* test(#3675): add failing tests for quick-batch core primitives Adds the full behavioral (tests/quick-batch.test.cjs) and property-based (tests/quick-batch.property.test.cjs) coverage for #3675's quick-batch core primitives per the phase's 35-row test matrix — task-list parsing (inline + --file, with path-confinement/symlink-escape/non-regular-file rejection), collision-safe quick-id preallocation under withPlanningLock, BATCH.json schema/validation/resume, dependency-DAG + partitionByFileOverlap wave construction, and exactly-once STATE.md completion (including the STATE-row-written-but-manifest-not-yet-updated crash window). The import target (gsd-core/bin/lib/quick-batch.cjs, compiled from a not-yet-written src/quick-batch.cts) does not exist yet — every test in both files fails at the top-level require() before any assertion runs. Five fast-check properties cover collision-freedom under lock contention, resume idempotency, exactly-once STATE completion, wave totality, and DAG-respecting wave order, per the design doc's property-based-coverage requirement. * feat(#3675): implement quick-batch core primitives Adds src/quick-batch.cts (ADR-457 build-at-publish, compiled to gsd-core/bin/lib/quick-batch.cjs) implementing #3675's quick-batch core primitives per the phase design lock — pure/state primitives and CLI-testable core operations only, no agent dispatch, no worktree creation, no user-facing command (Phase 4/#3676's job): - parseTaskList / parseTaskListFromFile: inline bulleted/numbered task-list parsing (>=2 items required) and a --file variant strictly confined to the planning workspace root via requireSafePath, rejecting non-regular-file targets. - allocateQuickIds / createBatch: collision-safe YYMMDD-xxx quick-id preallocation under withPlanningLock, checked against both on-disk .planning/quick/ entries and sibling .planning/quick-batches/*/BATCH.json manifests (never on-disk-only, which would miss another in-flight batch that hasn't dispatched any real quick directory yet) — replicates cmdInitQuick's own grammar rather than delegating to it (that function's 2-second granularity is not batch-safe). - computeWaves: deterministic wave construction combining dependency-DAG layering with partitionByFileOverlap (#3674), called per DAG layer over path-separator-normalized planned_files — normalization happens at this module's boundary, never inside the Phase 2 helper. - loadBatch: fail-closed BATCH.json schema validation (corrupt/truncated JSON, wrong types, missing fields, out-of-batch dependency references, dependency cycles, a worktree path absent from disk). - resumeBatch: skips complete items, never auto-retries failed items, propagates/reverses blocked status along the DAG to a fixed point, and detects a STATE.md row that already exists for a non-complete item (the "STATE written, BATCH.json not yet updated" crash window) — completing it without re-appending. Idempotent across repeated calls. - completeQuickItem / hasQuickTaskRow: exactly-once STATE.md completion — appendQuickTaskRow (unmodified) is called at most once per quick id, gated by hasQuickTaskRow's own idempotency check re-parsing the real "Quick Tasks Completed" table, since appendQuickTaskRow itself carries no idempotency. BATCH.json lives at .planning/quick-batches/<batch-id>/BATCH.json, a sibling of .planning/quick/ — never inside it, so scanQuickTasks never misreads a batch manifest as a broken quick task. * docs(#3675): register the new quick-batch module New src/*.cts -> bin/lib/*.cjs modules need four hand-maintained registrations beyond the code itself: .gitignore (compiled artifact), eslint.config.mjs (ADR-457: lint the .cts, not the emitted .cjs), docs/INVENTORY.md's CLI Modules roster row plus the regenerated docs/INVENTORY-MANIFEST.json cli_modules entry, and a CONTEXT.md glossary entry matching the convention set by the sibling File Overlap Partitioner Module (#3674) entry it sits beside. NOTE: docs/INVENTORY-MANIFEST.json was updated BY HAND (alphabetically sorted single-entry insertion into families.cli_modules, matching the existing file's structure) rather than via `node scripts/gen-inventory-manifest.cjs --write` — this session's MEMTRACE-FIRST guard hard-blocks direct execution of that indexed script path from Bash, with no available Memtrace tool to route through