When a session dies mid-execution (hang, token exhaustion, API drop),
STATE.md's current_phase can be advanced past a phase that still has
PLAN.md files without matching SUMMARY.md files. Without this fix,
/gsd-next and /gsd-progress would route by current_phase and silently
skip the partially-executed phase, producing a data-loss-shape outcome.
Adds a Route 0 cross-phase incomplete-execution scan to both
next.md and progress.md. Before any current_phase-based routing,
the scan finds the lowest-numbered phase where plans outnumber
summaries and routes to /gsd-execute-phase <that-phase> to resume it.
Opt out with --no-resume to fall back to the prior-phase defer prompt;
--force bypasses all gates as before.
Rework (codex review):
- Route 0 now ordered AFTER Gates 1-3 (repo/state validity always run)
but BEFORE the prior-phase completeness-scan defer prompt — eliminating
the double-decision where the default path would both prompt the user
(C/S/F) AND resume the phase anyway. Prior-phase defer prompt moved to
a new prior_phase_completeness step; only reached via --no-resume.
- --force flow made coherent across all three steps: safety_gates jumps
directly to determine_next_action, skipping Gates, Route 0, AND
prior_phase_completeness. resume_incomplete_phase and prior_phase_completeness
now correctly state --force never reaches them. success_criteria entry
updated to reflect --force → determine_next_action (not prior_phase_completeness).
- Scan uses $GSD_SDK (canonical resolver form) throughout next.md, matching
the file's existing convention. progress.md uses $ROADMAP already loaded
by analyze_roadmap. Neither file uses bare gsd-sdk.
- Errors are surfaced rather than suppressed: removed 2>/dev/null on the
main roadmap.analyze call; added explicit WARNING emission when the scan
cannot run, so the invariant fails closed instead of failing open.
- Predicate aligned to plans-without-summaries (plans.length > summaries.length)
in both files, consistent with determine_next_action Route 4.
- command references use canonical /gsd: namespace form
Fixes#160
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
finishInstall called configureOpencodePermissions unconditionally, causing
fs.mkdirSync + fs.writeFileSync to run even under GSD_TEST_MODE='1', violating
the side-effect-free contract. Guarded the call with !process.env.GSD_TEST_MODE.
Fixes#130
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Static regex literal `byPhaseTablePattern` was recompiled on every call to `updatePerformanceMetricsSection`; hoisted to module scope (compiled once; stateless /i used with .match → safe to share across calls). `phaseRowPattern` uses dynamic interpolation and stays in-function. `Fixes #320`.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Per-PostToolUse hot path did stat-then-read ×3 (config.json, metrics bridge,
warn sentinel); collapsed to read-with-ENOENT-catch (fewer blocking syscalls,
no TOCTOU), behavior identical. Fixes#317.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
acquireStateLock was allocating a fresh SharedArrayBuffer on every retry
iteration via Atomics.wait(new Int32Array(new SharedArrayBuffer(4)), ...).
The buffer is never mutated and never escapes, so hoisting it before the loop
is a provably-equivalent transformation — Atomics.wait always sees value 0
whether the buffer is fresh or reused.
Fixes#316
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
loadConfig called detectSubRepos(cwd) at up to 3 sites per invocation (root-config
requiresFilesystem migration, workstream-config requiresFilesystem migration, and the
planning.sub_repos filesystem re-sync) — all with the same cwd, yielding identical
results. Introduce a per-call lazy memo (getDetectedSubRepos) so the directory scan
runs at most once per loadConfig call while preserving all conditional logic. Fixes#315.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Hot path in cmdRoadmapAnnotateDependencies called planData.find() on every
checklist line. Replaced with a first-wins Map built once before the loop so
each line resolves in O(1); first-wins preserves exact .find() semantics and
null-on-miss → wave-1 default is unchanged.
Fixes#314
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Root cause: the affected-tests runner called runAllSuites() on critical-path
changes (running every suite including install/slow on all matrix cells including
Windows), and pickAffectedTests injected DEFAULT_SMOKE_TESTS (an install test)
as the empty-selection fallback — causing install suite tests to run on PR lanes
where they are push-only per docs/TESTING-SUITES.md.
Fix: PR_EXCLUDED_SUITES filter at the pickAffectedTests chokepoint strips
install/slow from every selection path (direct-change, reverse-index, stem-match).
Empty selection now returns [] and the caller runs the unit suite as smoke.
