cli: --reasoning-effort notice evaluates the EFFECTIVE model (works without --model) (MED); unsafe warning fires for --auto high too (MED); --auto value validated even with --skip-permissions-unsafe (LOW); DefaultRegistry built once per exec, not 3x (LOW); --profile kept as intentional legacy-compat accept-and-ignore (contested finding resolved, doc-commented). modelregistry: removed dead FallbackFor. tools: removed dead TaskNonInteractiveMessage + dead OnSandboxDecision hook (no setter anywhere); fixed the now-stale loop.go comment. TDD where behavioral; build/vet/full-suite green.
Refs #102
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Cancelled runs now flush in-flight session events incl. checkpoints so /rewind works + no orphaned blobs (H5); shift+tab cycles permission mode Auto/Ask/Unsafe (H7); interactive TUI defaults to Ask so write/edit/bash/apply_patch are advertised + gate through the permission prompt (H8 — VERIFY LIVE); zenline skin renders the ask_user questionnaire instead of a spinner when a prompt is pending (M11); applyExecMode runs before filter validation + --list-tools (low); mode-model deprecation notice surfaced via the shared --model path (low); zenline listed in --help (low). TDD incl. -race; build/vet/full-suite green.
Refs #102
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Proactive compaction at 80% of the model context window + reactive recovery on context-limit errors: summarizes the oldest middle messages into one user-role summary, preserves system + recent suffix. Pure Compact() (injectable Summarizer), low-water-mark guard against churn, suffix walked back to an assistant boundary so no consecutive-user/dangling-tool-result sequences, recover() gated on enabled. Disabled (no-op) when ContextWindow==0 — wired from the resolved model's context window in CLI+TUI. Build/vet/-race green.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Data-driven presets (smart/deep/fast/large/precise) bundling model + reasoning effort + max turns (+ optional tool filter), resolved through the registry. --mode seeds only unset exec options so explicit --model/-r/--max-turns still win; /mode lists/switches in the TUI. Build/vet/test green.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
New read-only ask_user tool (questions[] with options/multiSelect); the loop intercepts it like permissions and routes to Options.OnAskUser. TUI shows an async questionnaire (buffered channel + ctx.Done select, mirroring the permission flow — no deadlock); headless leaves OnAskUser nil so it degrades gracefully to 'proceed with assumptions'. Build/vet/-race green; 13 new tests.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Resolve()/ResolveWithFallback()/FallbackFor() + DefaultReasoningEffort/EffectiveReasoningEffort wired into CLI --model and TUI /model (deprecated models auto-redirect with a notice; unknown ids like minimax-m3 pass through unchanged). zero models renders an aligned table; catalog entries decorated with match patterns/default efforts/deprecation rules. Provider reasoning-effort propagation deferred (needs request-schema change). Build/vet/test green.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Local development snapshot of one session's intertwined workstreams. Committed together because changes span shared files (provider.go, model.go, loop.go, types.go, streamjson) and can't be cleanly split without interactive hunk staging. Build + vet + full -race suite green.
Zenline TUI skin: Zen home / Statusline chat / permission-modal surfaces, 5 themes (Phosphor/Cyan/Sage/Violet/Mono), boot splash, mouse support, and a headless --snapshot mode (internal/zenline, cli/zenline.go, tui/zenline_view.go).
OpenAI provider: streaming tool-call accumulation hardening plus an idle-timeout watchdog (reader goroutine + context cancel) so a stalled upstream surfaces an error instead of hanging the agent forever.
Agent reliability: update_plan schema makes id optional/auto-numbered with an advertised item schema; malformed (nameless) tool calls feed a retry to the model instead of silently ending the turn; secret scrubbing at the registry boundary (covers agent loop AND MCP); structured ToolResult (changedFiles/display/redacted) surfaced via stream-json.
Session file checkpoints + safe rewind: before-mutation content snapshots (content-addressed, deduped, size-capped, 0600) captured on write/edit/apply_patch in TUI and headless; RestoreToSequence + ApplyRewind (path-traversal-guarded, session-locked); /rewind in TUI and 'zero sessions rewind' headless; stream-json checkpoint/restore events.
Design specs under docs/superpowers/specs/.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
* feat: add runtime permission event contract
* feat: stream permission events from exec
* feat: add runtime command snapshots
* feat: add session rewind compaction plans
* feat: add verification event contracts
* fix: harden runtime contract review gaps
* test: assert approved exec stderr contract
* fix: address CodeRabbit CHANGES_REQUESTED items for #82
- Make OnSandboxDecision callback async + panic-safe (non-blocking, recovered) in tools/registry.go per CR.
- Add explicit regression test for ConfigSnapshot when ResolveRuntimeMetadata errors (no secret leak in Message/BaseURL).
- (Many prior items like shapedPayloadPreview for permission/tool events, dual-target conflict test, stderr empty on approved path, Kind/Framework preservation, redactCommand empty slice, etc. were already landed on this branch.)
All relevant tests now pass.
* fix(agent): make sandbox decision available synchronously on Result for permission events
The recent change to invoke OnSandboxDecision asynchronously (to satisfy non-blocking requirement) broke the agent's immediate capture of the decision for building PermissionEvent after registry.RunWithOptions.
- Add SandboxDecision *sandbox.Decision to tools.Result (not serialized).
- Populate it on all sandbox-involved return paths in registry.RunWithOptions (sync, before/while firing async callback for observers).
- In agent executeToolCall, use result.SandboxDecision for buildPermissionEvent (reliable ordering) instead of side-effect capture via the callback.
- The async callback is still supported for any external OnSandboxDecision consumers.
This fixes the two failing agent tests that were causing the cross-platform 'Smoke' CI jobs to fail on this PR:
- TestRunEmitsPermissionEventForPersistentSandboxGrant
- TestRunAppliesSandboxEvenInUnsafeMode
All go tests now pass.
---------
Co-authored-by: KRATOS <[email protected]>
* feat: advance Go-first TUI and npm wrapper
* fix: address wrapper and footer review feedback
* feat: deepen Go TUI model and exec controls
* fix: address coderabbit exec and model list feedback
