(t *testing.T) {
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	configPath := filepath.Join(t.TempDir(), "zero", "config.json")
+
+	// A differently-cased override must replace the catalog header in place, not
+	// leave both spellings behind to race when request construction canonicalizes.
+	exitCode := runWithDeps([]string{
+		"providers", "add", "aimlapi",
+		"--header", "x-aimlapi-partner-id=part_override",
+	}, &stdout, &stderr, providerSetupDeps(configPath))
+
+	if exitCode != exitSuccess {
+		t.Fatalf("expected exit code %d, got %d: %s", exitSuccess, exitCode, stderr.String())
+	}
+	profile := readFileConfig(t, configPath).Providers[0]
+	if _, ok := profile.CustomHeaders["x-aimlapi-partner-id"]; ok {
+		t.Fatalf("lowercase override left a duplicate key: %#v", profile.CustomHeaders)
+	}
+	if profile.CustomHeaders["X-AIMLAPI-Partner-ID"] != "part_override" {
+		t.Fatalf("override did not win on the canonical key: %#v", profile.CustomHeaders)
+	}
+}
+
+func TestRunProvidersAddAimlapiOverrideDropsCatalogHeaders(t *testing.T) {
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	configPath := filepath.Join(t.TempDir(), "zero", "config.json")
+
+	exitCode := runWithDeps([]string{
+		"providers", "add", "aimlapi",
+		"--base-url", "https://staging.example/v1",
+		"--header", "X-Environment=staging",
+	}, &stdout, &stderr, providerSetupDeps(configPath))
+	if exitCode != exitSuccess {
+		t.Fatalf("expected exit code %d, got %d: %s", exitSuccess, exitCode, stderr.String())
+	}
+	profile := readFileConfig(t, configPath).Providers[0]
+	if _, ok := profile.CustomHeaders["X-AIMLAPI-Partner-ID"]; ok {
+		t.Fatalf("partner attribution leaked to overridden endpoint: %#v", profile.CustomHeaders)
+	}
+	if _, ok := profile.CustomHeaders["X-AIMLAPI-Integration-Repo"]; ok {
+		t.Fatalf("integration attribution leaked to overridden endpoint: %#v", profile.CustomHeaders)
+	}
+	if profile.CustomHeaders["X-Environment"] != "staging" {
+		t.Fatalf("explicit user header was dropped: %#v", profile.CustomHeaders)
+	}
+}
+
 func TestRunProvidersAddWritesCustomHeaders(t *testing.T) {
 	var stdout bytes.Buffer
 	var stderr bytes.Buffer
@@ -118,7 +118,8 @@ func runSkillInfo(args []string, dir string, stdout io.Writer, stderr io.Writer)
 	if name == "" {
 		return writeExecUsageError(stderr, "usage: zero skill info <name> [--json]")
 	}
-	info, ok := skills.Info(dir, name)
+	// Resolve across primary + ~/.agents/skills; lock metadata only from primary.
+	info, ok := skills.InfoFromRoots(dir, skills.GlobalRoots(dir), name)
 	if !ok {
 		return writeAppError(stderr, fmt.Sprintf("skill %q not found", name), exitUsage)
 	}
@@ -25,6 +25,7 @@ import (
 	"github.com/Gitlawb/zero/internal/specmode"
 	"github.com/Gitlawb/zero/internal/streamjson"
 	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/usage"
 	"github.com/Gitlawb/zero/internal/worktrees"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
@@ -124,6 +125,12 @@ type execOptions struct {
 	// additional write roots for this run. Unioned with
 	// config.SandboxConfig.AdditionalWriteRoots at scope construction time.
 	addDirs []string
+	// tracePath, when set, writes a per-turn NDJSON trace (agenteval-compatible)
+	// to the given file path — or to stderr when the value is "-". Falls back to
+	// the ZERO_TRACE env var when the flag is absent. Off by default: a run
+	// without it leaves agent.Options.Trace nil and is byte-identical to before.
+	// The trace is pure observation — enabling it does not change agent behavior.
+	tracePath string
 }

 type execUsageError struct {
@@ -422,6 +429,12 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 	notify.MaybeAddWebhookSink(notifier, os.Getenv, func(format string, args ...any) {
 		fmt.Fprintf(stderr, "[notify] "+format+"\n", args...)
 	})
+	// Spec-draft runs synthesize a prompt offline and never drive a model turn, so
+	// there is no per-turn trace to emit. Reject --trace / ZERO_TRACE up front with
+	// a clear error rather than silently accepting and writing nothing.
+	if options.useSpec && resolveTracePath(options) != "" {
+		return writeExecFormatUsageError(stdout, stderr, options.outputFormat, "--trace / ZERO_TRACE are not supported for spec-draft runs")
+	}
 	if options.useSpec {
 		return runExecSpecDraft(execSpecDraftRun{
 			options:            options,
@@ -474,6 +487,24 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 	if err != nil {
 		return writeAppError(stderr, "failed to create run id: "+err.Error(), exitCrash)
 	}
+	// Per-turn tracing is opt-in (--trace <path> or ZERO_TRACE=<path>). The
+	// recorder is stamped throughout agent.Run and the providerio seam; the run
+	// itself is byte-identical to an untraced run. We finish the recorder at the
+	// agent.Run boundary (below) so the snapshot captures exactly one turn, then
+	// serialize the snapshot on every exit path via the defer. "-" writes NDJSON
+	// to stderr; otherwise the path is created/truncated. A failure to emit is
+	// logged to stderr, never fatal.
+	tracePath := resolveTracePath(options)
+	var traceRecorder *trace.Recorder
+	var traceSnapshot *trace.TurnTrace
+	if tracePath != "" {
+		traceRecorder = trace.NewRecorder(preparedSession.Session.SessionID, runID, "")
+		defer func() {
+			if err := writeTraceSnapshot(traceSnapshot, tracePath, stderr); err != nil {
+				fmt.Fprintf(stderr, "[zero] failed to write trace: %s\n", err)
+			}
+		}()
+	}
 	writer := execEventWriter{
 		stdout:       stdout,
 		stderr:       stderr,
@@ -556,6 +587,7 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 		Model:            resolved.Provider.Model,
 		ModelSwitcher:    modelSwitcher,
 		ReasoningEffort:  forwardEffort,
+		Trace:            traceRecorder,
 		Cwd:              workspaceRoot,
 		Images:           images,
 		Registry:         registry,
@@ -625,6 +657,12 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 			sessionRecorder.append(sessions.EventUsage, payload)
 		},
 	})
+	// Finish the trace now that the turn is done, so the snapshot captures exactly
+	// agent.Run's work and nothing the post-run cleanup stamps. The deferred
+	// writer serializes the snapshot on every exit path.
+	if traceRecorder != nil {
+		traceSnapshot = traceRecorder.Finish()
+	}
 	notifier.Notify(notify.Completion, notify.DefaultMessage(notify.Completion))
 	if writer.err != nil {
 		return exitCrash
@@ -815,14 +853,28 @@ func buildExecSandboxEngine(workspaceRoot string, resolved config.ResolvedConfig
 	policy := applyConfiguredSandboxPolicy(sandbox.DefaultPolicy(), resolved.Sandbox)
 	backend := deps.selectSandboxBackend(sandbox.BackendOptions{})
 	return sandbox.NewEngine(sandbox.EngineOptions{
-		WorkspaceRoot: workspaceRoot,
-		Policy:        policy,
-		Store:         store,
-		Backend:       backend,
-		Scope:         scope,
+		WorkspaceRoot:    workspaceRoot,
+		Policy:           policy,
+		Store:            store,
+		Backend:          backend,
+		Scope:            scope,
+		SensitiveEnvKeys: providerSensitiveEnvKeys(resolved),
 	}), nil
 }

+func providerSensitiveEnvKeys(resolved config.ResolvedConfig) []string {
+	keys := make([]string, 0, len(resolved.Providers)+1)
+	for _, profile := range resolved.Providers {
+		if key := strings.TrimSpace(profile.APIKeyEnv); key != "" {
+			keys = append(keys, key)
+		}
+	}
+	if key := strings.TrimSpace(resolved.Provider.APIKeyEnv); key != "" {
+		keys = append(keys, key)
+	}
+	return keys
+}
+
 // applyConfiguredSandboxPolicy overlays every config-sourced sandbox knob onto
 // the default policy.
 func applyConfiguredSandboxPolicy(policy sandbox.Policy, cfg config.SandboxConfig) sandbox.Policy {
@@ -1237,3 +1289,36 @@ func execNotifyMode(options execOptions, resolved config.ResolvedConfig) string
 	}
 	return resolved.Notify.Mode
 }
+
+// resolveTracePath returns the trace destination for this run: the --trace flag
+// value when set, else the ZERO_TRACE env var, else "" (tracing off). A value of
+// "-" means "write to stderr".
+func resolveTracePath(options execOptions) string {
+	if v := strings.TrimSpace(options.tracePath); v != "" {
+		return v
+	}
+	if v := strings.TrimSpace(os.Getenv("ZERO_TRACE")); v != "" {
+		return v
+	}
+	return ""
+}
+
+// writeTraceSnapshot writes an already-finished trace snapshot to dest ("-" =>
+// stderr, otherwise a file path created/truncated). A nil snapshot is a no-op
+// (tracing off, or the run exited before agent.Run produced a turn). The trace
+// is best-effort: a write error is returned to the caller, which logs it but
+// never fails the run.
+func writeTraceSnapshot(snapshot *trace.TurnTrace, dest string, stderr io.Writer) error {
+	if snapshot == nil || dest == "" {
+		return nil
+	}
+	if strings.TrimSpace(dest) == "-" {
+		return trace.WriteNDJSON(stderr, snapshot)
+	}
+	file, err := os.Create(dest)
+	if err != nil {
+		return err
+	}
+	defer file.Close()
+	return trace.WriteNDJSON(file, snapshot)
+}
@@ -32,6 +32,15 @@ func parseExecArgs(args []string) (execOptions, bool, error) {
 			options.noNotify = true
 		case arg == "--no-completion-gate":
 			options.noCompletionGate = true
+		case arg == "--trace":
+			value, next, err := nextFlagValue(args, index, arg)
+			if err != nil {
+				return options, false, err
+			}
+			options.tracePath = value
+			index = next
+		case strings.HasPrefix(arg, "--trace="):
+			options.tracePath = strings.TrimSpace(strings.TrimPrefix(arg, "--trace="))
 		case arg == "--notify":
 			value, next, err := nextFlagValue(args, index, arg)
 			if err != nil {
@@ -15,11 +15,11 @@ import (

 // pluginActivation holds what plugin activation contributed to a bootstrap so the
 // later dispatcher + skill wiring can consume it: the plugin hook definitions and
-// the plugin skill search roots. The zero value (no plugins) is inert — the
-// dispatcher gets no extra hooks and the skill tool keeps only the default dir.
-// The embedded trustSkip reports whether the project plugin layer was dropped for
-// an untrusted workspace, so the caller can fold it into the single combined
-// trust notice.
+// the plugin skill search roots. The zero value (no plugins) still overlays the
+// multi-root skill tool (primary + ~/.agents/skills) so discovery is identical
+// with or without plugins. The embedded trustSkip reports whether the project
+// plugin layer was dropped for an untrusted workspace, so the caller can fold it
+// into the single combined trust notice.
 type pluginActivation struct {
 	hooks      []hooks.Definition
 	skillRoots []string
@@ -52,6 +52,9 @@ func activatePlugins(workspaceRoot string, registry *tools.Registry, deps appDep
 	loaded, err := deps.loadPlugins(plugins.LoadOptions{Cwd: workspaceRoot, ExcludeProject: excludeProject})
 	if err != nil {
 		writePluginActivationWarning(stderr, "failed to load plugins: "+err.Error())
+		// Still overlay multi-root discovery (primary + agents) so a plugin load
+		// failure cannot hide shared skills.
+		registry.Register(plugins.NewSkillTool(deps.skillsDir(), nil))
 		return pluginActivation{trustSkip: skip}
 	}

@@ -67,14 +70,11 @@ func activatePlugins(workspaceRoot string, registry *tools.Registry, deps appDep
 		writePluginActivationWarning(stderr, warning)
 	}

-	// Re-register the skill tool to also resolve plugin-declared skills. The core
-	// skill tool only reads the default skills dir; the plugin-aware replacement
-	// merges the default dir with the plugin skill roots (default dir wins a name
-	// clash), so plugin skills appear in the agent's skill list. With no plugin
-	// skill roots this is byte-equivalent to the default skills surface.
-	if len(result.SkillRoots) > 0 {
-		registry.Register(plugins.NewSkillTool(deps.skillsDir(), result.SkillRoots))
-	}
+	// Always overlay the multi-root skill tool so discovery is identical with or
+	// without plugins: primary DefaultDir + optional ~/.agents/skills + plugin
+	// roots (earlier wins). With empty SkillRoots and no agents dir this is
+	// byte-equivalent to the core single-dir skill surface.
+	registry.Register(plugins.NewSkillTool(deps.skillsDir(), result.SkillRoots))

 	return pluginActivation{hooks: result.Hooks, skillRoots: result.SkillRoots, trustSkip: skip}
 }
@@ -91,8 +91,8 @@ func projectPluginsDirExists(workspaceRoot string) bool {
 }

 // skillInfos resolves the reusable skills the model can load via the skill tool —
-// the default skills dir merged with any plugin skill roots, the same set the
-// plugin-aware skill tool resolves against — as plain data for the agent's system
+// primary dir + optional ~/.agents/skills + plugin skill roots, the same set the
+// multi-root skill tool resolves against — as plain data for the agent's system
 // prompt. It returns nil when no skills are installed, so a skill-less run leaves
 // the prompt byte-identical.
 func (a pluginActivation) skillInfos(defaultDir string) []agent.SkillInfo {
@@ -69,6 +69,11 @@ func TestActivatePluginsRegistersToolAndCollectsHooks(t *testing.T) {
 }

 func TestActivatePluginsRegistersPluginSkillTool(t *testing.T) {
+	// Isolate host ~/.agents/skills so it cannot interfere with this test.
+	home := t.TempDir()
+	t.Setenv("HOME", home)
+	t.Setenv("USERPROFILE", home)
+
 	pluginDir := t.TempDir()
 	skillDir := filepath.Join(pluginDir, "skills", "demo-skill")
 	if err := os.MkdirAll(skillDir, 0o755); err != nil {
@@ -108,6 +113,42 @@ func TestActivatePluginsRegistersPluginSkillTool(t *testing.T) {
 	}
 }

+func TestActivatePluginsAlwaysRegistersMultiRootSkillTool(t *testing.T) {
+	// Even with zero plugins / zero plugin skill roots, activation must overlay
+	// the multi-root skill tool so ~/.agents/skills is loadable on bare runs.
+	home := t.TempDir()
+	t.Setenv("HOME", home)
+	t.Setenv("USERPROFILE", home)
+	agents := filepath.Join(home, ".agents", "skills")
+	if err := os.MkdirAll(agents, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	skillDir := filepath.Join(agents, "agents-skill")
+	if err := os.MkdirAll(skillDir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(skillDir, "SKILL.md"), []byte("---\nname: agents-skill\n---\nagents body"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewSkillTool(t.TempDir())) // core single-dir tool
+	var stderr bytes.Buffer
+	workspace := t.TempDir()
+	activation := activatePlugins(workspace, registry, fakePluginDeps(t, nil), &stderr, workspace)
+	if len(activation.skillRoots) != 0 {
+		t.Fatalf("expected no plugin skill roots, got %#v", activation.skillRoots)
+	}
+	skillTool, ok := registry.Get("skill")
+	if !ok {
+		t.Fatal("skill tool missing after activation")
+	}
+	res := skillTool.Run(context.Background(), map[string]any{"name": "agents-skill"})
+	if res.Status != tools.StatusOK || !bytes.Contains([]byte(res.Output), []byte("agents body")) {
+		t.Fatalf("agents skill must load with zero plugin roots: %q (%s)", res.Status, res.Output)
+	}
+}
+
 func TestActivatePluginsSurfacesLoadDiagnostics(t *testing.T) {
 	// Load succeeds (no top-level error) but reports a per-plugin diagnostic for a
 	// plugin it had to skip. activatePlugins must forward that diagnostic to stderr
@@ -3,8 +3,10 @@ package cli
 import (
 	"fmt"
 	"io"
+	"net/http"
 	"strings"

+	"github.com/Gitlawb/zero/internal/aimlapi"
 	"github.com/Gitlawb/zero/internal/config"
 	"github.com/Gitlawb/zero/internal/providercatalog"
 	"github.com/Gitlawb/zero/internal/providerhealth"
@@ -424,22 +426,39 @@ func providerProfileForAdd(options providerAddOptions) (config.ProviderProfile,
 	if apiKeyEnv == "" && len(descriptor.AuthEnvVars) > 0 {
 		apiKeyEnv = descriptor.AuthEnvVars[0]
 	}
+	baseURL := firstNonEmptyCLI(options.baseURL, descriptor.DefaultBaseURL)
+	var catalogHeaders map[string]string
+	if sameProviderSetupBaseURL(baseURL, descriptor.DefaultBaseURL) {
+		catalogHeaders = descriptor.CustomHeaders
+		if strings.EqualFold(strings.TrimSpace(descriptor.ID), "aimlapi") {
+			catalogHeaders = aimlapi.WithResolvedPartnerHeader(catalogHeaders)
+		}
+	}
 	profile := config.ProviderProfile{
 		Name:            name,
 		ProviderKind:    providerKindForDescriptor(descriptor),
 		CatalogID:       descriptor.ID,
-		BaseURL:         firstNonEmptyCLI(options.baseURL, descriptor.DefaultBaseURL),
+		BaseURL:         baseURL,
 		APIKeyEnv:       apiKeyEnv,
 		APIFormat:       firstAPIFormat(descriptor),
 		AuthHeader:      strings.TrimSpace(options.authHeader),
 		AuthScheme:      normalizeAuthScheme(options.authScheme),
 		AuthHeaderValue: strings.TrimSpace(options.authHeaderValue),
-		CustomHeaders:   copyProviderHeaders(options.customHeaders),
-		Model:           firstNonEmptyCLI(options.model, descriptor.DefaultModel),
+		// Catalog attribution belongs only to the catalog endpoint. Explicit
+		// user headers still apply to an override/staging/proxy endpoint.
+		CustomHeaders: mergeProviderHeaders(catalogHeaders, options.customHeaders),
+		Model:         firstNonEmptyCLI(options.model, descriptor.DefaultModel),
 	}
 	return profile, nil
 }

+func sameProviderSetupBaseURL(left string, right string) bool {
+	return strings.EqualFold(
+		strings.TrimRight(strings.TrimSpace(left), "/"),
+		strings.TrimRight(strings.TrimSpace(right), "/"),
+	)
+}
+
 func selectProviderForCheck(resolved config.ResolvedConfig, name string) (config.ProviderProfile, error) {
 	name = strings.TrimSpace(name)
 	if name == "" {
@@ -544,3 +563,43 @@ func copyProviderHeaders(headers map[string]string) map[string]string {
 	}
 	return copied
 }
+
+func mergeProviderHeaders(base map[string]string, overrides map[string]string) map[string]string {
+	merged := copyProviderHeaders(base)
+	if len(overrides) == 0 {
+		return merged
+	}
+	if merged == nil {
+		merged = map[string]string{}
+	}
+	for key, value := range overrides {
+		// HTTP header names are case-insensitive and request construction
+		// canonicalizes them (http.Header.Set), so an override that differs from a
+		// base key only in case must replace that key's value in place — otherwise
+		// both survive as distinct map entries and race non-deterministically at
+		// build time. Keep the existing spelling (the catalog's, which the resolver
+		// re-merges verbatim); only a genuinely new header is added as typed.
+		if existing, ok := headerKeyFold(merged, key); ok {
+			merged[existing] = value
+			continue
+		}
+		merged[key] = value
+	}
+	return merged
+}
+
+// headerKeyFold returns the key in headers whose canonical HTTP form matches key,
+// so a case-insensitive override lands on the existing entry instead of adding a
+// colliding duplicate.
+func headerKeyFold(headers map[string]string, key string) (string, bool) {
+	if _, ok := headers[key]; ok {
+		return key, true
+	}
+	canonical := http.CanonicalHeaderKey(key)
+	for existing := range headers {
+		if http.CanonicalHeaderKey(existing) == canonical {
+			return existing, true
+		}
+	}
+	return "", false
+}
@@ -0,0 +1,23 @@
+package cli
+
+import (
+	"reflect"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/config"
+)
+
+func TestProviderSensitiveEnvKeys(t *testing.T) {
+	resolved := config.ResolvedConfig{
+		Providers: []config.ProviderProfile{
+			{Name: "custom", APIKeyEnv: " COMPANY_LLM_SECRET "},
+			{Name: "catalog", APIKeyEnv: "OPENAI_API_KEY"},
+			{Name: "inline"},
+		},
+		Provider: config.ProviderProfile{Name: "active", APIKeyEnv: "ACTIVE_PROVIDER_SECRET"},
+	}
+	want := []string{"COMPANY_LLM_SECRET", "OPENAI_API_KEY", "ACTIVE_PROVIDER_SECRET"}
+	if got := providerSensitiveEnvKeys(resolved); !reflect.DeepEqual(got, want) {
+		t.Fatalf("providerSensitiveEnvKeys() = %v, want %v", got, want)
+	}
+}
@@ -4,6 +4,8 @@ import (
 	"context"
 	"fmt"
 	"io"
+	"os"
+	"path/filepath"
 	"strings"

 	"github.com/Gitlawb/zero/internal/mcp"
@@ -14,9 +16,21 @@ import (
 type serveOptions struct {
 	mcp              bool
 	cwd              string
+	addDirs          []string
 	allowUnsafeTools bool
 }

+// serveRootsScope is a PathScope used only for MCP serve: the workspace root
+// plus optional --add-dir roots, without the sandbox temp write roots that
+// would otherwise pollute resources/list.
+type serveRootsScope []string
+
+func (s serveRootsScope) Roots() []string {
+	roots := make([]string, len(s))
+	copy(roots, s)
+	return roots
+}
+
 func runServe(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
 	options, help, err := parseServeArgs(args)
 	if err != nil {
@@ -36,7 +50,11 @@ func runServe(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) i
 	if err != nil {
 		return writeExecUsageError(stderr, err.Error())
 	}
-	registry := newServeRegistry(workspaceRoot, options.allowUnsafeTools)
+	scope, err := buildServeScope(workspaceRoot, options.addDirs)
+	if err != nil {
+		return writeExecUsageError(stderr, err.Error())
+	}
+	registry := newServeRegistry(workspaceRoot, scope, options.allowUnsafeTools)
 	if options.allowUnsafeTools {
 		if _, err := fmt.Fprintln(stderr, "[zero] Unsafe MCP server tools enabled because --allow-unsafe-tools was passed."); err != nil {
 			return exitCrash
@@ -47,6 +65,8 @@ func runServe(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) i
 		Name:              "zero",
 		Version:           version,
 		PermissionGranted: options.allowUnsafeTools,
+		WorkspaceRoot:     workspaceRoot,
+		Scope:             scope,
 	})
 	if err != nil {
 		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
@@ -54,11 +74,11 @@ func runServe(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) i
 	return exitSuccess
 }

