Native ESM import() ignores NODE_PATH and resolves packages by walking up
the directory tree. Extension files synced to ~/.gsd/agent/extensions/ have
no ancestor node_modules, so imports of @gsd/* packages fail with "Cannot
find package" errors during report generation and other dynamic-import paths.
Create a symlink ~/.gsd/agent/node_modules -> GSD's node_modules after
resource sync so Node's standard resolution finds @gsd/* packages. Also
migrate the most critical dynamic imports in auto-loop, exit-command, and
commands to use importExtensionModule (jiti-based) as a belt-and-suspenders
fix.
Closes#1594
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Two bugs fixed:
1. recordHealthSnapshot counted ALL doctor issues including cross-milestone
stale errors, inflating consecutiveErrorUnits past the escalation threshold
from unfixable errors in other milestones. Now filters report.issues to
only the current milestone before summarizing for health tracking.
2. matchesScope used unitId.startsWith(scope) without a delimiter, so scope
"M004/S01" would false-match "M004/S010". Removed the redundant
delimiter-less startsWith branch — exact match and slash-delimited
startsWith are sufficient.
Closes#1579
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* refactor(auto-loop): hoist MAX_RECOVERY_CHARS to module level
Constant was defined inside the while loop body on every iteration.
Moved to module level next to MAX_LOOP_ITERATIONS.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* refactor(auto-loop): cache loadEffectiveGSDPreferences() once per iteration
Was called 9 times per loop iteration. Now called once at the top of the
try block and stored in `prefs`, used throughout the iteration.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
New UAT types skip human pause, enabling automated browser and script
verification by the engine.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
capPreamble() enforces MAX_PREAMBLE_CHARS via truncateAtSectionBoundary,
applied to all inlinedContext assembly points. Replaces deleted compression
subsystem with a simple deterministic cap.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace the crude sameUnitCount counter with a sliding window (size 6)
that detects three stuck patterns:
1. Same error repeated twice in a row → stuck immediately
2. Same unit derived 3 consecutive times → stuck (was 5, now faster)
3. Oscillation pattern A→B→A→B → stuck (previously undetected)
Graduated recovery preserved: first detection triggers cache invalidation
+ retry, second detection triggers hard stop.
Exported detectStuck() function with 8 unit tests covering all rules
plus edge cases (truncation, priority, non-triggers).
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
dispatchWorkflow now resolves per-phase model preferences (e.g.,
models.planning, models.execution) via resolveModelWithFallbacksForUnit
and applies them with pi.setModel before dispatching the workflow message.
All 22 call sites pass the appropriate unit type context so planning,
research, execution, and completion phases each use the configured model.
Closes#1582
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
closeoutUnit() ran at the start of the next loop iteration, creating a
window where a crash between runUnit() returning and the next iteration
would lose all telemetry (metrics, activity log, memory extraction).
completed-units.json was also never flushed to disk, causing severe
staleness (3 entries for 322 completed units in production).
Closes#1590
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
smartStage() ran `git add -A` on the entire repo then unstaged exclusions,
causing indefinite hangs on repos with large untracked artifact trees (57GB+).
autoCommitDirtyState() bypassed smartStage() entirely via direct nativeAddAll().
Add nativeAddAllWithExclusions() using `git add -A -- ':!pattern'` syntax so
excluded paths are never hashed. Route autoCommitDirtyState() through it with
RUNTIME_EXCLUSION_PATHS.
Closes#1605
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When a slice is marked [x] in ROADMAP but tasks are incomplete and no
summary exists, doctor detects slice_checked_missing_summary (declared
fixable) but had no shouldFix handler — creating an unrecoverable
deadlock. Add handler that unchecks the slice when tasks are incomplete,
and add markSliceUndoneInRoadmap to both doctor.ts and
roadmap-mutations.ts.
Closes#1591
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The MCP client passed raw "${VAR}" strings to child processes instead of
resolving them against process.env, breaking MCP servers that expect
resolved environment variable values.