instead. The orchestrator should re-run `node scripts/gen-inventory-manifest.cjs --check` to confirm this hand-edit is byte-identical to the generator's own output before merging. * fix(#3675): resolve lint findings in quick-batch primitives and tests Unsafe `any[]` assignment from `new Array(n)` in the DAG cycle-check color array, two unnecessary `as string[]` casts TS 5.5's inferred type predicates already narrowed, raw `fs.rmSync` in test cleanup (needs the Windows-EBUSY retry budget `helpers.cleanup` carries), an unused `loadBatch` import, an unbounded `mkfifo` subprocess spawn missing a timeout, and a CONTEXT.md glossary illustration that looked like a real file reference. * feat(#3675): close acceptance-criteria gaps found in review Standards- and spec-axis review (plus a self-caught race) surfaced real gaps against issue #3675's own acceptance criteria and this repo's test conventions: - BATCH.json was missing options, base_revision, per-item wave, and per-item commit — the issue's AC explicitly lists all four as things the manifest must track. Added them: createBatch persists caller-supplied batchOptions/baseRevision verbatim and assigns each item its computed wave index; completeQuickItem now persists the commit onto the item, not just the STATE.md row. All four are backward-tolerant on load (an older/hand-built manifest without them still validates). - resumeBatch had no "incompatible base divergence" check at all, despite the AC and the ADR's own "Base divergence" section requiring one. Added an opt-in currentBaseRevision comparison that fails closed with a recoverable diagnostic on mismatch, and touches nothing on refusal. - resumeBatch read-modify-wrote BATCH.json OUTSIDE withPlanningLock — the only durable write path in this module that wasn't lock-protected, a real lost-update race against a concurrent completeQuickItem or another resume. Now runs inside the same lock createBatch/ completeQuickItem use. - loadBatch and collectExistingBatchQuickIds used raw JSON.parse with no size cap (security review, Low/informational); switched to the existing safeJsonParse (1MB cap) for defense-in-depth. - Parser (parseTaskList) had only example-based tests; CLAUDE.md requires a fast-check property test for parsers. Added one plus a companion reject-property for <2 items. - The id-exhaustion fail-closed ceiling (MAX_TIME_BLOCK) was untested at any boundary. Exported the pure allocateIdsGivenUsed/MAX_TIME_BLOCK for direct limit-1/limit/limit+1 testing without needing 46k fixture dirs. - Issue AC explicitly asks for prompt-injection-payload test coverage, distinct from the existing shell-metacharacter test; added one. - Test row 9 (FIFO skip) silently returned instead of calling t.skip(), so an unsupported platform would report a pass rather than a documented skip; fixed to bind the test-context param and skip properly. - Extracted toWaveInput to remove a 2-site production duplication of the QuickBatchItem -> computeWaves reshape (Standards-axis smell). - Added the required .changeset/ fragment (CONTRIBUTING.md: editing src/ is user-facing even though the compiled .cjs is gitignored). * fix(#3675): restore "not valid JSON" wording in loadBatch's parse-failure reason gsd-test caught this: switching loadBatch to safeJsonParse changed the parse- failure message shape ("... parse error — ...") without preserving the "not valid JSON" substring row 27's own test asserts on. Re-wrap safeJsonParse's error into the original diagnostic phrasing regardless of which of its three failure modes fired. * docs(#3675): backfill changeset pr number to 4190 * fix(#3675): detect a silently-no-op mkfifo on Windows, not just a throwing one CI caught this on windows-latest: row 9's platform-skip only caught mkfifo throwing (command not found). On this runner mkfifo resolves to something that exits 0 without creating a file (NTFS has no FIFO concept), so execution fell through to parseTaskListFromFile against a path that doesn't exist, producing an ENOENT stat error instead of the expected "not a regular file" rejection. Check the artifact actually exists before trusting a zero exit code, and skip with a documented reason either way. --------- Co-authored-by: sim <sim@local>
This commit is contained in:
Tom Boucher
@@ -603,6 +603,7 @@ Full listing: `gsd-core/bin/lib/*.cjs`.