-func newServeRegistry(workspaceRoot string, allowUnsafeTools bool) *tools.Registry {
+func newServeRegistry(workspaceRoot string, scope tools.PathScope, allowUnsafeTools bool) *tools.Registry {
 	registry := tools.NewRegistry()
-	toolset := tools.CoreReadOnlyTools(workspaceRoot)
+	toolset := tools.CoreReadOnlyToolsScoped(workspaceRoot, scope)
 	if allowUnsafeTools {
-		toolset = tools.CoreTools(workspaceRoot)
+		toolset = tools.CoreToolsScoped(workspaceRoot, scope)
 	}
 	for _, tool := range toolset {
 		registry.Register(tool)
@@ -66,6 +86,58 @@ func newServeRegistry(workspaceRoot string, allowUnsafeTools bool) *tools.Regist
 	return registry
 }

+// buildServeScope returns nil when there are no extra roots (workspace-only),
+// otherwise a PathScope listing the workspace first followed by each --add-dir.
+func buildServeScope(workspaceRoot string, addDirs []string) (tools.PathScope, error) {
+	if len(addDirs) == 0 {
+		return nil, nil
+	}
+	roots := make([]string, 0, 1+len(addDirs))
+	workspaceRoot = filepath.Clean(workspaceRoot)
+	if abs, err := filepath.Abs(workspaceRoot); err == nil {
+		workspaceRoot = abs
+	}
+
+	// Keep lexical absolute paths in roots so path checks stay aligned with the
+	// un-evaluated WorkspaceRoot used when Scope is nil. Symlink-resolved forms
+	// are only keys for duplicate detection.
+	workspaceResolved := workspaceRoot
+	if resolved, err := filepath.EvalSymlinks(workspaceRoot); err == nil {
+		workspaceResolved = resolved
+	}
+	roots = append(roots, workspaceRoot)
+
+	seen := map[string]struct{}{workspaceResolved: {}}
+	for _, dir := range addDirs {
+		trimmed := strings.TrimSpace(dir)
+		if trimmed == "" {
+			return nil, execUsageError{"--add-dir requires a path"}
+		}
+		absolute, err := filepath.Abs(trimmed)
+		if err != nil {
+			return nil, execUsageError{fmt.Sprintf("--add-dir %q: %v", trimmed, err)}
+		}
+		info, err := os.Stat(absolute)
+		if err != nil {
+			return nil, execUsageError{fmt.Sprintf("--add-dir %q: %v", trimmed, err)}
+		}
+		if !info.IsDir() {
+			return nil, execUsageError{fmt.Sprintf("--add-dir %q is not a directory", trimmed)}
+		}
+
+		resolved := absolute
+		if r, err := filepath.EvalSymlinks(absolute); err == nil {
+			resolved = r
+		}
+		if _, ok := seen[resolved]; ok {
+			continue
+		}
+		seen[resolved] = struct{}{}
+		roots = append(roots, absolute)
+	}
+	return serveRootsScope(roots), nil
+}
+
 func parseServeArgs(args []string) (serveOptions, bool, error) {
 	options := serveOptions{}
 	for index := 0; index < len(args); index++ {
@@ -85,6 +157,18 @@ func parseServeArgs(args []string) (serveOptions, bool, error) {
 			options.cwd = args[index]
 		case strings.HasPrefix(arg, "--cwd="):
 			options.cwd = strings.TrimPrefix(arg, "--cwd=")
+		case arg == "--add-dir":
+			index++
+			if index >= len(args) {
+				return options, false, execUsageError{"--add-dir requires a path"}
+			}
+			options.addDirs = append(options.addDirs, args[index])
+		case strings.HasPrefix(arg, "--add-dir="):
+			value := strings.TrimPrefix(arg, "--add-dir=")
+			if strings.TrimSpace(value) == "" {
+				return options, false, execUsageError{"--add-dir requires a path"}
+			}
+			options.addDirs = append(options.addDirs, value)
 		default:
 			return options, false, execUsageError{fmt.Sprintf("unknown serve flag %q", arg)}
 		}
@@ -100,7 +184,8 @@ Starts Zero as an MCP stdio server.

 Flags:
      --mcp                   Run the MCP stdio server
-  -C, --cwd <path>            Set the workspace directory
+  -C, --cwd <path>            Set the workspace directory (resources + tools)
+      --add-dir <path>        Extra resource/tool root (repeatable)
      --allow-unsafe-tools    Expose write and shell tools to the MCP host
  -h, --help                  Show this help
 `)
@@ -4,6 +4,8 @@ import (
 	"bytes"
 	"encoding/json"
 	"fmt"
+	"os"
+	"path/filepath"
 	"strings"
 	"testing"
 )
@@ -80,6 +82,120 @@ func TestRunServeRequiresMCPMode(t *testing.T) {
 	}
 }

+func TestRunServeMCPWiresWorkspaceRootAndAddDirIntoResources(t *testing.T) {
+	workspace := t.TempDir()
+	extra := t.TempDir()
+	if err := os.WriteFile(filepath.Join(workspace, "workspace.txt"), []byte("ws\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(extra, "extra.txt"), []byte("ex\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	exitCode := runWithDeps(
+		[]string{"serve", "--mcp", "-C", workspace, "--add-dir", extra},
+		&stdout,
+		&stderr,
+		appDeps{stdin: bytes.NewReader(serveMCPResourcesInput(t))},
+	)
+	if exitCode != exitSuccess {
+		t.Fatalf("exit=%d stderr=%q", exitCode, stderr.String())
+	}
+	output := stdout.String()
+	if !strings.Contains(output, "workspace.txt") {
+		t.Fatalf("expected workspace resource, got %q", output)
+	}
+	if !strings.Contains(output, "extra.txt") {
+		t.Fatalf("expected --add-dir resource, got %q", output)
+	}
+}
+
+func TestRunServeMCPRejectsMissingAddDir(t *testing.T) {
+	workspace := t.TempDir()
+	missing := filepath.Join(workspace, "does-not-exist")
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	exitCode := runWithDeps(
+		[]string{"serve", "--mcp", "-C", workspace, "--add-dir", missing},
+		&stdout,
+		&stderr,
+		appDeps{stdin: bytes.NewReader(nil)},
+	)
+	if exitCode != exitUsage {
+		t.Fatalf("exit=%d want %d stderr=%q", exitCode, exitUsage, stderr.String())
+	}
+	if !strings.Contains(stderr.String(), "--add-dir") {
+		t.Fatalf("expected --add-dir error, got %q", stderr.String())
+	}
+}
+
+func TestParseServeArgsCollectsAddDirs(t *testing.T) {
+	options, help, err := parseServeArgs([]string{"--mcp", "--add-dir", "/one", "--add-dir=/two", "-C", "/ws"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if help {
+		t.Fatal("unexpected help")
+	}
+	if !options.mcp || options.cwd != "/ws" {
+		t.Fatalf("options=%#v", options)
+	}
+	if len(options.addDirs) != 2 || options.addDirs[0] != "/one" || options.addDirs[1] != "/two" {
+		t.Fatalf("addDirs=%v", options.addDirs)
+	}
+}
+
+func TestBuildServeScopeNilWithoutExtras(t *testing.T) {
+	scope, err := buildServeScope(t.TempDir(), nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if scope != nil {
+		t.Fatalf("expected nil scope, got %#v", scope)
+	}
+}
+
+func TestBuildServeScopeKeepsLexicalPaths(t *testing.T) {
+	base := t.TempDir()
+	realWorkspace := filepath.Join(base, "real-workspace")
+	realExtra := filepath.Join(base, "real-extra")
+	for _, dir := range []string{realWorkspace, realExtra} {
+		if err := os.MkdirAll(dir, 0o700); err != nil {
+			t.Fatal(err)
+		}
+	}
+	linkWorkspace := filepath.Join(base, "link-workspace")
+	linkExtra := filepath.Join(base, "link-extra")
+	if err := os.Symlink(realWorkspace, linkWorkspace); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.Symlink(realExtra, linkExtra); err != nil {
+		t.Fatal(err)
+	}
+
+	scope, err := buildServeScope(linkWorkspace, []string{linkExtra})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if scope == nil {
+		t.Fatal("expected non-nil scope")
+	}
+	roots := scope.Roots()
+	wantWorkspace, err := filepath.Abs(linkWorkspace)
+	if err != nil {
+		t.Fatal(err)
+	}
+	wantExtra, err := filepath.Abs(linkExtra)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(roots) != 2 || roots[0] != wantWorkspace || roots[1] != wantExtra {
+		t.Fatalf("roots=%v want [%q %q]", roots, wantWorkspace, wantExtra)
+	}
+}
+
 func serveMCPInput(t *testing.T) []byte {
 	t.Helper()

@@ -104,6 +220,30 @@ func serveMCPInput(t *testing.T) []byte {
 	return input.Bytes()
 }

+func serveMCPResourcesInput(t *testing.T) []byte {
+	t.Helper()
+
+	var input bytes.Buffer
+	writeServeMCPMessage(t, &input, map[string]any{
+		"jsonrpc": "2.0",
+		"id":      1,
+		"method":  "initialize",
+		"params":  map[string]any{},
+	})
+	writeServeMCPMessage(t, &input, map[string]any{
+		"jsonrpc": "2.0",
+		"method":  "notifications/initialized",
+		"params":  map[string]any{},
+	})
+	writeServeMCPMessage(t, &input, map[string]any{
+		"jsonrpc": "2.0",
+		"id":      2,
+		"method":  "resources/list",
+		"params":  map[string]any{},
+	})
+	return input.Bytes()
+}
+
 func writeServeMCPMessage(t *testing.T, buffer *bytes.Buffer, message map[string]any) {
 	t.Helper()
 	body, err := json.Marshal(message)
@@ -200,6 +200,54 @@ func TestSaveSetupProviderStoresPastedAPIKey(t *testing.T) {
 	}
 }

+func TestSaveSetupProviderAimlapiEnvReferenceAndOverrideHeaders(t *testing.T) {
+	configPath := filepath.Join(t.TempDir(), "zero", "config.json")
+	result, err := saveSetupProvider(appDeps{
+		userConfigPath: func() (string, error) { return configPath, nil },
+	}, tui.SetupSelection{
+		CatalogID: "aimlapi",
+		BaseURL:   "https://staging.example/v1",
+		Model:     "anthropic/claude-sonnet-5",
+	}, setupSaveOptions{})
+	if err != nil {
+		t.Fatalf("saveSetupProvider() error = %v", err)
+	}
+	if result.Provider.APIKey != "" || result.Provider.APIKeyEnv != "AIMLAPI_API_KEY" {
+		t.Fatalf("runtime profile did not retain AIMLAPI_API_KEY reference: %+v", result.Provider)
+	}
+	if len(result.Provider.CustomHeaders) != 0 {
+		t.Fatalf("catalog headers leaked to first-run override: %#v", result.Provider.CustomHeaders)
+	}
+	persisted := readFileConfig(t, configPath).Providers[0]
+	if persisted.APIKey != "" || persisted.APIKeyEnv != "AIMLAPI_API_KEY" || persisted.APIKeyStored {
+		t.Fatalf("persisted credential source changed: %+v", persisted)
+	}
+	if len(persisted.CustomHeaders) != 0 {
+		t.Fatalf("persisted override contains catalog headers: %#v", persisted.CustomHeaders)
+	}
+}
+
+func TestSaveSetupProviderAimlapiUsesResolvedPartnerOverride(t *testing.T) {
+	t.Setenv("AIMLAPI_PARTNER_ID", "part_override")
+	configPath := filepath.Join(t.TempDir(), "zero", "config.json")
+	result, err := saveSetupProvider(appDeps{
+		userConfigPath: func() (string, error) { return configPath, nil },
+	}, tui.SetupSelection{
+		CatalogID: "aimlapi",
+		BaseURL:   "https://api.aimlapi.com/v1",
+		Model:     "anthropic/claude-sonnet-5",
+	}, setupSaveOptions{})
+	if err != nil {
+		t.Fatalf("saveSetupProvider() error = %v", err)
+	}
+	if got := result.Provider.CustomHeaders["X-AIMLAPI-Partner-ID"]; got != "part_override" {
+		t.Fatalf("runtime partner header = %q, want part_override", got)
+	}
+	if got := readFileConfig(t, configPath).Providers[0].CustomHeaders["X-AIMLAPI-Partner-ID"]; got != "part_override" {
+		t.Fatalf("persisted partner header = %q, want part_override", got)
+	}
+}
+
 func TestSaveSetupProviderStoresCustomEndpointSelection(t *testing.T) {
 	configPath := filepath.Join(t.TempDir(), "zero", "config.json")

@@ -59,25 +59,21 @@ func runSkills(args []string, stdout io.Writer, stderr io.Writer, deps appDeps)
 }

 func runSkillsList(dir string, options skillListOptions, stdout io.Writer, stderr io.Writer) int {
-	discovered, err := skills.List(dir)
+	// Management CLI lists global roots only (primary + ~/.agents/skills); plugin
+	// skills stay out of `zero skills list` and surface via the agent skill tool.
+	roots := skills.GlobalRoots(dir)
+	discovered, dups, err := skills.ListFromRoots(roots)
 	if err != nil {
 		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
 	}
-	// Surface name collisions that List silently resolved (first directory wins),
-	// so a shadowed same-named skill is reported instead of just disappearing.
-	// Warnings go to stderr, keeping stdout (including --json) clean.
-	if dups, derr := skills.Duplicates(dir); derr == nil {
-		for _, dup := range dups {
-			fmt.Fprintf(stderr, "warning: duplicate skill %q: using %s, ignoring %s\n",
-				redaction.RedactString(dup.Name, redaction.Options{}),
-				redaction.RedactString(dup.Winner, redaction.Options{}),
-				redaction.RedactString(dup.Loser, redaction.Options{}))
-		}
-	} else {
-		// Don't silently swallow a scan failure: "no warnings" would then be
-		// ambiguous (no duplicates vs. detection broke). Surface it on stderr.
-		fmt.Fprintf(stderr, "warning: could not check for duplicate skills: %s\n",
-			redaction.ErrorMessage(derr, redaction.Options{}))
+	// Surface name collisions that ListFromRoots silently resolved (earlier root
+	// wins), so a shadowed same-named skill is reported instead of just
+	// disappearing. Warnings go to stderr, keeping stdout (including --json) clean.
+	for _, dup := range dups {
+		fmt.Fprintf(stderr, "warning: duplicate skill %q: using %s, ignoring %s\n",
+			redaction.RedactString(dup.Name, redaction.Options{}),
+			redaction.RedactString(dup.Winner, redaction.Options{}),
+			redaction.RedactString(dup.Loser, redaction.Options{}))
 	}
 	if options.json {
 		payload := struct {
@@ -88,16 +84,16 @@ func runSkillsList(dir string, options skillListOptions, stdout io.Writer, stder
 		}
 		return exitSuccess
 	}
-	output := redaction.RedactString(formatSkillList(discovered, dir), redaction.Options{})
+	output := redaction.RedactString(formatSkillList(discovered), redaction.Options{})
 	if _, err := fmt.Fprintln(stdout, output); err != nil {
 		return exitCrash
 	}
 	return exitSuccess
 }

-func formatSkillList(discovered []skills.Skill, dir string) string {
+func formatSkillList(discovered []skills.Skill) string {
 	if len(discovered) == 0 {
-		return fmt.Sprintf("No Zero skills found in %s.", dir)
+		return "No skills found."
 	}
 	lines := []string{"Zero Skills:"}
 	for _, skill := range discovered {
@@ -131,10 +127,13 @@ func writeSkillsHelp(w io.Writer) error {
  zero skills <command>

 Commands:
-  list                 List discovered Zero skills
-  add <git-url|path>   Install a skill (checksum-pinned in skills.lock)
+  list                 List discovered skills (Zero dir + ~/.agents/skills)
+  add <git-url|path>   Install a skill into the Zero skills dir (checksum-pinned in skills.lock)
  info <name>          Show a skill's frontmatter, source, and pinned hash
-  remove <name>        Remove an installed skill and its lockfile entry
+  remove <name>        Remove an installed skill and its lockfile entry from the Zero skills dir
+
+list and info also search ~/.agents/skills when present (read-only).
+add and remove always target the Zero-specific skills directory.
 `)
 	return err
 }
@@ -23,7 +23,15 @@ func writeSkillFixture(t *testing.T, dir string, name string, content string) {
 	}
 }

+func isolateCLIAgentsHome(t *testing.T) {
+	t.Helper()
+	home := t.TempDir()
+	t.Setenv("HOME", home)
+	t.Setenv("USERPROFILE", home)
+}
+
 func TestRunSkillsListText(t *testing.T) {
+	isolateCLIAgentsHome(t)
 	dir := t.TempDir()
 	writeSkillFixture(t, dir, "confirmation-policy", "---\nname: confirmation-policy\ndescription: Ask before risky actions.\n---\nbody")

@@ -43,6 +51,7 @@ func TestRunSkillsListText(t *testing.T) {
 }

 func TestRunSkillsListWarnsOnDuplicateNames(t *testing.T) {
+	isolateCLIAgentsHome(t)
 	dir := t.TempDir()
 	// Two directories declare the same frontmatter name; List keeps one and the
 	// other is shadowed. The command must warn instead of silently dropping it.
@@ -62,6 +71,7 @@ func TestRunSkillsListWarnsOnDuplicateNames(t *testing.T) {
 }

 func TestRunSkillsDefaultsToList(t *testing.T) {
+	isolateCLIAgentsHome(t)
 	dir := t.TempDir()
 	writeSkillFixture(t, dir, "demo", "body")

@@ -78,6 +88,7 @@ func TestRunSkillsDefaultsToList(t *testing.T) {
 }

 func TestRunSkillsListJSON(t *testing.T) {
+	isolateCLIAgentsHome(t)
 	dir := t.TempDir()
 	writeSkillFixture(t, dir, "demo", "---\nname: demo\ndescription: a demo\n---\nbody")

@@ -111,6 +122,7 @@ func TestRunSkillsListJSON(t *testing.T) {
 }

 func TestRunSkillsEmptyDir(t *testing.T) {
+	isolateCLIAgentsHome(t)
 	var stdout, stderr bytes.Buffer
 	exit := runWithDeps([]string{"skills", "list"}, &stdout, &stderr, appDeps{
 		skillsDir: func() string { return filepath.Join(t.TempDir(), "missing") },
@@ -118,11 +130,89 @@ func TestRunSkillsEmptyDir(t *testing.T) {
 	if exit != 0 {
 		t.Fatalf("exit = %d, stderr = %s", exit, stderr.String())
 	}
-	if !strings.Contains(strings.ToLower(stdout.String()), "no") {
+	if !strings.Contains(strings.ToLower(stdout.String()), "no skills found") {
 		t.Fatalf("expected a no-skills message, got:\n%s", stdout.String())
 	}
 }

+func TestRunSkillsListIncludesAgentsOnlySkill(t *testing.T) {
+	home := t.TempDir()
+	t.Setenv("HOME", home)
+	t.Setenv("USERPROFILE", home)
+	agents := filepath.Join(home, ".agents", "skills")
+	writeSkillFixture(t, agents, "agents-only", "---\nname: agents-only\ndescription: Shared multi-agent skill.\n---\nbody")
+
+	primary := t.TempDir()
+	var stdout, stderr bytes.Buffer
+	exit := runWithDeps([]string{"skills", "list"}, &stdout, &stderr, appDeps{
+		skillsDir: func() string { return primary },
+	})
+	if exit != 0 {
+		t.Fatalf("exit = %d, stderr = %s", exit, stderr.String())
+	}
+	out := stdout.String()
+	if !strings.Contains(out, "agents-only") {
+		t.Fatalf("list should include agents-only skill:\n%s", out)
+	}
+	if !strings.Contains(out, agents) {
+		t.Fatalf("list should show agents path:\n%s", out)
+	}
+}
+
+func TestRunSkillsListPrimaryShadowsAgents(t *testing.T) {
+	home := t.TempDir()
+	t.Setenv("HOME", home)
+	t.Setenv("USERPROFILE", home)
+	agents := filepath.Join(home, ".agents", "skills")
+	primary := t.TempDir()
+	writeSkillFixture(t, primary, "shared", "---\nname: shared\ndescription: From primary.\n---\nbody")
+	writeSkillFixture(t, agents, "shared", "---\nname: shared\ndescription: From agents.\n---\nbody")
+
+	var stdout, stderr bytes.Buffer
+	exit := runWithDeps([]string{"skills", "list"}, &stdout, &stderr, appDeps{
+		skillsDir: func() string { return primary },
+	})
+	if exit != 0 {
+		t.Fatalf("exit = %d, stderr = %s", exit, stderr.String())
+	}
+	if !strings.Contains(stdout.String(), "From primary") {
+		t.Fatalf("primary should win list description:\n%s", stdout.String())
+	}
+	if !strings.Contains(stdout.String(), primary) {
+		t.Fatalf("primary path should be shown:\n%s", stdout.String())
+	}
+	if !strings.Contains(stderr.String(), `duplicate skill "shared"`) {
+		t.Fatalf("expected cross-root duplicate warning, got: %q", stderr.String())
+	}
+}
+
+func TestRunSkillInfoResolvesAgentsOnly(t *testing.T) {
+	home := t.TempDir()
+	t.Setenv("HOME", home)
+	t.Setenv("USERPROFILE", home)
+	agents := filepath.Join(home, ".agents", "skills")
+	writeSkillFixture(t, agents, "agents-info", "---\nname: agents-info\ndescription: Agents info.\n---\nbody")
+
+	primary := t.TempDir()
+	var stdout, stderr bytes.Buffer
+	exit := runWithDeps([]string{"skill", "info", "agents-info"}, &stdout, &stderr, appDeps{
+		skillsDir: func() string { return primary },
+	})
+	if exit != 0 {
+		t.Fatalf("exit = %d, stderr = %s", exit, stderr.String())
+	}
+	out := stdout.String()
+	if !strings.Contains(out, "agents-info") || !strings.Contains(out, "Agents info.") {
+		t.Fatalf("info missing agents skill:\n%s", out)
+	}
+	if strings.Contains(out, "source:") || strings.Contains(out, "hash:") {
+		t.Fatalf("agents-only info must not invent lock metadata:\n%s", out)
+	}
+	if !strings.Contains(out, agents) {
+		t.Fatalf("info should show agents path:\n%s", out)
+	}
+}
+
 func TestRunSkillsDoesNotLaunchTUI(t *testing.T) {
 	var stdout, stderr bytes.Buffer
 	launchCalled := false
@@ -6,7 +6,7 @@ import (
 	"strings"
 )

-func mergeMCPConfig(dst *MCPConfig, src MCPConfig) {
+func mergeMCPConfig(dst *MCPConfig, src MCPConfig, canReenable bool) {
 	if len(src.Servers) == 0 {
 		return
 	}
@@ -14,10 +14,14 @@ func mergeMCPConfig(dst *MCPConfig, src MCPConfig) {
 		dst.Servers = map[string]MCPServerConfig{}
 	}
 	for name, server := range src.Servers {
-		dst.Servers[name] = mergeMCPServer(dst.Servers[name], server)
+		dst.Servers[name] = mergeMCPServer(dst.Servers[name], server, canReenable)
 	}
 }

