The forensics prompt was sent as a one-shot message via sendMessage()
with triggerTurn: true, causing context loss on follow-up turns. Now
writes an active-forensics.json marker to .gsd/runtime/ so that
buildBeforeAgentStartResult() can re-inject the forensics prompt on
subsequent turns, mirroring how guided task execution context works.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
The milestones list only refreshed on agent_end events, causing stale
milestone state during multi-turn agent execution. Add turn_end as a
workspace cache invalidation trigger so the UI reflects milestone
changes after each turn boundary.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Bug 1: orphaned worktree check now skips directories that only contain
doctor artifacts (.gsd/doctor-history.jsonl), preventing the circular
false positive where appendDoctorHistory recreates the dir it reports.
Bug 2: blocker_discovered_no_replan check now skips when all tasks are
done, treating the blocker as implicitly resolved and breaking the
deadlock with stale_replan_file.
Bug 3: parsePlan now scans the full body for task checkboxes after the
Tasks section, finding T02+ entries that appear after interleaved
detail headings (## Steps, ## Must-Haves).
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
The subprocess spawned by collectAuthoritativeAutoDashboardData always
starts with fresh module state (s.active === false), so the web UI
always showed "Start auto" even while auto mode was running. After
obtaining the subprocess result, reconcile active/paused state with
the on-disk session lock (.gsd/auto.lock) and paused-session metadata
(.gsd/runtime/paused-session.json).
Closes#2705
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Milestones with no registry entry (ghost directories with no planning
files) were falling through to eligible status due to the fallback
`entry?.status ?? "pending"` combined with empty deps. Now explicitly
classified as ineligible with "no planning data" reason before any
status/dep checks run.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
When `gsd auto` is run with piped stdout (e.g. `gsd auto | cat` or
`gsd auto > file`), the TUI cannot render on a non-terminal output
stream, causing the process to hang indefinitely.
This fix:
- Detects piped stdout before entering interactive mode and redirects
`gsd auto` to headless mode automatically
- Extends the interactive mode TTY gate to also check process.stdout.isTTY
(previously only checked stdin), with a descriptive error message
- Adds `gsd headless` to the non-interactive alternatives hint
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
When a file-backed database has a corrupted freelist, DDL operations
fail with "database disk image is malformed" even though integrity_check
passes. This adds VACUUM recovery to openDatabase() before re-throwing,
matching SQLite's documented recovery strategy for freelist corruption.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Three fixes for the auto-mode regression where db_unavailable causes
infinite artifact-retry re-dispatch loops:
1. resolveProjectRootDbPath now handles /.gsd/projects/<hash>/worktrees/
paths (symlink-resolved layout) in addition to /.gsd/worktrees/
2. ensureDbOpen emits structured diagnostics (resolvedPath, cwd, error)
instead of silently returning false
3. Post-unit artifact retry skips when isDbAvailable() is false, treating
DB infra failure as fatal instead of entering a retry loop
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
The searchWithOAuth() function sent a request body that the Cloud Code
Assist API rejected with 400 INVALID_ARGUMENT. Two issues:
1. URL was missing ?alt=sse query parameter (endpoint returns SSE format)
2. Request body was missing the required userAgent field
Also adds regression tests that capture the fetch call and assert the
request URL and body match the Cloud Code Assist wire contract.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Bug 1 -- UAT stuck-loop: syncProjectRootToWorktree used force:false for
all milestone files, which preserved stale ASSESSMENT files in the
worktree. When the project root had a passing verdict but the worktree
retained a FAIL copy (or lost it during DB rebuild), checkNeedsRunUat
found no passing verdict and re-dispatched run-uat indefinitely (x9).
Fix: after the additive-only safeCopyRecursive, walk ASSESSMENT files in
the project root and force-overwrite the worktree copy when the source
contains a verdict field. This is safe because ASSESSMENT verdicts are
only ever overwritten in a forward direction (FAIL -> PASS on retry).
Bug 2 -- Orphaned worktree: removeWorktree silently swallowed failures
from git worktree remove when untracked files (UAT-RESULT, ASSESSMENT)
blocked removal. The .git/worktrees/<name> internal directory held a
lock that also prevented the rmSync fallback from working.
Fix: after both native removal attempts fail, explicitly remove the git
internal .git/worktrees/<name> directory first, then retry rmSync on
the worktree filesystem directory. Log a warning with manual cleanup
instructions if the final attempt also fails.
