mirror of
https://github.com/Twigpine/zero.git
synced 2026-11-11 23:40:28 -10:00
feat/structural-code-query
13
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
|
8e266797a8 |
perf: improve harness context efficiency (#838)
* Improve harness context efficiency
* Address harness efficiency review findings
* Make recaps idle and cancellable
Delay recap generation until the session has been inactive, orient it around the active goal and next task, cancel stale work on activity, and render the result ephemerally above the composer.
Also strengthen the formatter safety regression to exercise formatter-modified edit output.
* Track every replace-all edit span
Translate each replacement into its post-edit line range, including cumulative shifts from earlier replacements, so follow-up edits remain authorized for every model-supplied change.
* Make review regressions cross-platform
Verify replace-all tracking through the canonical edit tool path instead of a platform-sensitive temp path, and cover cancellation when recaps are disabled during generation.
* fix(tools): derive SeenWhole from the recorded ranges
SeenWhole was a flag set only by a SINGLE read covering the file, while
SeenRange judged coverage from the accumulated ranges. The two then
disagreed about identical reads: after reading lines 1-50 and 51-100 of a
100-line file, SeenRange(1,100) was true and SeenWhole was false.
That matters because write_file gates overwrites on SeenWhole. A file past
the read byte budget cannot be read whole in one call — the read is
truncated and RecordSeenRange is correctly skipped, since the model did not
see those lines — so no sequence of reads could ever set the flag, and the
overwrite was refused permanently with:
the intended change depends on content that has not been read exactly
in this session. Read the affected lines with read_file, then retry.
Advice that could not change the answer. edit_file was unaffected because it
gates on SeenRange, so the escape was to edit rather than rewrite, but a
model following the error text would loop.
SeenWhole is now computed from the ranges against the file's recorded line
count, so scoped reads that together cover a file make it writable and the
error text becomes actionable. The observation carries that line count, and
a read reporting a different one clears the accumulated ranges: they
describe a shape the file no longer has.
Coverage is judged by merging sorted ranges rather than walking every line
against every range, which the previous SeenRange did — same answer, and it
no longer scans line-by-line over a large file.
Co-Authored-By: Claude Opus 5 <[email protected]>
* test(tools): key the large-file test on the resolved path
read_file resolves the workspace root through EvalSymlinks before recording
anything, so the tracker is keyed on the resolved path. The test asserted
against the raw t.TempDir() path, which is the same string on Windows and a
different one on macOS, where the temp dir sits under /var — a symlink to
/private/var. The SeenWhole assertions would have looked up a key that never
existed: passing here, failing on macOS CI.
Also switches the line generator to fmt.Fprintf (staticcheck QF1012).
Both from CodeRabbit on #842.
Co-Authored-By: Claude Opus 5 <[email protected]>
* test(tools): actually switch the line generator to fmt.Fprintf
The previous commit's message claimed this was done. It was not: the edit
silently did not apply and I reported it fixed without re-reading the file,
so builder.WriteString(fmt.Sprintf(...)) survived at the head — caught by
@anandh8x, not by me.
Verified this time by running the linter rather than by reading the diff:
golangci-lint --enable-only unused,ineffassign,staticcheck ./internal/tools/
reports 0 issues.
* fix: keep exact-read workflows recoverable
* test: cover permission normalization wiring
* fix: commit file observations after output budgeting
* fix: defer file observations across retries
* test: normalize patch line endings
---------
Co-authored-by: Vasanthdev2004 <[email protected]>
Co-authored-by: Claude Opus 5 <[email protected]>
|
||
|
Anandan
|
9242b9e037 |
Add transcript-grounded structured task state (#761)
* Add transcript-grounded structured task state Project existing plan, tool-result, changed-file, verification, and completion events into a deterministic per-run snapshot. Emit content-free aggregate snapshots to traces. Carry bounded objective context through compaction and ground completion checks only when the projection matches the transcript; mismatches retain transcript-derived behavior and drop stale carried state. Tested: make fmt-check Tested: go vet ./... Tested: go test ./... Tested: go test -race ./internal/agent ./internal/trace Tested: go run ./cmd/zero-release build Tested: go run ./cmd/zero-release smoke Tested: govulncheck ./... (no vulnerabilities found) Advisory: pinned golangci-lint reports only unrelated pre-existing findings Tested: git diff HEAD --check * Preserve task objective across repeated compaction Keep the immutable run objective when plan calls have been elided, while admitting mutable task fields only when plan parity still matches the transcript. Make plan parsing match update_plan rejection semantics, coalesce and cap task-state trace events, rename parity to its plan-only scope, use explicit nilable dependencies, and remove confirmable objective fingerprints from traces. Tested: make fmt-check Tested: go vet ./... Tested: go test ./... Tested: go test -race ./internal/agent ./internal/trace Tested: go run ./cmd/zero-release build Tested: go run ./cmd/zero-release smoke Tested: govulncheck ./... (no vulnerabilities found) Advisory: pinned golangci-lint reports only unrelated pre-existing findings Tested: git diff HEAD --check |
||
|
Vasanth T
|
abe24bcb10 |
Add per-turn tracing and a turn-benchmark harness (Phase 0 baseline) (#700)
* Add per-turn tracing and a turn-benchmark harness (Phase 0 baseline)
Introduce an opt-in `internal/trace` package that attributes a run's wall
time to named spans — prompt build, provider connect/queue, generation,
tool queue/execution, permission wait, verification, compaction,
persistence — plus counters for model requests, tool calls, retries,
reconnects, compactions, completion nudges, acceptance checks, model
switches, and input/cached/output tokens. The recorder is nil-safe (a nil
*Recorder is a no-op), so every stamp site can call it unguarded; an
untraced run is byte-identical to today.
The recorder threads into agent.Run via a new `Options.Trace` field and a
single `trace.WithContext` assignment, so the shared providerio seam
(connect around client.Do, queue around OAuth resolve, first token at the
first non-keepalive SSE payload) reaches it through `trace.FromContext`
without any provider-interface change. All instrumentation is opt-in: when
`Options.Trace` is nil the hot path is untouched.
Headless `zero exec` gains `--trace <path>` (and `ZERO_TRACE=<path>`),
writing an agenteval-contract-compatible NDJSON trace (`-` for stderr).
The trace is pure observation; enabling it does not change behavior.
A new per-turn benchmark harness extends `internal/perfbench` (reusing the
existing TaskSet/NumericStats/subprocess runner shapes): `RunTurnBench`
aggregates per-span median/P95 across iterations, ranks the top three
controllable latency sources by share of attributed time, rolls up per
task class, and totals tokens/counts. A checked-in 42-task manifest across
seven classes (nav, edit, fix, refactor, longproc, longctx, parallel) and
local offline fixtures back it, and `make baseline` runs it against the
built `zero`, writing `internal/perfbench/reports/baseline.json`. The
report is machine-specific and regenerated, not hand-edited.
The manifest is JSON rather than YAML to avoid adding a dependency;
per-task verification targets the specific test (`go test -run <name>`)
so fix tasks are independent. Each task is a fresh `zero exec` process, so
iterations are cold-start samples; a warm in-process path is left for
follow-up.
Verified: `go build ./...`, `go vet ./...`, `gofmt -l` clean on touched
files; `go test -race` green for internal/trace, internal/agent,
internal/perfbench, and cmd/zero-perf-bench.
* test(trace): make TestAttributionRatio clock-independent
The test recorded two synthetic 10ms spans with no real delay and asserted
the attribution ratio stayed <= 1.0. On a precise clock (Linux/macOS CI) the
wall is microseconds while attributed is 20ms, so attributed/wall blows past
1.0 and the test failed — even though the ratio legitimately exceeds 1.0
when spans overlap (parallel tool execution). On Windows the coarse timer
masked it, so the failure only surfaced in CI.
Assert the contract instead: attributed is the deterministic sum of span
durations, and the ratio equals attributed/wall (0 when wall is zero). Add a
zero-wall case for the divide-by-zero guard.
* Address review feedback on the per-turn tracer + benchmark harness
Benchmark runner (internal/perfbench):
- A nonzero agent exit code no longer counts as Passed. The runner set
VerifyErr but then fell through to outcome.Passed = true, so a crashed
run could be recorded as a pass. Return before the verification block
when VerifyErr is already set.
- A task now passes only when every iteration passes. The previous
"any iteration passes" rule let a flaky fail hide behind one good run.
- Each invocation copies its workspace fixture to a fresh temp dir and
runs the agent + verifier there, so mutating tasks (edit/fix/refactor)
can't dirty the checked-in fixtures or bleed into the next iteration.
- Trace load failures (missing/malformed trace file) are surfaced as
result warnings via a new TraceIssue field instead of being swallowed.
Tracer (internal/trace):
- Finish freezes the recorder: Counter, addSpan, and the first-event
stamps are dropped after Finish returns, so a snapshot captures exactly
one turn and late stamps can't mutate it.
- WriteText propagates the first write error instead of silently
swallowing sink failures.
Provider seam (internal/providers/providerio):
- FirstTokenAt is stamped only after a payload is accepted as real model
output, not before handle() decides to drop it as an error payload.
Agent loop (internal/agent):
- The compaction counter increments only when the history actually
shrinks, in both the proactive and reactive paths — paid no-ops no
longer inflate the compaction total.
- Document on Options.Trace that a fresh Recorder is required per Run.
CLI (internal/cli/exec):
- Finish the trace at the agent.Run boundary and defer only the
serialization, so the snapshot covers exactly the turn.
- Reject --trace / ZERO_TRACE for spec-draft runs with a clear error
instead of silently accepting and writing nothing.
Manifest + fixtures:
- Strengthen the edit verifiers: rename and value-bump tasks now assert
the old symbol is gone, and the getter task greps the method signature.
- Add the missing testdata/edit fixture (10 edit tasks referenced a
directory that was never committed, so the whole edit class errored
out at run time).
- The manifest test now stats each fixture so a missing dir fails at
load time instead of at every run.
- longproc benchmark keeps Process live across the loop; README code
fence gets a language tag and documents the per-invocation fixture copy.
- Add tests for the tracing wrapper (nil vs wired recorder, OnUsage
forwarding, token aggregation) and the frozen-recorder contract.
* Address review feedback: exclusive-time attribution + wired spans + fix-suite compile
Rewrite the trace attribution model so concurrent/nested spans no longer
double-count. Spans record wall intervals; the recorder derives a parent for
each by interval containment and computes exclusive time (duration minus the
union of its children). AttributedDuration sums only top-level spans; Coverage
is the union of intervals over wall, capped at 1, the honest ">=95%" metric.
Top-latency ranking uses exclusive time so shares sum to ~1 instead of being
inflated by overlapping provider_connect/generation and nested
permission_wait/tool_execution.
Wire the spans that were named but never stamped: verification around the
self-correct verify pass, and StampFirstVisibleEvent on the first forwarded
text event. Drop the tool_queue/process_wait/persistence constants that had no
stamps, and rename the misnamed near-noise prompt_build span to tool_partition.
Trim RequiredEventKeys to events a healthy short run actually emits; move
conditional phases (tools, permission, compaction, verification) to
OptionalEventKeys so a short trace does not false-fail.
Make ReadNDJSON fail loudly on a corrupt file: require a type:trace header,
error on span/counter lines that appear before it, and error when a header
yields no spans or counters.
Fix the blocking fix-suite compile error: Label's %d format verb on a string
tripped go test's vet before any fix task could verify. Make it a runtime
wrong-prefix bug ("admin-" vs "user-") that still type-checks, so each fix
task is independent again.
Manifest: correct refactor-04's stats.go to main.go, reword fix-07 to the
wrong-prefix bug, add a latency-only description (read-only tasks carry no
verification oracle by design), and frame the report README on exclusive time
and coverage.
Tests updated for the occurrence/exclusive/coverage model; trace, perfbench,
and agent green under -race.
* Fix exclusive-time round-trip, identical-interval nesting, and empty-trace parse
Three correctness bugs in the attribution model found by an adversarial review
of the prior commit, all undermining the no-double-count guarantee:
1. exclusive_ms:0 was overwritten with the inclusive Duration on re-parse. A
parent whose children tile its interval legitimately has exclusive time 0;
WriteNDJSON emits exclusive_ms:0, but ReadNDJSON's "<= 0 means missing"
fallback turned that back into the full duration, so every re-parsed trace
re-introduced the double-counting the model exists to prevent (the benchmark
harness, which ranks re-parsed traces, inherits it). Gate the fallback on
key ABSENCE, so a written 0 is preserved.