+// mergeProjectMCPConfig merges a project-scoped (lower-trust) MCP config into
+// dst. It never re-enables a server the user explicitly disabled or disabled a
+// server the user explicitly enabled: it merges with canReenable=false so the
+// trust-boundary guard in mergeMCPServer keeps the user's decision authoritative.
 func mergeProjectMCPConfig(dst *MCPConfig, src MCPConfig) error {
 	if len(src.Servers) == 0 {
 		return nil
@@ -27,7 +31,7 @@ func mergeProjectMCPConfig(dst *MCPConfig, src MCPConfig) error {
 	}
 	for name, server := range src.Servers {
 		base := dst.Servers[name]
-		candidate := mergeMCPServer(base, server)
+		candidate := mergeMCPServer(base, server, false)
 		if projectMCPServerTargetChanges(base, server) && hasInheritedMCPCredentialMaterial(server, candidate) {
 			return fmt.Errorf("project MCP server %q cannot override target while inheriting user credentials; set headers/env/oauth explicitly or use a new server name", name)
 		}
@@ -37,7 +41,11 @@ func mergeProjectMCPConfig(dst *MCPConfig, src MCPConfig) error {
 	return nil
 }

-func mergeMCPServer(base MCPServerConfig, next MCPServerConfig) MCPServerConfig {
+// mergeMCPServer merges a later config layer's MCP server into the base. The
+// canReenable flag marks the CLI-override scope, the only layer that is allowed
+// to lift a sticky user-level disable/enable decision (see the disabled-handling
+// block).
+func mergeMCPServer(base MCPServerConfig, next MCPServerConfig, canReenable bool) MCPServerConfig {
 	if strings.TrimSpace(next.Type) != "" {
 		base.Type = next.Type
 	}
@@ -62,8 +70,26 @@ func mergeMCPServer(base MCPServerConfig, next MCPServerConfig) MCPServerConfig
 	if next.OAuth != nil {
 		base.OAuth = next.OAuth
 	}
+	// Capture the higher-trust scope's prior decision before folding in
+	// next.disabledSet. The trust boundary must be evaluated against the state
+	// the higher-trust layer actually left behind, not against next's own flag.
+	baseDisabledSet := base.disabledSet
+	baseDisabled := base.Disabled
+	if next.disabledSet {
+		base.disabledSet = true
+	}
 	if next.disabledSet || next.Disabled {
-		base.Disabled = next.Disabled
+		// Trust boundary: once a higher-trust scope has explicitly set the
+		// disabled flag, a lower-trust layer (project config) cannot override
+		// that decision in either direction — the user's choice (whether to
+		// enable or disable a server) wins over the repo. This blocks a repo
+		// from re-enabling a server the user disabled, and from disabling a
+		// server the user explicitly enabled. The only layer allowed to
+		// override an explicit higher-scope decision is the CLI override scope
+		// (canReenable=true).
+		if canReenable || !(baseDisabledSet && baseDisabled != next.Disabled) {
+			base.Disabled = next.Disabled
+		}
 	}
 	if next.configured {
 		base.configured = true
@@ -75,7 +101,7 @@ func mergeMCPServer(base MCPServerConfig, next MCPServerConfig) MCPServerConfig
 }

 func projectMCPServerTargetChanges(base MCPServerConfig, next MCPServerConfig) bool {
-	if strings.TrimSpace(next.Type) != "" && mcpServerTransportKind(base) != mcpServerTransportKind(mergeMCPServer(base, next)) {
+	if strings.TrimSpace(next.Type) != "" && mcpServerTransportKind(base) != mcpServerTransportKind(mergeMCPServer(base, next, true)) {
 		return true
 	}
 	if strings.TrimSpace(next.URL) != "" && strings.TrimSpace(next.URL) != strings.TrimSpace(base.URL) {
@@ -9,6 +9,7 @@ import (
 	"strconv"
 	"strings"

+	"github.com/Gitlawb/zero/internal/aimlapi"
 	"github.com/Gitlawb/zero/internal/modelregistry"
 	"github.com/Gitlawb/zero/internal/notify"
 	"github.com/Gitlawb/zero/internal/providercatalog"
@@ -170,7 +171,10 @@ func ResolveMCP(options ResolveOptions) (MCPConfig, error) {
 		if err != nil {
 			return MCPConfig{}, err
 		}
-		mergeMCPConfig(&cfg.MCP, fileConfig.MCP)
+		// User config is higher-trust than project config: it may re-enable a
+		// server the user disabled (a user-level disable is sticky, but the user
+		// scope itself may lift it).
+		mergeMCPConfig(&cfg.MCP, fileConfig.MCP, true)
 	}
 	// Drop the project layer when the workspace is untrusted, so a cloned repo's
 	// ./.zero/config.json cannot register (and spawn) MCP servers. Fail-closed:
@@ -184,7 +188,7 @@ func ResolveMCP(options ResolveOptions) (MCPConfig, error) {
 			return MCPConfig{}, err
 		}
 	}
-	mergeMCPConfig(&cfg.MCP, options.Overrides.MCP)
+	mergeMCPConfig(&cfg.MCP, options.Overrides.MCP, true)
 	return cfg.MCP, nil
 }

@@ -211,7 +215,7 @@ func mergeConfig(dst *FileConfig, src FileConfig) {
 	for _, provider := range src.Providers {
 		mergeProvider(dst, provider)
 	}
-	mergeMCPConfig(&dst.MCP, src.MCP)
+	mergeMCPConfig(&dst.MCP, src.MCP, true)
 	if network := strings.TrimSpace(src.Sandbox.Network); network != "" {
 		dst.Sandbox.Network = network
 	}
@@ -709,7 +713,7 @@ func applyOverrides(cfg *FileConfig, overrides Overrides) {
 	if hasProviderFields(overrides.Provider) {
 		mergeProvider(cfg, overrides.Provider)
 	}
-	mergeMCPConfig(&cfg.MCP, overrides.MCP)
+	mergeMCPConfig(&cfg.MCP, overrides.MCP, true)
 }

 func mergeLocalControlConfig(dst *LocalControlConfig, src LocalControlConfig) {
@@ -1039,6 +1043,45 @@ func applyCatalogDescriptor(profile *ProviderProfile, descriptor providercatalog
 	if profile.APIKeyEnv == "" && len(descriptor.AuthEnvVars) > 0 && (!explicitBaseURL || sameBaseURL(profile.BaseURL, descriptor.DefaultBaseURL)) {
 		profile.APIKeyEnv = descriptor.AuthEnvVars[0]
 	}
+	canonicalCatalogEndpoint := !explicitBaseURL || sameBaseURL(profile.BaseURL, descriptor.DefaultBaseURL)
+	if len(descriptor.CustomHeaders) > 0 && canonicalCatalogEndpoint {
+		catalogHeaders := descriptor.CustomHeaders
+		if strings.EqualFold(strings.TrimSpace(descriptor.ID), "aimlapi") {
+			catalogHeaders = aimlapi.WithResolvedPartnerHeader(catalogHeaders)
+		}
+		merged := copyStringMap(catalogHeaders)
+		for key, value := range profile.CustomHeaders {
+			// Header names are case-insensitive. Preserve the catalog spelling while
+			// replacing its value so request construction cannot see two colliding
+			// map entries whose eventual winner depends on iteration order.
+			existingKey := ""
+			for candidate := range merged {
+				if strings.EqualFold(candidate, key) {
+					existingKey = candidate
+					break
+				}
+			}
+			if existingKey != "" {
+				merged[existingKey] = value
+				continue
+			}
+			merged[key] = value
+		}
+		profile.CustomHeaders = merged
+	} else if strings.EqualFold(strings.TrimSpace(descriptor.ID), "aimlapi") && !canonicalCatalogEndpoint {
+		// AIMLAPI attribution is owned by the catalog endpoint. A profile can retain
+		// those generated headers after its base URL is edited; strip their names
+		// before sending requests to an arbitrary staging/proxy host while preserving
+		// unrelated headers explicitly supplied by the user.
+		for profileKey := range profile.CustomHeaders {
+			for catalogKey := range descriptor.CustomHeaders {
+				if strings.EqualFold(profileKey, catalogKey) {
+					delete(profile.CustomHeaders, profileKey)
+					break
+				}
+			}
+		}
+	}
 }

 func sameBaseURL(left string, right string) bool {
@@ -9,6 +9,7 @@ import (
 	"testing"

 	"github.com/Gitlawb/zero/internal/modelregistry"
+	"github.com/Gitlawb/zero/internal/providercatalog"
 )

 func TestResolveAppliesLayerPrecedence(t *testing.T) {
@@ -554,7 +555,7 @@ func TestResolveReplacesMCPServerOverlayCollections(t *testing.T) {
 	}
 }

-func TestResolveMCPServerLayersCanClearAndReenable(t *testing.T) {
+func TestResolveMCPServerLayersCannotReenableUserDisabled(t *testing.T) {
 	userPath := writeConfig(t, `{
 		"mcp": {
 			"servers": {
@@ -588,9 +589,11 @@ func TestResolveMCPServerLayersCanClearAndReenable(t *testing.T) {
 	}

 	docs := resolved.MCP.Servers["docs"]
-	if docs.Disabled {
-		t.Fatalf("docs.Disabled = true, want project layer to re-enable")
+	// A user-level disable is sticky: project config must not re-enable it.
+	if !docs.Disabled {
+		t.Fatal("docs.Disabled = false, want user-level disable to remain sticky")
 	}
+	// Non-disable fields from the project layer still merge normally.
 	if len(docs.Args) != 0 {
 		t.Fatalf("docs.Args = %#v, want project layer to clear inherited args", docs.Args)
 	}
@@ -599,6 +602,125 @@ func TestResolveMCPServerLayersCanClearAndReenable(t *testing.T) {
 	}
 }

+func TestMergeMCPStickyDisableReenable(t *testing.T) {
+	disabled := MCPServerConfig{Disabled: true, disabledSet: true}
+
+	// A lower-trust (project) layer must not lift a sticky user disable.
+	project := MCPServerConfig{Disabled: false, disabledSet: true}
+	if got := mergeMCPServer(disabled, project, false); !got.Disabled {
+		t.Fatal("project layer re-enabled a sticky user disable")
+	}
+
+	// The user scope (explicit mcp enable command / CLI override) may re-enable.
+	override := MCPServerConfig{Disabled: false, disabledSet: true}
+	if got := mergeMCPServer(disabled, override, true); got.Disabled {
+		t.Fatal("user-scope override must be able to re-enable")
+	}
+
+	// A user-scope disable stacked on a project attempt still sticks.
+	if got := mergeMCPServer(disabled, project, false); !got.Disabled {
+		t.Fatal("sticky disable must survive a project re-enable attempt")
+	}
+
+	// Baseline: a project layer may disable a default-enabled server when no
+	// higher-trust scope has explicitly set disabled (the trust boundary must
+	// not block legitimate project disables). Regression guard for the
+	// disabledSet-before-check ordering bug.
+	defaultEnabled := MCPServerConfig{Disabled: false, disabledSet: false}
+	projectDisable := MCPServerConfig{Disabled: true, disabledSet: true}
+	if got := mergeMCPServer(defaultEnabled, projectDisable, false); !got.Disabled {
+		t.Fatal("project layer should be able to disable a default-enabled server")
+	}
+
+	// And an empty project layer must not flip an unconfigured server to disabled.
+	if got := mergeMCPServer(MCPServerConfig{}, MCPServerConfig{}, false); got.Disabled {
+		t.Fatal("an empty project layer must not disable an unconfigured server")
+	}
+}
+
+// TestResolveMCPCannotReenableUserDisabled exercises the trust boundary at the
+// actual ResolveMCP entry point: a project config (project scope) must not lift
+// a disable the user set in their higher-trust user config.
+func TestResolveMCPCannotReenableUserDisabled(t *testing.T) {
+	userPath := writeConfig(t, `{
+		"mcp": {
+			"servers": {
+				"docs": {
+					"type": "stdio",
+					"command": "docs-mcp",
+					"disabled": true
+				}
+			}
+		}
+	}`)
+	// Project config tries to re-enable it.
+	projectPath := writeConfig(t, `{
+		"mcpServers": {
+			"docs": {
+				"disabled": false
+			}
+		}
+	}`)
+
+	resolved, err := ResolveMCP(ResolveOptions{
+		UserConfigPath:    userPath,
+		ProjectConfigPath: projectPath,
+	})
+	if err != nil {
+		t.Fatalf("ResolveMCP() error = %v", err)
+	}
+
+	docs, ok := resolved.Servers["docs"]
+	if !ok {
+		t.Fatal("docs server missing from resolved config")
+	}
+	if !docs.Disabled {
+		t.Fatalf("docs.Disabled = false, want user-level disable to remain sticky across project re-enable")
+	}
+}
+
+// TestResolveMCPUserLiftsProjectDisabled asserts the reverse direction is open:
+// a user config (higher-trust scope) re-enabling a server the project config
+// disabled must be permitted, since user > project in the trust hierarchy.
+func TestResolveMCPUserLiftsProjectDisabled(t *testing.T) {
+	// Project config disables the server.
+	projectPath := writeConfig(t, `{
+		"mcpServers": {
+			"docs": {
+				"type": "stdio",
+				"command": "docs-mcp",
+				"disabled": true
+			}
+		}
+	}`)
+	// User config re-enables it.
+	userPath := writeConfig(t, `{
+		"mcp": {
+			"servers": {
+				"docs": {
+					"disabled": false
+				}
+			}
+		}
+	}`)
+
+	resolved, err := ResolveMCP(ResolveOptions{
+		UserConfigPath:    userPath,
+		ProjectConfigPath: projectPath,
+	})
+	if err != nil {
+		t.Fatalf("ResolveMCP() error = %v", err)
+	}
+
+	docs, ok := resolved.Servers["docs"]
+	if !ok {
+		t.Fatal("docs server missing from resolved config")
+	}
+	if docs.Disabled {
+		t.Fatalf("docs.Disabled = true, want user scope to lift the project-level disable")
+	}
+}
+
 func TestResolveRejectsDuplicateMCPRootAliases(t *testing.T) {
 	path := writeConfig(t, `{
 		"mcpServers": {
@@ -1374,6 +1496,71 @@ func TestResolveProviderProfileExtendedJSONAliases(t *testing.T) {
 	}
 }