Closes#2821
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
getServerConfig now trims whitespace and performs case-insensitive
matching so that names like "langgraph Code" resolve correctly.
getOrConnect uses config.name as the canonical cache key to prevent
duplicate connections from variant casing.
Closes#3029
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Two fixes for the state corruption chain reported in #2960:
1. extractVerdict() now detects verdicts in markdown body patterns
(e.g., **Verdict:** PASS) when YAML frontmatter is absent, preventing
the state machine from looping on validating-milestone when LLMs write
VALIDATION.md manually.
2. handlePlanMilestone() now refuses to re-plan a milestone that has
completed slices, preventing INSERT OR IGNORE from shadowing completed
work after worktree recreation or DB resync.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
The old STREAM_RE enumerated specific JSON error messages individually,
missing variants like "No number after minus sign", "Bad control character",
and "Expected ',' or '}' after property value" — causing auto-mode to
permanently pause instead of retrying.
Replace with a catch-all: all V8 JSON.parse errors end with
"in JSON at position \d+", so one pattern covers every current and
future variant. Keeps "Unexpected end of JSON" and "SyntaxError.*JSON"
for older runtimes.
Supersedes PR #3134. Refs #2882.
Adds a read-only command to display the current effective GSD
configuration — token profile, model assignments per phase, dynamic
routing tiers, git settings, budget, supervisor, workflow toggles,
parallel config, hooks, and preference file sources.
Renders as a themed TUI overlay (scrollable, esc/q to close) with
plain-text fallback for headless/RPC mode.
Comprehensive vscode extension redesign with sidebar reorganization,
new features, and enhanced agent integration:
- Redesign sidebar UI: reduce 6 panels to 3, declutter layout
- SCM provider for tracking agent-modified files
- Checkpoint system for saving/restoring agent state
- Diagnostic integration for surfacing errors in editor
- Line-level editor decorations for agent-modified lines
- Git integration for visualizing agent changes
- Execution plan viewer for live agent step visualization
- Approval/permissions mode system
- Auto-inject editor selection and diagnostics in chat
- Route workflow buttons through Chat panel
- Handle extension UI requests from agent (select, confirm, input)
- Session persistence, ISO timestamp support, descriptive checkpoints
- Bump to v0.3.0
Addresses 30+ issues found in a full review of the interactive TUI spanning
layout/visual, user flow, message rendering, and state management dimensions.
Critical (state/memory):
- Fix onBranchChange unsubscribe function being discarded; store and call in stop()
- Add onThemeChange cleanup in stop() to prevent stale callback retention
- Resolve getUserInput() Promise on shutdown so run() while-loop exits cleanly
- Serialize concurrent message_update event handlers via Promise chain to prevent
duplicate ToolExecutionComponent creation under rapid streaming
- Add cleanup of customFooter, customHeader, autocompleteProvider, and extension
widgets in stop() to prevent timer/watcher leaks
Major (UX/flow):
- Add two-step confirmation for provider auth removal (r key) — matches session
delete pattern; first press shows confirm hint, second press executes
- Normalize list navigation wrapping: oauth-selector and session-selector now
wrap at boundaries, consistent with all other selectors
- Ctrl+C in scoped-models-selector now always cancels modal immediately instead
of clearing search first
- Config-selector position indicator now counts only selectable items, excluding
non-selectable group headers from both numerator and denominator
- user-message-selector auto-dismiss replaced setTimeout(100) with
Promise.resolve().then() to eliminate 100ms flicker
- Add "Unknown command: /foo. Type /help for available commands." feedback for
unrecognized slash commands instead of silently submitting as chat
- Fix dead-end input path: submitPromptsDirectly=false now dispatches prompt
- Wrap session.prompt in isCompacting path with try/catch (was missing, other
path had it)
- Add Esc-to-close hint to provider-manager footer (was undocumented)
Rendering bugs:
- Remove identical dead-code else branch in assistant-message spacing logic
- Add 20-line truncation to generic/unknown tool JSON rendering (was unbounded)
- bash-execution updateDisplay() now uses stored _borderColorKey so
excludeFromContext dim styling is preserved on re-render
- Fix countdown-timer dispose race: _disposed flag prevents extra tick after
clearInterval
- extension-selector nextSelectable() guard prevents cursor landing on separator
- extension-input now rejects empty/whitespace-only submissions
- Normalize bordered-loader spacing: non-cancellable variant no longer adds
orphaned spacer before bottom border
Visual/theme:
- daxnuts.ts center() replaced naive ANSI regex with visibleWidth() from
@gsd/pi-tui for correct true-color sequence handling
- Remove incorrect mistral.ai URL from daxnuts component
- armin.ts now centers art using same visibleWidth approach as daxnuts
- Dark theme warning color: #ffff00 → #e6b800 (muted amber, less harsh)
- dynamic-border default color function wrapped in try/catch to guard against
undefined theme in jiti-loaded extension contexts
- Footer stats grouped with · separator; cache labels changed from R/W to cr:/cw:
- Replace raw \x1b[1m ANSI codes in custom-message, branch-summary-message,
compaction-summary-message, skill-invocation-message with theme.bold()
- welcome-screen visLen now uses strip-ansi instead of hand-rolled regex
Performance:
- diff.ts parseDiffLine regex: [+-\s] → [+\- ] (space only, not all whitespace)
- tab replacement width: 3 spaces → 4 spaces (standard) in both diff.ts and
tool-execution.ts
- chat-controller message_update: skip already-processed content blocks using
lastProcessedContentIndex to reduce O(n) scan per event
The validate-milestone prompt was rewritten to use 3 parallel reviewers.