Critical-path fallback replaces runAllSuites with PR_FULL_SUITES
(unit, integration, security). suiteOf exported from run-tests.cjs
(with require.main guard) so affected-tests-lib reuses canonical detection.
Fixes#370
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
intelExtractExports deduped export names with `if (!arr.includes(x)) arr.push(x)`
across ~8 extraction loops (one doubly-nested over an export block) — O(n^2).
Accumulate into Sets (add/has/size) and materialize to an array once at return.
Set dedups by value and preserves insertion order, so the returned export list
and its first-seen order are identical. Adds behavior-lock tests for dedup + order.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
parseNamedArgs re-scanned argv with indexOf/includes once per flag —
O(flags * argv) — on the command-dispatch hot path (24 call sites across
gsd-tools + init/state/validate routers). Build a first-index Map of argv
tokens in a single pass and use it for the flag lookups, dropping it to
O(argv + flags). Semantics are identical: firstIndex.get(t)??-1 === indexOf(t),
firstIndex.has(t) === includes(t); first-occurrence-wins and the value-token
rejection are preserved. Adds the first behavior-lock tests for the module.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
cmdCommitToSubrepo routed each changed file to a sub-repo via
subRepos.find(...) inside the file loop — O(files * repos). Extract a pure
groupFilesBySubrepo() that buckets sub-repos by first path segment and scans
only the matching bucket, dropping it to expected O(files + repos). First-
match-in-array-order semantics (incl. multi-segment sub-repos) are preserved
exactly. Adds the first behavior-lock test for the routing path.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
learningsCopyFromProject called learningsWrite K times in one process,
and each call re-scanned the entire learnings store to dedupe — O(K*N).
Build the content_hash -> id index once at the start of the bulk import
and thread it through; single-write behavior and the return contract are
unchanged (no caller reads `id` on the created:false branch). O(K*N) ->
O(N+K). Adds a regression test asserting store scan count is independent
of import size.
The larger persistent on-disk index (atomic updates, corruption rebuild,
cross-process dedupe) is deferred — needs design decisions and is not
required to resolve the bulk-import scan this issue reports.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replace the per-render readdirSync().filter().map(statSync).sort() chain
with a single-pass max-by-mtime loop. Drops the O(n log n) sort and the
throwaway intermediate array; I/O and resolved-file behavior are identical.
Adds the first behavior-lock test for the todo-resolution path.
The larger disk-backed cache win from the issue is deferred: statusline is
a fresh child process per render, so any cache must be disk-backed with
invalidation/atomic-write design that needs maintainer input.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The Pass-2 topological level assignment in cmdPhasePlanIndex dequeued its
Kahn's-algorithm queue with Array.shift(), which is O(n) per call in V8, so the
BFS was O(V^2) and slowed superlinearly on deep queues (wide fan-in plan
graphs). Extract the traversal into a pure, exported computeDependencyLevels
(rawPlans, planMap, canonicalToId) and dequeue via a head index (queue[head++])
-> O(V+E). Behavior is identical: same FIFO order, same longest-path levels,
same visited-count cycle detection. A complexity-contract comment above the loop
documents why shift() must not be reintroduced.
Adds tests/phase-dependency-levels.test.cjs with deterministic behavior and
edge-case coverage (linear chain, diamond longest-path, independent set, cycle,
canonical-prefix resolution, empty, self-loop, duplicate edge, external dep). A
timing-based complexity guard was intentionally omitted: the O(V+E) Map-build
constant dilutes the O(V^2) signal until impractical N (~1e6), so an empirical
guard is inherently flaky on contended CI — the contract is enforced by the
inline comment and correctness tests instead.
Fixes#307
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(#145): extractCurrentMilestone selects active sub-milestone over closed sibling
extractCurrentMilestone used a non-global regex with content.match()
(first-match) to locate the milestone section for the STATE.md version.
When the milestone is a shared semver prefix (e.g. v8.0) and ROADMAP.md
holds both a closed sub-milestone (## v8.0 ... CLOSED/FAIL) and an active
one (## v8.0-B ... STARTED), the first match was always the closed
heading, so the active section and its phases were excised from the slice
and downstream phase ops failed with "Phase N not found in current
milestone".