+func TestApplyCatalogDescriptorFoldsCustomHeadersCaseInsensitively(t *testing.T) {
+	t.Setenv("AIMLAPI_PARTNER_ID", "part_env")
+	descriptor, err := providercatalog.Require("aimlapi")
+	if err != nil {
+		t.Fatal(err)
+	}
+	profile := ProviderProfile{
+		BaseURL: descriptor.DefaultBaseURL,
+		CustomHeaders: map[string]string{
+			"x-aimlapi-partner-id": "part_override",
+		},
+	}
+
+	applyCatalogDescriptor(&profile, descriptor, true)
+
+	if got := profile.CustomHeaders["X-AIMLAPI-Partner-ID"]; got != "part_override" {
+		t.Fatalf("canonical header value = %q, want override", got)
+	}
+	if _, duplicate := profile.CustomHeaders["x-aimlapi-partner-id"]; duplicate {
+		t.Fatalf("case-colliding header survived merge: %#v", profile.CustomHeaders)
+	}
+}
+
+func TestApplyCatalogDescriptorUsesAimlapiPartnerEnvironmentOverride(t *testing.T) {
+	t.Setenv("AIMLAPI_PARTNER_ID", "part_env")
+	descriptor, err := providercatalog.Require("aimlapi")
+	if err != nil {
+		t.Fatal(err)
+	}
+	profile := ProviderProfile{BaseURL: descriptor.DefaultBaseURL}
+
+	applyCatalogDescriptor(&profile, descriptor, true)
+
+	if got := profile.CustomHeaders["X-AIMLAPI-Partner-ID"]; got != "part_env" {
+		t.Fatalf("partner header = %q, want part_env", got)
+	}
+}
+
+func TestApplyCatalogDescriptorStripsAimlapiAttributionFromRetargetedProfile(t *testing.T) {
+	descriptor, err := providercatalog.Require("aimlapi")
+	if err != nil {
+		t.Fatal(err)
+	}
+	profile := ProviderProfile{
+		BaseURL: "https://proxy.example.test/v1",
+		CustomHeaders: map[string]string{
+			"x-aimlapi-partner-id":          "persisted-partner",
+			"X-AIMLAPI-Integration-Repo":    "Gitlawb/zero",
+			"X-AIMLAPI-Integration-Version": "zero",
+			"X-Environment":                 "staging",
+		},
+	}
+
+	applyCatalogDescriptor(&profile, descriptor, true)
+
+	for key := range profile.CustomHeaders {
+		if strings.HasPrefix(strings.ToLower(key), "x-aimlapi-") {
+			t.Fatalf("catalog attribution survived retargeting: %#v", profile.CustomHeaders)
+		}
+	}
+	if profile.CustomHeaders["X-Environment"] != "staging" {
+		t.Fatalf("user header was removed: %#v", profile.CustomHeaders)
+	}
+}
+
 func TestResolveProviderProfileParseThinkTagsFalseAlias(t *testing.T) {
 	path := writeConfig(t, `{
 		"activeProvider": "custom",
@@ -0,0 +1,71 @@
+# Turn benchmark manifest
+
+The baseline manifest (`manifests/baseline.json`) is the per-turn benchmark's
+program keystone: it defines the tasks the harness runs, the workspace each
+starts in, and — critically — what "pass" means for each task. Because `make
+baseline` is re-run on every perf change, the manifest's pass/fail contract
+matters for months, so the contract is written down here.
+
+## Task count
+
+48 tasks across seven classes:
+
+| class     | count | oracle                      | tier        |
+|-----------|-------|-----------------------------|-------------|
+| nav       | 10    | none                        | latency     |
+| edit      | 10    | substring grep              | correctness |
+| fix       | 8     | scoped `go test -run <name>` | correctness |
+| refactor  | 6     | `go build ./...`            | build       |
+| longproc  | 4     | none                        | latency     |
+| longctx   | 4     | none                        | latency     |
+| parallel  | 6     | none                        | latency     |
+| **total** | **48**|                             |             |
+
+## Oracle tiers
+
+Pass/fail is reported per tier so the report cannot be misread as a blanket
+correctness verdict. The tier is decided per task from oracle presence and the
+manifest's `buildOnlyClasses` list:
+
+- **Correctness** (18 tasks: 10 edit + 8 fix) — a positive oracle (substring
+  grep that the requested change landed, or a scoped `go test -run` that the
+  bug fix passes). This is the only pass rate that can move with model quality:
+  `tasksVerified` / `tasksPassed` / `correctnessPassRate`.
+- **Build-only** (6 tasks: refactor) — `go build ./...` proves the edit
+  compiles, not that the refactor achieved its goal (a no-op refactor passes).
+  The manifest declares `buildOnlyClasses: ["refactor"]`. Reported as
+  `buildCheckedTasks` / `buildPassedTasks` / `buildPassRate`, never in
+  `correctnessPassRate`.
+- **Latency-only** (24 tasks: nav, longproc, longctx, parallel) — no
+  `verificationCommand`. An exit 0 only proves the turn ran, not that the answer
+  was right. They contribute to latency and span attribution only and are
+  counted in `latencyOnlyTasks`, never in any pass rate.
+
+A task's tier is driven by oracle **presence** first: a task with no
+`verificationCommand` is latency-only even if its class is listed in
+`buildOnlyClasses`, so a missing oracle can never silently pass on exit 0.
+
+## Known limitations (deferred)
+
+These are accepted for the Phase 0 baseline and tracked in a follow-up issue;
+they do not block the baseline because the tier split keeps the report honest:
+
+- The edit grep oracles are substring checks. A rename oracle asserts the new
+  name is present AND the old name is gone (compound `bash -c`), but an
+  add-field oracle (`grep -R Label .`) only proves the string appears, not that
+  it landed on the right struct. Strengthening these to `go vet`/`go build` +
+  structural assertions is follow-up work.
+- The refactor build oracle is non-positive (a no-op refactor passes). A
+  structural verifier that proves the refactor happened is follow-up work;
+  until then refactor is `buildPassRate`, not `correctnessPassRate`.
+- The 24 read-only tasks have no oracle. Adding deterministic oracles for nav
+  (diff against an expected answer) and documenting longproc/longctx/parallel
+  as permanently latency-only is follow-up work.
+
+## Fixtures
+
+Each task's `workspaceFixture` points at a small self-contained workspace under
+`testdata/` so the suite runs offline and repeatably. Mutating tasks (edit,
+fix, refactor) run against a per-invocation **copy** of their fixture, so the
+checked-in fixtures stay clean and one task's edits can't bleed into the next
+iteration or a later task.
@@ -0,0 +1,62 @@
+{
+  "id": "zero-baseline-turn",
+  "name": "Zero per-turn baseline (Phase 0)",
+  "description": "Latency-first baseline: measures where a turn spends wall time, with pass/fail reported per oracle tier so it cannot be misread as a blanket correctness verdict. Read-only classes (nav, longproc, longctx, parallel) carry no verification oracle — a zero-exit only proves the turn ran, so they count as latency-only and never in any pass rate. Correctness classes (edit, fix) carry a positive oracle (substring grep or a scoped `go test`) and are the only tasks in correctnessPassRate. Build-only classes (refactor) carry `go build ./...`, which proves the edit compiles but not that the refactor achieved its goal, so they are reported as buildPassRate and excluded from correctnessPassRate.",
+  "buildOnlyClasses": ["refactor"],
+  "tasks": [
+    {"id": "nav-01", "name": "list repo files", "class": "nav", "prompt": "List the files in this repository and report their count. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+    {"id": "nav-02", "name": "find a function", "class": "nav", "prompt": "Find the function named 'greet' and report the file and line it is defined on. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+    {"id": "nav-03", "name": "summarize module", "class": "nav", "prompt": "Read main.go and summarize what this program does in two sentences. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+    {"id": "nav-04", "name": "count tests", "class": "nav", "prompt": "Count the test functions (functions whose names start with Test) in this repository and report the number. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+    {"id": "nav-05", "name": "locate a string", "class": "nav", "prompt": "Find every occurrence of the string 'TODO' in the repository and report the files. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+    {"id": "nav-06", "name": "describe package", "class": "nav", "prompt": "Describe the exported API of this package by reading the source. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+    {"id": "nav-07", "name": "trace a call", "class": "nav", "prompt": "Trace which function calls which other function starting from main. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+    {"id": "nav-08", "name": "find imports", "class": "nav", "prompt": "List every third-party import used by this repository. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+    {"id": "nav-09", "name": "report config", "class": "nav", "prompt": "Read config.json and report the keys it contains. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+    {"id": "nav-10", "name": "doc lookup", "class": "nav", "prompt": "Find the doc comment on the greet function and quote it. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+
+    {"id": "edit-01", "name": "rename a constant", "class": "edit", "prompt": "Rename the constant MaxRetries to RetryLimit everywhere it appears.", "workspaceFixture": "../testdata/edit", "verificationCommand": ["bash", "-c", "grep -R RetryLimit . && ! grep -R MaxRetries ."]},
+    {"id": "edit-02", "name": "fix a typo", "class": "edit", "prompt": "Fix the typo in the comment above the greet function ('receieves' should be 'receives').", "workspaceFixture": "../testdata/edit", "verificationCommand": ["bash", "-c", "! grep -R receieves ."]},
+    {"id": "edit-03", "name": "add a field", "class": "edit", "prompt": "Add a string field named Label to the Config struct in main.go.", "workspaceFixture": "../testdata/edit", "verificationCommand": ["grep", "-R", "Label", "."]},
+    {"id": "edit-04", "name": "bump a version", "class": "edit", "prompt": "Change the version string in version.go from 1.0.0 to 1.1.0.", "workspaceFixture": "../testdata/edit", "verificationCommand": ["bash", "-c", "grep -R 1.1.0 . && ! grep -R 1.0.0 ."]},
+    {"id": "edit-05", "name": "remove a log line", "class": "edit", "prompt": "Remove the line that prints 'debug: starting'.", "workspaceFixture": "../testdata/edit", "verificationCommand": ["bash", "-c", "! grep -R 'debug: starting' ."]},
+    {"id": "edit-06", "name": "add a license header", "class": "edit", "prompt": "Add a comment line '// SPDX-License-Identifier: MIT' at the top of main.go.", "workspaceFixture": "../testdata/edit", "verificationCommand": ["grep", "-R", "SPDX-License-Identifier", "."]},
+    {"id": "edit-07", "name": "change a default", "class": "edit", "prompt": "Change the default value of the Port constant from 8080 to 9090.", "workspaceFixture": "../testdata/edit", "verificationCommand": ["bash", "-c", "grep -R 9090 . && ! grep -R 8080 ."]},
+    {"id": "edit-08", "name": "add a getter", "class": "edit", "prompt": "Add a method func (c *Config) GetLabel() string that returns c.Label.", "workspaceFixture": "../testdata/edit", "verificationCommand": ["bash", "-c", "grep -R 'func (c *Config) GetLabel' ."]},
+    {"id": "edit-09", "name": "wrap an error", "class": "edit", "prompt": "In the load function, wrap the returned error with fmt.Errorf so it includes 'load failed'.", "workspaceFixture": "../testdata/edit", "verificationCommand": ["grep", "-R", "load failed", "."]},
+    {"id": "edit-10", "name": "update a message", "class": "edit", "prompt": "Change the printed greeting from 'hello' to 'hello, world'.", "workspaceFixture": "../testdata/edit", "verificationCommand": ["grep", "-R", "hello, world", "."]},
+
+    {"id": "fix-01", "name": "off-by-one", "class": "fix", "prompt": "In bugs.go the Sum function has an off-by-one error that skips the last element; fix the loop so it sums all elements.", "workspaceFixture": "../testdata/fix", "verificationCommand": ["go", "test", "-run", "TestSumIncludesLastElement", "./..."]},
+    {"id": "fix-02", "name": "nil deref", "class": "fix", "prompt": "In bugs.go DefaultConfig returns nil and Port() dereferences it; make DefaultConfig return a valid Config with Port 8080.", "workspaceFixture": "../testdata/fix", "verificationCommand": ["go", "test", "-run", "TestPortReturnsDefault", "./..."]},
+    {"id": "fix-03", "name": "wrong operator", "class": "fix", "prompt": "In bugs.go Max uses the wrong inequality and returns the smaller value; fix it to return the larger.", "workspaceFixture": "../testdata/fix", "verificationCommand": ["go", "test", "-run", "TestMaxReturnsLarger", "./..."]},
+    {"id": "fix-04", "name": "truncated read", "class": "fix", "prompt": "In bugs.go ReadFirst returns only a single character instead of the full content; fix it to return the whole string.", "workspaceFixture": "../testdata/fix", "verificationCommand": ["go", "test", "-run", "TestReadFirstReturnsFullContent", "./..."]},
+    {"id": "fix-05", "name": "swapped args", "class": "fix", "prompt": "In bugs.go Divide swaps its operands; fix the order so Divide(10,2) returns 5.", "workspaceFixture": "../testdata/fix", "verificationCommand": ["go", "test", "-run", "TestDivideOrder", "./..."]},
+    {"id": "fix-06", "name": "missing return", "class": "fix", "prompt": "In bugs.go Classify is missing a return for the high case; add it so Classify(500) returns \"high\".", "workspaceFixture": "../testdata/fix", "verificationCommand": ["go", "test", "-run", "TestClassifyHigh", "./..."]},
+    {"id": "fix-07", "name": "wrong label prefix", "class": "fix", "prompt": "In bugs.go Label returns 'admin-<name>' but the test expects 'user-<name>'; change the prefix from 'admin-' to 'user-'.", "workspaceFixture": "../testdata/fix", "verificationCommand": ["go", "test", "-run", "TestLabelUsesString", "./..."]},
+    {"id": "fix-08", "name": "race on counter", "class": "fix", "prompt": "In bugs.go the global counter is incremented without synchronization; guard it with a mutex.", "workspaceFixture": "../testdata/fix", "verificationCommand": ["go", "test", "-race", "-run", "TestCounterRace", "./..."]},
+
+    {"id": "refactor-01", "name": "extract helper", "class": "refactor", "prompt": "Extract the duplicated greeting logic into a single helper function and call it from both sites.", "workspaceFixture": "../testdata/refactor", "verificationCommand": ["go", "build", "./..."]},
+    {"id": "refactor-02", "name": "split a file", "class": "refactor", "prompt": "Split main.go into two files: one for Config, one for the rest, keeping the package building.", "workspaceFixture": "../testdata/refactor", "verificationCommand": ["go", "build", "./..."]},
+    {"id": "refactor-03", "name": "rename package", "class": "refactor", "prompt": "Rename the package from app to zeroapp everywhere and keep it building.", "workspaceFixture": "../testdata/refactor", "verificationCommand": ["go", "build", "./..."]},
+    {"id": "refactor-04", "name": "introduce a type", "class": "refactor", "prompt": "Introduce a named type for the map[string]int used in main.go and update its users.", "workspaceFixture": "../testdata/refactor", "verificationCommand": ["go", "build", "./..."]},
+    {"id": "refactor-05", "name": "inline a wrapper", "class": "refactor", "prompt": "Inline the thin wrapper function into its single caller and remove it.", "workspaceFixture": "../testdata/refactor", "verificationCommand": ["go", "build", "./..."]},
+    {"id": "refactor-06", "name": "consolidate errors", "class": "refactor", "prompt": "Consolidate the repeated error-construction blocks into a single helper.", "workspaceFixture": "../testdata/refactor", "verificationCommand": ["go", "build", "./..."]},
+
+    {"id": "longproc-01", "name": "long build", "class": "longproc", "prompt": "Run the build command and report whether it succeeds. Do not modify anything.", "workspaceFixture": "../testdata/longproc"},
+    {"id": "longproc-02", "name": "long test", "class": "longproc", "prompt": "Run go test ./... and report the result. Do not modify anything.", "workspaceFixture": "../testdata/longproc"},
+    {"id": "longproc-03", "name": "bench run", "class": "longproc", "prompt": "Run go test -bench=. -benchtime=1x and report the benchmark numbers. Do not modify anything.", "workspaceFixture": "../testdata/longproc"},
+    {"id": "longproc-04", "name": "vet run", "class": "longproc", "prompt": "Run go vet ./... and report the output. Do not modify anything.", "workspaceFixture": "../testdata/longproc"},
+
+    {"id": "longctx-01", "name": "summarize large file", "class": "longctx", "prompt": "Read the large generated file big.go and summarize its structure. Do not modify anything.", "workspaceFixture": "../testdata/longctx"},
+    {"id": "longctx-02", "name": "find in large file", "class": "longctx", "prompt": "In big.go, find every function that returns an error and list them. Do not modify anything.", "workspaceFixture": "../testdata/longctx"},
+    {"id": "longctx-03", "name": "cross-file question", "class": "longctx", "prompt": "Read all .go files and report which two functions share the same name across files. Do not modify anything.", "workspaceFixture": "../testdata/longctx"},
+    {"id": "longctx-04", "name": "trace through large file", "class": "longctx", "prompt": "Starting from Handle, trace the call path through big.go and report the chain. Do not modify anything.", "workspaceFixture": "../testdata/longctx"},
+
+    {"id": "parallel-01", "name": "read six files", "class": "parallel", "prompt": "Read a.txt, b.txt, c.txt, d.txt, e.txt, and f.txt, then report the first line of each. Do not modify anything.", "workspaceFixture": "../testdata/parallel"},
+    {"id": "parallel-02", "name": "grep six patterns", "class": "parallel", "prompt": "Search for 'alpha', 'beta', 'gamma', 'delta', 'epsilon', and 'zeta' across the repo and report which files contain each. Do not modify anything.", "workspaceFixture": "../testdata/parallel"},
+    {"id": "parallel-03", "name": "list six dirs", "class": "parallel", "prompt": "List the contents of dir1, dir2, dir3, dir4, dir5, and dir6. Do not modify anything.", "workspaceFixture": "../testdata/parallel"},
+    {"id": "parallel-04", "name": "stat six files", "class": "parallel", "prompt": "Report the line count of a.txt, b.txt, c.txt, d.txt, e.txt, and f.txt. Do not modify anything.", "workspaceFixture": "../testdata/parallel"},
+    {"id": "parallel-05", "name": "read configs", "class": "parallel", "prompt": "Read config1.json, config2.json, config3.json, config4.json, config5.json, and config6.json and report each top-level key. Do not modify anything.", "workspaceFixture": "../testdata/parallel"},
+    {"id": "parallel-06", "name": "summarize six", "class": "parallel", "prompt": "Read the six .txt files and give a one-line summary of each. Do not modify anything.", "workspaceFixture": "../testdata/parallel"}
+  ]
+}
@@ -0,0 +1,84 @@
+# Per-turn benchmark reports
+
+This directory holds generated per-turn benchmark reports. It is **not** a
+checked-in source of truth for performance — the numbers are machine- and
+model-specific, so a single snapshot here would mislead rather than inform.
+
+## Generating a baseline
+
+Run the harness over the checked-in manifest:
+
+```sh
+make baseline ZERO_BENCH_MODEL=<model>              # uses ./zero
+make baseline ZERO_BENCH_MODEL=<model> ZERO_BENCH_BINARY=/path/to/zero
+```
+
+This builds `zero`, then runs `zero-perf-bench turn` over
+`internal/perfbench/manifests/baseline.json`, capturing each turn's trace
+(`zero exec --trace <tmpfile>`) and writing the aggregated result to
+`reports/baseline.json`.
+
+The JSON report is self-describing: model, mode, self-correct flag, version,
+commit, date, per-span median/P95, the **top three controllable latency sources**
+ranked by **exclusive** time, per-class roll-ups, and token/count totals.
+That top-three list is the Phase 0 baseline's "do not proceed until" criterion —
+it names where a turn actually spends time so later optimization work is
+targeted, not guessed.
+
+### Attribution model (honest by construction)
+
+Spans record wall intervals and are **not** summed into each other. Each span's
+**exclusive** time is its own duration minus the union of its nested children's
+intervals, derived at finish by interval containment. So a `provider_connect`
+that runs concurrently inside `generation`, or a `permission_wait` nested inside
+`tool_execution`, each contributes only its own exclusive time — they no longer
+double-count the same wall. The top-latency shares therefore sum to ~1 for a
+well-instrumented run, and the ranking reflects where wall time is actually
+spent.
+
+**Coverage** is the fraction of wall covered by the union of all span
+intervals (capped at 1.0) — the honest "≥95% of wall accounted for" metric. A
+run with `coverage < 0.95` has uninstrumented gaps, not an inflated attribution.
+
+### Pass/fail is reported per oracle tier
+
+Pass/fail is split into three tiers so it cannot be misread as a blanket
+correctness verdict (see `MANIFEST.md` for the class breakdown):
+
+- **Correctness** (`tasksVerified` / `tasksPassed` / `correctnessPassRate`):
+  tasks with a positive oracle — edit's substring grep, fix's scoped `go test`.
+  This is the only pass rate that can move with model quality.
+- **Build-only** (`buildCheckedTasks` / `buildPassedTasks` / `buildPassRate`):
+  refactor's `go build ./...`, which proves the edit compiles but not that the
+  refactor achieved its goal. Reported separately, never in `correctnessPassRate`.
+- **Latency-only** (`latencyOnlyTasks`): the read-only classes (nav, longproc,
+  longctx, parallel) carry no oracle — an exit 0 only proves the turn ran. They
+  contribute to latency and span attribution and are excluded from every pass
+  rate.
+
+`tasksAttempted` is still the total across all three tiers. The tier class lists
+are echoed in the report so a consumer can see exactly which classes each rate
+is computed over.
+
+> **Do not average the tier pass rates. Do not report a single "pass rate".**
+> `correctnessPassRate` and `buildPassRate` measure different things over
+> different task sets and must never be combined — a weighted or arithmetic
+> mean of the two is a number that means nothing. The schema offers no
+> headline pass rate on purpose: a consumer must name the tier it is reporting
+> (`correctness`, `build`, or `latency-only`) and quote that tier's fields. A
+> read-only run that exits 0 is a latency sample, not a pass.
+
+## What to commit
+
+Commit the manifest and fixtures (`manifests/`, `../testdata/`), not a generated
+`baseline.json`. A generated report belongs in a PR description or a shared
+dashboard as evidence for one configuration, not in the tree as a durable
+expectation. `.gitkeep` keeps the directory present between runs.
+
+## Caveats
+
+- Each task is a fresh `zero exec` process, so iterations are **cold-start**
+  samples. A warm path (reusing an in-process agent) is future work.
+- Mutating tasks (edit/fix/refactor) run against a per-invocation **copy** of
+  their fixture, so the checked-in fixtures stay clean and one task's edits
+  can't bleed into the next iteration.
@@ -25,17 +25,29 @@ const TaskSchemaVersion = 1
 // ZERO must satisfy. The set is recorded by ID with every result so a published
 // number is traceable to the exact tasks that produced it.
 type TaskSet struct {
-	ID    string      `json:"id"`
-	Name  string      `json:"name,omitempty"`
-	Tasks []BenchTask `json:"tasks"`
+	ID          string      `json:"id"`
+	Name        string      `json:"name,omitempty"`
+	Description string      `json:"description,omitempty"`
+	Tasks       []BenchTask `json:"tasks"`
+	// BuildOnlyClasses lists task classes whose verificationCommand is a
+	// non-positive build check (e.g. refactor's `go build ./...`): it proves the
+	// edit compiles, not that the refactor achieved its goal. The turn benchmark
+	// reports these separately from correctness oracles so a build-pass cannot be
+	// misread as a correctness pass. Classes with a verificationCommand that are
+	// NOT listed here are treated as correctness classes; classes whose tasks
+	// carry no verificationCommand are latency-only regardless of this list.
+	BuildOnlyClasses []string `json:"buildOnlyClasses,omitempty"`
 }