Adds a resolveEnv() helper that interpolates ${VAR} patterns in env
config values before passing them to StdioClientTransport.
Closes#1599
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
After dispatchDoctorHeal fires pi.sendMessage({ triggerTurn: true }),
the function fell through to return "continue". The auto-loop treated
"continue" as "proceed to next unit", called newSession() while the
session manager was still processing the heal turn, and the 30s timeout
killed auto-mode.
Returning "dispatched" causes the auto-loop to break, letting the heal
turn complete and trigger its own handleAgentEnd to resume the loop.
Closes#1580
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
In manual sessions (no auto-mode), bootstrapAutoSession never runs, so the
GSD database is never opened. This causes gsd_save_decision,
gsd_update_requirement, and gsd_save_summary tools to always fail with
'GSD database is not available'.
Add ensureDbOpen() helper that checks isDbAvailable() first, then tries to
open the DB from the expected .gsd/gsd.db path if it exists. All three tool
handlers now use this helper instead of the check-only pattern.
The fix is backward-compatible: in auto-mode the DB is already open, so
ensureDbOpen() returns true immediately on the isDbAvailable() check.
Move `pendingResolve` and `sessionSwitchInFlight` from AutoSession to
module-level variables in auto-loop.ts (`_currentResolve`,
`_sessionSwitchInFlight`). Remove `pendingAgentEndQueue` entirely —
agent_end events arriving with no pending resolver are now dropped
(with a debug warning) instead of queued.
This eliminates the `_activeSession` singleton, the queue drain logic
in `runUnit`, and three properties from `AutoSession.reset()`.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Extract validateProjectId() and validate at startup in
bootstrapAutoSession() so users get immediate feedback on invalid
values. repoIdentity() returns the custom ID directly when set.
Delete prompt-compressor, summary-distiller, and semantic-chunker modules
plus all associated tests. Replace all compression/distillation/chunking
call sites with section-boundary truncation via truncateAtSectionBoundary.
Remove compression_strategy preference, validation, and documentation.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Each cleanup group in stopAuto is wrapped in its own try/catch so a
failure in one step (e.g., worktree exit, DB close, model restore)
cannot abort remaining cleanup. Critical invariants (s.active=false,
s.paused=false, UI reset, pendingResolve=null) are moved into a
finally block that executes unconditionally.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(gsd): tighten prompt automation contracts
* fix(gsd): restore confirmation gates for reflection/requirements/roadmap, scope workflow autonomy by complexity
Amends PR #1556 to address two behavioral risks:
1. discuss.md: Remove "treat continuation as confirmation" fallthrough —
elaboration is not confirmation. Restore explicit confirmation gates
for requirements and roadmap preview.
2. workflow-start.md: Gate autonomy on {{complexity}} — low/medium
workflows keep moving by default, high complexity workflows confirm
at phase transitions.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Address six convergent audit findings in the auto-mode agent loop:
1. Move rewriteAttemptCount to AutoSession — eliminates module-level state
that leaked across stop/start cycles in auto-dispatch.ts
2. Add unit correlation to agent_end queue — tag events with unitId so late
completions from unit A cannot falsely resolve unit B
3. Split post-unit into heavy/light paths — sidecars skip settle delay,
doctor, state rebuild, and worktree sync; reduce sleep 500ms→100ms
4. Data-driven budget thresholds — consolidate 75/80/90% copy-pasted
notification blocks into BUDGET_THRESHOLDS array lookup
5. Fix session teardown — stopAuto() restores model first then calls
s.reset() replacing 36 lines of manual field clearing
6. Add debugLog to 12 silent catch blocks in auto-post-unit.ts
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Centralise all ~/.gsd path construction through app-paths.ts (compiled
code) or a module-level `gsdHome` const (runtime extensions that cannot
import app-paths). When GSD_HOME is set, every path that previously
resolved under ~/.gsd now resolves under the override.
Existing overrides (GSD_STATE_DIR, GSD_CODING_AGENT_DIR) continue to
take precedence when set.