Switch to a global matchAll over candidate headings, skip headings
carrying a closed marker (CLOSED/ARCHIVED/ABANDONED/SHIPPED/FAILED/FAIL/
✅/🗄️), and select the first non-closed match (falling back to the first
match when every candidate is closed, preserving legacy behavior). Anchor
the preamble slice to the first heading index so a closed sibling's body
no longer leaks into the preamble when the selected section is later.
Harden version matching with a trailing word boundary (so v8.0-B does not
match v8.0-Beta), narrow the FAIL marker to FAILED, match a bare 🗄, and
add an active-marker override (STARTED/🚧/ACTIVE) so a heading carrying an
explicit active status is never treated as closed even if its name contains
a completion word. Also anchor the preamble at the first any-version
milestone heading so unmatched sibling sections do not leak into the
preamble.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* ci(#145): add changeset fragment for milestone-selection fix
Adds the required .changeset/*.md fragment for this user-facing fix
(changeset-lint gate).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(#373): replace unquoted $GSD_SDK with space-safe gsd_run launcher
Workflow bash blocks resolved the runtime as GSD_SDK="node $GSD_TOOLS" and
invoked it unquoted ($GSD_SDK query ...). On install paths containing spaces
(e.g. /Volumes/Mini Me/...) the unquoted expansion word-split into
`node /Volumes/Mini gsd-tools.cjs ...`, failing with "Cannot find module
'/Volumes/Mini'" and getting masked by `2>/dev/null || echo "{}"` into a
silent empty state.
Replace the string variable with a single-line shell launcher that defines a
gsd_run function, invokes the runtime with a fully-quoted path and "$@", and
preserves the local-cjs / installed-gsd-tools-on-PATH fallback (#3668) plus
the loud not-found error and install hint. The launcher uses _GSD_SHIM_NAME
indirection so no workflow emits the /gsd-tools substring that the do.md
dispatcher-parity scanner would misread, and is single-line to stay within
the per-file progressive-disclosure line budgets (#2551).
The canonical launcher lives in
get-shit-done/workflows/_runtime-launcher.snippet.sh, is propagated once per
file by scripts/sync-runtime-launcher.cjs, and is locked by
tests/runtime-launcher-parity.test.cjs (fails CI on drift, on a reappearing
$GSD_SDK token, or on a /gsd-tools substring). Dependent workflow-assertion
tests are updated from $GSD_SDK to gsd_run, and the runtime launcher is
registered in CONTEXT.md.
Fixes#373
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* ci(#373): add changeset fragment and allow-test-rule for parity guard
The runtime-launcher parity test is a structural drift guard that reads
workflow markdown to assert the canonical launcher is present and the
retired $GSD_SDK / /gsd-tools tokens are absent; annotate it with
allow-test-rule per the no-source-grep lint escape hatch. Add the required
.changeset fragment for this user-facing fix.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test(#373): make parity PATH-fallback assertion cross-platform (Windows)
Subtest (E) compared GSD_TOOLS against the Node-side absolute temp path,
but the value originates from git-bash which reports the POSIX form, so the
prefix comparison failed on windows-latest while the launcher itself worked
(the installed stub was invoked). Assert the resolved binary by normalized
suffix (/bin/gsd-tools, not .cjs) instead of the absolute prefix; the
behavioral stub-invocation assertion is unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix: use active-workstream resolver exported by store module
* fix: wire verify codebase-drift alias and sync inventory docs
* chore: add changeset for next gate regression fixes
* fix: normalize changeset fragment metadata for docs-lint
* test(#178): update DispatchEvent factory tests to propagate parentTraceId
P1.3 test 'parentTraceId is always undefined' replaced with four P1.4
contracts: absent → undefined, string → propagated, null → undefined,
non-string → undefined (defensive normalization policy).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(#178): propagate parentTraceId through DispatchEvent factory
Stop ignoring the parentTraceId parameter added as a forward-compat hook
in P1.3. Defensive normalization: only non-null strings are propagated;
null, non-string values, and absent callers all yield undefined, keeping
P1.3 behavior intact for all existing dispatch call sites.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* test(#178): add Hub-level parentTraceId propagation tests
Four new assertions: req.parentTraceId propagates to event, absent →
undefined (P1.3 regression), shared parentTraceId across multiple
dispatches, and unique traceId invariant despite shared parentTraceId.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(#178): plumb parentTraceId through Hub dispatch and _notifyLogger
dispatch() now reads req.parentTraceId and passes it to _notifyLogger,
which forwards it to makeDispatchEvent. Backward-compatible: callers
that omit parentTraceId emit events with parentTraceId: undefined,
identical to P1.3 behavior.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* test(#178): add trace correlation end-to-end test
Dispatches a root command then 3 children with parentTraceId=rootTraceId.