2. deriveNesting formed a 2-cycle for two spans with identical [start, end]
intervals: the symmetric containment check made each parent the other,
dropping both from top-level and zeroing AttributedDuration. Break the tie
deterministically — strict containment is an unambiguous parent; for
identical intervals, only the lower-indexed span may parent the other, so
one stays top-level.
3. ReadNDJSON returned a valid empty TurnTrace for empty/blank-only input, so a
run that crashed before emitting a trace (or where --trace was not honored)
masqueraded as a clean zero-attribution sample. Empty input now errors, so
the harness records a TraceIssue.
Also: decode counters via json.Number so int64 values above 2^53 round-trip
exactly instead of losing precision through float64; and drop the harness's
own "exclusive <= 0 -> Duration" ranking fallback, which was the same
double-count on the re-parsed path. Added regression tests covering the
zero-exclusive round-trip, the identical-interval nesting, empty-input
rejection, and counter precision.
* perf(turn-bench): split pass/fail into oracle tiers so the pass rate can't be misread as correctness
The turn benchmark counted any exit-0 task as passed, including the 24
read-only tasks (nav/longproc/longctx/parallel) that carry no verification
oracle — so the JSON tasksPassed / "N/M passed" headline could read as a
correctness verdict it never was. Split pass/fail into three oracle tiers:
- correctness (edit grep + fix go test): the only pass rate that can move
with model quality (tasksVerified / tasksPassed / correctnessPassRate)
- build-only (refactor go build): a non-positive oracle, reported as
buildPassRate and excluded from correctnessPassRate; the manifest declares
buildOnlyClasses: ["refactor"]
- latency-only (no oracle): counted in latencyOnlyTasks, never in any pass
rate
NewTurnExecRunner no longer sets Passed=true for tasks with no
verificationCommand, and the summary headline names each tier explicitly so
an exit-0 read-only run cannot inflate a correctness number. Bumps
TurnSchemaVersion to 2. Class breakdown and tier contract documented in
internal/perfbench/MANIFEST.md and reports/README.md.
Also install the trace first-token/visible-event forwarding handlers in
agent.Run whenever options.Trace is set, not only when a UI OnText /
OnReasoning callback is set, so a headless traced run (zero exec --trace)
still captures FirstTokenAt. forwardedVisibleText stays tied to the user
callback so stall-retry semantics are unchanged. Adds a regression
assertion to the existing traced-run test.
* Address CodeRabbit review on
|
||
|
Anandan
|
afb5010a46 |
polish v2: smoothness & delight (#484)
* feat(tui): actionable hints on provider/model errors Classify a failed turn's provider error into auth / rate-limit / connectivity / model-not-found / context-overflow and render a one-line next step below the red error row, instead of dumping an identical raw blob for every failure mode. The classifier lives in a new internal/errhint package so both the TUI error row and the CLI exec path can share it (TUIHint references slash commands, CLIHint references zero subcommands). Classification is a conservative string heuristic keyed off providerio.ClassifiedError's prefixes plus lower-level DNS/TLS/timeout/context-length signatures, since the numeric HTTP status is gone by the time the error reaches a UI surface. * feat(tools): budget bash output with head+tail truncation bash was the one tool that returned command output with no byte cap — 'cat large.log' or a verbose test run could dump megabytes into context and force compaction, while every other read/search tool already applies a budget. Cap stdout and stderr at 96KB each, keeping the head and tail of an oversized stream (build/test failures usually land at the tail) and dropping the middle behind a marker that points at redirect-to-file + read_file. Record raw_bytes/emitted_bytes/estimated_tokens/truncated in Meta like the other tools. Sandbox-denial and shell-issue detection still run on the full raw output before budgeting. * fix(gemini): retry 401 with refreshed OAuth token The Gemini provider used providerio.SendWithRetry (429/503/529 only), so an expired OAuth token surfaced as a terminal 'API key not valid' auth error while OpenAI and Anthropic silently recovered via SendWithAuthRetry's force-refresh-and-replay on 401. Plumb OAuthResolver through gemini.Options/Provider and the factory (exactly as the other two providers already do) and switch the stream send to SendWithAuthRetry. With a nil resolver (API-key users) behavior is unchanged. * fix(errhint): gate classification on provider-origin marker Address CodeRabbit review on #1: agentResponseMsg.err can also carry local failures (a tool's 'permission denied', a 'file does not exist', a config error), and broad substrings would have attached a bogus /provider or /model hint to those. Classify now returns Unknown unless the message carries a provider marker the provider layer always attaches (auth error: / rate limit error: / provider error: / provider request error: / provider stream error:). Local errors never draw a provider hint; provider errors sub-classify as before. Added local-failure test cases proving they stay Unknown. * feat(cli): actionable hint on text-mode exec provider errors zero exec printed provider failures as a bare '[zero] <raw error>' with no next step. Append a one-line hint (reusing the shared errhint classifier) for recognized provider failures — 'run `zero auth`', 'run `zero doctor`', etc. The provider-origin gate means non-provider codes (sandbox_error, mcp_error) and JSON/stream-json output are untouched. * fix(tools): set Result.Truncated for budgeted bash output Address CodeRabbit review on #3: budgetBashOutput recorded truncation in meta["truncated"] but never set the Result.Truncated struct field, so consumers reading the struct (like glob/web_fetch/read_minified do) would miss truncated bash results. Return the bool from budgetBashOutput and set Result.Truncated on every return path. * perf(tui): reuse one LSP manager across prompts runAgentWithOptions built a fresh lsp.NewManager per run and shut it down when the run returned, so gopls (and every other language server) cold-started on the first edit of every turn — 200ms-2s of latency the manager's own design (long-lived, reused servers) exists to avoid. Build one manager per session in newModel (cheap: servers start lazily on first Check) and reuse it across runs; the SelfCorrector still gets a fresh checker wrapping the shared manager each run. Torn down with a short deadline in quit(). Runs fall back to a per-run manager when the session manager is absent (cwd unknown, or a directly-built test model). * perf(tui): accumulate streamingText as []byte (O(1) append) m.streamingText += msg.delta was O(len) per delta -> O(n²) across a long generation (a 10k-token code gen allocates hundreds of MB of intermediate strings and the UI gets progressively laggier). Accumulate into a []byte with append (O(1) amortized) and read via streamingTextString(). A []byte rather than strings.Builder because the TUI model is copied by value on every Update, which would trip strings.Builder's copy check; []byte is nil-safe and copies cleanly like the other slice fields. * perf(tui): coalesce streamed text deltas to one frame Every OnText delta was its own tea.Msg, so a fast provider (100+ tok/s) drove 100+ full Update->View cycles per second, each re-parsing the growing markdown — visible stutter over SSH and wasted CPU. Batch agentTextMsg deltas at the runtime sink over a ~16ms frame and forward them as a single message, decoupling render rate from token rate. Any non-text message flushes pending text first so ordering with tool-call / reasoning / row messages is preserved; the final agentResponseMsg (a tea.Cmd return, not a sink message) is safe because the model already drops deltas for an inactive runID. * fix(agent): raise tool-failure stop threshold 4->6 guardrails.go halted a run after 4 consecutive same-error tool failures. A corrective hint fires at 2, and a model iterating on a genuinely tricky edit can legitimately fail a couple more times after the hint while converging — stopping at 4 cut those runs short. Raise to 6; the streak still resets the instant the tool succeeds or hits a different error, so only true same-error loops are affected. * fix(agent): more reconnect retries with jitter; narrow to transport errors #19: bump maxStreamReconnects 2->4 and add up-to-50% jitter on the exponential backoff (capped at 8s), so a multi-second network blip is ridden out instead of killing the run on a 2s hiccup, and concurrent runs (swarms, cron fleets) don't reconnect in lockstep. #34: drop the '502'/'503' substring matches from shouldReconnect. 503 already exhausted providerio.SendWithRetry (retrying here is a redundant double-retry) and 502 is non-idempotent by providerio's rule (the POST may have been processed). Only genuine transport failures — where no response was received — reconnect now. * fix(agent): reduce stall retries 2->1 to bound stuck-session time A no-output stream stall is detected only after the full stream idle timeout (~5min) elapses, and each retry can idle again — so 2 retries left an interactive session frozen for ~15min. Drop to 1 retry: keeps the common single-hiccup recovery while bounding the worst case to ~2x the idle timeout. * fix(tui): serialize coalescer forwarding to preserve message order Address CodeRabbit review on #7: flush() and the run-switch path drained the text buffer under c.mu but called forward() after releasing it, so a timer-fired text flush could race a concurrent non-text send() and land after the tool-call/reasoning message it should precede. Hold c.mu across drain AND forward (drainAndForwardLocked), so whoever holds the lock delivers atomically and the other caller blocks until it is done — text can never overtake a following non-text message. Added a concurrency stress test (timer racing an inline send) that passes under -race. * feat(tui): accept /compact now Bare /compact already triggers compaction, but /compact now — what users reach for when the context gauge climbs — hit the usage-error branch. Accept the 'now' keyword and advertise it in the command usage. * feat(tui): add /retry, /edit, /copy, /export commands /retry resends the last prompt; /edit recalls it into the composer to tweak and resend — both read a new lastPrompt field captured verbatim (pre-expansion) in launchPrompt. /copy puts the last answer on the clipboard via the existing copy machinery; /export writes a plain role-prefixed transcript to a file (timestamped default, or a given path). Real gaps for SSH users with no mouse select. * feat(tools): separator-insensitive tool_search matching tool_search ranked deferred tools by exact substring, so a model that typed 'webfetch' (dropping the underscore) matched nothing and looped. Add a separator-squashing fallback ('web_fetch' -> 'webfetch') scored below an exact substring match, so precise queries still rank first. * feat(tools): signal ask_user dismissal distinctly from a blank field Empty answers all rendered as '(no answer provided)', so the model couldn't tell a wholesale dismissal (user closed the prompt without answering anything) from one field left blank amid real answers — and might invent a default. FormatAskUserAnswers now flags a full dismissal up front as a skip and marks individual empties '(left blank)' vs '(skipped)'. * feat(tui): fuzzy ranking in model/session pickers The pickers filtered by flat strings.Contains with no ranking, so finding 'sonnet 4.5' among 50+ models meant an exact prefix or scrolling. Rank matches (exact < prefix < contains < subsequence, reusing fuzzySubsequenceGap) so the closest match lands on top, and 'snt45' now matches 'Sonnet 4.5'. Groups stay contiguous — ordered by their best match, never split — so the grouped model picker still renders one header per provider. Covers both /model and /resume (shared commandPicker). * perf(agent): cache per-tool schema render across turns partitionTools re-ran the recursive schema->map conversion (schemaToRuntimeMap) for every tool on every turn — redundant, since a tool's advertised name/description/schema is stable for the run. Add a per-run definition cache keyed by tool name (nil-safe; the plain partitionTools entrypoint and tests pass nil for fresh renders). The partitioning itself (visibility, deferral, ordering) still recomputes every turn — it must, because a tool's deferred state can flip mid-run (swarm tools un-defer once a swarm is active). Only the expensive schema render is memoized. tool_search is excluded by its callers (dynamic description) so it never poisons the cache. * perf(agent): calibrate compaction estimate against real prompt tokens The byte/4 heuristic (ApproxTextTokens) over-counts code-heavy content ~15-20%, so with triggerRatio=0.7 compaction fired at ~60% of true capacity — premature summarizer calls that degrade quality. Each turn now folds (rawEstimate, provider InputTokens) into an EMA calibration ratio, clamped per-sample to a sane band; maybeCompact and the reactive path scale their estimates by it. Turn 1 uses the raw estimate (no data yet); later turns compact near real capacity. * style: gofmt the new agent test files The CI gofmt check flagged the #22/#25 test files (struct-field and switch alignment). No behavior change. * feat(agent): turn-based plan staleness reminder The stale-plan nudge only fired after 10 tool calls since the last update_plan, so a plan that drifts stale across many low-tool-call turns (one call per turn) took many turns to catch — the 'plan set turn 1 stays pending forever' case. Add a turn-count complement: after stalePlanTurnThreshold (8) turns without an update AND with items still pending, fire the same one-shot reminder. Gated on pending items so a