Update the prompt contract test to verify the new structure instead of
the old gsd_validate_milestone tool approach.
Add getMissingAnnotationTasks() to surface which tasks lack IO
annotations and prevent parallel dispatch. Also add subagent_model
to ReactiveExecutionConfig for overriding the model used by
subagents during parallel task execution.
- getMissingAnnotationTasks() with 4 tests
- subagent_model field on ReactiveExecutionConfig type
- Validation for reactive_execution.subagent_model preference
Add dispatch rule that detects when multiple slices need research
simultaneously and dispatches them in parallel via subagents. Also
rewrite validate-milestone prompt to use 3 parallel reviewers
(requirements, integration, UAT) for faster validation.
- New dispatch rule: planning → parallel-research-slices (2+ ready)
- buildParallelResearchSlicesPrompt with per-slice subagent prompts
- Parallel research slices prompt template
- Validate-milestone rewritten for 3 parallel reviewers
Add parallel.worker_model preference so coordinators can assign a
cheaper model to parallel workers (e.g. Haiku for execution) instead
of inheriting the coordinator's model. The override is applied via
GSD_WORKER_MODEL env var during worker bootstrap.
- Add worker_model to ParallelConfig type and validation
- Inject GSD_WORKER_MODEL env in spawnWorker when configured
- Apply override in bootstrapAutoSession for parallel workers
- Document in preferences-reference.md
Move plan-milestone and plan-slice from standard to heavy tier so
planning uses the best configured model (e.g. Opus) and is not
downgraded by dynamic routing. Milestone-level planning analysis
also returns heavy instead of standard.
Add configured remote channel (Discord/Slack/Telegram) as a checkmark
in the tools row alongside Brave/Answers/Jina. Remove verbose remote
status lines and duplicate display from header-renderer and register-hooks.
Previously, headless --verbose mode accumulated text_delta events into a
buffer and displayed a single truncated 120-char [thinking] line before
tool calls. The model's actual text responses between tool calls were
effectively invisible.
Changes:
- Stream text_delta and thinking_delta events directly to stderr in
verbose mode with [text] and [thinking] block markers
- No truncation — full model output is visible
- Fix non-verbose fallback: read from ame.delta (correct field) instead
of ame.text (always undefined for text_delta events)
- Track inTextBlock/inThinkingBlock state to properly close streaming
blocks before tool calls
- Expand summarizeToolArgs with support for async_bash, await_job,
cancel_job, find, ls, lsp, hashline_edit, subagent, browser_navigate,
and gsd_* tools
- Add streaming formatter functions: formatTextStart, formatTextEnd,
formatThinkingStart, formatThinkingEnd
- Update tests for new tool arg summarization and path field handling
Saves in-progress daemon work from M005-m138xe that was sitting uncommitted.
Includes orchestrator expansion, event bridge/formatter enhancements,
message batcher tweaks, and discord bot additions.
checkAutoStartAfterDiscuss() fire-and-forgets startAuto() when a
milestone is ready. The headless runner then chains `/gsd auto`,
calling startAuto() a second time. Two concurrent auto-loops on the
same AutoSession singleton corrupt shared state (counters, dispatch
maps), causing planning/execution to never run after research.