PR #1527 fixed metrics.ts but missed several other paths that still
reach shared/mod.js → ui.js → @gsd/pi-tui during report generation
via native dynamic import() (which bypasses jiti alias resolution).
Remaining chains fixed:
- preferences.ts, preferences-validation.ts, export.ts, forensics.ts,
migrate/parsers.ts: import from shared/format-utils.js directly
- state.ts, visualizer-data.ts, files.ts: import from milestone-ids.js
instead of guided-flow.js (which pulls in shared/mod.js)
- files.ts: import checkExistingEnvKeys from new env-utils.ts instead
of get-secrets-from-user.ts (which imports @gsd/pi-tui)
New file: env-utils.ts extracts the pure checkExistingEnvKeys function.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
New detections:
- Circular dependency detection (DFS cycle check on slice depends:[])
- Orphaned slice directories (dirs not in roadmap)
- Duplicate task IDs in plan files
- Task summary files on disk not in plan (info)
- Stale REPLAN.md when all tasks are done (info)
- Metrics ledger corruption (version != 1 or units not array)
- Large planning files >100KB (warning)
- Future completed_at timestamps >24h ahead (warning)
New modes and output:
- --dry-run flag: reports [dry-run] would fix entries without writing
- --json flag: formatDoctorReportJson() for CI/tooling integration
- --build / --test flags: opt-in slow checkBuildHealth/checkTestHealth
- Per-check timing: timing.{git,runtime,environment,gsdState} on DoctorReport
- Doctor history: appends compact JSONL entry to .gsd/doctor-history.jsonl;
exports readDoctorHistory() for programmatic access
Tests: 27 new test scenarios in doctor-enhancements.test.ts covering all features
* fix: apply pi manifest opt-out to extension-discovery.ts (#1537 follow-up)
The cmux fix in #1537 patched resolveExtensionEntries() in
packages/pi-coding-agent/src/core/extensions/loader.ts to honor
"pi": {} as an opt-out from auto-discovery. However, there is a
second copy of resolveExtensionEntries() in src/extension-discovery.ts
that was not updated. This is the version actually used at startup
by loader.js via discoverExtensionEntryPaths().
As a result, cmux/index.js is still discovered and loaded as an
extension on startup, producing:
Extension does not export a valid factory function: .../cmux/index.js
Fix: Apply the same authoritative-manifest logic to the
extension-discovery.ts copy. When a package.json has a "pi" field,
treat it as authoritative and return early — either with declared
extension paths or an empty array for library opt-out.
Tests: 7 new tests covering resolveExtensionEntries and
discoverExtensionEntryPaths behavior for opt-out, declared
extensions, and fallback discovery.
* fix: apply pi manifest opt-out to package-manager.ts (third copy)
There are THREE copies of resolveExtensionEntries():
1. packages/pi-coding-agent/src/core/extensions/loader.ts (fixed in #1537)
2. src/extension-discovery.ts (fixed in previous commit)
3. packages/pi-coding-agent/src/core/package-manager.ts (THIS commit)
Copy #3 is used by collectAutoExtensionEntries() which is called from
addAutoDiscoveredResources() during DefaultPackageManager.resolve().
This is the actual code path that discovers ~/.gsd/agent/extensions/cmux
and passes it to loadExtensions(), producing the factory function error.
* fix: rewrite pi.extensions .ts paths to .js during resource copy
copy-resources.cjs compiles .ts → .js via tsc but copies package.json
files verbatim. Extensions with pi.extensions: ["./index.ts"] end up
in dist/ pointing to a .ts file that doesn't exist (only .js does).
This causes resolveExtensionEntries() to find no valid entry points,
silently skipping the extension. Affected: gsd, browser-tools, context7,
google-search, universal-config — all extensions with pi manifests.
Fix: When copying package.json files, rewrite .ts/.tsx extensions in
pi.extensions arrays to .js so they match the compiled output.
