Wanneer een TASK_IMPLEMENTATION-job FAILED wordt, cancelt
cancelPbiOnFailure alle queued/claimed/running siblings binnen
dezelfde PBI (over alle stories heen) en draait gepushte commits
ongedaan:
- Open PR → gh pr close --delete-branch (PR-close + remote-branch-
delete in één).
- Gemergde PR → revert-PR via git revert -m 1 <mergeSha> in een
korte worktree, gepusht naar revert/<orig>-<jobid>, gh pr create
zonder auto-merge (mens reviewed).
- Branch zonder PR → best-effort git push origin --delete.
Race-protectie: update_job_status weigert nu een statuswijziging op
een job die al CANCELLED is met een specifieke JOB_CANCELLED-error,
zodat een parallelle worker zijn lokale werk weggooit ipv een DONE
te forceren. Idempotent — een tweede cascade voor dezelfde PBI is
een no-op. Non-blocking — alle fouten worden warnings in de trace
op de oorspronkelijke failed job zijn error-veld; cascade throwt
nooit naar de caller.
Niet in scope: per-product opt-out, sprint-niveau cascade,
idea-job cascade.
11 nieuwe vitest-cases dekken DB-cascade, branch-grouping, open/
merged/no-PR paden, repo-root-mismatch en de never-throws-garantie.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
update_job_status accepts optionele model_id + 4 token-velden conform het
runbook-contract (mcp-integration.md:42). De waarden komen niet van de agent
zelf maar van scripts/persist-job-usage.ts, een PostToolUse-hook die het
lokale Claude Code transcript (~/.claude/projects/.../*.jsonl) leest en de
usage tussen de laatste wait_for_job en update_job_status optelt.
Geen Anthropic API-key nodig — alle data staat al lokaal op disk omdat
Claude Code per assistant-message het API usage-blok logt
(input_tokens, output_tokens, cache_creation_input_tokens,
cache_read_input_tokens + message.model).
Robustness:
- Subagent (isSidechain: true) lines worden geskipt om double-counting
te voorkomen tegen subagents/-subdirectory transcripts.
- Lines worden gededupliceerd op uuid (branching/resumption).
- model_id wordt genormaliseerd: claude-opus-4-7[1m] -> claude-opus-4-7-1m
zodat de [1m]-variant op een aparte model_prices-rij kan matchen.
- Hook is non-blocking: elke fout logt een warning en exit 0.
Hook-config in .claude/settings.json met SCRUM4ME_MCP_DIR-fallback zodat
de agent vanuit een product-worktree (andere cwd) ook werkt mits de user
de hook in ~/.claude/settings.json kopieert.
16 nieuwe vitest-cases voor parseTranscript, computeUsageFromTranscript,
normalizeModelId en persistJobUsage.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
T-505 in v0.6.0 wired the idea-failure side-effects but missed the
'skip verify-gate for IDEA_*-kinds on done' branch from the M12 plan.
Reproduced live on IDEA-002: agent answered 5 questions, called
update_idea_grill_md (status → GRILLED, grill_md persisted), but
update_job_status('done') was rejected by the verify-gate because
idea-jobs have no task → no plan_snapshot → verify_task_against_plan
cannot run. Job got marked FAILED + idea reverted to GRILL_FAILED
even though the grill itself succeeded.
Fix: in update_job_status, when status='done' AND kind in
[IDEA_GRILL, IDEA_MAKE_PLAN]: skip checkVerifyGate AND
prepareDoneUpdate (no git push, no branch). The idea-status was
already moved to GRILLED/PLAN_READY by update_idea_*_md; the job
just needs to flip to DONE.
Tests: 153/153 still green.
Bump 0.6.1 → 0.6.2.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds the 4 new MCP-tools for the Scrum4Me M12 Idea-entity flow + extends
3 existing tools to handle the new ClaudeJobKind discriminator.
New tools:
- get_idea_context: full idea + product + open questions + recent logs
- update_idea_grill_md: save grill-result + status → GRILLED + IdeaLog
- update_idea_plan_md: server-side yaml parser validates frontmatter;
ok → PLAN_READY, fail → PLAN_FAILED + line-info errors
- log_idea_decision: DECISION/NOTE entries on the timeline
Extended tools:
- ask_user_question: xor schema (story_id | idea_id); idea-questions are
user-private with productId derived from idea.product_id
- wait_for_job: returns \`kind\` discriminator; IDEA_* payloads include
idea + prompt_text (from src/prompts/idea/) and skip worktree creation
- update_job_status: failed on IDEA_* auto-transitions idea-status to
GRILL_FAILED / PLAN_FAILED + IdeaLog{JOB_EVENT}; auto-PR + worktree-
cleanup skipped for idea-jobs
Other changes:
- Health version now read dynamically from package.json (was hardcoded
'0.1.0' which caused deploy-sync confusion)
- Schema synced to Scrum4Me M12 (Idea + IdeaLog + enums + ClaudeJob/
Question nullable-FKs + check-constraints + pg_notify-trigger update)
- New @scrum4me-mcp/lib/idea-plan-parser duplicates Scrum4Me's parser
(drift detected by vendor schema-watchdog)
- Embedded grill+make-plan prompts copied to src/prompts/idea/
- New userOwnsIdea access helper
Tests: 153/153 green; tsc + build clean.