 // BenchTask is one benchmark task. WorkspaceFixture is the relative path of the
 // task's starting workspace; VerificationCommand (optional) is the command the
-// default runner executes to decide pass/fail after ZERO finishes.
+// default runner executes to decide pass/fail after ZERO finishes. Class groups
+// the task for the turn-benchmark's per-group latency breakdown (e.g. "nav",
+// "edit", "fix"); it is optional and ignored by the pass/fail runner.
 type BenchTask struct {
 	ID                  string   `json:"id"`
 	Name                string   `json:"name,omitempty"`
+	Class               string   `json:"class,omitempty"`
 	Prompt              string   `json:"prompt"`
 	WorkspaceFixture    string   `json:"workspaceFixture,omitempty"`
 	VerificationCommand []string `json:"verificationCommand,omitempty"`
@@ -0,0 +1,42 @@
+// Package edit is a small workspace for the edit-class benchmark tasks: each
+// task asks the agent to make one targeted edit, verified by a grep/build
+// command in the manifest.
+package edit
+
+import (
+	"fmt"
+	"os"
+)
+
+// Port is the default listen port.
+const Port = 8080
+
+// MaxRetries is the maximum number of retries before giving up.
+const MaxRetries = 3
+
+// Config holds runtime configuration.
+type Config struct {
+	Name string
+}
+
+// greet returns a greeting for the named user.
+//
+// This receieves a name and formats a hello message.
+func greet(name string) string {
+	return "hello"
+}
+
+// load reads a config file and returns its bytes, or an error on failure.
+func load(path string) ([]byte, error) {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return nil, err
+	}
+	return data, nil
+}
+
+func main() {
+	fmt.Println("debug: starting")
+	cfg := Config{Name: "demo"}
+	fmt.Println(greet(cfg.Name))
+}
@@ -0,0 +1,4 @@
+package edit
+
+// Version is the build version string.
+const Version = "1.0.0"
@@ -0,0 +1,89 @@
+// Package fix holds deliberately buggy functions for the bug-fix benchmark
+// tasks. Each function is covered by its own test (see bugs_test.go) and each
+// task verifies with `go test -run <TestName>`, so the tasks are independent:
+// the initial package compiles and the targeted test fails until the named bug
+// is fixed. Every bug is fixable without changing a function signature, so the
+// test files compile in both the buggy and the fixed state.
+package fix
+
+import "fmt"
+
+// Sum returns the sum of all elements. BUG: off-by-one — skips the last element.
+func Sum(xs []int) int {
+	total := 0
+	for i := 0; i < len(xs)-1; i++ {
+		total += xs[i]
+	}
+	return total
+}
+
+// DefaultConfig returns the default configuration. BUG: returns nil, which
+// Port() then dereferences.
+func DefaultConfig() *Config {
+	return nil
+}
+
+// Config holds demo settings.
+type Config struct {
+	Port int
+}
+
+// Port returns the configured port. BUG: panics when DefaultConfig() is nil.
+func Port() int {
+	return DefaultConfig().Port
+}
+
+// Max returns the larger of a and b. BUG: uses the wrong inequality and returns
+// the smaller value.
+func Max(a, b int) int {
+	if a < b {
+		return a
+	}
+	return b
+}
+
+// ReadFirst returns the first byte of the given file's content as a string. BUG:
+// returns only a single character instead of the full content.
+func ReadFirst(content string) string {
+	if len(content) == 0 {
+		return ""
+	}
+	return string(content[0])
+}
+
+// Divide returns a divided by b. BUG: the operands are swapped.
+func Divide(a, b int) int {
+	return b / a
+}
+
+// Classify returns "low", "mid", or "high". BUG: the "high" branch is missing a
+// return, so high values fall through to the default "".
+func Classify(n int) string {
+	if n < 10 {
+		return "low"
+	}
+	if n < 100 {
+		return "mid"
+	}
+	// BUG: missing return for the high case
+	return ""
+}
+
+// Label returns a display label for the given name. BUG: uses the wrong prefix
+// ("admin-" instead of "user-"), producing the wrong text.
+func Label(name string) string {
+	return fmt.Sprintf("admin-%s", name)
+}
+
+// counter is a shared, unsynchronized counter. BUG: Inc is not goroutine-safe.
+var counter int
+
+// Inc increments the shared counter.
+func Inc() {
+	counter++
+}
+
+// Count returns the current counter value.
+func Count() int {
+	return counter
+}
@@ -0,0 +1,68 @@
+package fix
+
+import (
+	"sync"
+	"testing"
+)
+
+func TestSumIncludesLastElement(t *testing.T) {
+	if got := Sum([]int{1, 2, 3, 4}); got != 10 {
+		t.Fatalf("Sum = %d, want 10", got)
+	}
+}
+
+func TestPortReturnsDefault(t *testing.T) {
+	if got := Port(); got != 8080 {
+		t.Fatalf("Port() = %d, want 8080 (DefaultConfig returned nil)", got)
+	}
+}
+
+func TestMaxReturnsLarger(t *testing.T) {
+	if got := Max(3, 7); got != 7 {
+		t.Fatalf("Max(3,7) = %d, want 7", got)
+	}
+}
+
+func TestReadFirstReturnsFullContent(t *testing.T) {
+	if got := ReadFirst("hello"); got != "hello" {
+		t.Fatalf("ReadFirst = %q, want %q", got, "hello")
+	}
+}
+
+func TestDivideOrder(t *testing.T) {
+	if got := Divide(10, 2); got != 5 {
+		t.Fatalf("Divide(10,2) = %d, want 5", got)
+	}
+}
+
+func TestClassifyHigh(t *testing.T) {
+	if got := Classify(500); got != "high" {
+		t.Fatalf("Classify(500) = %q, want %q", got, "high")
+	}
+}
+
+func TestLabelUsesString(t *testing.T) {
+	if got := Label("alice"); got != "user-alice" {
+		t.Fatalf("Label = %q, want %q", got, "user-alice")
+	}
+}
+
+// TestCounterRace exercises Inc concurrently. Run with `go test -race`:
+// the unsynchronized counter trips the race detector until a mutex is added.
+func TestCounterRace(t *testing.T) {
+	counter = 0
+	var wg sync.WaitGroup
+	for i := 0; i < 200; i++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			for j := 0; j < 100; j++ {
+				Inc()
+			}
+		}()
+	}
+	wg.Wait()
+	if got := Count(); got != 20000 {
+		t.Fatalf("Count = %d, want 20000", got)
+	}
+}
@@ -0,0 +1,21 @@
+// Package longproc is a small buildable workspace for the long-running-process
+// benchmark tasks, which run build/test/vet/bench commands against it.
+package longproc
+
+import "fmt"
+
+// Version is the demo build's version string.
+const Version = "1.0.0"
+
+// Process simulates a bounded unit of work for the benchmark task.
+func Process(n int) int {
+	total := 0
+	for i := 0; i < n; i++ {
+		total += i
+	}
+	return total
+}
+
+func main() {
+	fmt.Println("longproc fixture", Version, Process(100))
+}
@@ -0,0 +1,21 @@
+package longproc
+
+import "testing"
+
+// benchSink keeps Process(100) live across the benchmark loop so the compiler
+// can't optimize the call away and report an unrealistically fast benchmark.
+var benchSink int
+
+func TestProcess(t *testing.T) {
+	if got := Process(100); got != 4950 {
+		t.Fatalf("Process(100) = %d, want 4950", got)
+	}
+}
+
+func BenchmarkProcess(b *testing.B) {
+	var sink int
+	for i := 0; i < b.N; i++ {
+		sink = Process(100)
+	}
+	benchSink = sink
+}
@@ -0,0 +1,3 @@
+# nav fixture
+
+Small read-only workspace for navigation/classification benchmark tasks.
@@ -0,0 +1,5 @@
+{
+  "port": 8080,
+  "name": "nav-fixture",
+  "retries": 3
+}
@@ -0,0 +1,24 @@
+// Package nav is a small read-only fixture for navigation tasks.
+package nav
+
+import "fmt"
+
+// MaxRetries is the configured retry limit for the demo client.
+const MaxRetries = 3
+
+// Config holds the demo client's settings.
+type Config struct {
+	Port int
+	Name string
+}
+
+// greet returns a greeting for the given name.
+// receieves is a deliberate typo present in the fixture (see edit tasks).
+func greet(name string) string {
+	return fmt.Sprintf("hello, %s", name)
+}
+
+// main is the fixture entry point.
+func main() {
+	fmt.Println(greet("world"))
+}
@@ -0,0 +1,4 @@
+first line of a.txt
+second line of a.txt
+third line of a.txt
+alpha alpha alpha
@@ -0,0 +1,4 @@
+first line of b.txt
+second line of b.txt
+third line of b.txt
+beta beta beta
@@ -0,0 +1,4 @@
+first line of c.txt
+second line of c.txt
+third line of c.txt
+gamma gamma gamma
@@ -0,0 +1 @@
+{ "id": 1, "key": "value-1", "enabled": true }
@@ -0,0 +1 @@
+{ "id": 2, "key": "value-2", "enabled": true }
@@ -0,0 +1 @@
+{ "id": 3, "key": "value-3", "enabled": true }
@@ -0,0 +1 @@
+{ "id": 4, "key": "value-4", "enabled": true }
@@ -0,0 +1 @@
+{ "id": 5, "key": "value-5", "enabled": true }
@@ -0,0 +1 @@
+{ "id": 6, "key": "value-6", "enabled": true }
@@ -0,0 +1,4 @@
+first line of d.txt
+second line of d.txt
+third line of d.txt
+delta delta delta
@@ -0,0 +1 @@
+contents of dir1/notes.md
@@ -0,0 +1 @@
+contents of dir2/notes.md
@@ -0,0 +1 @@
+contents of dir3/notes.md
@@ -0,0 +1 @@
+contents of dir4/notes.md
@@ -0,0 +1 @@
+contents of dir5/notes.md
@@ -0,0 +1 @@
+contents of dir6/notes.md
@@ -0,0 +1,4 @@
+first line of e.txt
+second line of e.txt
+third line of e.txt
+epsilon epsilon epsilon
@@ -0,0 +1,4 @@
+first line of f.txt
+second line of f.txt
+third line of f.txt
+zeta zeta zeta
@@ -0,0 +1,55 @@
+// Package refactor is the starting workspace for the multi-file refactor
+// benchmark tasks. It is intentionally a little messy (duplicated logic, a
+// bare map type, a thin wrapper) so the refactor prompts have real work to do.
+// It builds cleanly at all times; the tasks keep it building.
+package refactor
+
+import "fmt"
+
+// Config holds demo settings.
+type Config struct {
+	Name string
+	Port int
+}
+
+// GreetFromConfig returns a greeting using the config's name.
+func GreetFromConfig(c Config) string {
+	return fmt.Sprintf("hello, %s", c.Name)
+}
+
+// GreetByName returns a greeting using a bare name. BUG-note: this duplicates
+// GreetFromConfig's formatting — refactor-01 extracts a shared helper.
+func GreetByName(name string) string {
+	return fmt.Sprintf("hello, %s", name)
+}
+
+// Wrapper is a thin wrapper around GreetByName — refactor-05 inlines it.
+func Wrapper(name string) string {
+	return GreetByName(name)
+}
+
+// stats is a bare map used across the file — refactor-04 introduces a named type.
+var stats = map[string]int{}
+
+// Record bumps a stat by name.
+func Record(name string) {
+	stats[name]++
+}
+
+// Lookup returns a stat by name.
+func Lookup(name string) int {
+	return stats[name]
+}
+
+// load failed wrapped error — refactor-06 consolidates these.
+func firstError() error {
+	return fmt.Errorf("load failed: missing input")
+}
+
+func secondError() error {
+	return fmt.Errorf("load failed: bad format")
+}
+
+func buildError(ctx string) error {
+	return fmt.Errorf("load failed: %s", ctx)
+}
@@ -0,0 +1,679 @@
+package perfbench
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"io/fs"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"sort"
+	"strings"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/trace"
+)
+
+// TurnSchemaVersion is the schema version of a published turn-benchmark result.
+// Bump when the TurnBenchResult shape changes so consumers can detect drift.
+//
+// v2 splits pass/fail into three oracle tiers so the report cannot be misread as
+// a blanket correctness verdict: tasksVerified/tasksPassed/correctnessPassRate
+// cover only the positive-oracle classes (edit/fix); buildCheckedTasks/
+// buildPassedTasks/buildPassRate cover the non-positive build-check classes
+// (refactor); latencyOnlyTasks covers the no-oracle classes (nav/longproc/
+// longctx/parallel). tasksAttempted is still every task run.
+const TurnSchemaVersion = 2
+
+// TurnRunner runs one benchmark task and reports its outcome plus the captured
+// per-turn trace. A non-nil Err means the run failed to execute (process crash);
+// Passed reflects the verification result. Trace is the parsed NDJSON trace
+// (nil when the run errored before emitting one).
+type TurnRunner func(ctx context.Context, task BenchTask, rc RunContext) TurnTaskOutcome
+
+// TurnTaskOutcome is what a TurnRunner reports for one task iteration.
+type TurnTaskOutcome struct {
+	Passed    bool
+	VerifyErr string
+	WallMs    float64
+	Trace     *trace.TurnTrace
+	// TraceIssue, when non-empty, explains why the per-turn trace could not be
+	// parsed (file missing or malformed). The run still has a pass/fail verdict,
+	// but an incomplete measurement is surfaced as a result warning so it can't
+	// masquerade as a clean attribution sample.
+	TraceIssue string
+	Err        error
+}
+
+// TurnBenchConfig configures a turn-benchmark run.
+type TurnBenchConfig struct {
+	Model       string
+	Mode        string
+	SelfCorrect bool
+	Version     string
+	Commit      string
+	// Iterations is how many times each task is run. The per-process `zero exec`
+	// runner is inherently cold-start, so this is the sample count for per-span
+	// median/P95 — a genuine warm path needs an in-process runner (future work).
+	Iterations int
+	// Runner executes one task iteration. Required.
+	Runner TurnRunner
+	// Now overrides the clock for the recorded date (tests inject a fixed time).
+	Now func() time.Time
+}
+
+// SpanStats summarizes one span's duration across all measured task iterations.
+type SpanStats struct {
+	Count    int     `json:"count"`
+	TotalMs  float64 `json:"totalMs"`
+	MedianMs float64 `json:"medianMs"`
+	P95Ms    float64 `json:"p95Ms"`
+	MaxMs    float64 `json:"maxMs"`
+}
+
+// LatencySource is one of the top controllable latency sources, ranked by total
+// exclusive time across the whole run. Share is its fraction of total exclusive
+// span time. Because exclusive time subtracts nested children, concurrent and
+// nested spans no longer double-count, so shares of the top sources sum to ~1
+// for a well-instrumented run.
+type LatencySource struct {
+	Span    string  `json:"span"`
+	TotalMs float64 `json:"totalMs"`
+	Share   float64 `json:"share"`
+}
+
+// ClassSummary is the per-class (task group) roll-up.
+//
+// Passed is the count that passed THIS class's oracle: a correctness pass for
+// edit/fix, a build pass for refactor, and always 0 for the no-oracle classes
+// (nav/longproc/longctx/parallel). Verified is how many tasks in the class
+// carry an oracle (correctness or build); LatencyOnly is how many carry none.
+// A latency-only class therefore reports Passed=0, Verified=0, LatencyOnly=Tasks.
+type ClassSummary struct {
+	Tasks       int                `json:"tasks"`
+	Verified    int                `json:"verified"`
+	Passed      int                `json:"passed"`
+	LatencyOnly int                `json:"latencyOnly"`
+	WallMs      NumericStats       `json:"wallMs"`
+	SpanTotals  map[string]float64 `json:"spanTotals"`
+}
+
+// TurnBenchResult is the publishable turn-benchmark record.
+//
+// Pass/fail is split into three oracle tiers so it cannot be misread as a
+// blanket correctness verdict:
+//   - Correctness (tasksVerified / tasksPassed / correctnessPassRate): tasks
+//     with a positive oracle (edit/fix). This is the only pass rate that can
+//     move with model quality.
+//   - Build-only (buildCheckedTasks / buildPassedTasks / buildPassRate): tasks
+//     whose oracle is a non-positive build check (refactor `go build`). A pass
+//     means it compiles, not that the refactor is correct.
+//   - Latency-only (latencyOnlyTasks): tasks with no oracle (nav/longproc/
+//     longctx/parallel). They ran for latency and span attribution only and are
+//     excluded from every pass rate.
+//
+// tasksAttempted is still the total number of tasks run across all three tiers.
+// The tier class lists are echoed so a consumer can see exactly which classes
+// each rate is computed over.
+type TurnBenchResult struct {
+	SchemaVersion       int                     `json:"schemaVersion"`
+	Suite               string                  `json:"suite"`
+	Model               string                  `json:"model"`
+	Mode                string                  `json:"mode,omitempty"`
+	SelfCorrect         bool                    `json:"selfCorrect"`
+	Version             string                  `json:"version,omitempty"`
+	Commit              string                  `json:"commit,omitempty"`
+	Date                string                  `json:"date"`
+	TasksAttempted      int                     `json:"tasksAttempted"`
+	TasksVerified       int                     `json:"tasksVerified"`
+	TasksPassed         int                     `json:"tasksPassed"`
+	LatencyOnlyTasks    int                     `json:"latencyOnlyTasks"`
+	BuildCheckedTasks   int                     `json:"buildCheckedTasks"`
+	BuildPassedTasks    int                     `json:"buildPassedTasks"`
+	CorrectnessPassRate float64                 `json:"correctnessPassRate"`
+	BuildPassRate       float64                 `json:"buildPassRate"`
+	CorrectnessClasses  []string                `json:"correctnessClasses,omitempty"`
+	BuildOnlyClasses    []string                `json:"buildOnlyClasses,omitempty"`
+	LatencyOnlyClasses  []string                `json:"latencyOnlyClasses,omitempty"`
+	Iterations          int                     `json:"iterations"`
+	PerSpan             map[string]SpanStats    `json:"perSpan"`
+	TopLatency          []LatencySource         `json:"topLatency"`
+	PerClass            map[string]ClassSummary `json:"perClass"`
+	Totals              TurnBenchTotals         `json:"totals"`
+	Warnings            []Warning               `json:"warnings,omitempty"`
+}
+
+// TurnBenchTotals aggregates token and count totals across the whole run.
+type TurnBenchTotals struct {
+	InputTokens       int64 `json:"inputTokens"`
+	CachedInputTokens int64 `json:"cachedInputTokens"`
+	OutputTokens      int64 `json:"outputTokens"`
+	ModelRequests     int64 `json:"modelRequests"`
+	ToolCalls         int64 `json:"toolCalls"`
+	Retries           int64 `json:"retries"`
+	Reconnects        int64 `json:"reconnects"`
+	Compactions       int64 `json:"compactions"`
+}
+
+// RunTurnBench executes every task in the set with the configured runner and
+// returns a self-describing per-turn result. It never aborts on a single task
+// failure — every task is attempted and recorded. Per-span stats aggregate
+// across iterations; the top three controllable latency sources are ranked by
+// total attributed time.
+func RunTurnBench(ctx context.Context, set TaskSet, cfg TurnBenchConfig) (TurnBenchResult, error) {
+	if len(set.Tasks) == 0 {
+		return TurnBenchResult{}, errors.New("task set has no tasks")
+	}
+	if strings.TrimSpace(cfg.Model) == "" {
+		return TurnBenchResult{}, errors.New("turn benchmark requires a model")
+	}
+	if cfg.Runner == nil {
+		return TurnBenchResult{}, errors.New("turn benchmark requires a runner")
+	}
+	iterations := cfg.Iterations
+	if iterations < 1 {
+		iterations = 1
+	}
+	now := cfg.Now
+	if now == nil {
+		now = time.Now
+	}
+	rc := RunContext{Model: cfg.Model, Mode: cfg.Mode, SelfCorrect: cfg.SelfCorrect}
+
+	perSpanSamples := map[string][]float64{}
+	classWalls := map[string][]float64{}
+	classSpanTotals := map[string]map[string]float64{}
+	classTasks := map[string]int{}
+	classVerified := map[string]int{}
+	classPassed := map[string]int{}
+	classLatencyOnly := map[string]int{}
+	correctnessClasses := map[string]bool{}
+	buildOnlyClasses := map[string]bool{}
+	latencyOnlyClasses := map[string]bool{}
+	var totals TurnBenchTotals
+
+	// A class is build-only when the manifest declares it in BuildOnlyClasses.
+	// A task's tier is then decided per-task: no verificationCommand => latency-
+	// only; otherwise build-only if its class is declared, else correctness. The
+	// latency-only tier is always driven by oracle presence, never by the
+	// declared list, so a declared build-only class with a missing oracle still
+	// counts as latency-only rather than silently passing on exit 0.
+	buildOnly := map[string]bool{}
+	for _, c := range set.BuildOnlyClasses {
+		buildOnly[strings.TrimSpace(c)] = true
+	}
+
+	result := TurnBenchResult{
+		SchemaVersion: TurnSchemaVersion,
+		Suite:         strings.TrimSpace(set.ID),
+		Model:         strings.TrimSpace(cfg.Model),
+		Mode:          strings.TrimSpace(cfg.Mode),
+		SelfCorrect:   cfg.SelfCorrect,
+		Version:       strings.TrimSpace(cfg.Version),
+		Commit:        strings.TrimSpace(cfg.Commit),
+		Date:          now().UTC().Format(time.RFC3339),
+		Iterations:    iterations,
+		PerSpan:       map[string]SpanStats{},
+		PerClass:      map[string]ClassSummary{},
+	}
+
+	for _, task := range set.Tasks {
+		if err := ctx.Err(); err != nil {
+			return result, err
+		}
+		class := strings.TrimSpace(task.Class)
+		if class == "" {
+			class = "default"
+		}
+		classTasks[class]++
+		// A task counts as passed only when every iteration passed. The per-process
+		// runner is cold-start, so a flaky pass on one iteration and a fail on
+		// another is a real regression signal, not noise to average away.
+		passedForTask := true
+		for iter := 0; iter < iterations; iter++ {
+			outcome := cfg.Runner(ctx, task, rc)
+			if outcome.TraceIssue != "" {
+				result.Warnings = append(result.Warnings, Warning{
+					Metric:  "trace",
+					Message: fmt.Sprintf("task %s: %s", task.ID, outcome.TraceIssue),
+				})
+			}
+			if outcome.Err != nil {
+				// A crashed run must not look like a normal sample that's merely
+				// absent — surface it as a warning so a run that died every
+				// iteration can't pass as "fewer measurements."
+				result.Warnings = append(result.Warnings, Warning{
+					Metric:  "run",
+					Message: fmt.Sprintf("task %s: %v", task.ID, outcome.Err),
+				})
+				passedForTask = false
+				continue
+			}
+			if !outcome.Passed {
+				passedForTask = false
+			}
+			wall := outcome.WallMs
+			if wall <= 0 && outcome.Trace != nil {
+				wall = float64(outcome.Trace.WallDuration().Microseconds()) / 1000
+			}
+			if wall > 0 {
+				classWalls[class] = append(classWalls[class], wall)
+			}
+			if outcome.Trace != nil {
+				aggregateTotals(&totals, outcome.Trace)
+				for _, span := range outcome.Trace.Spans {
+					// Rank by exclusive time (duration minus nested children) so
+					// concurrent/nested spans do not double-count: provider_connect
+					// inside generation and permission_wait inside tool_execution
+					// each contribute their own exclusive time, not their parent's.
+					// A span whose exclusive is legitimately zero (a parent fully
+					// covered by its children) contributes zero on purpose — do
+					// NOT fall back to Duration, which would re-introduce the
+					// double-count. ReadNDJSON preserves a written exclusive_ms:0
+					// as 0 and only falls back to Duration when the key is absent,
+					// so span.Exclusive is always populated here.
+					ms := float64(span.Exclusive.Microseconds()) / 1000
+					perSpanSamples[span.Name] = append(perSpanSamples[span.Name], ms)
+					if classSpanTotals[class] == nil {
+						classSpanTotals[class] = map[string]float64{}
+					}
+					classSpanTotals[class][span.Name] += ms
+				}
+			}
+		}
+		result.TasksAttempted++
+
+		// Classify the task into an oracle tier and update only that tier's
+		// counters. A latency-only task (no verificationCommand) is never counted
+		// in any pass rate even when the runner reports Passed — an exit-0
+		// read-only run proves the turn ran, not that the answer was right.
+		hasOracle := len(task.VerificationCommand) > 0
+		switch {
+		case !hasOracle:
+			result.LatencyOnlyTasks++
+			classLatencyOnly[class]++
+			latencyOnlyClasses[class] = true
+		case buildOnly[class]:
+			result.BuildCheckedTasks++
+			classVerified[class]++
+			buildOnlyClasses[class] = true
+			if passedForTask {
+				result.BuildPassedTasks++
+				classPassed[class]++
+			}
+		default:
+			result.TasksVerified++
+			classVerified[class]++
+			correctnessClasses[class] = true
+			if passedForTask {
+				result.TasksPassed++
+				classPassed[class]++
+			}
+		}
+	}
+
+	for name, samples := range perSpanSamples {
+		result.PerSpan[name] = summarizeSpan(samples)
+	}
+	result.TopLatency = topLatencySources(result.PerSpan, 3)
+	for class := range classTasks {
+		walls := classWalls[class]
+		var wallStats NumericStats
+		if len(walls) > 0 {
+			wallStats = SummarizeSamples(walls)
+		}
+		result.PerClass[class] = ClassSummary{
+			Tasks:       classTasks[class],
+			Verified:    classVerified[class],
+			Passed:      classPassed[class],
+			LatencyOnly: classLatencyOnly[class],
+			WallMs:      wallStats,
+			SpanTotals:  classSpanTotals[class],
+		}
+	}
+	result.CorrectnessClasses = sortedSetKeys(correctnessClasses)
+	result.BuildOnlyClasses = sortedSetKeys(buildOnlyClasses)
+	result.LatencyOnlyClasses = sortedSetKeys(latencyOnlyClasses)
+	result.CorrectnessPassRate = passRate(result.TasksPassed, result.TasksVerified)
+	result.BuildPassRate = passRate(result.BuildPassedTasks, result.BuildCheckedTasks)
+	result.Totals = totals
+	return result, nil
+}
+
+// passRate is passed/total rounded to the benchmark's metric precision, or 0
+// when the denominator is zero (no tasks in that tier were run).
+func passRate(passed, total int) float64 {
+	if total <= 0 {
+		return 0
+	}
+	return RoundMetric(float64(passed) / float64(total))
+}
+
+func sortedSetKeys(set map[string]bool) []string {
+	if len(set) == 0 {
+		return nil
+	}
+	keys := make([]string, 0, len(set))
+	for k := range set {
+		keys = append(keys, k)
+	}
+	sort.Strings(keys)
+	return keys
+}
+
+func summarizeSpan(samples []float64) SpanStats {
+	if len(samples) == 0 {
+		return SpanStats{}
+	}
+	stats := SummarizeSamples(samples)
+	total := 0.0
+	for _, s := range samples {
+		total += s
+	}
+	return SpanStats{
+		Count:    len(samples),
+		TotalMs:  RoundMetric(total),
+		MedianMs: stats.Median,
+		P95Ms:    stats.P95,
+		MaxMs:    stats.Max,
+	}
+}
+
+func topLatencySources(perSpan map[string]SpanStats, top int) []LatencySource {
+	sources := make([]LatencySource, 0, len(perSpan))
+	totalAttributed := 0.0
+	for _, s := range perSpan {
+		totalAttributed += s.TotalMs
+	}
+	for name, s := range perSpan {
+		share := 0.0
+		if totalAttributed > 0 {
+			share = s.TotalMs / totalAttributed
+		}
+		sources = append(sources, LatencySource{
+			Span:    name,
+			TotalMs: s.TotalMs,
+			Share:   RoundMetric(share),
+		})
+	}
+	sort.SliceStable(sources, func(i, j int) bool {
+		if sources[i].TotalMs != sources[j].TotalMs {
+			return sources[i].TotalMs > sources[j].TotalMs
+		}
+		return sources[i].Span < sources[j].Span
+	})
+	if top > 0 && len(sources) > top {
+		sources = sources[:top]
+	}
+	return sources
+}
+
+func aggregateTotals(totals *TurnBenchTotals, tr *trace.TurnTrace) {
+	totals.InputTokens += tr.Counter(trace.CounterInputTokens)
+	totals.CachedInputTokens += tr.Counter(trace.CounterCachedInputTokens)
+	totals.OutputTokens += tr.Counter(trace.CounterOutputTokens)
+	totals.ModelRequests += tr.Counter(trace.CounterModelRequests)
+	totals.ToolCalls += tr.Counter(trace.CounterToolCalls)
+	totals.Retries += tr.Counter(trace.CounterRetryCount)
+	totals.Reconnects += tr.Counter(trace.CounterReconnectCount)
+	totals.Compactions += tr.Counter(trace.CounterCompactionCount)
+}
+
+// FormatTurnBenchSummary renders a human-readable turn-benchmark summary that
+// names the top controllable latency sources — the baseline's "do not proceed
+// until" criterion.
+func FormatTurnBenchSummary(result TurnBenchResult) string {