fully-completed plan is never nagged; the existing tool-call trigger is unchanged. (The roadmap's fuzzy 'tool calls unrelated to plan items' relevance match is deliberately omitted — too false-positive-prone.) * feat(agent): preserve recent-edits summary across compaction Compaction preserved the plan, loaded skills, tool schemas, and project instructions, but not file edits — so after the editing turns were elided the model knew it had changed a file only vaguely, and would re-read to rediscover its own footprint. Extract write_file/edit_file targets from the elided middle with a one-line note from each tool result, merge them across repeated compactions (newer note per path wins), and carry them in the preserved-state JSON (capped at 20 paths / 160 bytes per note). * feat(tui): workspace-scoped /resume + fewer event reads /resume listed sessions from every project globally, and read the full event file for each of them to test for resumable content. Filter the picker (and /resume latest) by the current workspace Cwd, checked BEFORE the per-session read so a large global history no longer pays N full file reads to build one workspace's list; also skip zero-event sessions via metadata without any read. Sessions with no recorded Cwd (older runs) and an unknown current workspace stay visible so nothing is hidden that can't be confidently placed. Explicit /resume <id> still resolves any session. * feat(tui): warn on model switch that drops staged images Staged images were silently dropped at submit when the active model had no vision support — the user only learned after sending. Chips already render above the composer (renderAttachmentChips); add an immediate amber warning line to the /model switch status (both same-provider and cross-provider paths) when images are staged and the newly selected model can't accept them, so the drop is surfaced at switch time. * feat(tui): inline recovery affordance on setup errors A first-run setup error rendered as a bare red 'error: <text>' with no pointer to the fix. Append a faint, stage-tailored recovery line (edit the endpoint/name/key above then Enter, or pick another model) that names only keys the stage's footer confirms, so the guidance is always accurate. Empty for stages whose footer already makes recovery obvious. * fix: address CodeRabbit review (5 findings) - loop.go: recompute the compaction calibration estimate at calibration time, not request-build time, so a reactive-compaction reissue that shrinks the request feeds the EMA the sample that actually completed. - errhint: match bare HTTP status codes (401/403/429/529) only at digit boundaries, so an incidental number ('4290ms', 'id 14015') isn't mis-bucketed as auth/rate-limit. - bash: drop the stray unmatched ']' in the output-truncation marker. - tui /retry: apply the same exiting/compactInFlight guards commandPrompt has, so a retry can't race compactResultMsg's rewrite of session state. - tui /export: write the transcript 0o600 (it may contain echoed secrets), not world/group-readable 0o644. * test(tui): skip export file-mode assertion on Windows os.WriteFile ignores Unix permission bits on Windows and Stat reports 0666, so the 0o600 assertion failed there. The production 0o600 is honored on Linux/macOS (where it matters); guard the check with runtime.GOOS so Windows smoke passes. * fix(tui): align /resume latest filters + case-insensitive workspace match on Windows Address CodeRabbit review on #21: - latestResumableInWorkspace now applies the SAME filters as the picker (skip zero-event metadata and empty/failed runs), so /resume latest can't land on a session the picker intentionally hides. - sessionMatchesWorkspace compares paths case-insensitively on Windows (its filesystem is), so the same workspace spelled with different casing no longer hides resumable sessions; other platforms stay case-sensitive. * fix(tui): /retry resends the last prompt's image/PDF attachments /retry relaunched only lastPrompt's text, but launchPrompt clears the pending image/document queues once a turn is sent. A vision- or PDF-backed prompt that failed would silently retry as text-only and answer a different task. Snapshot the consumed attachments at launch and re-stage them on /retry so launchPrompt rebuilds an identical request (document preamble, images, and the submit-time vision re-check). startNewSession drops the snapshot so a post-/new /retry cannot leak the previous session's attachments. Also correct the OAuthResolver factory comment, which omitted the Gemini provider it is now passed to. * fix: address jatmn review (reconnect 5xx, /edit attachments, edit-cap, coalescer test) - reconnect: exclude HTTP 5xx (500/502/503/504) before the transport substring match so "504 Gateway Timeout" no longer slips through on the generic "timeout" needle. A gateway timeout is non-idempotent — the completion POST may already have reached the model — so replaying the connect risked duplicate billable work. Reuses a digit-boundary status matcher exported from errhint (HasStatusCode). Test asserts 504/500 stay non-reconnectable. - tui /edit: re-stage the remembered image/PDF snapshot alongside the recalled prompt text, mirroring the /retry fix — editing a vision- or document-backed prompt no longer silently resends a text-only version. - compaction recent-edits: order edits by LAST edit, not first, and move a re-touched path to the newest position on merge, so the maxRecentEdits tail cap keeps the file the model most recently edited instead of dropping it. Adds mergeRecentEdits (edit-specific) and a capped-list regression. - coalescer test: inject the frame timer via afterFunc so TestCoalescerBatchesDeltas proves buffering deterministically instead of racing the real 16ms wall-clock timer. * fix(tools): bound bash output capture in memory, not just model-visible text Previously stdout/stderr streamed into unbounded bytes.Buffers and were only truncated to the 96 KiB budget AFTER the command finished, so a runaway command (cat huge.log, yes) could grow Zero's memory until it stalled or OOMed before budgetBashOutput ever ran. Replace the capture with boundedBuffer, an io.Writer that keeps only the head+tail each stream will surface and discards the middle as it arrives, while counting the true total for the truncation marker. Memory is now bounded to ~head+2×tail per stream regardless of output size. truncateHeadTail is refactored onto a total-aware core so the string path (and its tests) are unchanged, and budgetBashCapture reports the true raw_bytes even though only a bounded slice was held. Adds boundedBuffer/capture unit tests and an end-to-end test that streams ~5× the budget and asserts bounded emission with an accurate raw_bytes. |
||
|
Anandan
|
f401c66e81 |
polish: cut per-turn overhead, stabilize cache, add /new (#478)
* polish: cut per-turn overhead, stabilize cache, add /new Context economy - Compaction: preserveLast 8->6, triggerRatio 0.8->0.7 (both copies in sync). - System prompt: kill the "ELABORATE summary" directive; scale summary to the work. Comment discipline moved into the core prompt for all model families (was Anthropic-only); Minimax/DeepSeek/Qwen/etc. now get it. - Token-budget regression test ratchets the fixed per-turn overhead (base prompt + eager tool schemas) using the runtime's own estimator, so silent prompt/schema growth fails CI. Response quality - defaultResponseStyle balanced->concise. Reversible via /style. Tool discovery - <available_skills> block in the system prompt lists installed skills (name + one-line description) so the model loads the right one on the first call instead of guessing and reading the failure. Budget-capped at 640 bytes with an "...and N more" fallback; renders nothing when no skills are installed (byte-identical to prior prompt). Session lifecycle - /new starts a fresh session in place; previous session stays on disk and is named in the confirmation for /resume. Guards against mid-run reset. - /clear now notes that context persists and points to /new. Performance & caching - partitionTools no longer alpha-sorts the whole tools array every turn, which inserted a loaded deferred tool mid-array and shifted the cached prefix on every load. It now builds append-only regions [eager sorted] + [tool_search] + [loaded deferred sorted] so a mid-session load grows the tail instead of busting the eager prefix. Regression test asserts the eager block precedes any loaded tool. - /context now shows cache hit rate (flags low rate over >5 turns as a possible unstable prefix) and rolling average turn latency with time-to-first-token. Both reset on /new. Measured baselines (auto-mode, no workspace): base prompt 3162 tokens, eager tool schemas 2433 across 12 tools. Budget ceilings set ~10% above. Deliberately deferred (need a decision): inverted safety-policy rewrite (sandbox doesn't catch ~half the B-tier rules), core-tool/web_fetch deferral (activation-semantics blocker), aggressive prompt slim. * fix(tui): clear staged attachments + queued message on /new CodeRabbit: startNewSession reset conversation/usage/latency state but not the staged input (pendingImages, pendingImageLabels, pendingDocuments, queuedMessage), which is only consumed at prompt-submit. Without clearing it, the fresh session's first prompt silently inherited the previous session's images/documents/queued text. Test now stages all four and asserts they clear. |
||
|
KRATOS
|
7fb0092de7 |
fix: agent runtime resilience — stream stalls, weak-model tool args, and sub-agent reliability (#349)
* fix(local): defeat the macOS stale-pooled-connection hang (transport + stall-retry)
Root cause (you confirmed it: both chatgpt AND ollama hung, only on macOS): providers
fell back to http.DefaultClient, whose pooled keep-alive connections get reused on a
later turn after the server/NAT silently dropped them. The model call is a POST
(non-idempotent), so Go won't auto-retry it on a fresh conn — it blocks forever.
macOS keeps dead pooled conns around longer than Linux, so it reproduced on macOS only.
1. providerio.HTTPClient now returns a shared, tuned transport (ResponseHeaderTimeout
60s bounds a hung POST that gets no response; IdleConnTimeout 30s stops dead conns
from lingering to be reused). Routed openai/codex through it too (anthropic/gemini
already used HTTPClient). Slow first tokens / long reasoning are unaffected — the
header timeout only bounds time-to-first-response-header, not the stream body.
2. Agent loop now retries a stream idle/stall timeout that produced NO output yet, on
a fresh connection (capped + backoff). A partial-output turn is NOT retried (would
duplicate). Recovers any stall that slips past the transport timeout.
Local only — not pushed.
* fix(local): tolerate concatenated tool-arg JSON + widen response-header timeout
1) Weak models (minimax-m3, glm) pack MULTIPLE concatenated top-level JSON objects
into one tool-call arguments slot ({A}{B}), which json.Unmarshal rejects with
'invalid character {' after top-level value' (~9 failed tool calls/session).
New recoverableToolArguments returns the first JSON value ONLY when the payload is
one value optionally followed by more WHOLE JSON values; it rejects a bad/truncated
first value OR trailing NON-JSON garbage ({"x":1}xyz) so genuine corruption still
errors (adversarial-review fix). decodeToolArguments uses it at the central dispatch
(executeToolCall) so {A}{B} runs the first object; historySafeToolCalls uses the
SAME helper so the replayed transcript records the recovered first object (not {}),
keeping dispatch and history consistent. Truly-malformed JSON still hits the
existing sanitize-to-{} + retry path. Trailing objects are dropped (the model
re-issues on a later turn).
2) Raise the shared transport ResponseHeaderTimeout 60s -> 120s so a slow cloud proxy
(ollama *:cloud) that withholds its 200 header until first token isn't wrongly
aborted; still bounds a truly dead reused connection.
Local only — not pushed.
* feat(local): add /turns command + raise default maxTurns 30->50
Agentic multi-step runs (build + delegate + swarm + MCP + self-verify) routinely hit
the 30-turn ceiling mid-task and never reached the later steps; removing /mode earlier
also left NO interactive way to raise the budget.
- /turns [n]: per-session tool-turn budget setter (commandGroupSession), mirroring
/selfcorrect. Empty/status shows the current budget; a positive int sets
agentOptions.MaxTurns (read by the run at model.go); clamped to maxTurnsCeiling=500
to guard against typos; invalid/non-positive shows usage without changing it.
- defaultMaxTurns 30 -> 50 (matches the old 'deep' preset). Raise per-run with /turns
for long tasks (e.g. /turns 150 for a full acceptance run).
Local only — not pushed.
* fix(local): swarm reports failed members as failed + sub-agents inherit /turns budget
Issue B — a swarm member whose child exited non-zero (e.g. exit 4 / max-turns) was
recorded [done] (swarm_status '0 failed'), so the orchestrator — which can't see the
AGENTS panel — trusted incomplete work. The specialist launcher ignored
res.Result.Status; it now surfaces a StatusError result as a member FAILURE
(errors.New(report)) so the swarm marks it [failed]. Added Coordinator.FailWithSession
and switched the member watch path to it so a failed member keeps its session id
(still drillable) and its report rides along as the failure message.
Issue A — a member hit its OWN turn limit because /turns raised only the parent's
budget; children re-resolved config.json's default. New config.SetMaxTurnsEnv exports
ZERO_MAX_TURNS, read in applyEnv -> cfg.MaxTurns; /turns now exports it so spawned
sub-agents / swarm members (which inherit the env) run with the same budget. A
non-numeric/zero env value never clobbers the configured budget.
Tests: launcher non-zero-exit -> failed (+ session kept); SetMaxTurnsEnv round-trip
+ garbage-env guard; /turns exports + clamps the env. Full gate green (73 pkg).