* fix: add missing commands to /gsd description and rate sub-completions
- Add 9 missing commands to the description string: widget, rate, park,
unpark, init, setup, logs, inspect, extensions
- Add sub-completions for /gsd rate (over/ok/under)
* feat: grid layout for parallel cmux splits and completion trailing-space fix
CmuxClient.createGridLayout(count) pre-creates a tiled grid of surfaces
before launching parallel agents, instead of the previous approach of
creating splits per-agent with alternating right/down directions.
Grid layout strategy:
1 agent: [gsd | A]
2 agents: [gsd | A] (A split down)
[ | B]
3 agents: [gsd | A] (2x2 grid)
[ C | B]
4 agents: [gsd | A] (additional splits from bottom-right)
[ C | B]
[ | D]
Changes:
- Add CmuxClient.createSplitFrom(sourceSurfaceId, direction) to split
from a specific surface rather than always the gsd surface
- Add CmuxClient.createGridLayout(count) that builds the grid and
returns surface IDs in order
- Update runSingleAgentInCmuxSplit to accept a pre-created surface ID
(string) or a direction for backward compatibility
- Parallel dispatch pre-creates grid, assigns each agent a surface
- Fix getArgumentCompletions trailing-space handling so sub-completions
work (e.g., /gsd cmux <tab> now shows status/on/off/etc.)
- 5 new tests for grid layout logic
* feat(ui): add GSD welcome screen on interactive startup
Renders a two-panel boxed welcome screen to stderr before the TUI
takes over, mirroring the style of the Claude Code welcome screen.
Left panel — personalized greeting, GSD ASCII logo, active model + cwd
Right panel — getting-started tips, recent session activity
The screen is printed to stderr immediately before InteractiveMode.run(),
so it appears on launch and reappears when the TUI exits (alternate-screen
buffer swap). It silently skips when not a TTY or terminal < 60 cols.
Files:
src/welcome-screen.ts — printWelcomeScreen() implementation
src/cli.ts — call site before interactiveMode.run()
src/tests/welcome-screen.test.ts — 11 unit tests (all passing)
* refactor(ui): minimal welcome screen — logo + metadata, no box
Replace two-panel boxed layout with a minimal design:
logo block with version/model/cwd alongside it, dim hint below.
No box borders, no tips panel. Clean and fast.
* feat(ui): show tool status line (Brave/Jina/Tavily) when keys are configured
When .gsd is a symlink (e.g., openclip/.gsd -> ~/.gsd/projects/<hash>),
worktrees resolve to ~/.gsd/projects/<hash>/worktrees/<name> instead of
the expected <repo>/.gsd/worktrees/<name>. All worktree detection
functions used the marker /.gsd/worktrees/ which did not match the
resolved path /.gsd/projects/<hash>/worktrees/.
This caused three cascading failures:
1. escapeStaleWorktree failed to detect stale worktree CWD
2. isUnderGsdWorktrees returned false, causing nested worktrees
3. Empty registry was conflated with "all milestones complete"
Changes:
- Add findWorktreeSegment helper matching both direct and symlink layouts
- Refactor detectWorktreeName and resolveProjectRoot to use the helper
- Fix escapeStaleWorktree in auto-worktree-sync.ts for symlink paths
- Fix isUnderGsdWorktrees in auto-start.ts for symlink paths
- Fix resolveCapturesPath in captures.ts for symlink paths
- Distinguish empty registry from all-complete in auto-loop.ts
- Add tests for symlink-resolved path detection
* fix(gsd): batch-specific artifact verification for reactive-execute
The reactive-execute artifact verifier previously checked only that
'at least one task summary exists' in the slice. This meant the unit
could report success even when none of the dispatched tasks actually
completed — a pre-existing T01 summary would satisfy the check.
Fix:
- Encode dispatched task IDs in the unitId: M001/S01/reactive+T02,T03
- Persist dispatched batch in ReactiveExecutionState before dispatch
- Verify each dispatched task's summary file exists individually
- Legacy unitId format (no +batch suffix) falls back to old behavior
The verifier now answers 'did the tasks we dispatched actually finish?'
instead of 'does any summary exist?'