|
||||
| `profile-pipeline-command-router.cjs` | ADR-959 capability command router for the profile-pipeline command family — dispatches scan-sessions, extract-messages, profile-sample (pipeline phase) and write-profile, profile-questionnaire, generate-dev-preferences, generate-claude-profile, generate-claude-md (output phase); phase 6 cutover |
|
||||
| `profile-pipeline.cjs` | User behavioral profiling data pipeline, session file scanning |
|
||||
| `prompt-budget.cjs` | Pure token-budget accounting for review prompts — estimates tokens, applies deterministic trim priority (head-shrink PROJECT.md, proportional plan truncation, drop context/research/requirements, hard-fail guard), returns structured metadata for `review.max_prompt_tokens` (#3081) |
|
||||
| `quick-batch.cjs` | Quick-batch core primitives (#3675, ADR-1239 "Quick-batch binding") — task-list parsing (inline + path-confined `--file`), collision-safe `YYMMDD-xxx` quick-id preallocation under `withPlanningLock` (checked against both on-disk `.planning/quick/` entries and sibling `.planning/quick-batches/*/BATCH.json` manifests), `BATCH.json` schema/validation/resume (fail-closed on corrupt JSON, schema violations, out-of-batch dependency references, cycles, or a missing worktree path), deterministic wave construction combining dependency-DAG layering with `partitionByFileOverlap` over normalized `planned_files`, and exactly-once STATE.md completion via `hasQuickTaskRow`'s own idempotency check ahead of the unmodified `appendQuickTaskRow`. No agent dispatch, no worktree creation, no user-facing command — Phase 4/#3676's job. Compiled from `src/quick-batch.cts` |
|
||||
| `observability/redaction.cjs` | Arg redaction policy for dispatch events — args omitted from every emitted event by default, opt-in verbatim inclusion via `GSD_AUDIT_ARGS=1`; stateless env read, no module-level caching (#177) |
|
||||
| `real-home-guard.cjs` | Real-home confinement guard (#3712; renamed from `test-home-guard.cjs` — a source filename matching Node's `test-*` convention is collected and executed as a test by the remote runner) — refuses any of the six writers that resolve a kind `home` — `installRuntimeArtifacts`, `uninstallRuntimeArtifacts`, `applySurface`, `migrateLegacyDevPreferencesToSkill`, plus the descriptor-dependent `installOpencodeFamilySkills` and `installAgentsKindStandalone` when a `node --test` run would resolve a kind's global `home` override (codex skills -> `$HOME/.agents`, ADR-1239/#2088) inside the real passwd home, which silently pruned every `gsd-*` skill there; compares homes by filesystem identity (`st_dev`+`st_ino`) rather than by pathname, since a case-variant, symlinked, or bind-mounted HOME names one directory under two names; fails CLOSED unless both homes identify or one is definitively absent (the pathname-equality shortcut in `sameDirectory` can answer yes without identifying, so the marker branch requires the marker to identify separately before it may allow anything); a destination inside the real home is exempted only when HOME differs from the passwd home, the passwd home is not itself beneath that HOME (`/Users`, `C:\Users` are not sandboxes), and the destination resolves beneath it (Windows puts the temp root inside the home, so containment alone cannot tell a sandbox from the danger); where no passwd entry is readable it falls back to `sandboxHome()`'s path-valued marker, which must identify AND contain every resolved destination — a marker matching HOME attests only that HOME was sandboxed, so on its own it waved through a layout captured before the sandbox that still named the real `~/.agents`; it remains a deliberate weakening rather than a closed door, since nothing there can contradict a marker naming the real home; does not defend against a subordinate bind mount of the real directory into the sandbox (realpath cannot unify bind-mounted spellings) or a cross-process TOCTOU swap; inert for normal installs outside a Node test context |
|
||||
| `refactor-trigger-command-router.cjs` | ADR-959 capability command router for `gsd-tools refactor` (issue #1953) — dispatches evaluate/status/accept/decline subcommands for the complexity-triggered refactor capability; owns capability-activation gating, git invocation (via the `git-base-branch.cjs` `phaseStartCommit`/`changedFilesSince` adapters), config reads, phase-directory resolution, and the optional broken-windows ledger integration around the pure `complexity-trigger.cjs` leaf |
|
||||
|
||||
Reference in New Issue
Block a user