Local only — not pushed.
* fix(local): corrective error for unknown specialist name + clamp ZERO_MAX_TURNS
1) A model that invents a specialist name (validator-runner, file-creator,
file-writer — none registered) got an opaque 'specialist not found' and kept
spawning doomed sub-agents. loadManifest now lists the available specialists and
notes that 'worker' has shell/edit while explorer/code-review are read-only, so the
model self-corrects to a capable specialist. New availableSpecialistList helper.
2) Defense-in-depth (review nit): applyEnv now clamps an inherited ZERO_MAX_TURNS to
MaxTurnsCeiling, so a raw shell 'export ZERO_MAX_TURNS=999999' can't bypass the
/turns ceiling. Ceiling promoted to config.MaxTurnsCeiling (single source of truth,
referenced by the TUI /turns clamp).
Tests: unknown-name error lists worker/explorer/code-review; applyEnv clamps an
over-ceiling env value. Full gate green (73 pkg).
Local only — not pushed.
* fix(local): surface sub-agent kill signal instead of opaque exit -1
When a sub-agent child is terminated by a signal, command.Wait returns an ExitError
whose ExitCode() is -1, losing the cause. Capture ProcessState.String() (portable,
e.g. "signal: killed") into ChildRunResult.Signal and have BuildFinalResult render
it — "Subagent terminated by a signal (signal: killed) — killed before it finished.
Likely causes: the OS out-of-memory killer ... a timeout, or cancellation." — instead
of an opaque "Subagent failed (exit -1)". The message lists causes rather than
asserting OOM, since timeout/cancellation also SIGKILL the child. A captured signal
always surfaces even if a late run_end reported a clean exit. The background task
path (which can't carry the signal name) emits the same hint for a negative exit code.
* address CodeRabbit review on #349
- loop.go: gate stall-retry on a turn-level forwardedAnything flag (wrapped
callbacks), so a retry never re-streams on top of output already forwarded earlier
in the turn (e.g. across the reactive-compaction retry). New test covers a
reasoning-then-stall (collected.Text empty but output forwarded).
- loop.go: route a reissued stream's non-stall error through the SAME recovery as
the initial stream (image-rejection wrapping / context-limit compaction) via a
shared recoverStreamError closure, instead of returning it raw.
- exec.go: drop the hardcoded worker/explorer/code-review guidance from the
unknown-specialist error; rely on the dynamically-rendered available list (a
custom registry may not have those names).
- streamer.go: kill-signal message lists OOM/timeout/cancellation evenhandedly
instead of asserting OOM (the branch also covers timeouts/cancellations).
- launcher_specialist.go + exec.go: preserve the child SessionID on a post-start
failure too, so FailWithSession keeps those failed members drillable.
- model.go: block /turns while a run is active (mid-run change would make the
inherited ZERO_MAX_TURNS budget inconsistent for later-spawned sub-agents).
|
||
|
Vasanth T
|
b2bb6c4b41 |
feat: five capability features from the reference-agent comparison (#276)
* feat(commands): user-defined slash commands from markdown files
Drop a markdown file at `.zero/commands/<name>.md` (project, shared with the
team) or `<userConfigDir>/zero/commands/<name>.md` (personal); typing
`/<name> args` expands its body template and submits it as a normal prompt.
Optional YAML frontmatter (`description:`, `model:`, `agent:`) feeds help text
and routing. This turns Zero's model-pulled skills into user-invokable,
repo-checked-in team workflows (`/release v1.2`, `/fix-flaky <test>`).
New internal/usercommands package: Load() scans the project + user dirs
(project shadows user on name collision, mirroring specialist scope
precedence), validates names to a safe slash-command shape, and parses
frontmatter with a self-contained parser. Expand() substitutes $ARGUMENTS,
$1..$9, and $$ (literal $); a placeholder-free template gets the raw args
appended so a trivial command still receives input.
Wiring: model loads commands at construction; the commandUnknown dispatch
checks the user set before reporting "unknown"; autocomplete lists user
commands alongside builtins. Builtins win on collision (listed first, shadow at
dispatch). Tests cover load/precedence/name-validation, the placeholder
substitution table, and the TUI dispatch + autocomplete + still-unknown paths.
* feat(init): guided /init that generates AGENTS.md seeded with repo facts
`/init` (TUI) and `zero init` (headless) investigate the repo and write a
concise AGENTS.md so future runs start with project context. Zero already has
the AGENTS.md loader and the repoinfo scanner, but the file only existed if a
user hand-authored it — this activates the machinery in one command.
New internal/agentinit.BuildPrompt seeds the bootstrap prompt with
repoinfo.Collect() output (primary language, build/test tools, CI, workspace
layout, git) formatted as a fact block, then instructs the agent to investigate
the gaps and write a tight AGENTS.md (build/test commands, conventions, things
to avoid). Collection failure is non-fatal — the agent investigates from
scratch. This is one better than a pure-prompt /init: the agent starts from
known facts instead of re-deriving them.
TUI: /init command + handleInitCommand reuse launchPrompt (standard run, tools,
AGENTS.md write path). Headless: `zero init` builds the prompt and forwards it
through the existing exec machinery, so all exec flags (--model, -w, --add-dir)
work unchanged. Tests cover FormatFacts (rendered/omitted fields), BuildPrompt,
and the TUI dispatch.
* feat(tools): lsp_navigate — semantic code navigation via the language server
A read-only, model-callable tool exposing LSP navigation: jump-to-definition,
find-all-references, find-implementations, and workspace_symbol search. The
agent can answer "where is X defined / who calls X / what implements this
interface" precisely, where grep collides on common names or misses indirect
references — cutting wasted read/grep turns and wrong-edit risk on large repos.
Zero already runs language servers for diagnostics, so this adds the
request/response navigation methods on top: Manager.Navigate syncs the file
(didOpen) then issues textDocument/definition|references|implementation or
workspace/symbol via the existing client.Call, decoding all three result shapes
(Location, []Location, []LocationLink). 1-based file:line:col in and out
(converting LSP's 0-based positions); the tool has its own lazily-started
manager (servers start on first use, reused across calls).
Opportunistic: a file type with no installed server, or a server lacking the
capability, degrades to a clear "unavailable — fall back to grep" message, not
an error. Registered in CoreReadOnlyToolsScoped. Tests cover the response-shape
decoding (incl. LocationLink + null), symbol decoding/kinds, unknown-op and
unsupported-extension degradation, and the tool's arg validation + read-only
safety.
* feat(agent): prune stale tool output before paid summarization
Long, dump-heavy sessions accumulate large read_file/grep/glob/bash tool
results the model has already acted on; their bodies are dead weight the loop
otherwise keeps paying an LLM summarizer to compress. Add a zero-cost prune
stage inside maybeCompact, before the summarizer: when over threshold, replace
the bodies of OLD, large tool results with a compact placeholder, recompute
size, and only fall through to the paid Compact() if still over.
pruneStaleToolOutput walks newest-first, protects the last preserveLast
messages plus a trailing 40k-token window of tool output (most likely still in
use), and prunes only tool results older than that and larger than 200 tokens —
gated so the whole pass is a no-op unless reclaimable output exceeds 20k tokens.
It keeps the tool message and its ToolCallID (provider replay stays valid; the
model knows what was there and can re-run the tool), never touches non-tool
messages, never re-prunes a placeholder (idempotent), and copies-on-write so it
never mutates the caller's slice. Recent turns are preserved verbatim — no
paraphrase loss, no token or latency cost.
Tests cover the reclaim gate, recent-window protection, idempotence,
non-tool-message safety, ToolCallID/role preservation, and no-input-mutation.
* feat(agent): reconnect on a transient model disconnect instead of ending the run
A long autonomous task (a big refactor, a swarm member, a headless/cron run)
should survive a single transient upstream hiccup rather than dying and
re-burning every token on a restart. When the initial StreamCompletion connect
fails with a disconnect-shaped error, streamWithReconnect re-issues the SAME
request with exponential backoff up to 2 times.
It retries only the connect — before any content is forwarded — so no
already-streamed OnText is ever duplicated. A context-limit error (the
compactor's job), an image rejection, a non-disconnect error, or a cancelled
context is returned immediately without retry, so the reconnect never fights
the existing handlers. Disconnect classification matches EOF / connection
reset|refused|closed / broken pipe / timeout / 502 / 503 / server-closed.
Reconnect attempts surface through OnReasoning ("connection lost —
reconnecting N/max"), a non-content channel that never corrupts the answer
text; silent when there's no reasoning sink.
Tests: recover-from-transient, give-up-after-max, no-retry-on-context-limit,
stop-on-context-cancel, the disconnect-vs-not classifier, backoff growth, and
the notice routing.
* fix(specialist): add lsp_navigate to knownToolNames
The new lsp_navigate tool was added to the core read-only registry but not to
specialist.knownToolNames, which mirrors the core tool set — so
TestKnownToolNamesMatchCoreRegistry failed (the list drifted from the registry).
Add it to the read-only group to match.
* fix(tools): confine lsp_navigate to the workspace (close arbitrary-file-read)
Review caught that lsp_navigate dropped the PathScope every sibling read-only
tool (read_file/glob/grep) carries, and its readFile joined or used the
model-supplied path verbatim — so `path:"/etc/passwd"` (absolute) or
`path:"../../etc/..."` (escaping) read arbitrary files off disk AND opened them
into the language-server subprocess, reachable via indirect prompt injection.
Add NewScopedLSPNavigateTool(workspaceRoot, scope) (NewLSPNavigateTool now
delegates with nil scope), resolve the path through resolveScopedReadPath BEFORE
reading or building the LSP request — the same confinement the neighbouring
tools enforce — and hand the LSP manager the resolved absolute path so it never
opens an out-of-workspace file. Errors echo only the relative path, never an
absolute one. Registered via the scoped constructor in CoreReadOnlyToolsScoped.
Regression test asserts `/etc/passwd` and `../../` are rejected for both the
position and workspace_symbol ops; verified by mutation (bypassing the scope
resolution makes the test fail).
|
||
|
KRATOS
|
b903825139 |
perf(tools): lower default deferThreshold to stop eager MCP-schema token waste (#241)
* perf(tools): lower default deferThreshold 10 -> 3 to stop eager MCP-schema waste MCP tool schemas are only collapsed into compact tool_search reminder lines once the deferred-eligible tool count reaches tools.deferThreshold. The default of 10 meant a typical single MCP server (Exa-style, a handful of tools) stayed BELOW it, so every full MCP schema — 300-600 tokens each — was re-sent on every turn, even for a one-word message. A small MCP set could add several thousand input tokens per message. Lower the default to 3 so a normal MCP server's toolset defers and stays cheap. Capability is unchanged: every tool is still reachable; the model just loads a deferred tool's schema via tool_search before its first use, after which it is loaded for the rest of the run. Fully overridable via tools.deferThreshold (0 keeps the old always-eager behavior for models without tool_search). * test(agent): measure the deferral token saving (66% smaller tool payload) * fix(tokens): count non-whitespace bytes so estimates match the real tokenizer The token estimate was naive len/4, which overcounted by ~15-20%: `zero context` reported ~6k for a prompt the provider's tokenizer actually counted as 5,203. BPE tokenizers (the ones zero targets) fold a run's leading space into the following token rather than emitting whitespace separately, so counting NON-whitespace bytes / 4 tracks the real count closely (now ~5.2k, ~1% off the measured 5,203) while staying dependency-free. Add agent.ApproxTextTokens and route the context-budget preview (contextreport), the compaction threshold (agent), and the TUI transcript estimate through it so every surface uses one accurate scale. The provider's exact usage is still authoritative on each request; this only sharpens the pre-request preview and compaction timing. |
||
|
KRATOS
|
474f494b93 |
Deep-audit remediation + sandbox hardening (batches 1–35) (#189)
* audit batch 1: redact nested secrets fully; count image tokens in compaction estimate
- secrets.Redact: replace longest matches first so a containing PRIVATE KEY
block is redacted before a nested shorter match corrupts its exact string
(audit medium internal/secrets/scanner.go:81) + regression test.
- agent compaction estimateTokens: add a flat per-image token cost so an
image-heavy context trends toward compaction instead of reading as ~0
(audit medium internal/agent/compaction.go:58) + test.
- ledger: re-verify against current main; mark scanner.go:35 and
redaction.go:362 already-fixed-on-main, record this batch's fixes.