Migration: requires Scrum4Me M12 migration (20260504172747_add_ideas_and_grill_jobs)
applied on the target DB. See vendor/scrum4me/docs/runbooks/mcp-integration.md
for the updated batch-loop with kind-switch.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Sluit story 'Verify-gate uitbreiden' in PBI 'Agent verify-flow hardening' af.
The previous gate weighed only EMPTY: any PARTIAL or DIVERGENT verify
slipped through. The Insights batch (2 May 2026) showed why that's
weak — agent-jobs claiming DONE while only delivering helpers, not
the requested UI components, with verify=DIVERGENT/PARTIAL accepted.
New decision matrix:
null → reject (run verify_task_against_plan)
EMPTY + !verify_only → reject
EMPTY + verify_only → allowed
ALIGNED → always allowed
PARTIAL/DIVERGENT
required=ALIGNED → reject (strict task)
required=ALIGNED_OR_PARTIAL (default) → allowed only if summary
≥20 chars (acknowledge drift)
required=ANY → allowed (refactor escape hatch)
`update_job_status('done')` now reads `task.verify_required` from the DB
(field added in Scrum4Me PR #53) and passes it + `summary` to the gate.
Tool description updated with the new rules.
Vendor submodule synced to pick up the schema enum.
Tests: 129/129 (was 120 + 9 new combinatorial gate tests).
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Implementeert vier open stories uit PBI 'Veilige Claude-agent-workflow':
**Branch per story (cmon11tbe001zbortx35n155c)**
- `resolveBranchForJob`: zoek sibling-job in dezelfde story; reuse z'n
branch (1 PR per story i.p.v. per task).
- Branch-naam: `feat/story-<8-char>` voor nieuwe stories.
- `createWorktreeForJob` kent nu `reuseBranch=true`: detecteert stale
sibling-worktree die de branch nog vasthoudt en verwijdert die eerst.
- `attachWorktreeToJob` neemt `storyId` mee.
**PR-hergebruik (zelfde story)**
- `maybeCreateAutoPr`: als sibling-job in story al een pr_url heeft,
hergebruik die zonder nieuwe `gh pr create`-call. PR-titel komt nu
van de story (was task) zodat het als 'story-PR' leest.
**Worktree-cleanup uitgesteld bij actieve siblings**
- `cleanupWorktreeForTerminalStatus`: count active sibling-jobs in
dezelfde story; defer als > 0 (volgende sub-task gebruikt branch).
**Worktree-cleanup logging (cmon0jc14001ubortjxf2a2ck)**
- Warning bij ontbrekende repoRoot, met productId + jobId in message.
- Warning bij removeWorktreeForJob-failure met keepBranch in message.
**resolveRepoRoot fallback (cmon0jc14001ubortjxf2a2ck)**
- Convention-based fallback: `~/Projects/<repo-name>` afgeleid uit
`product.repo_url` als noch env-var noch config-bestand iets oplevert.
- `repoNameFromUrl` helper geëxporteerd voor herbruikbaarheid.
**Verify EMPTY-detection edge-case (cmon0kdq6001xbort2kgbcqmr)**
- `classifyDiffAgainstPlan`: na file-paths-check ook content-lines
checken; als alle +/- regels alleen headers of whitespace zijn,
return EMPTY met duidelijke reasoning.
Tests: 120/120 groen (3 nieuwe), tsc clean, build clean.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
New src/git/pr.ts helper wraps 'gh pr create'; returns { url } or { error }.
maybeCreateAutoPr() in update-job-status checks product.auto_pr, builds title
from story.code + task.title, writes pr_url to DB. Non-fatal: gh failure logs
a warning and leaves DONE status intact. Also syncs schema: auto_pr on Product,
pr_url on ClaudeJob.
On DONE/FAILED, resolves repoRoot and calls removeWorktreeForJob (best-effort).
keepBranch=true when status=done and agent reported a branch (push assumed);
false otherwise. Cleanup failures are logged as warnings — DB status is preserved.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- wait_for_job: blocks ≤600s, claims QUEUED job atomically via FOR UPDATE
SKIP LOCKED, resets stale CLAIMED jobs (>30min), registers ClaudeWorker
presence with heartbeat, emits worker_connected/disconnected via NOTIFY
- update_job_status: agent reports running|done|failed, validates token
ownership (claimed_by_token_id), emits claude_job_status via NOTIFY
- auth.ts extended with tokenId so tools can set claimed_by_token_id
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