+	lines := []string{
+		"Zero turn benchmark: " + displayOrUnknown(result.Suite),
+		"model: " + displayOrUnknown(result.Model),
+		// The headline separates the three oracle tiers so an exit-0 read-only
+		// task can never inflate a "pass rate" that reads as correctness:
+		// correctness (positive oracle, edit/fix), build (non-positive `go build`,
+		// refactor), and latency-only (no oracle, nav/longproc/longctx/parallel).
+		fmt.Sprintf("tasks: %d total | correctness %d/%d (%.0f%%) | build %d/%d (%.0f%%) | latency-only %d | %d iter",
+			result.TasksAttempted,
+			result.TasksPassed, result.TasksVerified, result.CorrectnessPassRate*100,
+			result.BuildPassedTasks, result.BuildCheckedTasks, result.BuildPassRate*100,
+			result.LatencyOnlyTasks, result.Iterations),
+	}
+	if result.Mode != "" {
+		lines = append(lines, "mode: "+result.Mode)
+	}
+	if len(result.TopLatency) > 0 {
+		lines = append(lines, "top latency sources:")
+		for _, src := range result.TopLatency {
+			lines = append(lines, fmt.Sprintf("  %-18s %10s  %5.1f%%", src.Span, FormatMetric(src.TotalMs, "ms"), src.Share*100))
+		}
+	}
+	lines = append(lines, fmt.Sprintf("totals: in=%d (cached %d) out=%d | requests=%d tools=%d retries=%d reconnects=%d compactions=%d",
+		result.Totals.InputTokens, result.Totals.CachedInputTokens, result.Totals.OutputTokens,
+		result.Totals.ModelRequests, result.Totals.ToolCalls, result.Totals.Retries,
+		result.Totals.Reconnects, result.Totals.Compactions))
+	for _, class := range sortedClasses(result.PerClass) {
+		summary := result.PerClass[class]
+		median := FormatMetric(summary.WallMs.Median, "ms")
+		tier := classTier(result, class)
+		lines = append(lines, fmt.Sprintf("  [%s/%s] %d/%d passed, %d latency-only, wall median %s",
+			class, tier, summary.Passed, summary.Verified, summary.LatencyOnly, median))
+	}
+	return strings.Join(lines, "\n")
+}
+
+// classTier returns the oracle tier label a class was classified into, so the
+// per-class roll-up states explicitly which oracle (if any) its "passed" count
+// is measured against.
+func classTier(result TurnBenchResult, class string) string {
+	for _, c := range result.BuildOnlyClasses {
+		if c == class {
+			return "build"
+		}
+	}
+	for _, c := range result.LatencyOnlyClasses {
+		if c == class {
+			return "latency"
+		}
+	}
+	return "correctness"
+}
+
+func sortedClasses(perClass map[string]ClassSummary) []string {
+	classes := make([]string, 0, len(perClass))
+	for c := range perClass {
+		classes = append(classes, c)
+	}
+	sort.Strings(classes)
+	return classes
+}
+
+// WriteTurnBenchJSON writes the indented JSON form of a turn-benchmark result.
+func WriteTurnBenchJSON(w io.Writer, result TurnBenchResult) error {
+	encoder := json.NewEncoder(w)
+	encoder.SetIndent("", "  ")
+	return encoder.Encode(result)
+}
+
+// NewTurnExecRunner builds the production turn-benchmark runner: it invokes
+// headless `zero exec` with stream-json output AND `--trace <tmpfile>`, then
+// parses the emitted NDJSON trace into a *trace.TurnTrace. binary is the path to
+// the `zero` binary; extraArgs are appended to every invocation. Pass/fail is
+// decided from the stream-json run_end exit code (and the task's
+// VerificationCommand when present), exactly like NewExecRunner.
+func NewTurnExecRunner(binary string, extraArgs ...string) TurnRunner {
+	return func(ctx context.Context, task BenchTask, rc RunContext) TurnTaskOutcome {
+		// Isolate the workspace: copy the fixture into a fresh temp dir so a
+		// mutating task (edit/fix/refactor) can't dirty the shared, checked-in
+		// fixture or bleed into a later iteration of the same task. When no
+		// fixture is configured the agent runs in the caller's cwd as before.
+		if fixture := strings.TrimSpace(task.WorkspaceFixture); fixture != "" {
+			copyDir, cerr := copyFixture(fixture)
+			if cerr != nil {
+				return TurnTaskOutcome{Err: fmt.Errorf("isolate fixture: %w", cerr)}
+			}
+			defer os.RemoveAll(copyDir)
+			task.WorkspaceFixture = copyDir
+		}
+
+		traceFile, err := os.CreateTemp("", "zero-turn-trace-*.ndjson")
+		if err != nil {
+			return TurnTaskOutcome{Err: fmt.Errorf("create trace file: %w", err)}
+		}
+		_ = traceFile.Close()
+		tracePath := traceFile.Name()
+		defer os.Remove(tracePath)
+
+		args := buildTurnExecArgs(task, rc, tracePath, extraArgs)
+		cmd := exec.CommandContext(ctx, binary, args...)
+		cmd.Env = appendNoColor(os.Environ())
+		if dir := strings.TrimSpace(task.WorkspaceFixture); dir != "" {
+			cmd.Dir = dir
+		}
+		var outBuf, errBuf bytes.Buffer
+		cmd.Stdout = &outBuf
+		cmd.Stderr = &errBuf
+		start := time.Now()
+		runErr := cmd.Run()
+		wallMs := float64(time.Since(start).Microseconds()) / 1000
+
+		exitCode, haveExit := streamJSONExitCode(outBuf.Bytes())
+		outcome := TurnTaskOutcome{WallMs: wallMs}
+		if haveExit && exitCode != 0 {
+			outcome.VerifyErr = fmt.Sprintf("agent run_end exit code %d", exitCode)
+		} else if !haveExit {
+			detail := strings.TrimSpace(errBuf.String())
+			// If the process never produced a run_end event, the actual failure
+			// is in runErr (binary missing, exec permission denied, etc.) —
+			// prefer it over the generic fallback so the real reason survives.
+			if detail == "" && runErr != nil {
+				detail = runErr.Error()
+			}
+			if detail == "" {
+				detail = "missing terminal run_end event"
+			}
+			outcome.Err = fmt.Errorf("zero exec failed: %s", detail)
+			return outcome
+		}
+
+		// Parse the captured trace. The trace is the attribution layer, so a
+		// missing/malformed file is not fatal — but it IS surfaced as a warning
+		// (via outcome.TraceIssue) so an incomplete measurement can't look valid.
+		if f, ferr := os.Open(tracePath); ferr != nil {
+			outcome.TraceIssue = fmt.Sprintf("trace open failed: %v", ferr)
+		} else {
+			tr, perr := trace.ReadNDJSON(f)
+			_ = f.Close()
+			if perr != nil {
+				outcome.TraceIssue = fmt.Sprintf("trace parse failed: %v", perr)
+			} else {
+				outcome.Trace = tr
+			}
+		}
+
+		// A nonzero agent exit already decided failure; don't run verification or
+		// mark the task passed.
+		if outcome.VerifyErr != "" {
+			return outcome
+		}
+		if len(task.VerificationCommand) > 0 {
+			if vOutcome := runVerification(ctx, task); !vOutcome.Passed {
+				outcome.VerifyErr = strings.TrimSpace(vOutcome.Detail)
+				return outcome
+			}
+			// A positive oracle (grep/test/build) passed: this is the only path
+			// that sets Passed. A task with no verificationCommand is latency-only
+			// (read-only nav/longproc/longctx/parallel): its exit 0 proves the turn
+			// ran, not that the answer was right, so it never reports Passed and the
+			// harness counts it in latencyOnlyTasks rather than any pass rate.
+			outcome.Passed = true
+		}
+		return outcome
+	}
+}
+
+// copyFixture copies the fixture directory at src into a fresh temp dir and
+// returns the copy's path. Used to give each benchmark invocation an isolated
+// workspace so mutating tasks (edit/fix/refactor) can't dirty the checked-in
+// fixture or a later iteration.
+func copyFixture(src string) (string, error) {
+	info, err := os.Stat(src)
+	if err != nil {
+		return "", err
+	}
+	if !info.IsDir() {
+		return "", fmt.Errorf("fixture %q is not a directory", src)
+	}
+	dst, err := os.MkdirTemp("", "zero-turn-fixture-*")
+	if err != nil {
+		return "", err
+	}
+	err = filepath.WalkDir(src, func(path string, d fs.DirEntry, werr error) error {
+		if werr != nil {
+			return werr
+		}
+		rel, rerr := filepath.Rel(src, path)
+		if rerr != nil {
+			return rerr
+		}
+		target := filepath.Join(dst, rel)
+		if d.IsDir() {
+			return os.MkdirAll(target, 0o755)
+		}
+		data, rerr := os.ReadFile(path)
+		if rerr != nil {
+			return rerr
+		}
+		return os.WriteFile(target, data, 0o644)
+	})
+	if err != nil {
+		os.RemoveAll(dst)
+		return "", err
+	}
+	return dst, nil
+}
+
+func buildTurnExecArgs(task BenchTask, rc RunContext, tracePath string, extraArgs []string) []string {
+	args := []string{"exec", "--output-format", "stream-json", "--trace", tracePath}
+	if model := strings.TrimSpace(rc.Model); model != "" {
+		args = append(args, "--model", model)
+	}
+	if mode := strings.TrimSpace(rc.Mode); mode != "" {
+		args = append(args, "--mode", mode)
+	}
+	if rc.SelfCorrect {
+		args = append(args, "--self-correct")
+	}
+	args = append(args, extraArgs...)
+	args = append(args, task.Prompt)
+	return args
+}
+
+// ResolveBinary locates the zero binary for a benchmark run: an explicit path
+// when provided, else a `zero` (or zero.exe) on PATH, else a binary built into
+// the repo root. Returns an error when none is found.
+func ResolveBinary(explicit string) (string, error) {
+	if v := strings.TrimSpace(explicit); v != "" {
+		if _, err := os.Stat(v); err != nil {
+			return "", fmt.Errorf("trace binary not found: %w", err)
+		}
+		return v, nil
+	}
+	if path, err := exec.LookPath("zero"); err == nil {
+		return path, nil
+	}
+	if path, err := exec.LookPath("zero.exe"); err == nil {
+		return path, nil
+	}
+	cwd, err := os.Getwd()
+	if err == nil {
+		for _, name := range []string{"zero", "zero.exe"} {
+			candidate := filepath.Join(cwd, name)
+			if _, err := os.Stat(candidate); err == nil {
+				return candidate, nil
+			}
+		}
+	}
+	return "", errors.New("zero binary not found; build it first or pass an explicit path")
+}
@@ -0,0 +1,455 @@
+package perfbench
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/trace"
+)
+
+// fakeTurnRunner returns a canned *trace.TurnTrace per task so the harness's
+// aggregation logic can be exercised without a model or a binary. Each task
+// yields one generation span (the dominant cost), one tool_execution span, and
+// the counters the totals aggregate.
+func fakeTurnRunner(canned map[string]*trace.TurnTrace) TurnRunner {
+	return func(ctx context.Context, task BenchTask, rc RunContext) TurnTaskOutcome {
+		tr, ok := canned[task.ID]
+		if !ok {
+			return TurnTaskOutcome{Err: errNoCanned}
+		}
+		wallMs := float64(tr.WallDuration().Microseconds()) / 1000
+		return TurnTaskOutcome{Passed: true, WallMs: wallMs, Trace: tr}
+	}
+}
+
+var errNoCanned = &cannedError{}
+
+type cannedError struct{}
+
+func (*cannedError) Error() string { return "no canned trace for task" }
+
+// cannedTrace builds a deterministic *trace.TurnTrace with the given spans and
+// counters. Spans are recorded as fixed durations (no real timing) so the
+// aggregation math is predictable in assertions.
+func cannedTrace(genMs, toolMs int, tokens int64) *trace.TurnTrace {
+	r := trace.NewRecorder("sess", "run-1", "test")
+	r.Start()
+	r.RecordSpan(trace.SpanGeneration, time.Duration(genMs)*time.Millisecond)
+	r.RecordSpan(trace.SpanToolExecution, time.Duration(toolMs)*time.Millisecond)
+	r.Counter(trace.CounterInputTokens, tokens)
+	r.Counter(trace.CounterOutputTokens, tokens/2)
+	r.Counter(trace.CounterModelRequests, 1)
+	r.Counter(trace.CounterToolCalls, 1)
+	r.StampFirstToken()
+	return r.Finish()
+}
+
+func TestRunTurnBenchAggregation(t *testing.T) {
+	set := TaskSet{
+		ID: "fake-suite",
+		Tasks: []BenchTask{
+			{ID: "t1", Class: "nav", Prompt: "p1"},                                             // latency-only
+			{ID: "t2", Class: "nav", Prompt: "p2"},                                             // latency-only
+			{ID: "t3", Class: "edit", Prompt: "p3", VerificationCommand: []string{"true"}},     // correctness
+			{ID: "t4", Class: "refactor", Prompt: "p4", VerificationCommand: []string{"true"}}, // build-only
+		},
+		BuildOnlyClasses: []string{"refactor"},
+	}
+	canned := map[string]*trace.TurnTrace{
+		"t1": cannedTrace(100, 10, 1000),
+		"t2": cannedTrace(300, 10, 1000),
+		"t3": cannedTrace(200, 50, 2000),
+		"t4": cannedTrace(150, 20, 1500),
+	}
+	cfg := TurnBenchConfig{
+		Model:      "fake-model",
+		Iterations: 1,
+		Runner:     fakeTurnRunner(canned),
+		Now:        func() time.Time { return time.Date(2026, 1, 2, 3, 4, 5, 0, time.UTC) },
+	}
+	result, err := RunTurnBench(context.Background(), set, cfg)
+	if err != nil {
+		t.Fatalf("RunTurnBench: %v", err)
+	}
+	// 4 tasks attempted; the tier split is 2 latency-only, 1 correctness, 1 build.
+	if result.TasksAttempted != 4 {
+		t.Fatalf("attempted=%d, want 4", result.TasksAttempted)
+	}
+	if result.TasksVerified != 1 || result.TasksPassed != 1 {
+		t.Fatalf("correctness verified=%d passed=%d, want 1/1", result.TasksVerified, result.TasksPassed)
+	}
+	if result.LatencyOnlyTasks != 2 {
+		t.Fatalf("latencyOnly=%d, want 2", result.LatencyOnlyTasks)
+	}
+	if result.BuildCheckedTasks != 1 || result.BuildPassedTasks != 1 {
+		t.Fatalf("build checked=%d passed=%d, want 1/1", result.BuildCheckedTasks, result.BuildPassedTasks)
+	}
+	if result.CorrectnessPassRate != 1.0 {
+		t.Fatalf("correctnessPassRate=%v, want 1.0", result.CorrectnessPassRate)
+	}
+	if result.BuildPassRate != 1.0 {
+		t.Fatalf("buildPassRate=%v, want 1.0", result.BuildPassRate)
+	}
+	if len(result.CorrectnessClasses) != 1 || result.CorrectnessClasses[0] != "edit" {
+		t.Fatalf("correctnessClasses=%v, want [edit]", result.CorrectnessClasses)
+	}
+	if len(result.BuildOnlyClasses) != 1 || result.BuildOnlyClasses[0] != "refactor" {
+		t.Fatalf("buildOnlyClasses=%v, want [refactor]", result.BuildOnlyClasses)
+	}
+	if len(result.LatencyOnlyClasses) != 1 || result.LatencyOnlyClasses[0] != "nav" {
+		t.Fatalf("latencyOnlyClasses=%v, want [nav]", result.LatencyOnlyClasses)
+	}
+	if result.SchemaVersion != TurnSchemaVersion {
+		t.Fatalf("schemaVersion = %d, want %d", result.SchemaVersion, TurnSchemaVersion)
+	}
+	if result.Date != "2026-01-02T03:04:05Z" {
+		t.Fatalf("date = %q", result.Date)
+	}
+
+	// Per-span: generation appears in all four (100+300+200+150=750ms), tool in
+	// all four (10+10+50+20=90ms). Count must equal the number of tasks * iterations.
+	gen := result.PerSpan[trace.SpanGeneration]
+	if gen.Count != 4 {
+		t.Fatalf("generation count = %d, want 4", gen.Count)
+	}
+	if gen.TotalMs != 750 {
+		t.Fatalf("generation totalMs = %v, want 750", gen.TotalMs)
+	}
+	tool := result.PerSpan[trace.SpanToolExecution]
+	if tool.TotalMs != 90 {
+		t.Fatalf("tool totalMs = %v, want 90", tool.TotalMs)
+	}
+
+	// Top latency: generation (750) ranks above tool (90). Exactly two spans
+	// here, so both appear and generation is first.
+	if len(result.TopLatency) != 2 || result.TopLatency[0].Span != trace.SpanGeneration {
+		t.Fatalf("topLatency = %+v", result.TopLatency)
+	}
+	if result.TopLatency[0].Share <= result.TopLatency[1].Share {
+		t.Fatalf("top latency not ranked by share: %+v", result.TopLatency)
+	}
+
+	// Totals: 4 model requests, 4 tool calls, input tokens 1000+1000+2000+1500=5500.
+	if result.Totals.ModelRequests != 4 {
+		t.Fatalf("modelRequests = %d, want 4", result.Totals.ModelRequests)
+	}
+	if result.Totals.ToolCalls != 4 {
+		t.Fatalf("toolCalls = %d, want 4", result.Totals.ToolCalls)
+	}
+	if result.Totals.InputTokens != 5500 {
+		t.Fatalf("inputTokens = %d, want 5500", result.Totals.InputTokens)
+	}
+	if result.Totals.OutputTokens != 2750 {
+		t.Fatalf("outputTokens = %d, want 2750", result.Totals.OutputTokens)
+	}
+
+	// Per-class tier roll-up: nav is latency-only (0 verified, 2 latency-only),
+	// edit is correctness (1/1 verified passed), refactor is build (1/1 passed).
+	nav := result.PerClass["nav"]
+	if nav.Tasks != 2 || nav.Verified != 0 || nav.Passed != 0 || nav.LatencyOnly != 2 {
+		t.Fatalf("nav class = %+v", nav)
+	}
+	edit := result.PerClass["edit"]
+	if edit.Tasks != 1 || edit.Verified != 1 || edit.Passed != 1 || edit.LatencyOnly != 0 {
+		t.Fatalf("edit class = %+v", edit)
+	}
+	refactor := result.PerClass["refactor"]
+	if refactor.Tasks != 1 || refactor.Verified != 1 || refactor.Passed != 1 || refactor.LatencyOnly != 0 {
+		t.Fatalf("refactor class = %+v", refactor)
+	}
+}
+
+// TestRunTurnBenchLatencyOnlyNeverPassed asserts the honesty gate: a task with
+// no verificationCommand reports Passed=true from the (stub) runner, yet the
+// harness counts it ONLY in latencyOnlyTasks — never in tasksPassed or any pass
+// rate — so an exit-0 read-only run cannot inflate a correctness number.
+func TestRunTurnBenchLatencyOnlyNeverPassed(t *testing.T) {
+	set := TaskSet{
+		ID: "lo-suite",
+		Tasks: []BenchTask{
+			{ID: "n1", Class: "nav", Prompt: "p"}, // no oracle — runner still says Passed=true
+		},
+	}
+	cfg := TurnBenchConfig{
+		Model:  "fake-model",
+		Runner: fakeTurnRunner(map[string]*trace.TurnTrace{"n1": cannedTrace(50, 5, 100)}),
+		Now:    func() time.Time { return time.Date(2026, 1, 2, 3, 4, 5, 0, time.UTC) },
+	}
+	result, err := RunTurnBench(context.Background(), set, cfg)
+	if err != nil {
+		t.Fatalf("RunTurnBench: %v", err)
+	}
+	if result.TasksPassed != 0 || result.TasksVerified != 0 {
+		t.Fatalf("latency-only leaked into pass: passed=%d verified=%d, want 0/0",
+			result.TasksPassed, result.TasksVerified)
+	}
+	if result.CorrectnessPassRate != 0 || result.BuildPassRate != 0 {
+		t.Fatalf("pass rates should be 0 with no oracle tasks: c=%v b=%v",
+			result.CorrectnessPassRate, result.BuildPassRate)
+	}
+	if result.LatencyOnlyTasks != 1 {
+		t.Fatalf("latencyOnlyTasks=%d, want 1", result.LatencyOnlyTasks)
+	}
+	if result.PerClass["nav"].Passed != 0 || result.PerClass["nav"].LatencyOnly != 1 {
+		t.Fatalf("nav class = %+v, want passed=0 latencyOnly=1", result.PerClass["nav"])
+	}
+}
+
+func TestRunTurnBenchIterationsAggregates(t *testing.T) {
+	set := TaskSet{
+		ID:    "iter-suite",
+		Tasks: []BenchTask{{ID: "t1", Class: "nav", Prompt: "p1"}},
+	}
+	canned := map[string]*trace.TurnTrace{"t1": cannedTrace(100, 10, 500)}
+	cfg := TurnBenchConfig{
+		Model:      "fake-model",
+		Iterations: 3,
+		Runner:     fakeTurnRunner(canned),
+		Now:        func() time.Time { return time.Date(2026, 1, 2, 3, 4, 5, 0, time.UTC) },
+	}
+	result, err := RunTurnBench(context.Background(), set, cfg)
+	if err != nil {
+		t.Fatalf("RunTurnBench: %v", err)
+	}
+	if result.PerSpan[trace.SpanGeneration].Count != 3 {
+		t.Fatalf("generation count = %d, want 3 (one per iteration)", result.PerSpan[trace.SpanGeneration].Count)
+	}
+	if result.PerSpan[trace.SpanGeneration].TotalMs != 300 {
+		t.Fatalf("generation totalMs = %v, want 300", result.PerSpan[trace.SpanGeneration].TotalMs)
+	}
+}
+
+func TestRunTurnBenchRequiresModelAndRunner(t *testing.T) {
+	set := TaskSet{ID: "s", Tasks: []BenchTask{{ID: "t", Prompt: "p"}}}
+	if _, err := RunTurnBench(context.Background(), set, TurnBenchConfig{Runner: fakeTurnRunner(nil)}); err == nil {
+		t.Fatal("expected error for missing model")
+	}
+	if _, err := RunTurnBench(context.Background(), set, TurnBenchConfig{Model: "m"}); err == nil {
+		t.Fatal("expected error for missing runner")
+	}
+	if _, err := RunTurnBench(context.Background(), TaskSet{ID: "empty"}, TurnBenchConfig{Model: "m", Runner: fakeTurnRunner(nil)}); err == nil {
+		t.Fatal("expected error for empty task set")
+	}
+}
+
+func TestTopLatencySourcesRanksByTotalAndCapsTopN(t *testing.T) {
+	perSpan := map[string]SpanStats{
+		"a": {TotalMs: 100},
+		"b": {TotalMs: 500},
+		"c": {TotalMs: 300},
+		"d": {TotalMs: 50},
+	}
+	top := topLatencySources(perSpan, 3)
+	if len(top) != 3 {
+		t.Fatalf("len = %d, want 3", len(top))
+	}
+	wantOrder := []string{"b", "c", "a"}
+	for i, w := range wantOrder {
+		if top[i].Span != w {
+			t.Fatalf("top[%d] = %q, want %q", i, top[i].Span, w)
+		}
+	}
+	// Shares sum to 1 across all four (100+500+300+50=950); the top-3 retain
+	// their global share (not renormalized to the top-3).
+	// Share is rounded to 2 decimals by RoundMetric (500/950 -> 0.53), so compare
+	// against the rounded value with a small tolerance.
+	if got, want := top[0].Share, RoundMetric(500.0/950.0); !approxEqual(got, want, 0.001) {
+		t.Fatalf("top[0] share = %v, want %v", got, want)
+	}
+}
+
+func TestWriteTurnBenchJSONRoundTrip(t *testing.T) {
+	set := TaskSet{ID: "json-suite", Tasks: []BenchTask{
+		{ID: "t1", Class: "edit", Prompt: "p1", VerificationCommand: []string{"true"}},
+	}}
+	canned := map[string]*trace.TurnTrace{"t1": cannedTrace(150, 20, 800)}
+	result, err := RunTurnBench(context.Background(), set, TurnBenchConfig{
+		Model:  "fake-model",
+		Runner: fakeTurnRunner(canned),
+		Now:    func() time.Time { return time.Date(2026, 1, 2, 3, 4, 5, 0, time.UTC) },
+	})
+	if err != nil {
+		t.Fatalf("RunTurnBench: %v", err)
+	}
+	var buf bytes.Buffer
+	if err := WriteTurnBenchJSON(&buf, result); err != nil {
+		t.Fatalf("WriteTurnBenchJSON: %v", err)
+	}
+	var decoded TurnBenchResult
+	if err := json.Unmarshal(buf.Bytes(), &decoded); err != nil {
+		t.Fatalf("unmarshal: %v\n%s", err, buf.String())
+	}
+	if decoded.SchemaVersion != TurnSchemaVersion {
+		t.Fatalf("schemaVersion = %d, want %d", decoded.SchemaVersion, TurnSchemaVersion)
+	}
+	if decoded.TasksVerified != 1 || decoded.TasksPassed != 1 || decoded.LatencyOnlyTasks != 0 {
+		t.Fatalf("decoded tier counts = verified=%d passed=%d latency=%d, want 1/1/0",
+			decoded.TasksVerified, decoded.TasksPassed, decoded.LatencyOnlyTasks)
+	}
+	if decoded.CorrectnessPassRate != 1.0 {
+		t.Fatalf("decoded correctnessPassRate = %v, want 1.0", decoded.CorrectnessPassRate)
+	}
+	if decoded.PerSpan[trace.SpanGeneration].TotalMs != 150 {
+		t.Fatalf("decoded generation totalMs = %v, want 150", decoded.PerSpan[trace.SpanGeneration].TotalMs)
+	}
+}
+
+func TestFormatTurnBenchSummaryNamesTopSources(t *testing.T) {
+	set := TaskSet{ID: "fmt-suite", Tasks: []BenchTask{{ID: "t1", Class: "nav", Prompt: "p1"}}}
+	canned := map[string]*trace.TurnTrace{"t1": cannedTrace(150, 20, 800)}
+	result, err := RunTurnBench(context.Background(), set, TurnBenchConfig{
+		Model:  "fake-model",
+		Runner: fakeTurnRunner(canned),
+		Now:    func() time.Time { return time.Date(2026, 1, 2, 3, 4, 5, 0, time.UTC) },
+	})
+	if err != nil {
+		t.Fatalf("RunTurnBench: %v", err)
+	}
+	summary := FormatTurnBenchSummary(result)
+	if !strings.Contains(summary, "top latency sources") {
+		t.Fatalf("summary missing top-latency header:\n%s", summary)
+	}
+	if !strings.Contains(summary, trace.SpanGeneration) {
+		t.Fatalf("summary missing top span name %q:\n%s", trace.SpanGeneration, summary)
+	}
+}
+
+func TestLoadBaselineManifest(t *testing.T) {
+	path := filepath.Join("manifests", "baseline.json")
+	set, err := LoadTaskSet(path)
+	if err != nil {
+		t.Fatalf("LoadTaskSet: %v", err)
+	}
+	if set.ID == "" {
+		t.Fatal("manifest has no id")
+	}
+	// The baseline must clear the "do not proceed until ≥30 tasks" gate.
+	if len(set.Tasks) < 30 {
+		t.Fatalf("baseline has %d tasks, want >= 30", len(set.Tasks))
+	}
+	// The six required classes must all be present and non-empty.
+	wantClasses := map[string]bool{
+		"nav": false, "edit": false, "fix": false,
+		"refactor": false, "longproc": false, "longctx": false, "parallel": false,
+	}
+	counts := map[string]int{}
+	for _, task := range set.Tasks {
+		class := strings.TrimSpace(task.Class)
+		if class == "" {
+			t.Fatalf("task %q has no class", task.ID)
+		}
+		if _, ok := wantClasses[class]; !ok {
+			t.Fatalf("unexpected class %q on task %q", class, task.ID)
+		}
+		wantClasses[class] = true
+		counts[class]++
+	}
+	for class, present := range wantClasses {
+		if !present {
+			t.Fatalf("manifest missing required class %q", class)
+		}
+		if counts[class] == 0 {
+			t.Fatalf("class %q has zero tasks", class)
+		}
+	}
+	// Every task must have a prompt and a workspace fixture pointing under testdata,
+	// and that fixture must actually exist on disk — a manifest referencing a
+	// missing fixture would make every task in its class error out at run time, so
+	// catch it at load time instead.
+	seen := map[string]bool{}
+	for _, task := range set.Tasks {
+		if strings.TrimSpace(task.Prompt) == "" {
+			t.Fatalf("task %q has empty prompt", task.ID)
+		}
+		if strings.TrimSpace(task.WorkspaceFixture) == "" {
+			t.Fatalf("task %q has no workspace fixture", task.ID)
+		}
+		if !strings.Contains(task.WorkspaceFixture, "testdata") {
+			t.Fatalf("task %q fixture %q not under testdata", task.ID, task.WorkspaceFixture)
+		}
+		if seen[task.WorkspaceFixture] {
+			continue
+		}
+		seen[task.WorkspaceFixture] = true
+		info, err := os.Stat(task.WorkspaceFixture)
+		if err != nil {
+			t.Fatalf("task %q fixture %q does not exist: %v", task.ID, task.WorkspaceFixture, err)
+		}
+		if !info.IsDir() {
+			t.Fatalf("task %q fixture %q is not a directory", task.ID, task.WorkspaceFixture)
+		}
+	}
+}
+
+func approxEqual(a, b, tol float64) bool {
+	d := a - b
+	if d < 0 {
+		d = -d
+	}
+	return d < tol
+}
+
+// TestCopyFixtureIsolatesSourceFromMutation asserts the property that lets a
+// mutating task run twice (or across iterations) without poisoning the next
+// sample: the runner operates on a per-invocation copy, so mutating the copy
+// leaves the checked-in source fixture byte-identical. This is verified by
+// reading the code today; a test makes it durable against future refactors of
+// the runner's isolation path.
+func TestCopyFixtureIsolatesSourceFromMutation(t *testing.T) {
+	src, err := os.MkdirTemp("", "zero-fixture-src-*")
+	if err != nil {
+		t.Fatalf("mkdtemp src: %v", err)
+	}
+	defer os.RemoveAll(src)
+	orig := "package main\n\nfunc main() {}\n"
+	if err := os.WriteFile(filepath.Join(src, "main.go"), []byte(orig), 0o644); err != nil {
+		t.Fatalf("write fixture: %v", err)
+	}
+	if err := os.MkdirAll(filepath.Join(src, "sub"), 0o755); err != nil {
+		t.Fatalf("mkdir sub: %v", err)
+	}
+	if err := os.WriteFile(filepath.Join(src, "sub", "a.go"), []byte("package sub\n"), 0o644); err != nil {
+		t.Fatalf("write sub: %v", err)
+	}
+
+	// Snapshot the source tree before any mutation of the copy.
+	wantMain, err := os.ReadFile(filepath.Join(src, "main.go"))
+	if err != nil {
+		t.Fatalf("read source main.go: %v", err)
+	}
+
+	copyDir, err := copyFixture(src)
+	if err != nil {
+		t.Fatalf("copyFixture: %v", err)
+	}
+	defer os.RemoveAll(copyDir)
+
+	// The copy must be a real copy, not a symlink/alias of the source, and
+	// mutating it must not touch the source.
+	if copyDir == src {
+		t.Fatalf("copyFixture returned the source dir, not a copy: %s", copyDir)
+	}
+	if err := os.WriteFile(filepath.Join(copyDir, "main.go"), []byte("package main\n\n// mutated\nfunc main() {}\n"), 0o644); err != nil {
+		t.Fatalf("mutate copy: %v", err)
+	}
+	if err := os.WriteFile(filepath.Join(copyDir, "new.go"), []byte("package main\n"), 0o644); err != nil {
+		t.Fatalf("add file to copy: %v", err)
+	}
+
+	gotMain, err := os.ReadFile(filepath.Join(src, "main.go"))
+	if err != nil {
+		t.Fatalf("re-read source main.go: %v", err)
+	}
+	if string(gotMain) != string(wantMain) {
+		t.Fatalf("source fixture mutated by copy: got %q, want %q", gotMain, wantMain)
+	}
+	if _, err := os.Stat(filepath.Join(src, "new.go")); !os.IsNotExist(err) {
+		t.Fatalf("file added to copy appeared in source: %v", err)
+	}
+}
@@ -271,72 +271,55 @@ func skillSearchRoot(path string) string {
 // recognize a manifest path that points at the file rather than the directory.
 const skillFileName = "SKILL.md"