* audit batch 2: worktrees defaultRunGit separates stdout/stderr
Capture git stdout and stderr into separate buffers instead of CombinedOutput,
so parsed rev-parse output is not polluted by git warnings and CommandResult.Stderr
carries the real error text (audit medium internal/worktrees/worktrees.go:247) + test.
* audit: re-verify all 87 pending findings against current main
Independent agent per file re-checked each open/partial finding against the
actual current code (the 06-11 ledger was verified on a since-diverged branch),
with an adversarial second pass on every already_fixed verdict. Authoritative
result: 78 genuinely pending (66 still_open + 12 partial); 9 rows were already
fixed on main and are reclassified. Full table in 2026-06-13-reverification.md.
* audit batch 3 (agent): cancellation identity, always-allow w/o sandbox, compaction usage, dedup
- loop.go: check ctx.Err() before collected.Error so a cancelled run returns
context.Canceled (errors.Is holds) instead of a plain stringified error.
- loop.go: 'Always allow' with no sandbox engine now allows THIS call instead of
denying it (persistence is skipped when there's nowhere to persist).
- loop.go: remove a duplicate schema["items"] assignment in propertyToRuntimeMap.
- compaction.go: forward OnUsage (token accounting) through the summarizer while
still omitting OnText, so compaction's token cost is counted in usage/budgets.
- tests for all three behavioral fixes (each fails pre-fix).
* audit batch 4 (cli): list-tools -o json, interrupted -o json terminal, surface session-record failures
- exec --list-tools -o json now emits a JSON tool listing instead of falling
through to the human-readable text (audit medium exec.go:182) + test.
- interrupted run with -o json now emits a terminal error+done on stdout instead
of only printing to stderr, so a JSON consumer sees a clean end (medium exec.go:500) + test.
- execSessionRecorder: surface a latched session-append failure to stderr once at
run end (deferred) instead of silently dropping it (medium exec_sessions.go:114/115) + test.
- left app.go trailing-arg ignore as-is: documented-intentional, TUI has no argv prompt.
* audit batch 5 (zerogit): status -z porcelain (renames + non-ASCII paths); document env-drop on plain Runner
- parseStatus: switch `git status` to `--porcelain -z` and parse the
NUL-delimited form (audit medium zerogit.go:212). Previously the default
--short output was split on newlines, so a rename `R old -> new` was kept as
one bogus path and a non-ASCII/whitespace filename arrived C-quoted+escaped
(e.g. `"caf\303\251.txt"`). The -z form never quotes paths and emits a
rename/copy as `XY <dest>\0<src>`, so we now record the destination verbatim
and consume the source field. + TestParseStatusZHandlesRenamesAndSpecialPaths.
- resolveRunners: document that the plain-Runner EnvRunner adapter intentionally
drops env (audit low zerogit.go:417). The branch is reached only when a caller
supplies a custom Runner without a RunGitEnv — in practice fake-Runner tests.
Production leaves both nil and gets defaultRunGit/defaultRunGitEnv, which honor
GIT_INDEX_FILE, so snapshot-diff isolation holds. Routing env calls to real git
instead would break the fake-Runner test pattern and fix nothing in production.
- update Inspect/Commit mock tests to the -z status format and the new command.
* audit batch 6 (config): reject negative maxTurns; drop dead contract-gap helpers
- FileConfig.UnmarshalJSON now rejects a negative maxTurns instead of letting the
`MaxTurns > 0` merge gates silently drop it and fall back to the default,
hiding the misconfiguration (audit medium resolver.go:136). 0 is left as the
"unset" sentinel (omitempty) and still falls back to the default; the CLI
--max-turns flag continues to reject 0 too since there it is distinguishable
from absent. + TestResolveRejectsNegativeMaxTurns / ...ZeroFallsBackToDefault.
- remove dead config/contracts.go (ContractGap, DefaultContractGaps,
FindContractGapsByMilestone, ContractOwnerRuntime) and its test — referenced
only by that test, never by production (audit low contracts.go:12).
- NOT changed: ResolveMCP intentionally does not run the provider command
(TestResolveMCPDoesNotRunProviderCommand guards this), so the "ZERO_PROVIDER_COMMAND
MCP servers dropped" row is by-design, not a defect.
* audit batch 7 (mcp): bound captured MCP server stderr
connectStdio attached a plain bytes.Buffer as cmd.Stderr that is only ever read
on initialize failure, yet it kept growing for the entire process lifetime as a
long-lived server logged to stderr (audit medium client.go:98). Replace it with
a concurrency-safe boundedBuffer capped at 64 KiB that keeps the head (enough for
init-error diagnostics) and discards the rest, reporting full write lengths so
os/exec never sees a short write. + TestBoundedBufferCapsRetainedBytes.
Left as-is (not defects worth the risk in the MCP client):
- client.go writer.write under client.mu: already designed to release the lock
before the unbounded response wait; the residual concern is a server that stops
draining stdin, an involved async rewrite for a theoretical deadlock.
- readLoop not answering server-initiated requests (ping): responding safely
needs re-locking around the shared writer; no configured server relies on it.
* audit batch 8 (tools/providers): read_file truncation marker; drop dead provider branch
- read_file: when a max_lines cut occurs, append an explicit "[truncated: N more
line(s)... set start_line=X to continue]" marker to the output. Previously only
the Result.Truncated flag was set, which is invisible in the rendered text, so
the model could not distinguish a cut read from a complete one (audit low
read_file.go:89). + TestReadFileToolMarksTruncation.
- providers/factory: simplify `if !explicitProvider || isImplicitOpenAI(...)` to
`if !explicitProvider` and delete the now-unused isImplicitOpenAI. The clause
was dead: explicitProvider==true implies ProviderKind or Provider is set, but
isImplicitOpenAI required both empty, so it could never contribute a case
(audit low factory.go:118). Behavior-preserving; providers tests still green.
* audit batch 9 (sessions): Tree builds from a single List() snapshot (no N+1 disk scans)
Store.Tree recursed via ListChildren per node, and each ListChildren ran a full
store.List() that re-read every metadata.json from disk — O(nodes * total-sessions)
reads for one tree (audit medium lineage.go:140). Snapshot all sessions once, index
children by parent in memory, and recurse over that. Child ordering is factored into
sortChildSessions so ListChildren and Tree stay identical; path-scoped cycle
detection (delete(seen,...)) is preserved. Existing tree/lineage tests still pass.
Left as-is: store.go timestamp() second precision (RFC3339) — switching to
RFC3339Nano would break lexical UpdatedAt ordering against existing second-precision
records ("...00.5Z" sorts before "...00Z"), so it is not a safe one-line change.
* audit batch 10 (specialist): serialize accounting dedup-then-append
recordSpecialistStop and appendSpecialistUsageRollup checked for an existing event
(specialistEventExists → ReadEvents) and then appended under SEPARATE store locks,
so two concurrent finishers — a foreground onExit racing a TaskOutput poll, or a
background reaper — could both pass the dedup check and write duplicate stop/usage
events (audit medium accounting.go:72). Guard each check-then-append pair with a
package-level accountingMu (Executor is used by value, so a struct field would not
be shared). + TestRecordSpecialistStopDedupesUnderConcurrency (passes under -race).
* audit batch 11 (zeroruntime/gemini): O(1) text accumulation; gemini cached-token counts
- zeroruntime CollectStreamWithOptions accumulated assistant text with
`collected.Text += event.Content`, reallocating the whole string on every chunk
(O(n^2) over a long streamed response). Accumulate in a strings.Builder on the
collector and materialize it once in flush, the single finalization point hit
before every return (audit low helpers.go:104). Existing collect tests still pass.
- gemini decoded usageMetadata without cachedContentTokenCount, so a cached-prompt
response reported zero cached input tokens while the other providers report them
(audit medium gemini/types.go:88). Decode the field and thread it through
streamState into NormalizeUsage as CachedInputTokens. + extended
TestStreamCompletionEmitsTextUsageAndReasoningTokens.
* audit batch 12 (specialist): parse child stream without a line-length cap; drop dead helper
- ParseStream used a bufio.Scanner capped at 1 MiB, so one large stream-json line
from a child (a big tool result or final answer) hit bufio.ErrTooLong and aborted
the entire specialist run (audit medium streamer.go:43). Switch to a bufio.Reader
with ReadString, which has no per-line limit; the child is our own trusted
subprocess, so its output length is the legitimate bound. Line numbering and
blank-line/EOF handling are preserved. + TestParseStreamHandlesLineLargerThan1MiB.
- remove dead formatTaskOutput (output_tool.go:244): zero callers and zero tests;
it only forwarded to formatTaskOutputSummary, which is the one actually used.
* audit batch 13 (background/specialist): kill background tasks as a process group
Background children were launched in the parent's process group and terminated by
signalling only the single PID, so any process the task forked was orphaned and
outlived a cancel; the signal also risked hitting a recycled PID after the child
was reaped (audit medium process_posix.go:38/50).
- ConfigureChildProcessGroup (Setpgid on POSIX, no-op on Windows where taskkill /T
already tree-kills) is applied at launch so the child leads its own group.
- terminateProcess now SIGTERM/SIGKILLs the negative PID (the whole group), with a
pid>1 guard so a bogus 0/1 can never expand to "our own group" or "everything",
and treats ESRCH as already-gone. Liveness polls the group. Targeting a group
(rarely recycled vs. a bare PID) also shrinks the reaped-then-recycled window.
- specialist launchBackgroundProcess opts every child into its own group; on a
SetPID failure it now kills the child + group, marks the task errored, and records
the stop (deduped) instead of leaking an untracked orphan.
- tests launch as group leaders; new TestTerminateProcessKillsForkedChildren proves
a forked grandchild dies with the group.
* audit batch 14 (sandbox): quote-aware shell segment splitting
splitShellSegments used a quote-unaware strings.Replacer, so a shell operator
inside quotes was treated as a separator: `git commit -m "use top | less"` split
into a `less` segment and `echo "a; vim b"` into a `vim` segment, both falsely
flagged as interactive (audit medium safe_command.go:138).
Replace it with a quote-aware scanner that mirrors the shell: single quotes make
everything literal; double quotes keep $(...)/`...` substitution active but treat
|, ;, && as literal; unquoted text splits on the same operator set as before
(;, |, ||, &&) plus substitution boundaries. Real unquoted operators still split,
so an interactive program behind a separator — or inside a live `"$(...)"`
substitution — is still caught (no new false negatives, the security-relevant
direction). + TestDetectInteractiveQuoteAwareSeparators covering both directions.
* audit batch 15 (cron): collapse DST fall-back repeated hour to one fire
On a fall-back day the clock repeats an hour (e.g. 01:30 EDT then 01:30 EST), so
Next("30 1 * * *", <01:30 EDT>) returned the same-day 01:30 EST — firing a daily
job twice (audit medium schedule.go:171, which previously guarded only the
spring-forward gap).
Next now skips a match whose local wall-clock minute equals `after`'s but whose
absolute time is later (the repeated-hour duplicate) and keeps searching. Normal
forward search always advances the wall-clock, so this only triggers on the
fall-back repeat; when the scheduler passes the last fire time as `after`, the
repeated hour collapses to a single fire. + TestNextDSTFallBackDoesNotRepeatHour
(spring-forward termination test still passes).
Note: Next is stateless, so it dedupes relative to `after`; a caller that passes
an arbitrary mid-hour "now" instead of the last fire time is out of scope here.
* audit batch 16 (mcp): skip SSE notifications before the response; larger SSE buffer
- decodeSSERPCMessage returned the first non-empty `message` event unconditionally,
so a server-initiated notification/request arriving before the response made the
caller see an id mismatch and fail (audit medium network_client.go:621). It now
skips any message carrying a method (notifications/requests have one; a response
does not) and returns the first real response; the caller's rpcIDMatches still
validates the id. + TestDecodeSSERPCMessageSkipsNotifications.
- raise the per-event SSE scanner cap from 1 MiB to a bounded 8 MiB so a large but
legitimate MCP message no longer hits bufio.ErrTooLong and fails the request
(audit medium network_client.go:637); still bounded against an unbounded server.
Out of scope (kept): full reconnect of a persistent SSE stream after a fatal read
error — a larger change than this remediation.