Added ReactiveExecutionState.dispatched field to track the batch.
5 new tests covering: all-pass, partial-fail, pre-existing-irrelevant,
legacy fallback, and unitId round-trip encoding.
* fix(gsd): dependency-based carry-forward for reactive task execution
In reactive mode, each subagent task was getting order-based carry-forward
(all prior task summaries by number), not dependency-based. T05 depending
only on T02 would still receive T01, T03, T04 summaries — noise context
that wastes tokens and could confuse execution.
Fix:
- Add getDependencyTaskSummaryPaths() — returns only summaries for tasks
in the derived dependsOn set, falling back to order-based for root tasks
with no dependencies (preserves continuity)
- Add ExecuteTaskPromptOptions with carryForwardPaths override
- buildExecuteTaskPrompt accepts optional override, sequential callers
unchanged (no options = order-based, backward compatible)
- buildReactiveExecutePrompt now passes dependency-scoped paths per task
Sequential execute-task dispatch is completely unchanged — the new code
path only activates when carryForwardPaths is explicitly provided.
3 new tests: dependency-only filtering, root task fallback, missing
dependency summary handling.
* fix(gsd): enforce backtick file paths in task plan IO sections
The reactive task graph (ADR-004) derives dependencies from backtick-wrapped
file paths in ## Inputs and ## Expected Output sections. Without concrete
paths, the graph is ambiguous and falls back to sequential execution.
Changes:
- task-plan.md template: add comments explaining paths are machine-parsed
- plan-slice.md prompt: explicitly instruct planner to write backtick file
paths in IO sections, add self-audit check for path presence
- observability-validator.ts: new validation rules missing_output_file_paths
(warning) and missing_input_file_paths (info) catch plans without paths
- plan-quality-validator.test.ts: 4 new test cases for IO path validation
* fix(ci): increase max_tokens and add JSON parse error handling in ai-triage
max_tokens: 300 was too low, causing truncated JSON responses from Claude
that failed to parse. Bumped to 1024 and added try/catch with raw text
logging for easier debugging.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add reactive (graph-derived parallel) task execution within slices.
When enabled via preferences, the dispatch table derives a task dependency
graph from IO annotations in task plans and dispatches multiple ready,
non-conflicting tasks in parallel via subagent.
Architecture:
- Graph derivation happens at dispatch time (auto-dispatch.ts)
- A new reactive-execute prompt instructs the agent to use subagent
parallel mode to dispatch all currently-ready tasks
- The auto-loop treats reactive-execute as a single unit type
- After agent_end, the orchestrator checks which tasks completed and loops
New files:
- reactive-graph.ts: pure graph derivation, ready-set resolution,
conflict detection, deadlock detection, IO loader, state persistence
- prompts/reactive-execute.md: prompt template for parallel dispatch
- tests/reactive-graph.test.ts: 22 unit tests for graph functions
- tests/reactive-executor.test.ts: 11 integration tests for dispatch
rules, preferences validation, state persistence, re-entry
Modified files:
- types.ts: TaskIO, DerivedTaskNode, ReactiveExecutionConfig,
ReactiveExecutionState interfaces
- files.ts: parseTaskPlanIO() extracts IO from task plan sections
- preferences-types.ts: reactive_execution config + known keys
- preferences-validation.ts: validation with range checks
- auto-dispatch.ts: new reactive-execute dispatch rule
- auto-prompts.ts: buildReactiveExecutePrompt()
- auto-recovery.ts: artifact verification for reactive-execute
- auto-post-unit.ts: reactive state cleanup on slice completion
Backward compatible: disabled by default, falls through to sequential
execution when disabled, ambiguous, or only 1 task is ready.
* feat: add anthropic-vertex provider for Claude models on Google Vertex AI
Add a new anthropic-vertex provider that enables using Claude models
(Opus 4.6, Sonnet 4.6, Haiku 4.5) through Google Vertex AI using the
@anthropic-ai/vertex-sdk package. Follows the same pattern as the
existing google/google-vertex provider split.