Add an early `s.active` check at the top of startAuto() so the second
call no-ops. Add source-scanning test to enforce the guard exists.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Three defects in the completing-milestone dispatch guard caused false
positive blocks on valid validation output:
1. Single-line constraint: [^\n]* stopped at newlines, missing verdicts
on subsequent lines. Fixed with [\s\S]{0,500}? (bounded lazy match).
2. Missing keywords: 'satisfied' and 'partially' were absent from the
alternation. LLMs commonly write 'PARTIALLY SATISFIED' or 'FULLY
SATISFIED'. Added both.
3. Markdown bold delimiters: **Operational** blocked [\s:] after the
word. The new [\s\S] class handles any character including *.
Also adds SATISFIED to the structuredMatch includes check, and ✅ to
the prose regex (overlaps with #2862).
Includes 8 regression test cases covering multi-line formats, satisfied
keyword variants, markdown bold tables, and checkmark emoji.
Bug 1 — Workers exit immediately (#2792):
spawnWorker() used `--print "/gsd auto"` which calls session.prompt()
that returns immediately when ctx.newSession() resets the session inside
the auto-loop. Changed to `headless --json auto` which uses an RPC
client that keeps the process alive until auto-mode completes.
Bug 2 — Dispatch guard blocks parallel workers (#2797):
getPriorSliceCompletionBlocker() checked ALL milestones in queue order,
blocking M012 when M011 had incomplete slices. When GSD_MILESTONE_LOCK
is set, the guard now only checks intra-milestone slice dependencies.
Added test covering cross-milestone bypass + intra-milestone preservation.
Bug 3 — Orphaned RPC children on stop (#2798):
stopParallel() gave only 750ms for SIGTERM before SIGKILL. The headless
parent needs ~1500ms to cascade shutdown to its RPC child via
client.stop(). Increased to 3000ms to prevent orphaned processes holding
auto.lock.
Updated tests:
- dispatch-guard.test.ts: new test for GSD_MILESTONE_LOCK bypass
- parallel-worker-monitoring.test.ts: updated spawn args assertion
The dashboard reads elapsed time, total cost, and tokens used
exclusively from AutoDashboardData. When auto-mode is not active
(e.g. manual /gsd next), auto is null and all three metrics show 0
— even though the status bar displays real values via /api/visualizer.
Add the same projectTotals polling pattern (30s interval via
/api/visualizer) that status-bar.tsx already uses, and wire it into
the fallback chain: projectTotals ?? auto ?? 0.
Closes#2709
When worktrees use shared-WAL mode (R012), the worktree DB path resolves
to the same physical file as the project root DB via symlink. Calling
reconcileWorktreeDb() ATTACHes this WAL-mode file to itself, corrupting
the database with 'database disk image is malformed'.
Fix 1 — auto-worktree.ts mergeMilestoneToMain(): skip reconciliation
when isSamePath() confirms both DB paths resolve to the same file.
Fix 2 — gsd-db.ts reconcileWorktreeDb(): defence-in-depth realpathSync
guard inside the function itself, before the ATTACH statement.
Fix 3 — auto/infra-errors.ts: classify 'database disk image is
malformed' as SQLITE_CORRUPT infrastructure error so the auto-loop
stops immediately instead of burning 3 retries on a guaranteed failure.
Regression tests verify:
1. Same-file via symlink returns zero (no ATTACH)
2. Identical string paths return zero
3. Genuinely different DBs still reconcile normally
4. Malformed DB message classified as infra error
5. Transient SQLITE_BUSY is not falsely classified
Closes#2823
When GSD_WEB_DAEMON_MODE=1 is set, scheduleShutdown() becomes a no-op.
The /api/shutdown endpoint still returns { ok: true } so the client
beacon fires without a network error, but process.exit() is never
called. This allows gsd --web to run as a persistent daemon behind a
reverse proxy without exiting on every browser tab close or refresh.
Closes#2835
Multi-turn commands (auto, next) have their own completion signals via
isTerminalNotification ("Auto-mode stopped..."/"Step-mode stopped...").
The execution_complete event fires after command setup before any real
work begins, causing these commands to exit immediately with zero work done.
Closes#2917
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
mapStatusToExitCode only handled "complete" but RPC v2 emits "completed",
causing all headless sessions to falsely timeout and restart.
Also emits milestone-ready notification in checkAutoStartAfterDiscuss so
headless parent can detect and chain into auto-mode.
Closes#2914
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