-// MergedSkills loads the default skills directory plus the supplied plugin skill
-// roots and returns one merged, name-deduplicated list (Content stripped, like
-// skills.List) alongside the duplicate-name collisions across all roots. Earlier
-// roots win a name clash, matching skills.Load's first-wins rule; the default dir
-// is always considered first so a user skill shadows a same-named plugin skill.
-// A bad root simply yields no skills rather than failing the merge.
+// MergedSkills loads the default skills directory, the shared ~/.agents/skills
+// root when present, plus the supplied plugin skill roots and returns one
+// merged, name-deduplicated list (Content stripped, like skills.List) alongside
+// the duplicate-name collisions across all roots. Earlier roots win a name
+// clash, matching skills.Load's first-wins rule; the default dir is always
+// considered first so a user skill shadows agents and plugins. A bad root simply
+// yields no skills rather than failing the merge.
 func MergedSkills(defaultDir string, pluginRoots []string) ([]skills.Skill, []skills.DuplicateName) {
 	return mergeSkills(defaultDir, pluginRoots, false)
 }

-// mergeSkills merges the default skills dir and plugin roots into one sorted,
-// name-deduplicated list. keepContent retains each skill's body (skills.Load)
-// versus stripping it (skills.List). Roots are considered in order with the
-// default dir first, so an earlier root wins a name clash; collisions are recorded
-// rather than crashing, preserving the skills package's Duplicates behaviour.
+// mergeSkills merges the primary skills dir, optional ~/.agents/skills, and
+// plugin roots into one sorted, name-deduplicated list via skills.LoadFromRoots /
+// ListFromRoots. keepContent retains each skill's body versus stripping it.
+// Roots are considered in order with the default dir first, so an earlier root
+// wins a name clash; collisions are recorded rather than crashing.
 func mergeSkills(defaultDir string, pluginRoots []string, keepContent bool) ([]skills.Skill, []skills.DuplicateName) {
-	merged := []skills.Skill{}
-	dups := []skills.DuplicateName{}
-	byName := map[string]skills.Skill{}
-
-	roots := append([]string{defaultDir}, pluginRoots...)
-	for _, root := range roots {
-		if strings.TrimSpace(root) == "" {
-			continue
-		}
-		var loaded []skills.Skill
-		var err error
-		if keepContent {
-			loaded, err = skills.Load(root)
-		} else {
-			loaded, err = skills.List(root)
-		}
-		if err != nil {
-			continue
-		}
-		for _, skill := range loaded {
-			if winner, clash := byName[skill.Name]; clash {
-				dups = append(dups, skills.DuplicateName{Name: skill.Name, Winner: winner.Path, Loser: skill.Path})
-				continue
-			}
-			byName[skill.Name] = skill
-			merged = append(merged, skill)
+	// Keep defaultDir injectable for tests rather than always calling DefaultDir.
+	roots := skills.GlobalRoots(defaultDir)
+	// GlobalRoots already includes AgentsDir; append plugin roots only.
+	for _, root := range pluginRoots {
+		if root = strings.TrimSpace(root); root != "" {
+			roots = append(roots, root)
 		}
 	}
-
-	sort.Slice(merged, func(left int, right int) bool {
-		return merged[left].Name < merged[right].Name
-	})
+	if keepContent {
+		merged, dups, _ := skills.LoadFromRoots(roots)
+		return merged, dups
+	}
+	merged, dups, _ := skills.ListFromRoots(roots)
 	return merged, dups
 }

 // skillTool is a drop-in replacement for the core skill tool that resolves a
-// named skill across the default skills directory PLUS the plugin skill roots, so
-// plugin-declared skills surface in the agent's skill list. It keeps the core
-// tool's name, schema, and read-only/allow safety so registering it simply
-// overlays plugin skills onto the existing surface.
+// named skill across the default skills directory, the shared ~/.agents/skills
+// root, and plugin skill roots, so agents- and plugin-declared skills surface in
+// the agent's skill list. It keeps the core tool's name, schema, and
+// read-only/allow safety so registering it simply overlays multi-root discovery
+// onto the existing surface.
 type skillTool struct {
 	defaultDir  string
 	pluginRoots []string
 }

-// NewSkillTool builds the plugin-aware skill tool. defaultDir is the standard
+// NewSkillTool builds the multi-root skill tool. defaultDir is the standard
 // skills directory (skills.DefaultDir); pluginRoots are the plugin skill search
-// roots from an ActivationResult. The returned tool merges both, deterministically
-// deduplicating by name (default dir wins a clash) and listing all available
-// skills when an unknown name is requested.
+// roots from an ActivationResult. The returned tool merges primary + agents +
+// plugins, deterministically deduplicating by name (default dir wins a clash)
+// and listing all available skills when an unknown name is requested.
 func NewSkillTool(defaultDir string, pluginRoots []string) tools.Tool {
 	return skillTool{defaultDir: defaultDir, pluginRoots: append([]string{}, pluginRoots...)}
 }
@@ -345,7 +328,7 @@ func (tool skillTool) Name() string { return "skill" }

 func (tool skillTool) Description() string {
 	return "Load a named Zero skill and return its instructions as the tool output. " +
-		"Skills are reusable, on-demand instruction sets (including any contributed by plugins). " +
+		"Skills are reusable, on-demand instruction sets (including shared ~/.agents/skills and any contributed by plugins). " +
 		"Call this when a relevant skill exists; an unknown name returns the list of available skills."
 }

@@ -350,6 +350,7 @@ func TestActivateResolvesManifestRelativeSkillRoot(t *testing.T) {
 }

 func TestNewSkillToolSurfacesPluginSkillInList(t *testing.T) {
+	isolateAgentsHome(t)
 	defaultDir := t.TempDir()
 	writeTestSkill(t, defaultDir, "core-skill", "core body")
 	pluginRoot := t.TempDir()
@@ -384,6 +385,7 @@ func TestNewSkillToolSurfacesPluginSkillInList(t *testing.T) {
 }

 func TestMergeSkillRootsListsPluginSkillInAgentList(t *testing.T) {
+	isolateAgentsHome(t)
 	defaultDir := t.TempDir()
 	writeTestSkill(t, defaultDir, "core-skill", "core")

@@ -404,6 +406,9 @@ func TestMergeSkillRootsListsPluginSkillInAgentList(t *testing.T) {
 }

 func TestMergedSkillsRecordsDuplicatesWithoutCrashing(t *testing.T) {
+	// Isolate AgentsDir so a host ~/.agents/skills cannot add extra collisions.
+	isolateAgentsHome(t)
+
 	defaultDir := t.TempDir()
 	writeTestSkill(t, defaultDir, "shared", "default copy")

@@ -426,6 +431,90 @@ func TestMergedSkillsRecordsDuplicatesWithoutCrashing(t *testing.T) {
 	}
 }

+// isolateAgentsHome points HOME/USERPROFILE at an empty temp home so AgentsDir
+// does not pick up the developer's real ~/.agents/skills during unit tests.
+func isolateAgentsHome(t *testing.T) {
+	t.Helper()
+	home := t.TempDir()
+	t.Setenv("HOME", home)
+	t.Setenv("USERPROFILE", home)
+}
+
+func TestMergedSkillsIncludesAgentsWithoutPlugins(t *testing.T) {
+	home := t.TempDir()
+	t.Setenv("HOME", home)
+	t.Setenv("USERPROFILE", home)
+	agents := filepath.Join(home, ".agents", "skills")
+	if err := os.MkdirAll(agents, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	writeTestSkill(t, agents, "agents-skill", "from agents")
+
+	defaultDir := t.TempDir()
+	listed, dups := MergedSkills(defaultDir, nil)
+	if len(dups) != 0 {
+		t.Fatalf("unexpected dups: %#v", dups)
+	}
+	if len(listed) != 1 || listed[0].Name != "agents-skill" {
+		t.Fatalf("agents skill should surface with no plugins, got %#v", listed)
+	}
+}
+
+func TestMergedSkillsPriorityPrimaryAgentsPlugins(t *testing.T) {
+	home := t.TempDir()
+	t.Setenv("HOME", home)
+	t.Setenv("USERPROFILE", home)
+	agents := filepath.Join(home, ".agents", "skills")
+	if err := os.MkdirAll(agents, 0o755); err != nil {
+		t.Fatal(err)
+	}
+
+	primary := t.TempDir()
+	pluginRoot := t.TempDir()
+	writeTestSkill(t, primary, "shared", "primary body")
+	writeTestSkill(t, agents, "shared", "agents body")
+	writeTestSkill(t, agents, "agents-plugin", "agents body")
+	writeTestSkill(t, pluginRoot, "shared", "plugin body")
+	writeTestSkill(t, pluginRoot, "agents-plugin", "plugin body")
+	writeTestSkill(t, pluginRoot, "plugin-only", "plugin only")
+
+	listed, dups := MergedSkills(primary, []string{pluginRoot})
+	byName := map[string]string{}
+	for _, skill := range listed {
+		// Content is stripped by MergedSkills; use path origin instead.
+		byName[skill.Name] = skill.Path
+	}
+	if !strings.Contains(byName["shared"], primary) {
+		t.Fatalf("primary should win shared, path=%q", byName["shared"])
+	}
+	if !strings.Contains(byName["agents-plugin"], agents) {
+		t.Fatalf("agents should win over plugin for agents-plugin, path=%q", byName["agents-plugin"])
+	}
+	if !strings.Contains(byName["plugin-only"], pluginRoot) {
+		t.Fatalf("plugin-only missing, path=%q", byName["plugin-only"])
+	}
+	if len(dups) < 2 {
+		t.Fatalf("expected cross-layer dups, got %#v", dups)
+	}
+}
+
+func TestNewSkillToolLoadsAgentsSkillWithoutPluginRoots(t *testing.T) {
+	home := t.TempDir()
+	t.Setenv("HOME", home)
+	t.Setenv("USERPROFILE", home)
+	agents := filepath.Join(home, ".agents", "skills")
+	if err := os.MkdirAll(agents, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	writeTestSkill(t, agents, "agents-skill", "agents body")
+
+	tool := NewSkillTool(t.TempDir(), nil)
+	got := tool.Run(context.Background(), map[string]any{"name": "agents-skill"})
+	if got.Status != tools.StatusOK || !strings.Contains(got.Output, "agents body") {
+		t.Fatalf("expected agents skill body with zero plugin roots, got %q (%s)", got.Status, got.Output)
+	}
+}
+
 func TestActivateSkipsMalformedPluginAndContinues(t *testing.T) {
 	registry := tools.NewRegistry()
 	good := toolPlugin(t.TempDir(), ToolExtension{Name: "good", Command: "good", Permission: PermissionPrompt})
@@ -34,6 +34,9 @@ const (

 var ErrUnknownProvider = errors.New("unknown provider")

+// AIMLAPIID is the provider catalog identifier for the aimlapi.com preset.
+const AIMLAPIID = "aimlapi"
+
 type Descriptor struct {
 	ID                  string
 	Name                string
@@ -41,6 +44,7 @@ type Descriptor struct {
 	DefaultBaseURL      string
 	DefaultModel        string
 	AuthEnvVars         []string
+	CustomHeaders       map[string]string
 	RequiresAuth        bool
 	UsesAmbientAuth     bool
 	Public              bool
@@ -67,9 +71,9 @@ type Descriptor struct {
 	// headless / SSH use) in addition to the browser flow.
 	OAuthDeviceFlow bool

-	// Recommended marks the provider that should be surfaced first and badged
-	// (★ … (recommended)) in every catalog-ordered list and picker. At most one
-	// descriptor should set this.
+	// Recommended marks a provider surfaced at the top and badged in
+	// catalog-ordered lists and pickers. The recommended descriptors are the first
+	// entries in the catalog, in order.
 	Recommended bool
 }

@@ -100,6 +104,11 @@ var descriptors = []Descriptor{
 	// minimax, qwen, google, nvidia, tencent, z-ai). Flat /v1/chat/completions
 	// with a Bearer ogw_live_… key; listed first and badged in every picker.
 	recommended(openAICompat("gitlawb-opengateway", "GitLawb OpenGateway", "https://opengateway.gitlawb.com/v1", "mimo-v2.5-pro", []string{"GITLAWB_OPENGATEWAY_API_KEY"}, "gitlawb opengateway", "opengateway")),
+	// aimlapi.com — OpenAI-compatible aggregating gateway, also badged as
+	// recommended (second, right after the OpenGateway default). The built-in TUI
+	// onboarding can register/top up an account and save the issued key; the partner
+	// headers attribute usage for the rebate.
+	recommended(aimlapi()),
 	openAI("openai", "OpenAI", "https://api.openai.com/v1", "gpt-4.1", []string{"OPENAI_API_KEY"}),
 	anthropic("anthropic", "Anthropic", "https://api.anthropic.com", "claude-sonnet-4.5", []string{"ANTHROPIC_API_KEY"}),
 	google("google", "Google", "https://generativelanguage.googleapis.com", "gemini-2.5-pro", []string{"GEMINI_API_KEY", "GOOGLE_API_KEY"}, "gemini"),
@@ -298,8 +307,21 @@ func google(id string, name string, baseURL string, model string, env []string,
 	}
 }

-// recommended marks a descriptor as the recommended default so list/picker
-// surfaces sort it first and render the ★ … (recommended) badge.
+func aimlapi() Descriptor {
+	descriptor := openAICompat(AIMLAPIID, "aimlapi.com", "https://api.aimlapi.com/v1", "anthropic/claude-sonnet-5", []string{"AIMLAPI_API_KEY"}, "aimlapi", "aiml api", "ai/ml api", "ai ml api")
+	descriptor.CustomHeaders = map[string]string{
+		"X-AIMLAPI-Partner-ID":          "part_62yQoGYDq4Yqnrj2R1iGrDNJ",
+		"X-AIMLAPI-Integration-Repo":    "Gitlawb/zero",
+		"X-AIMLAPI-Integration-Version": "zero",
+	}
+	// Onboarding is key-based only: the TUI sub-flow takes an existing key or an
+	// email top-up (internal/aimlapi Paths A/B). Not flagged
+	// OAuth-capable, so it does not appear in the wizard's "Sign in with OAuth" list.
+	return descriptor
+}
+
+// recommended marks a descriptor as a recommended default so list/picker
+// surfaces sort it to the top and render a provider-specific badge.
 func recommended(descriptor Descriptor) Descriptor {
 	descriptor.Recommended = true
 	return descriptor
@@ -380,5 +402,19 @@ func cloneDescriptor(descriptor Descriptor) Descriptor {
 	descriptor.AuthEnvVars = append([]string{}, descriptor.AuthEnvVars...)
 	descriptor.SupportedAPIFormats = append([]APIFormat{}, descriptor.SupportedAPIFormats...)
 	descriptor.Aliases = append([]string{}, descriptor.Aliases...)
+	if descriptor.CustomHeaders != nil {
+		descriptor.CustomHeaders = copyStringMap(descriptor.CustomHeaders)
+	}
 	return descriptor
 }
+
+func copyStringMap(values map[string]string) map[string]string {
+	if values == nil {
+		return nil
+	}
+	copied := make(map[string]string, len(values))
+	for key, value := range values {
+		copied[key] = value
+	}
+	return copied
+}
@@ -9,6 +9,7 @@ import (

 var expectedCatalogIDs = []string{
 	"gitlawb-opengateway",
+	"aimlapi",
 	"openai",
 	"anthropic",
 	"google",
@@ -68,25 +69,31 @@ func TestAllHasStableUniqueIDs(t *testing.T) {
 	}
 }

-func TestRecommendedProviderIsFirstAndUnique(t *testing.T) {
+func TestRecommendedProvidersAreTopOfCatalog(t *testing.T) {
 	descriptors := All()
-	if len(descriptors) == 0 {
-		t.Fatal("All() returned no descriptors")
+	// The recommended providers are badged and pinned to the top of the catalog,
+	// in this order (OpenGateway remains the default; aimlapi.com is also badged).
+	wantTop := []string{"gitlawb-opengateway", "aimlapi"}
+	if len(descriptors) < len(wantTop) {
+		t.Fatalf("All() returned %d descriptors, want at least %d", len(descriptors), len(wantTop))
 	}
-	if !descriptors[0].Recommended {
-		t.Fatalf("All()[0] = %q, want it to be the recommended provider", descriptors[0].ID)
-	}
-	if descriptors[0].ID != "gitlawb-opengateway" {
-		t.Fatalf("recommended provider = %q, want %q", descriptors[0].ID, "gitlawb-opengateway")
+	for index, id := range wantTop {
+		if descriptors[index].ID != id {
+			t.Fatalf("descriptors[%d] = %q, want %q", index, descriptors[index].ID, id)
+		}
+		if !descriptors[index].Recommended {
+			t.Fatalf("descriptors[%d] (%q) should be recommended", index, id)
+		}
 	}
+	// Exactly those are recommended, and they are contiguous at the top.
 	recommended := 0
 	for _, descriptor := range descriptors {
 		if descriptor.Recommended {
 			recommended++
 		}
 	}
-	if recommended != 1 {
-		t.Fatalf("recommended descriptor count = %d, want exactly 1", recommended)
+	if recommended != len(wantTop) {
+		t.Fatalf("recommended descriptor count = %d, want %d", recommended, len(wantTop))
 	}
 }