Detection uses ANTHROPIC_VERTEX_PROJECT_ID (same env var as Claude Code)
with CLOUD_ML_REGION for region selection, falling back to us-central1.
Extracts shared Anthropic utilities into anthropic-shared.ts (message
conversion, tool conversion, param building, stream processing) to
avoid duplication between anthropic.ts and anthropic-vertex.ts.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat: add full Claude model set for anthropic-vertex provider
Add 200K context window variants for Opus 4.6 and Sonnet 4.6, plus
older models (Sonnet 4.5, Sonnet 4, Opus 4.5, Opus 4.1, Opus 4, Haiku 4.5).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: add @anthropic-ai/vertex-sdk to root dependencies
Required for the published package to resolve the vertex SDK at runtime.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* chore: remove unnecessary comments to match codebase style
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: remove duplicate stream functions after rebase
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Nathan Roe <nathan.roe@carvana.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Replace the simple notifyRemoteAutoActive notification with an interactive
guardRemoteSession menu that shows session details and offers actionable
choices (view status, steer, stop, or force start). Guards all auto-mode
entry points: bare /gsd, /gsd next, and /gsd auto.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(#1526): auto-mode worktree commits land on main instead of milestone branch
GitServiceImpl.getMainBranch() was designed to detect manual /worktree worktrees
(worktree/<name> branches) but incorrectly applied the same logic to auto-mode
worktrees (milestone/<MID> branches). When no worktree/<name> branch existed,
it fell back to the current branch, which in certain contexts could be main,
causing slice commits to land on main instead of the milestone branch.
Fix: Detect if currently on a milestone/* branch first (auto-mode case) and
return it, before checking for worktree/* branches (manual worktree case).
- Modify getMainBranch() to detect milestone branches first
- Add test verifying getMainBranch() returns correct branch in auto-worktree
- All tests pass, build succeeds
Fixes#1526
* fix: cmux library directory incorrectly loaded as extension
The extension auto-discovery in resolveExtensionEntries() finds
index.js files in subdirectories and treats them as extensions.
The cmux directory has an index.js but it's a utility library
(imported by gsd and subagent extensions), not an extension itself.
Two changes:
1. When a package.json has a "pi" manifest, treat it as authoritative
and don't fall through to index.ts/index.js auto-detection. This
lets library directories opt out with "pi": {}.
2. Add package.json to cmux directory with empty pi manifest.
Report generation in auto-loop uses native dynamic import() which
bypasses jiti's alias resolution. The import chain
metrics.js → mod.js → ui.js → @gsd/pi-tui failed because Node
cannot resolve @gsd/pi-tui from ~/.gsd/agent/extensions/.
Split ANSI-aware layout helpers (padRight, joinColumns, centerLine,
fitColumns) into layout-utils.ts and keep format-utils.ts pure so
report modules can import formatting functions without pulling in
the @gsd/pi-tui dependency.
GitServiceImpl.getMainBranch() was designed to detect manual /worktree worktrees
(worktree/<name> branches) but incorrectly applied the same logic to auto-mode
worktrees (milestone/<MID> branches). When no worktree/<name> branch existed,
it fell back to the current branch, which in certain contexts could be main,
causing slice commits to land on main instead of the milestone branch.
Fix: Detect if currently on a milestone/* branch first (auto-mode case) and
return it, before checking for worktree/* branches (manual worktree case).
- Modify getMainBranch() to detect milestone branches first
- Add test verifying getMainBranch() returns correct branch in auto-worktree
- All tests pass, build succeeds
Fixes#1526
The Anthropic API's max_uses resets per request — when pause_turn triggers
a resubmit, the model gets a fresh budget each time. This allowed unlimited
total searches across a research unit, overwhelming the TUI render buffer.