* audit batch 17 (sessions): fsync session metadata before the atomic rename
writeMetadata wrote the temp file with os.WriteFile (no sync) and renamed it into
place, so a crash right after the rename could expose a metadata file whose bytes
were never flushed — a torn or empty file that corrupts the whole session (audit
medium store.go:632). Write+fsync the temp via a new writeFileSync helper before
renaming, closing that window. + TestWriteFileSyncRoundTrips.
Deliberately scoped:
- Event-log appends (appendEventLocked) are NOT fsynced per append: that would add
a disk flush to every event (every tool call), and a hard-crash loss of the log's
unsynced tail is acceptable, whereas torn METADATA is not. Metadata is the
integrity-critical part and is what this syncs.
- Directory fsync is omitted for cross-platform simplicity (it errors on Windows);
the file fsync already prevents the torn-contents failure, the serious case.
Left for follow-up (in this bucket but not safe/clear wins here):
- hooks append sequence still re-reads the file per append; the in-memory cache is
cross-process-unsafe and the safe last-line read is fiddly for a low-volume log.
- Fork event re-append/de-dup is nuanced usage-accounting and needs its own change.
* audit batch 18 (agent): count tool-defs in compaction estimate; don't deny an always-allow on persist failure
- maybeCompact's estimate omitted the tool definitions, which ride on every
request, so the real input could exceed the model limit while the message-only
estimate stayed under threshold and compaction never fired (audit medium
compaction.go:58). partitionTools is now computed before maybeCompact (it does
not depend on the messages) and the exposed tools' estimated tokens are added to
both the threshold check and the post-compaction shrink check. + estimateToolDefTokens
and TestEstimateToolDefTokensCountsDefinitions.
- PermissionDecisionAlwaysAllow denied the call when persistPermissionGrant failed,
overriding what the user explicitly allowed (audit medium loop.go:468). Persisting
is now best-effort: a failure (or no sandbox engine) just means the grant is not
remembered and the user is re-prompted next time, not denied now. This also drops
the dead requestEvent.GrantMatched/Grant writes (audit low loop.go:473) — the
emitted event derives them from the sandbox decision via buildPermissionEvent.
Left intentional (documented, not defects): the reactive-retry path omits OnText to
avoid duplicating mid-stream text (loop.go:183); OnContext/MeasureContext is a
complete, tested context-breakdown scaffold awaiting a consumer (types.go:183).
* audit batch 19 (usage): price escalated usage from the event's own model
exec persists payload["model"] on --allow-escalation runs (the model can change
mid-run), but BuildReport priced every event from the session's Metadata.ModelID
and never read the payload, so usage from an escalated model was mis-priced at the
base model (audit medium report.go:17). usageEventPayload now carries Model and
BuildReport prefers it, falling back to the session model when absent (older
events stay byte-identical and behave as before). + TestBuildReportPricesFromEventModelWhenPresent.
* audit batch 20 (sessions): don't copy usage events into a fork (no double-count)
Fork re-appended every parent event, including provider_usage, into the child, so
a usage report aggregating the parent and the fork double-counted the parent's
usage (audit medium store.go:387). Skip EventUsage when copying — it already
counted against the parent and is not part of the conversation history a fork
replays — and record the actual copied count in the fork marker. + TestStoreForkSkipsUsageEvents.
* audit batch 21 (cli/skills): warn on duplicate skill names instead of silently dropping
skills.Load deduplicates same-named skills (first directory wins) and records the
dropped collisions via skills.Duplicates, but nothing surfaced them, so a shadowed
skill just vanished with no signal — and Duplicates had no production caller (audit
low skills.go:91). `skills list` now reports each collision to stderr (keeping
stdout/--json clean), using Duplicates. + TestRunSkillsListWarnsOnDuplicateNames.
Note: skills.Get is kept — it is a by-name accessor used throughout the package's
tests, not dead code.
* audit batch 22 (cli/cron): capture the run error from stream-json stdout
Cron jobs run exec with --output-format stream-json, so a failure is reported as an
`error` event on STDOUT, but the run record's Error was set only from stderr — so a
failed job often recorded an empty/uninformative error while the real message sat
unread in outBuf (audit low cron_run.go:150). On a non-zero exit, scan the stdout
stream-json for the error event's message and prefer it, falling back to stderr.
+ extractStreamJSONError and TestExtractStreamJSONError.
* audit batch 23 (sandbox): richer macOS seatbelt profile (signal + mach-lookup)
The generated sandbox-exec profile was default-deny with only process/sysctl/file
allowances, so two real gaps remained on macOS (sandbox quality report H1):
- A sandboxed script could not signal the children it spawned — `kill` returned
"Operation not permitted" under seatbelt, breaking common `cmd & ... kill`,
test-runner, and timeout patterns. Add `(allow signal (target self) (target pgrp))`,
which stays scoped to the command's own process group.
- XPC to system daemons was denied, so keychain (securityd/trustd), user/group
lookup (opendirectoryd), preferences (cfprefsd), network config
(SystemConfiguration), launch services, and the pasteboard all failed. Add a
curated `(allow mach-lookup ...)` allowlist (`sandboxMachServices`).
Neither touches the file-write or network rules, so the workspace boundary is
unchanged — verified on a real macOS host: `security`/`scutil`/`pbpaste`/self-kill
now succeed while writes to $HOME and /etc are still denied. iokit-open was tried
and dropped as unnecessary. + unit assertions (TestSandboxExecProfileGrantsSignalAndMachLookup)
and extended the darwin integration test with a self-kill case and an
outside-workspace deny check.
* audit batch 24 (sandbox/tools): remove dead OnSandboxDecision hook
RunOptions.OnSandboxDecision had no setter anywhere, yet registry.RunWithOptions
spawned a goroutine (with panic recovery) for it on every sandboxed tool call —
pure per-call overhead for an unreachable callback (sandbox report M1). The sandbox
decision is already surfaced to observers via the permission event
(buildPermissionEvent reads the decision), so this hook is redundant; delete the
field and its call site.
* audit batch 25 (sandbox/cli): exact-path grants — create and revoke by path
The engine already matched ScopeFile grants and DeriveScope already produced them
from path/file tool args, but the CLI could neither create one explicitly nor
revoke a single one (sandbox report L3 / prompt 7). Add:
- GrantStore.RevokePath(tool, path): revokes only the grant whose scope matches the
given file/dir, leaving tool-wide and other-path grants intact (path canonicalized
to absolute like stored scopes). + TestGrantStoreRevokePathRemovesOnlyMatchingScope.
- `zero sandbox grants allow <tool> --path <p>`: persists an exact-file (ScopeFile)
grant, path resolved to absolute so it matches lookups and revokes.
- `zero sandbox grants revoke <tool> --path <p>`: revokes just that grant (default
still revokes all grants for the tool). + TestRunSandboxGrantsCreateAndRevokeByPath.
The TUI permission-card "Allow always (exact path)" option is intentionally left for
the TUI feature branches (it overlaps #187/#188); the engine + CLI now cover the
full create/match/revoke round-trip for file-scoped grants.
* audit batch 26 (sandbox): Engine.Precheck — report violations before execution
Add Engine.Precheck(ctx, request) []Violation, which reports the sandbox
violations that would block a tool request before it runs (sandbox report / prompt 6).
It reuses Evaluate so policy is not duplicated: an allowed or merely-prompted
request yields no violations; a denied request yields its violation (via the shared
violationsFromDecision helper, with a ViolationPolicyDenied fallback for a deny that
carries no structured violation). A nil engine yields nothing.
+ TestEnginePrecheckReportsViolationsBeforeExecution.
Wiring Precheck into the batch-confirmation prompt is intentionally left to the TUI
branch: it touches prompt rendering, and short-circuiting the agent loop's deny path
would skip the before/after-tool hooks that currently fire around a policy-denied
call. The engine API is in place for that consumer.
* audit batch 27 (sandbox): scoped-egress domain prompt with timeout + caching
The scoped-egress proxy failed closed on any host outside the allowlist. Add an
opt-in DomainPrompt callback (prompt 8): an UNKNOWN host (neither allowed nor
explicitly denied) is authorized at request time via the callback instead of being
refused outright. An explicitly denied host is still refused without prompting; the
callback is bounded by PromptTimeout (default 60s, timeout/error => deny); decisions
are cached for the proxy's lifetime so a host is prompted at most once. With no
callback set the strict fail-closed behavior is unchanged. Wired through both the
CONNECT and HTTP handlers via authorizeTarget.
+ TestEgressProxyAuthorizeDomainPrompt / ...TimeoutDenies / ...NilPromptFailsClosed.
The user-facing prompt that supplies the callback is left to the permission UI (TUI)
branch; the egress mechanism (callback + timeout + cache) is in place and tested.
* audit batch 28 (sandbox): AST-based shell analysis (AnalyzeCommand)
Add AnalyzeCommand(script) AnalysisResult, a static shell analyzer built on the
mvdan.cc/sh/v3 parser (prompt 12). Walking the parsed command tree makes it a more
precise second opinion than the regex detector: a program name only counts when it
is an actual command, so `echo "vim ..."`, `printf 'open with vim'`, and
`git commit -m "rm -rf /"` are clean, while real commands are flagged —
interactive (editors/pagers + REPLs, with the same -e/script suppression as
nonInteractiveREPLFlags), destructive (rm -rf/-fr, dd, mkfs, shred, ...), and
network (curl/wget/ssh/nc/...). An unparseable script yields TooComplex so callers
can treat it as higher-risk. + TestAnalyzeCommand (16 cases) / ...EmptyIsClean.
Dependency: mvdan.cc/sh/v3 pinned to v3.11.0 — compatible with the existing go 1.24
directive (the latest v3.13.1 would have bumped the module to go 1.25). It is kept
as a standalone analyzer for callers to consult; feeding it into Engine.Evaluate is
deliberately not done here, since that would change allow/deny classification.
* audit batch 29 (sandbox): unify proxy-env generation (ProxyEnv)
Export ProxyEnv as the single source of truth for proxy-env injection (prompt 9
req 1), replacing the unexported scopedProxyEnv used by the sandboxed-shell path.
+ TestProxyEnv.
Investigation that bounds the rest of prompt 9 (documented in the ProxyEnv doc):
- web_fetch already honors HTTP_PROXY/HTTPS_PROXY — its transport clones
http.DefaultTransport, whose Proxy is http.ProxyFromEnvironment.
- MCP children already inherit the parent env (mergeProcessEnv appends os.Environ),
so they pick up any proxy vars the agent has.
So the only missing link to scope FetchUrl/MCP is a SESSION-level egress proxy whose
address reaches the agent process. That is deliberately NOT wired here: doing it
safely requires allowlisting the active LLM provider's own domain first, or the
agent's provider calls would route through the scoped proxy and be denied — a
correctness pitfall that also can't be verified end-to-end on this host. Flagged
rather than shipped blind.
* audit batch 30 (sandbox): Linux seccomp Unix-socket block — UNVERIFIED, needs Linux CI
Adds a pure-Go (no cgo, no vendored binary) seccomp BPF filter that denies
socket(AF_UNIX) with EPERM on x86-64/arm64, closing the Unix-socket gap bubblewrap's
filesystem/network isolation leaves open (prompt 11):
- internal/sandbox/seccomp.go: the classic-BPF program as a platform-neutral
[]sockFilter, built and UNIT-TESTED (structure + every jump offset in range — the
classic way a hand-written BPF filter goes silently wrong).
- seccomp_linux.go: ApplyUnixSocketBlock installs it via prctl(NO_NEW_PRIVS) +
prctl(SET_SECCOMP, MODE_FILTER). seccomp_other.go is a no-op everywhere else.
- cmd/zero-seccomp: a tiny linux-only exec wrapper that applies the filter then
execs the real command (the standard way to install seccomp in a child with
os/exec); degrades gracefully (warns + runs) if the kernel lacks seccomp.
⚠️ HONEST CAVEAT: I developed this on macOS and CANNOT run the actual AF_UNIX block
here. darwin build, `GOOS=linux` cross-compile, vet, and the structure test all pass,
but the unit test verifies the program's SHAPE, not that the kernel actually blocks
AF_UNIX — that MUST be verified on Linux CI (create a Unix socket under the wrapper
and assert EPERM). Safe to ship as-is: it is linux-only, no-op on every other
platform, and not wired into the bwrap command yet (opt-in prefixing the sandboxed
command with zero-seccomp is the remaining integration), so it cannot regress any
current path.