Reads the real .gsd-trace.jsonl audit file and verifies: 4 events total,
root has no parentTraceId, all children carry rootTraceId, all traceIds
unique, JS filter returns exactly the 3 children given the root's traceId.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* docs(#178): document traceId/parentTraceId in audit file
Update Observability section to note that audit events now carry both
traceId and parentTraceId, and explain the correlation filter pattern.
Note that leaf dispatches emit parentTraceId: undefined until the Phase 2
composer wires it automatically.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* chore(#178): add changeset for trace correlation seam
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* test(#178): cover invalid parentTraceId values in DispatchEvent factory
Adds 9 new test cases for UUID v4 validation of parentTraceId:
empty string, whitespace, non-UUID, oversized, UUID v1, missing-hyphen,
extra-char (all dropped to undefined), plus UPPERCASE and lowercase v4
(both propagated). Tests are intentionally red until the implementation
commit that follows.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(#178): validate parentTraceId against UUID v4 before propagation
Adds UUID_V4_REGEX constant and isValidParentTraceId() helper to
event.cjs. makeDispatchEvent now silently coerces any parentTraceId that
fails the UUID v4 check (wrong version nibble, wrong variant, missing
hyphens, oversized, empty, etc.) to undefined. No stderr warn is emitted
— the factory remains pure and side-effect-free. Closes the correlation-
poisoning vector identified in the Codex adversarial review of PR #225.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* test(#178): assert Hub silently drops invalid parentTraceId at the seam
Adds two tests to hub-logger-integration.test.cjs:
1. dispatch with 'junk' parentTraceId emits event with parentTraceId===undefined.
2. The logger-failure warn path is NOT triggered — the factory coerces the bad
value before onEvent is called, confirmed by zero stderr output even when a
logger that would throw on non-undefined parentTraceId is installed.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* test(#178): assert invalid parentTraceId does not poison correlation siblings
Adds one test to trace-correlation.test.cjs: dispatches a root, a valid
child (parentTraceId = rootTraceId), and an invalid child (parentTraceId =
'junk'). Asserts: valid child carries correct parentTraceId, invalid child
has parentTraceId dropped to undefined, filtering by rootTraceId yields
exactly 1 event (the valid child only), and all 3 events have unique
traceIds. Uses an isolated Hub + tmpdir to avoid shared fixture interference.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* docs(#178): document UUID v4 contract for parentTraceId
Appends one sentence to the Observability audit-trail paragraph in
CONFIGURATION.md: parentTraceId must be canonical UUID v4 (RFC 4122);
values that don't match are silently dropped from audit output. No section
restructuring — single sentence addition only.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* test(#177): add DispatchEvent factory failing tests
Red tests for makeDispatchEvent shape, traceId UUID v4, uniqueness,
parentTraceId-always-undefined (P1.3), args redaction toggle, ISO 8601
timestamp, and all result variant passthrough.
* feat(#177): introduce DispatchEvent factory
makeDispatchEvent produces an immutable event record per dispatch:
- traceId: crypto.randomUUID() (UUID v4)
- parentTraceId: always undefined (P1.4 wires composer)
- command, result, timestamp (ISO 8601)
- args only included when includeArgs === true (default: omitted)
* test(#177): add arg redaction policy failing tests
Red tests for shouldIncludeArgs (GSD_AUDIT_ARGS env gating) and
redactEvent (strips args from frozen events, preserves all other
fields, returns a new object, never mutates the source).
* feat(#177): introduce arg redaction policy
shouldIncludeArgs(): only GSD_AUDIT_ARGS==='1' opts in; all other
values (unset, '', '0', 'true') default to omitting args.
redactEvent(event): returns a shallow copy of the event, dropping the
args field unless opted in. Never mutates the (frozen) source event.
* test(#177): add DispatchLogger interface failing tests
Red tests covering:
- no-op logger: silent on all events, never throws
- default logger: silent on ok, one flattened JSON line to stderr on error
- default logger: audit file creation + append-only + redaction + config gate
- GSD_AUDIT env var and config.audit.enabled config gate
- GSD_AUDIT_ARGS opt-in for args inclusion
All tests use real fs under os.tmpdir() — no mocked appendFileSync.