@@ -103,6 +110,28 @@ func TestRecommendedProviderEndpoint(t *testing.T) {
 	}
 }

+func TestAIMLAPIDescriptor(t *testing.T) {
+	descriptor, err := Require("aimlapi")
+	if err != nil {
+		t.Fatalf("Require(aimlapi) error = %v", err)
+	}
+	if descriptor.Name != "aimlapi.com" {
+		t.Fatalf("Name = %q, want aimlapi.com", descriptor.Name)
+	}
+	if descriptor.DefaultBaseURL != "https://api.aimlapi.com/v1" {
+		t.Fatalf("DefaultBaseURL = %q, want aimlapi.com endpoint", descriptor.DefaultBaseURL)
+	}
+	if descriptor.DefaultModel != "anthropic/claude-sonnet-5" {
+		t.Fatalf("DefaultModel = %q, want anthropic/claude-sonnet-5", descriptor.DefaultModel)
+	}
+	if descriptor.Transport != TransportOpenAICompatible {
+		t.Fatalf("Transport = %q, want %q", descriptor.Transport, TransportOpenAICompatible)
+	}
+	if !reflect.DeepEqual(descriptor.AuthEnvVars, []string{"AIMLAPI_API_KEY"}) {
+		t.Fatalf("AuthEnvVars = %#v, want AIMLAPI_API_KEY", descriptor.AuthEnvVars)
+	}
+}
+
 func TestLongCatDescriptor(t *testing.T) {
 	descriptor, err := Require("longcat")
 	if err != nil {
@@ -293,7 +322,7 @@ func TestListByTransportPreservesCatalogOrder(t *testing.T) {
 		TransportBedrock:         {"bedrock"},
 		TransportVertex:          {"vertex"},
 		TransportAnthropicCompat: {"minimax", "minimaxi-cn", "opencode-go-anthropic-compatible", "custom-anthropic-compatible"},
-		TransportOpenAICompat:    {"gitlawb-opengateway", "ollama-cloud", "ollama", "lmstudio", "openrouter", "huggingface", "chatgpt", "groq", "deepseek", "together", "dashscope", "moonshot", "longcat", "nvidia-nim", "mistral", "github", "xai", "venice", "xiaomi-mimo", "bankr", "zai", "zai-cn", "kilocode", "opencode", "opencode-go", "atomic-chat", "chatgpt-proxy", "custom-openai-compatible"},
+		TransportOpenAICompat:    {"gitlawb-opengateway", "aimlapi", "ollama-cloud", "ollama", "lmstudio", "openrouter", "huggingface", "chatgpt", "groq", "deepseek", "together", "dashscope", "moonshot", "longcat", "nvidia-nim", "mistral", "github", "xai", "venice", "xiaomi-mimo", "bankr", "zai", "zai-cn", "kilocode", "opencode", "opencode-go", "atomic-chat", "chatgpt-proxy", "custom-openai-compatible"},
 	}

 	for transport, wantIDs := range cases {
@@ -61,6 +61,17 @@ var zaiCuratedModels = []Model{
 }

 var curatedModels = map[string][]Model{
+	// aimlapi.com aggregates every upstream provider behind one OpenAI-compatible
+	// endpoint, so the curated defaults are a cross-provider "best of" flagship
+	// list. The live /v1/models discovery (when a key is present) replaces this;
+	// these are the offline/pre-key fallback and the picker's default order.
+	"aimlapi": {
+		{ID: "anthropic/claude-sonnet-5", Description: "catalog default (Anthropic Claude Sonnet 5)"},
+		{ID: "google/gemini-3.5-flash", Description: "Google Gemini Flash 3.5"},
+		{ID: "openai/gpt-5.5-2026-04-23", Description: "OpenAI GPT-5.5"},
+		{ID: "qwen/qwen-3.7-max", Description: "Qwen 3.7 Max"},
+		{ID: "deepseek/deepseek-v4-pro", Description: "DeepSeek V4"},
+	},
 	"ollama-cloud": {
 		{ID: "qwen3-coder:480b", Description: "catalog default"},
 		{ID: "gpt-oss:120b", Description: "agentic coding model"},
@@ -33,6 +33,11 @@ func TestModelsAreProviderScoped(t *testing.T) {
 			want:     []string{"gpt-5.5", "gpt-5.4", "gpt-5.4-mini", "gpt-5.3-codex-spark"},
 			notWant:  []string{"gpt-5", "gpt-4.1", "openai/gpt-4.1"},
 		},
+		{
+			provider: "aimlapi",
+			want:     []string{"anthropic/claude-sonnet-5", "google/gemini-3.5-flash", "openai/gpt-5.5-2026-04-23", "qwen/qwen-3.7-max", "deepseek/deepseek-v4-pro"},
+			notWant:  []string{"gpt-4o", "openai/gpt-4o", "anthropic/claude-opus-4.8", "x-ai/grok-4-5"},
+		},
 		{
 			provider: "mistral",
 			want:     []string{"mistral-large-latest", "codestral-latest"},
@@ -195,7 +195,7 @@ func discoverOpenAIModels(ctx context.Context, profile config.ProviderProfile, o
 		AuthHeader:        profile.AuthHeader,
 		AuthScheme:        profile.AuthScheme,
 		AuthHeaderValue:   profile.AuthHeaderValue,
-		CustomHeaders:     profile.CustomHeaders,
+		CustomHeaders:     providerio.CopyHeaders(profile.CustomHeaders),
 	}, nil)
 }

@@ -52,6 +52,44 @@ func TestDiscoverOpenAICompatibleModelsFetchesModelsEndpoint(t *testing.T) {
 	}
 }

+func TestDiscoverAIMLAPIModelsSendsAuthAndCustomHeadersWithoutAttribution(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		for header, want := range map[string]string{
+			"Authorization": "Bearer test-key",
+			"X-Trace":       "test",
+		} {
+			if got := r.Header.Get(header); got != want {
+				t.Errorf("%s = %q, want %q", header, got, want)
+			}
+		}
+		// No first-party referral/attribution headers are injected for catalog
+		// presets; aimlapi rides through CopyHeaders like every other provider.
+		for _, header := range []string{
+			"X-AIMLAPI-Partner-ID",
+			"X-AIMLAPI-Integration-Repo",
+			"X-AIMLAPI-Integration-Version",
+		} {
+			if got := r.Header.Get(header); got != "" {
+				t.Errorf("%s = %q, want no attribution header", header, got)
+			}
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"data":[{"id":"openai/gpt-5-chat"}]}`))
+	}))
+	defer server.Close()
+
+	_, err := Discover(context.Background(), config.ProviderProfile{
+		CatalogID:     "aimlapi",
+		ProviderKind:  config.ProviderKindOpenAICompatible,
+		BaseURL:       server.URL + "/v1",
+		APIKey:        "test-key",
+		CustomHeaders: map[string]string{"X-Trace": "test"},
+	}, Options{HTTPClient: server.Client()})
+	if err != nil {
+		t.Fatalf("Discover returned error: %v", err)
+	}
+}
+
 func TestDiscoverOpenAICompatibleModelsHonorsAuthHeaderValue(t *testing.T) {
 	// A profile can authenticate via a raw auth-header value instead of APIKey;
 	// discovery must send it rather than probe unauthenticated.
@@ -67,7 +67,7 @@ func New(profile config.ProviderProfile, options Options) (zeroruntime.Provider,
 			AuthHeader:            profile.AuthHeader,
 			AuthScheme:            profile.AuthScheme,
 			AuthHeaderValue:       profile.AuthHeaderValue,
-			CustomHeaders:         profile.CustomHeaders,
+			CustomHeaders:         providerio.CopyHeaders(profile.CustomHeaders),
 			OAuthResolver:         options.OAuthResolver,
 			MaxTokens:             resolved.maxOutputTokens,
 			HTTPClient:            options.HTTPClient,
@@ -83,7 +83,7 @@ func New(profile config.ProviderProfile, options Options) (zeroruntime.Provider,
 			AuthHeader:      profile.AuthHeader,
 			AuthScheme:      profile.AuthScheme,
 			AuthHeaderValue: profile.AuthHeaderValue,
-			CustomHeaders:   profile.CustomHeaders,
+			CustomHeaders:   providerio.CopyHeaders(profile.CustomHeaders),
 			OAuthResolver:   options.OAuthResolver,
 			MaxTokens:       resolved.maxOutputTokens,
 			HTTPClient:      options.HTTPClient,
@@ -97,7 +97,7 @@ func New(profile config.ProviderProfile, options Options) (zeroruntime.Provider,
 			AuthHeader:      profile.AuthHeader,
 			AuthScheme:      profile.AuthScheme,
 			AuthHeaderValue: profile.AuthHeaderValue,
-			CustomHeaders:   profile.CustomHeaders,
+			CustomHeaders:   providerio.CopyHeaders(profile.CustomHeaders),
 			OAuthResolver:   options.OAuthResolver,
 			MaxTokens:       resolved.maxOutputTokens,
 			HTTPClient:      options.HTTPClient,
@@ -220,6 +220,55 @@ func TestNewThreadsCustomProviderHeaders(t *testing.T) {
 	}
 }

+func TestNewAIMLAPIProviderSendsEndpointAndAuthWithoutAttribution(t *testing.T) {
+	transport := &captureTransport{responseBody: "data: [DONE]\n\n"}
+	provider, err := New(config.ProviderProfile{
+		Name:          "aimlapi",
+		CatalogID:     "aimlapi",
+		ProviderKind:  config.ProviderKindOpenAICompatible,
+		APIKey:        "aimlapi-test-key",
+		Model:         "openai/gpt-5-chat",
+		CustomHeaders: map[string]string{"X-Trace": "test"},
+	}, Options{HTTPClient: &http.Client{Transport: transport}})
+	if err != nil {
+		t.Fatalf("New() error = %v", err)
+	}
+	stream, err := provider.StreamCompletion(context.Background(), zeroruntime.CompletionRequest{
+		Messages: []zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: "hello"}},
+	})
+	if err != nil {
+		t.Fatalf("StreamCompletion() error = %v", err)
+	}
+	for range stream {
+	}
+
+	if transport.request == nil {
+		t.Fatal("HTTP client was not used")
+	}
+	if got := transport.request.URL.String(); got != "https://api.aimlapi.com/v1/chat/completions" {
+		t.Fatalf("request URL = %q, want AI/ML API endpoint", got)
+	}
+	for header, want := range map[string]string{
+		"Authorization": "Bearer aimlapi-test-key",
+		"X-Trace":       "test",
+	} {
+		if got := transport.request.Header.Get(header); got != want {
+			t.Fatalf("%s = %q, want %q", header, got, want)
+		}
+	}
+	// No first-party referral/attribution headers are injected for catalog
+	// presets; aimlapi rides through CopyHeaders like every other provider.
+	for _, header := range []string{
+		"X-AIMLAPI-Partner-ID",
+		"X-AIMLAPI-Integration-Repo",
+		"X-AIMLAPI-Integration-Version",
+	} {
+		if got := transport.request.Header.Get(header); got != "" {
+			t.Fatalf("%s = %q, want no attribution header", header, got)
+		}
+	}
+}
+
 func TestNewSupportsOpenAIProviderKind(t *testing.T) {
 	provider, err := New(config.ProviderProfile{
 		Name:         "openai",
@@ -4,6 +4,8 @@ import (
 	"context"
 	"net/http"
 	"strings"
+
+	"github.com/Gitlawb/zero/internal/trace"
 )

 // TokenResolver yields a fresh OAuth credential for one request, or ok=false to
@@ -42,7 +44,12 @@ func SendWithAuthRetry(
 		// request (leaking the path/body) before we return the error.
 		headers := base
 		if resolver != nil {
+			// ProviderQueue captures pre-send auth-wait (OAuth token resolve). No
+			// send-side semaphore exists today; a future request queue would also
+			// accumulate here. nil recorder (untraced) is a no-op.
+			queueSpan := trace.FromContext(ctx).Span(trace.SpanProviderQueue)
 			header, value, ok, rerr := resolver(ctx, forceRefresh)
+			queueSpan.End()
 			if rerr != nil {
 				return nil, rerr
 			}
@@ -16,6 +16,7 @@ import (
 	"strings"
 	"time"

+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )

@@ -371,9 +372,15 @@ func ScanSSEDataWithContext(
 				// The provider asked to stop (e.g. it already emitted an error
 				// for this payload). Abort the read and end like ScanSSEData:
 				// return nil so callers fall through to their post-scan checks.
+				// Do not stamp FirstToken — this payload was not accepted model
+				// output, so counting it as first-token time would be misleading.
 				cancel()
 				return nil
 			}
+			// First accepted non-keepalive payload = first real model output. Stamp
+			// once; later real payloads are no-ops. nil recorder (untraced run) is a
+			// no-op.
+			trace.FromContext(ctx).StampFirstToken()
 		}
 	}
 }
@@ -7,6 +7,8 @@ import (
 	"strconv"
 	"strings"
 	"time"
+
+	"github.com/Gitlawb/zero/internal/trace"
 )

 // Transient-failure retry, shared by every provider.
@@ -67,7 +69,9 @@ func SendWithRetry(
 			setHeader(request)
 		}

+		connectSpan := trace.FromContext(ctx).Span(trace.SpanProviderConnect)
 		response, err := client.Do(request)
+		connectSpan.End()
 		if err != nil {
 			// A transport failure on a POST does NOT mean the server didn't receive
 			// it — the request may have arrived and be generating a (billable,
@@ -82,6 +86,9 @@ func SendWithRetry(
 		}

 		if ShouldRetryStatus(response.StatusCode) && attempt < maxAttempts {
+			if r := trace.FromContext(ctx); r != nil {
+				r.Counter(trace.CounterRetryCount, 1)
+			}
 			wait := RetryAfter(response)
 			_ = response.Body.Close()
 			if Backoff(ctx, attempt, wait) {
@@ -14,18 +14,22 @@ type EngineOptions struct {
 	Store         *GrantStore
 	Backend       Backend
 	Scope         *Scope
+	// SensitiveEnvKeys adds config-derived credential variable names to the
+	// catalog and namespace secrets scrubbed from sandboxed commands.
+	SensitiveEnvKeys []string
 }

 type Engine struct {
-	workspaceRoot   string
-	policy          Policy
-	store           *GrantStore
-	backend         Backend
-	scope           *Scope
-	sessionGrants   *memoryGrantSet
-	sessionProfiles *permissionProfileGrantSet
-	turnProfiles    *permissionProfileGrantSet
-	commandPrefixes *commandPrefixGrantSet
+	workspaceRoot    string
+	policy           Policy
+	store            *GrantStore
+	backend          Backend
+	scope            *Scope
+	sensitiveEnvKeys []string
+	sessionGrants    *memoryGrantSet
+	sessionProfiles  *permissionProfileGrantSet
+	turnProfiles     *permissionProfileGrantSet
+	commandPrefixes  *commandPrefixGrantSet
 }

 func NewEngine(options EngineOptions) *Engine {
@@ -49,15 +53,16 @@ func NewEngine(options EngineOptions) *Engine {
 		scope = newScopeBestEffort(workspaceRoot)
 	}
 	return &Engine{
-		workspaceRoot:   workspaceRoot,
-		policy:          policy,
-		store:           options.Store,
-		backend:         options.Backend,
-		scope:           scope,
-		sessionGrants:   newMemoryGrantSet(),
-		sessionProfiles: newPermissionProfileGrantSet(),
-		turnProfiles:    newPermissionProfileGrantSet(),
-		commandPrefixes: newCommandPrefixGrantSet(),
+		workspaceRoot:    workspaceRoot,
+		policy:           policy,
+		store:            options.Store,
+		backend:          options.Backend,
+		scope:            scope,
+		sensitiveEnvKeys: normalizeSensitiveEnvKeys(options.SensitiveEnvKeys),
+		sessionGrants:    newMemoryGrantSet(),
+		sessionProfiles:  newPermissionProfileGrantSet(),
+		turnProfiles:     newPermissionProfileGrantSet(),
+		commandPrefixes:  newCommandPrefixGrantSet(),
 	}
 }

@@ -29,6 +29,9 @@ type CommandSpec struct {
 	Args []string
 	Dir  string
 	Env  []string
+	// sensitiveEnvKeys is populated by Engine from config-derived credential
+	// variable names and carried through SandboxManager to the platform plan.
+	sensitiveEnvKeys []string
 }

 type CommandPlan struct {
@@ -122,6 +125,7 @@ func (engine *Engine) BuildCommandPlan(spec CommandSpec) (CommandPlan, error) {
 		return CommandPlan{}, errors.New("sandbox command name is required")
 	}
 	spec.Dir = commandDir
+	spec.sensitiveEnvKeys = engine.sensitiveEnvKeys

 	backend := engine.backend
 	if backend.Name == "" {
@@ -207,7 +211,7 @@ func linuxSandboxHelperCommandPlan(execRequest SandboxExecutionRequest, policy P
 	if err != nil {
 		return CommandPlan{}, err
 	}
-	env := sandboxEnvironmentForCommand(spec.Env, policy, BackendLinuxBwrap, "")
+	env := sandboxEnvironmentForCommandWithSensitiveEnv(spec.Env, policy, BackendLinuxBwrap, "", spec.sensitiveEnvKeys)
 	planDir := spec.Dir
 	if helper.Dir != "" {
 		planDir = helper.Dir
@@ -253,10 +257,27 @@ func directCommandPlan(spec CommandSpec, backend Backend, policy Policy, workspa
 		Name:              spec.Name,
 		Args:              cloneStrings(spec.Args),
 		Dir:               spec.Dir,
-		Env:               cloneStrings(spec.Env),
+		Env:               directCommandEnv(spec),
 	}
 }

+// directCommandEnv scrubs sensitive credentials from the environment for a
+// direct (unwrapped) command plan: the platform sandbox backend is
+// unavailable, disabled, or not required, so this is the actual environment
+// the child process inherits. Without scrubbing here, config-derived
+// apiKeyEnv and dynamic OAuth client-secret variables would leak into
+// commands that fall back to this path (e.g. EnforcementDegraded).
+func directCommandEnv(spec CommandSpec) []string {
+	env := cloneStrings(spec.Env)
+	if spec.Env == nil {
+		// Match the wrapped-plan behavior: an unset spec.Env means "inherit the
+		// caller's environment," so scrub a snapshot of it rather than passing
+		// nil through to exec.Cmd, which would skip scrubbing entirely.
+		env = os.Environ()
+	}
+	return scrubSensitiveEnv(env, spec.sensitiveEnvKeys...)
+}
+
 func (engine *Engine) resolveCommandDir(dir string, policy Policy) (string, string, error) {
 	workspaceRoot := strings.TrimSpace(engine.workspaceRoot)
 	if workspaceRoot == "" {
@@ -318,7 +339,7 @@ func seatbeltCommandPlanWithProfile(spec CommandSpec, workspaceRoot string, prof
 	if envBackend == "" {
 		envBackend = BackendMacOSSeatbelt
 	}
-	env := sandboxEnvironmentForCommand(spec.Env, policy, envBackend, "")
+	env := sandboxEnvironmentForCommandWithSensitiveEnv(spec.Env, policy, envBackend, "", spec.sensitiveEnvKeys)
 	plan := CommandPlan{
 		Backend:           backend,
 		TargetBackend:     backend.TargetBackend(),
@@ -377,6 +398,10 @@ func sandboxEnvironment(policy Policy, backend BackendName, workspaceRoot string
 }

 func sandboxEnvironmentForCommand(specEnv []string, policy Policy, backend BackendName, workspaceRoot string) []string {
+	return sandboxEnvironmentForCommandWithSensitiveEnv(specEnv, policy, backend, workspaceRoot, nil)
+}
+
+func sandboxEnvironmentForCommandWithSensitiveEnv(specEnv []string, policy Policy, backend BackendName, workspaceRoot string, sensitiveEnvKeys []string) []string {
 	env := cloneStrings(specEnv)
 	if specEnv == nil {
 		// Preserve the caller environment for sandboxed commands. The sandbox
@@ -384,7 +409,7 @@ func sandboxEnvironmentForCommand(specEnv []string, policy Policy, backend Backe
 		// command env values still replace inherited values below.
 		env = os.Environ()
 	}
-	env = scrubSensitiveEnv(env)
+	env = scrubSensitiveEnv(env, sensitiveEnvKeys...)
 	pathValue := envListValue(env, "PATH", defaultPath())
 	if runtime.GOOS == "darwin" {
 		// Preserve standard user tool locations so a bare `python3`/`node`
@@ -967,7 +992,7 @@ func regexpQuoteMeta(value string) string {
 	return replacer.Replace(value)
 }

-func scrubSensitiveEnv(env []string) []string {
+func scrubSensitiveEnv(env []string, additionalKeys ...string) []string {
 	// Secrets not covered by the provider catalog: cloud/VCS credentials and
 	// providers Zero talks to through generic OpenAI-compatible endpoints.
 	sensitiveKeys := []string{
@@ -997,6 +1022,7 @@ func scrubSensitiveEnv(env []string) []string {
 			sensitiveKeys = append(sensitiveKeys, key)
 		}
 	}
+	sensitiveKeys = append(sensitiveKeys, normalizeSensitiveEnvKeys(additionalKeys)...)

 	out := make([]string, 0, len(env))
 	for _, kv := range env {
@@ -1005,7 +1031,7 @@ func scrubSensitiveEnv(env []string) []string {
 			out = append(out, kv)
 			continue
 		}
-		drop := false
+		drop := isDynamicSensitiveEnvKey(key)
 		for _, sensitive := range sensitiveKeys {
 			if strings.EqualFold(key, sensitive) {
 				drop = true
@@ -1018,3 +1044,37 @@ func scrubSensitiveEnv(env []string) []string {
 	}
 	return out
 }
+
+func normalizeSensitiveEnvKeys(keys []string) []string {
+	out := make([]string, 0, len(keys))
+	seen := make(map[string]struct{}, len(keys))
+	for _, key := range keys {
+		key = strings.TrimSpace(key)
+		// A misconfigured value like "COMPANY_LLM_SECRET=..." would never
+		// match a real env key during scrubbing; keep only the name part.
+		if name, _, found := strings.Cut(key, "="); found {
+			key = strings.TrimSpace(name)
+		}
+		folded := strings.ToUpper(key)
+		if key == "" {
+			continue
+		}
+		if _, ok := seen[folded]; ok {
+			continue
+		}
+		seen[folded] = struct{}{}
+		out = append(out, key)
+	}
+	return out
+}
+
+func isDynamicSensitiveEnvKey(key string) bool {
+	const (
+		prefix = "ZERO_OAUTH_"
+		suffix = "_CLIENT_SECRET"
+	)
+	key = strings.ToUpper(strings.TrimSpace(key))
+	return strings.HasPrefix(key, prefix) &&
+		strings.HasSuffix(key, suffix) &&
+		len(key) > len(prefix)+len(suffix)
+}
				`@@ -0,0 +1 @@`
				`{ "id": 1, "key": "value-1", "enabled": true }`
				`@@ -0,0 +1 @@`
				`{ "id": 2, "key": "value-2", "enabled": true }`
				`@@ -0,0 +1 @@`
				`{ "id": 3, "key": "value-3", "enabled": true }`
				`@@ -0,0 +1 @@`
				`{ "id": 4, "key": "value-4", "enabled": true }`
				`@@ -0,0 +1 @@`
				`{ "id": 5, "key": "value-5", "enabled": true }`
				`@@ -0,0 +1 @@`
				`{ "id": 6, "key": "value-6", "enabled": true }`