* audit batch 31 (sandbox): macOS denial log monitor (opt-in, default off)
Re-adds the macOS seatbelt-denial monitor as an opt-in feature (prompt 10). When
Policy.MonitorDenials is set, the sandbox-exec profile tags its default-deny with a
marker and bash tails `log stream` for that tag during the command, appending any
captured denials as a <sandbox_violations> block to the command's stderr so the
model can see what was blocked.
- DenialMonitor (darwin) tails/parses tagged denials with a capped, deduped ring
buffer and noise filtering; no-op stub on other platforms. parseSandboxDenyLine is
unit-tested. + profile-tag test + appendSandboxViolations test.
- Default OFF: with MonitorTag empty, StartDenialMonitor is a no-op and the command
path is byte-for-byte unchanged, so there is no regression or per-command overhead
unless enabled.
CAVEAT: on at least one macOS host (verified live) seatbelt denials are NOT delivered
to the `log stream`-queryable unified log, so the monitor captures nothing there. It
is the standard mechanism and works on macOS versions that do log denials; shipping
it default-off means it never affects whether a command runs.
* audit batch 32 (sandbox): wire seccomp Unix-socket block into bubblewrap (opt-in)
Completes prompt 11: the zero-seccomp helper (which installs the AF_UNIX seccomp
filter) is now actually invoked by the sandbox instead of only existing as a
standalone binary.
- bubblewrapCommandPlan, when Policy.BlockUnixSockets is set, ro-binds the
discovered zero-seccomp helper into the sandbox at its host path and prefixes the
command with it so the filter is installed before the real argv runs.
- findSeccompHelper discovers the helper next to the running executable first, then
on PATH; it returns "" when absent, so the sandbox degrades gracefully (runs
without the extra filter — bubblewrap's fs/net isolation still applies) rather
than failing the command. Exposed as a package var so the wiring is unit-tested
on any OS.
- Reachable from config: SandboxConfig.BlockUnixSockets and .MonitorDenials now map
onto the policy in applyConfiguredSandboxPolicy (opt-in only — config can turn a
hardening flag on, never off). Both default OFF.
- Tests: helper-prefix present when enabled / absent by default; config flags apply
independently and stay off when omitted.
Default OFF and Linux-only effect; the runtime seccomp enforcement still needs a
Linux host/CI to exercise end-to-end (the wiring and BPF program are unit-tested).
* audit batch 33 (sandbox): enforce scoped/deny network policy in web_fetch & web_search
Completes prompt 9: the built-in network tools now honor the SAME allow/deny/scoped
policy the sandboxed shell egress proxy enforces, instead of bypassing it. Before
this, web_fetch/web_search would reach any public host even when the policy was
NetworkDeny or NetworkScoped — the egress proxy only constrained shell commands.
- New shared gate Engine.NetworkHostAllowed(host) reuses the egress domain matcher
(domainAllowed/normalizeDomains/effectiveNetwork) as the single source of truth:
allow → all hosts; deny → none; scoped → only AllowedDomains minus DeniedDomains
(empty allowlist collapses to deny). Disabled policy / nil engine impose nothing.
- web_fetch is now sandbox-aware: the host is checked before the request AND on every
redirect, so a denied/unlisted host fails closed before any dial. Plumbed via a
networkPolicyGate so the no-sandbox path is byte-for-byte unchanged.
- web_search is sandbox-aware: the configured backend's endpoint host is checked
before the query leaves the machine; blocked under deny, and under scoped unless the
backend host is allowlisted.
- The registry already routes sandbox-aware tools through RunWithSandbox when an
engine is present, so this activates automatically wherever the sandbox is wired.
- Tests: engine gate matrix (deny/allow/scoped/denylist/empty/disabled/nil + host:port);
web_fetch & web_search block denied/unlisted hosts without dialing and allow listed
ones; endpoint-host parsing.
* audit batch 34 (sandbox): prove network-policy wiring end-to-end + document opt-in flags
Verification follow-up to batches 31-33; no behaviour change.
- Adds a registry-level test that exercises the FULL chain for the scoped-network
enforcement: registry.RunWithOptions -> engine.Evaluate (passes a scoped,
egress-capable sandbox-exec backend) -> routes web_fetch through RunWithSandbox ->
per-host allowlist blocks an unlisted host before any dial. This proves the
Tool-interface value actually satisfies sandboxAwareTool at the registry boundary,
not just in a direct method call.
- Documents the two opt-in hardening flags (sandbox.blockUnixSockets,
sandbox.monitorDenials) and the scoped web_fetch/web_search behaviour in the README
so the config keys are discoverable. Both remain off by default.
Confirmed wiring: agent loop forwards options.Sandbox to the registry; the registry
routes sandbox-aware tools (web_fetch, web_search) through RunWithSandbox when an
engine is present; MonitorTag is set only on the macOS sandbox-exec plan; the
seccomp helper prefix is added only on the bubblewrap plan when BlockUnixSockets is
set; both config flags map onto the policy via applyConfiguredSandboxPolicy.
* audit batch 35 (sessions): fix Windows CI failure in TestWriteFileSyncRoundTrips
The exact-permission assertion (perm == 0o600) added in batch 17 fails on
windows-latest because Windows does not honor Unix permission bits: Go reports
0o666 for a writable file regardless of the mode passed to OpenFile. This was the
sole cause of the failing Smoke (windows-latest) CI job; macOS and Ubuntu passed.
Guard the perm check with runtime.GOOS != "windows" so it still verifies the mode
on the platforms that enforce it, while the round-trip (write + read-back) is still
asserted on every platform.
* audit batch 36: address CodeRabbit review (5 findings)
- background/process_posix: after SIGKILL, poll until the process group is
actually gone (SIGKILL is async) and return an error if it survives past a
second deadline, instead of returning nil while descendants still race.
- sandbox/safe_command: only treat ')' as a substitution boundary when it closes
an active $(...). A literal ')' (e.g. echo "a) less") no longer splits into a
fake `less` segment and is no longer misflagged as interactive. Depth-tracked so
nested $(...) still isolate their inner command. Direct splitShellSegments test
(fails pre-fix: yields ["echo \"a", "less\""]).
- cron/schedule: the DST fall-back collapse now only applies when `after` sits
exactly on the minute boundary (its "last fire time" form). With sub-minute
precision (e.g. 01:30:30 EDT) the first 01:30 fire already passed, so the repeated
01:30 EST is the legitimate next fire and is returned — restoring strictly-after
semantics. New test fails pre-fix (skipped to next day).
- config/types: maxTurns error message now reads "must be >= 0" to match the check
(0 is accepted; only < 0 is rejected). Negative-maxTurns test asserts the message.
- docs/audit: header date/branch metadata now matches the body (2026-06-13,
fix/audit-batch) instead of the stale 2026-06-11/fix/audit-critical-high snapshot.
Verified: gofmt clean; go vet clean; host + linux + windows builds; affected suites
green. Each behavioural fix ships with a test confirmed to fail without the change.
* audit batch 37: address CodeRabbit re-review (12 findings)
Findings on the audit code from batches 21-36; each behavioural fix ships with a
regression test (several confirmed to fail without the change).
sandbox/analyzer (AnalyzeCommand):
- Unwrap launcher prefixes (sudo/env/nice/bash -c/...) and recurse into `sh -c`
payloads so `sudo rm -rf`, `env curl …`, `bash -c 'vim x'` classify on the real
command, not the launcher. Mirrors DetectInteractiveCommand.
- Honor rm's `--` end-of-options marker so `rm -- -rf` (delete a file named -rf) is
not flagged destructive.
sandbox/engine: NetworkHostAllowed now shares one backend-aware effective-mode
helper with Evaluate, so scoped collapses to deny when the backend can't enforce
scoped egress — the per-tool gate can no longer permit traffic the engine-level
decision would fail closed.
sandbox/egress: cache domain-prompt decisions by host:port, not host, so one
"allow" can't authorize other ports on the same host.
sandbox/safe_command: the segment splitter is now escape-aware (\| , \" ) so an
escaped operator/quote can't manufacture a fake interactive segment.
sandbox/log_monitor (darwin): Wait() the `log stream` child after draining stdout
so the cancelled process isn't leaked as a zombie.
tools/web_search: confine the search backend to same-host redirects (with a limit)
so a redirect chain can't egress to a host the scoped/deny policy never authorized.
cli/sandbox: an explicit empty --path (`--path=` / `--path ""`) now fails closed
instead of silently widening an allow to tool-wide or a revoke to all-grants.
cli/skills: warn on stderr when duplicate-skill detection fails instead of
swallowing the error.
specialist/exec: don't mask a failed orphan-kill — join the TerminateProcess error
with the SetPID error so a surviving process is visible.
background/process_posix: resolve the PGID and only group-signal when pid leads its
own group; otherwise signal the individual PID. Avoids both a false-success on a
non-leader ESRCH and the hazard of signalling an unrelated (possibly our own) group.
Verified: gofmt + go vet clean; host + linux + windows builds; full go test ./...
green.
* audit batch 38: address CodeRabbit re-review round 2 (3 findings)
Follow-ups to batch 37, each with a regression test:
- safe_command (splitter): a ')' is now only a substitution boundary when it
closes an active, UNQUOTED $(...). A quote-state stack saved at each '$(' lets a
substitution run in a fresh quoting context and restores the outer quoting on
its matching ')'. Fixes `echo $(printf "a) less")` splitting into a fake `less`
segment, while still isolating the inner command of `echo "$(vim x)"`.
- egress: deduplicate in-flight domain prompts per host:port. Concurrent requests
for the same unknown target now wait on the first prompt and reuse its cached
decision instead of each firing a prompt (a prompt storm) and racing to write
the final decision. Race-tested with 25 concurrent authorizers → one prompt.
- analyzer + safe_command: wrapper option-value consumption is now wrapper-specific
(wrapperValueOptionsByProg) instead of a global option set. A valueless flag like
`sudo -n` no longer swallows the real payload, so `sudo -n rm -rf` /
`sudo -n curl …` are correctly classified destructive/network; `sudo -u root vim`
still consumes its value. Shared by the AST analyzer and the regex detector so
they can't diverge.
Verified: gofmt + go vet clean; host + linux + windows builds; go test ./... and
-race on sandbox/background all green.
* audit batch 39: address CodeRabbit re-review round 3 (12 findings)
All verified against current code; each behavioural fix has a regression test.
Security / correctness:
- tools/web_search: redirect policy now also refuses a scheme change, blocking a
same-host https→http downgrade that would send the query + bearer token over
plaintext.
- sandbox/safe_command: a ')' is only a substitution boundary when it closes an
active UNQUOTED $(...), and backticks now get the same quote-state save/restore
as $( — so `echo "`a | less`"` and `echo "$(true | less)"` no longer hide the
inner pager from interactive detection.
- sandbox/analyzer + safe_command: wrapper option-value consumption stays
wrapper-specific, and unwrapping no longer aborts on a dynamic ($x) wrapper arg
once a wrapper is active, so `env "$opts" curl` / `sudo "$x" rm -rf` are still
classified.
- sandbox/egress: in-flight prompts deduped per host:port (carried from batch 38);
the prompt callback is now cancellable (context cancelled on timeout) so a stuck
prompt can abort instead of leaking; signature threaded through.
- sandbox/runner: per-plan unique denial tag (pid+counter) so concurrent monitored
sandbox-exec runs can't ingest each other's denials; helper selection now
requires an executable regular file so a non-exec sibling degrades gracefully.
- mcp/network_client: aggregate SSE event-size cap (not just per-line) so a server
can't grow memory with many unterminated data: lines.
- specialist/accounting: a previously-recorded stop/usage with an empty runId is a
catch-all and now suppresses a later runId-bearing duplicate.
- sessions/lineage: Tree fetches the root via Get() first so a corrupt root
surfaces its real error instead of degrading to "not found".
Docs/tests:
- background/terminate: docstring now states the leader-only group-kill (PID-only
fallback for non-leaders); test child-liveness probe uses ESRCH specifically.
- cli/sandbox: grant usage strings include --path.
Verified: gofmt + go vet clean; host + linux + windows builds; go test ./... and
-race on sandbox/background all green.
* audit batch 40: address CodeRabbit re-review round 4 (8 fixed, 1 skipped)
Verified each against current code.
Fixed:
- sessions/lineage: PANIC FIX — store.Get returns (nil, nil) for a missing
session, so the batch-39 root prefetch could nil-deref. Guard root == nil and
return a not-found error. Regression test added (was a panic).