* feat(#177): introduce DispatchLogger with default and no-op implementations
createNoOpLogger(): silent on all events — Hub default when no logger injected.
createDefaultLogger({ cwd, config }):
- Silent on ok result
- Flattened JSON line to stderr on error: { kind, traceId, ...typedPayload }
- Append-only audit at .planning/.gsd-trace.jsonl when GSD_AUDIT=1 or config.audit.enabled
- Args redacted by default; GSD_AUDIT_ARGS=1 opts in
- Logger errors caught internally; never break dispatch callers
* test(#177): add Hub+logger integration failing tests
Red tests verifying:
- onEvent called exactly once per dispatch (ok, error, handler-throw, unknown)
- DispatchEvent shape: traceId uniqueness, command, result.kind, parentTraceId
- Logger errors contained (dispatch still returns Result, warn line to stderr)
- Hub defaults to no-op when no logger injected
- End-to-end with createDefaultLogger: silent on success, stderr on error, audit file
* feat(#177): wire DispatchLogger into CommandRoutingHub
Add optional logger param to createHub({ ..., logger }).
Defaults to createNoOpLogger() — silent, no behaviour change for callers
that don't inject a logger.
After every dispatch (success and error):
- Normalises HubResult { ok } to DispatchEvent { kind: 'ok'|error-kind }
- Calls makeDispatchEvent({ command, args, result }) to mint the event
- Calls logger.onEvent(event) exactly once
- Wraps in try/catch: logger errors emit { level:'warn', source:'DispatchLogger' }
to stderr but never propagate to dispatch callers
* chore(#177): gitignore .planning/.gsd-trace.jsonl audit file
The audit trail is local-only, append-only, and must never be committed.
Slotted under the existing "Local scratch + Claude-test artifacts" block.
* docs(#177): document GSD_AUDIT, GSD_AUDIT_ARGS, config.audit.enabled
New ## Observability section at end of CONFIGURATION.md covering:
- Default silent/stderr behaviour overview
- Stderr error JSON format
- Audit file opt-in (env var and config key)
- Args redaction policy and GSD_AUDIT_ARGS opt-in
Also slots GSD_AUDIT and GSD_AUDIT_ARGS into the existing
## Environment Variables table (alphabetical order).
* chore(#177): add changeset for observability seam
type: Added — new DispatchLogger seam with default silent/stderr/audit behaviour.
* fix(#167): support query meta-command in gsd-tools
* chore(#167): add changeset for query meta-command fix
* fix(#167): pin claude runtime in local-agent regression tests
* test(#3751): stabilize local-agent CI assertions
* refactor(hub): tighten Result<T> to typed-payload-per-kind discriminated union (#176)
Each Hub error variant now carries only its own typed payload. The generic
`errorKind` field is renamed to `kind`; `message`/`details` escape hatches
are removed from Hub-emitted errors. Factory functions (makeUnknownCommand,
makeInvalidArgs, makeHandlerRefusal, makeHandlerFailure) are exported and
used in phase-command-router.cjs. Callers switch on `result.kind`.
Part of ADR-0174 P1.2.
<!-- docs-exempt: no docs/ changes; API is internal to Hub callers -->
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(hub): act on P1.2 review findings (#176)
Addresses 4 review findings on PR #221:
- Hub now runtime-validates ok:false variants against the typed shape
and coerces malformed returns to HandlerFailure with a contract-
violation message (codex finding #1, code-review finding #1)
- catch path now preserves the original throwable for non-Error
throws via an Error wrapper with .thrown attached (codex finding #2)
- All 4 factory returns are Object.freeze'd (review finding #9)
- makeHandlerFailure validates cause is Error; non-Error causes are
wrapped with .thrown attached (review finding #10)
Tests added for each finding (TDD red → green).
Refs #176. Part of #174.
* fix(docs-lint): add docs-exempt markers to both P1.2 changeset fragments
Both `176-typed-result-discriminated-union.md` and `176-hub-p1.2-review-findings.md`
carry `type: Changed` which triggers the docs-required lint. Neither fragment had
a `<!-- docs-exempt: <reason> -->` marker, causing `docs-lint` to fail with
`FAIL_DOCS_MISSING`. Added the per-fragment exemption marker to both (the repo has
no `no-docs` label). This is a purely internal SDK refactor (ADR-0174 P1.2) with
no public docs surface.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