Fix:
- Count web_search_tool_result blocks in conversation history on each
before_provider_request to track cumulative searches per session
- Cap total native searches at 15 per session (3 full turns of 5)
- Dynamically set max_uses to min(5, remaining) — preserves per-turn cap
while enforcing session ceiling
- When budget exhausted, omit web_search tool entirely instead of letting
the model hit max_uses_exceeded repeatedly
- Reset counter on session_start (new agent unit)
- Add web search budget guidance to research prompts (defense in depth)
Tests: 5 new tests covering budget tracking, exhaustion, and reset.
All 35 native-search tests pass.
* feat(dashboard): two-column layout with redesigned widget
- Two-column layout: progress bar left, task checklist right
- 4 widget modes: full → small → min → off (cycle with /gsd widget)
- Health indicator and ETA in header line for immediate visibility
- Simplified stats: 3 items (hit rate, cost, context %) instead of 7
- Short PWD (last 2 segments), git worktree name with ⎇ prefix
- Last commit time + message in footer (cached every 15s)
- Preview script with mock data for all modes
* docs: add dashboard widget screenshots for PR #1530
* docs: update dashboard screenshots with wider renders
* docs: wider full-width dashboard screenshots
* feat(dashboard): persist widget_mode in preferences
- Add widget_mode to GSDPreferences and KNOWN_PREFERENCE_KEYS
- Load saved widget_mode from preferences on first access
- Persist to global PREFERENCES.md on /gsd widget change
- Default remains "full" when no preference is set
Remove the bundled SwiftUI skill which had 13+ broken references to a
non-existent `../macos-apps/references/` directory. Add a CI script
that validates all relative .md file references in bundled skills,
preventing this class of bug from shipping again. Fix 5 additional
pre-existing broken references in other skills.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When a unit fails and is retried, the model tier is now escalated
(light -> standard -> heavy) if dynamic routing is enabled and
escalate_on_failure is not explicitly disabled. This connects the
existing escalateTier() function, which was fully implemented and
tested but never called at runtime.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Bare /gsd and /gsd next now check for a remote auto-mode session via
readSessionLockData before attempting to start step-mode. If another
process holds the lock, a steering menu is shown instead of competing
for the lock and killing the running session.
Also fixes the guided-flow "all slices discussed" message to detect
active auto-mode and direct users to /gsd status instead of bare /gsd.
Closes#1507
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Wire resolveProfileDefaults into loadEffectiveGSDPreferences so
token_profile: budget actually sets cheaper models and skips phases
- Add /gsd rate <over|ok|under> command to submit user feedback on
model tier assignments, completing the adaptive routing feedback loop
- Document that models config is required for dynamic routing activation
- Document ceiling behavior when dynamic routing is active
- Document reassess_after_slice as required for reassessment
Closes#1505 (partial — escalateTier wiring deferred to follow-up)
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The onCompromised callback in the retry acquisition path lacked the
elapsed-time suppression that the primary path had, causing unconditional
_lockCompromised=true on benign mtime drift. Additionally, validateSessionLock
now attempts PID-based recovery and re-acquisition before declaring the lock
lost, preventing sessions from stopping when no other process has taken over.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace loadAgentInstructions() with a deprecation warning that fires when
legacy agent-instructions.md files are detected. Pi core already supports
AGENTS.md (with CLAUDE.md fallback) per directory, making the custom GSD
mechanism redundant.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Directories under .gsd/milestones/ that don't match the M\d+ pattern
(e.g. slices/, temp-backup/) are now excluded instead of being returned
with their raw name. This prevents rogue directories from blocking
auto-mode milestone discovery.
Closes#1494
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
isValidationTerminal() now normalizes verdict: passed → pass before
comparison. LLM-generated validation files that write "passed" instead
of "pass" are accepted as terminal, preventing milestones from being
treated as incomplete.
Closes#1429
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add missing `parseRoadmap` import in `auto-dispatch.ts`
- Add missing `unlinkSync` import in `auto.ts`
- Add missing `syncGsdStateToWorktree` import in `worktree-sync-milestones.test.ts`
All three were dropped during the PR #1419 merge.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