- agent/compaction: the reactive recover() now records lowWaterMark in the SAME
combined (messages + tool-defs) token domain maybeCompact uses, so the proactive
shrink-guard compares like with like. recover() takes tools; callers pass
request.Tools.
- mcp/network_client: SSE aggregate counter no longer counts a joining newline for
the first data line (only when prior content exists), avoiding premature rejects.
- sessions/store: fsync the parent directory after the metadata rename so the
rename is durable across a crash (no-op on Windows, which can't fsync a dir).
- sandbox/egress: parse the target port with strconv.Atoi instead of
net.LookupPort (no context-unaware resolver use).
- sandbox/runner: escape the per-plan denial tag through sandboxProfileString when
embedding it in the seatbelt profile.
- sandbox/seccomp_linux: guard against an empty filter program before taking
&kernelFilters[0] (defensive; the program is fixed non-empty today).
- cli/cron_run_test: assert last-error-wins with two error events.
- background/terminate, cli/sandbox usage, process_posix_test ESRCH (carried).
Skipped (with reason):
- zerogit worktree-rename (Y='R') test: git porcelain v1 -z only emits the
rename-source field when the INDEX column is R/C; the worktree column never
carries 'R' with a source field in v1, so the fixture would test output git
doesn't produce, and making parseStatus consume on code[1]=='R' would risk
mis-consuming the following entry for a plain ` R` status. Left as-is.
Verified: gofmt + go vet clean; host + linux + windows builds; go test ./... and
-race on sandbox/background all green.
* audit batch 41: address CodeRabbit re-review round 5 (4 findings)
All verified against current code; code fixes have regression tests.
- agent/compaction: summarizeWithFallback's reduce pass no longer swallows ALL
errors. A context-limit error still falls back to the joined partial summaries
(documented "extreme" case), but a non-context failure (auth/network/provider)
now surfaces unchanged, per the function's contract. Two tests (propagate vs.
joined-fallback); the propagate test fails without the fix.
- sandbox/safe_command + analyzer: wrapper value-consuming options now include the
long spellings (sudo --user root, env --unset NAME, timeout --signal …), so a
long flag's value is no longer mis-read as the program (interactive/destructive/
network detection bypass). A `--flag=value` token carries its own value and never
consumes the next token. Tests for space and = forms in both the AST analyzer and
the regex detector.
- docs/audit: the status-ledger summary table's open/partial cells (64/14) were the
lone outlier — the ledger's own narrative and 2026-06-13-reverification.md both
state 66/12. Corrected to 66/12 (66+12+97 = 175); derivation: body 69 open + 18
partial minus the 9 already-fixed-on-main (3 open + 6 partial) = 66/12 = 78
pending, fixed 88+9 = 97. Added a comment documenting the summary-vs-body
relationship. The reverification doc already states 66/12, so both now agree.
Verified: gofmt + go vet clean; host + linux + windows builds; full go test ./...
green.
* audit batch 42: address CodeRabbit re-review round 6 (1 fixed, 1 kept w/ proof)
CodeRabbit re-surfaced two "duplicate" refinements; CI was already green (its 6th
pass was a COMMENT with no new findings).
- cli/sandbox_test (TestRunSandboxGrantsRejectsEmptyPath): now seeds a tool-wide
grant and asserts grant-store immutability after EACH rejected empty-`--path`
call (reflect.DeepEqual before/after), so a mutation in one call can't be masked
by a later one — and a buggy "revoke all" is actually observable (it's a no-op on
an empty store). Stronger than the suggested patch.
- zerogit.parseStatus: KEPT consuming the rename source on the index column only
(code[0]), with an explanatory comment. Verified empirically that git porcelain
v1 -z reports a rename/copy (and emits the extra source field) only in the index
column: `git mv` → "R new\0old\0"; a worktree-only rename → " D old\0?? new\0"
(delete + untracked, never "R" in code[1]). Consuming on code[1]=='R'/'C' would
match no real git output and would only risk mis-consuming the next entry on
malformed input, so the suggestion is declined as incorrect for v1.
Verified: gofmt + go vet clean; host + linux + windows builds; affected suites green.
|
||
|
Vasanth T
|
0f103174d0 |
Add session context compaction flow (#166)
* Add session context compaction flow * Animate manual compaction flow * Simplify compact progress copy * Render compact progress as compression card * Render compact completion as success card * Fix compaction review findings * Fix compaction resume fallback |
||
|
KRATOSandKRATOS
|
98c27f81da |
Model harness: per-model system prompts + summarizer fallback (#135)
* [LOCAL PR-X] Agent: per-model system-prompt specialization Appends a family-specific guidance block (OpenAI: markdown+native-tool preference+persistence; Gemini: tool preference+conciseness; Anthropic: comment discipline) chosen by model id. Builds on the #129 system-prompt builder. Tests for classification + addendum selection. * [LOCAL PR-X] Agent: chunked summarization fallback on context-limit If the elided middle is itself too large to summarize in one call, summarizeWithFallback splits it in half recursively and joins the partial summaries, so compaction keeps working past the summarizer's own input window. Non-context errors and unsplittable single messages surface unchanged. Builds on #131. Tests cover split-and-join, non-context propagation, and the single-message edge. * Summarizer: re-summarize partial summaries instead of persisting a blob CodeRabbit major: summarizeWithFallback joined the two partial summaries and Compact stored that as one message; if that single message later exceeded the summarizer window, the guard couldn't split it (a single message) and compaction failed again for long sessions. After summarizing the halves, the partials are now re-summarized into ONE compressible unit. If even the combined partials don't fit (extreme), it falls back to the joined text (still the already-compacted halves, better than failing). Adds TestSummarizeWithFallbackReSummarizesPartialsIntoOne. build/-race/full-suite + GOOS=windows green. --------- Co-authored-by: KRATOS <[email protected]> |
||
|
KRATOSandKRATOS
|
413c323bf7 |
Agent: preserve active plan + loaded skills across compaction (#131)
* Agent: preserve active plan + loaded skills across compaction Compaction summarized the elided middle into prose, which silently dropped two pieces of structured state the model needs to keep working: the active plan (update_plan) and any loaded skill instructions (skill tool). Now Compact extracts them from the elided middle and appends them VERBATIM to the injected summary, so they survive a compaction exactly instead of being paraphrased away. - internal/agent/compaction_preserve.go (new): extractLatestPlan (most recent update_plan, rendered as status-tagged bullets), extractLoadedSkills (skill tool calls matched to their result bodies by id, latest-per-name, 2KiB-capped per skill so a large skill can't defeat the compaction), and appendPreservedState which folds both into the summary under clear labels. - internal/agent/compaction.go: Compact appends the preserved state; summaryInstructions now tell the summarizer to integrate any earlier summary block (delta-aware summarization) so re-compaction never drops older facts. Tests: plan + skill preservation through Compact, no sections when absent, latest-plan-wins, and skill calls without a result body are skipped. build/vet/-race/full-suite + GOOS=windows build green; no new deps. * Fix capBody to honor the byte budget (UTF-8 safe) capBody sliced by RUNE count against maxPreservedSkillBytes (a BYTE budget), so multibyte UTF-8 could preserve far more than 2 KiB, and it could append the truncation note without actually truncating. Now it cuts to the byte budget on a UTF-8 rune boundary (never splitting a rune), reserves room for the note within the budget, and only adds the note on real truncation. Adds tests: short-body unchanged, multibyte over-budget stays <= cap and valid UTF-8, and the rune-count<cap-but-bytes>cap case. * Compaction: carry preserved plan/skills across repeated compactions Vasanthdev P1: after the first compaction the preserved plan/skill sections live only as text in the injected summary message; on a SECOND compaction that summary is in the elided middle with no real update_plan/skill tool calls left, so the structured state was lost unless the summarizer happened to copy it (non-deterministic). appendPreservedState now carries forward existing '## Active plan' / '## Loaded skills' sections parsed from the most recent prior summary in middle, with fresh tool calls overriding per name (skills merged by name; plan: fresh update_plan wins, else carried forward). Adds TestCompactCarriesPreservedStateAcrossRepeatedCompaction — a second compaction whose summarizer drops the sections still keeps the plan + skill body. build/vet/-race/full-suite + GOOS=windows build green. * Compaction: serialize preserved state as JSON (lossless for arbitrary bodies) CodeRabbit major: the carry-forward format parsed markdown delimiters (section ends at the next '## ', skills split on '### '), so a verbatim skill body containing headings (## Usage, ### Examples) or code fences was truncated or split into bogus extra skills on the next compaction — breaking the repeated-compaction guarantee. The preserved plan + skills are now written as a single labelled line of JSON (formatPreservedState) and recovered with json.Unmarshal (parsePreservedState). JSON escapes everything, so any body — markdown headings, fences, quotes, multibyte — round-trips losslessly. Removed the markdown section/skill parsers; fresh tool calls still override the carried-forward copy by name. Adds TestCompactPreservesSkillBodyWithMarkdownHeadings: a body with ## / ### / a code fence stays byte-identical across two compactions. build/vet/-race/full-suite + GOOS=windows green. * Compaction test: guard compacted shape before indexing out[1] The round-trip helper indexed out[1] without checking len(out)>=2 / role first; add the guard so a regression in Compact fails cleanly instead of panicking. --------- Co-authored-by: KRATOS <[email protected]> |
||
|
|
73eddf758b |
Agent runtime: ask_user interception, context compaction, guardrails (#125)
* Agent runtime: ask_user interception, context compaction, guardrails Ports the agent-runtime slices from the reskin branch onto main, MINUS the blueprint's own sub-agent/task interception (main already has the merged specialist Task implementation, which this preserves). - ask_user (S6): AskUserRequest/AskUserResponse/AskUserQuestion types + Options.OnAskUser; the loop intercepts ask_user and routes to the interactive front-end, degrading gracefully (non-interactive fallback) when no handler is set. Answers are scrubbed through the redaction boundary. - compaction (S8): Options.ContextWindow + CompactionPreserveLast; proactive compaction when the estimated history crosses a fraction of the window, and reactive recovery on context-limit errors (summarizes the old middle, keeps system + last N). ContextWindow=0 disables it, so every existing caller/test behaves identically. - guardrails: empty-turn stop, repeated-tool-failure hint/stop, and a stale plan reminder; an aborted-placeholder is appended for unexecuted tool calls on guard stop. confirmation_policy.md is embedded into the system prompt. Preserved from main's specialist work: the RunWithOptions tool-context fields (ToolCallID/SessionID/Model/ReasoningEffort/Depth/Cwd), propertyToRuntimeMap nested Items, and ToolAdvertised AdvertiseInAuto. Dropped Options.Provider (it existed only for the blueprint task; compaction uses Run's provider arg). build/vet/-race/full-suite + GOOS=windows build green; no new deps. Unblocks the TUI shell + cli wiring PRs. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]> * Address #125 review (CodeRabbit + @Vasanthdev2004): scope failure hint, fix ask_user fallback/cancel - Repeated-failure guard now fires only on RETRIABLE tool errors (bad args / execution failures), not policy refusals (disabled tool, permission denial, sandbox violation) — a 'match this schema' hint there misdirects the model or nudges it to retry blocked behavior. New isRetriableToolError classifier (keys off permission_action meta + the known policy-error messages). Test: TestIsRetriableToolError. - Headless ask_user fallback now goes through registry.RunWithOptions with the full run context (ToolCallID/SessionID/Model/ReasoningEffort/Depth/Cwd, sandbox/permission) and copies the full result fields (Meta/ChangedFiles/Display), so it behaves like every other tool path instead of the bare registry.Run that dropped them. - A canceled / timed-out ask_user prompt (OnAskUser returns context.Canceled/DeadlineExceeded) now ABORTS the run and returns that error instead of fabricating a synthetic 'no interactive user' answer and continuing to mutate the transcript. executeToolCall/executeAskUser return an abort error that the loop honors (closing out remaining advertised calls for replay validity). Test: TestRunAskUserCancellationAbortsRun. - compaction_test: corrected the comment to match the assertion (retried text must NOT be re-streamed to OnText). build/vet/-race/full-suite + GOOS=windows build green. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]> * gofmt internal/agent/guardrails_test.go The file carried over from the reskin branch was not gofmt-clean; format it so the gofmt/lint check is green. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]> --------- Co-authored-by: KRATOS <[email protected]> Co-authored-by: Claude Opus 4.8 (1M context) <[email protected]> |