@bonesofspring/ai-rules 0.2.19 → 0.2.21

package/presets/claude/java/hooks/chain-team-phases.sh

@@ -134,10 +134,17 @@ const DEFAULT_ARTIFACT_OUTPUT_IDS = {
   'build-verifier': 'validation',
   'code-reviewer': 'review',
   'security-reviewer': 'security-review',
-  'accessibility-reviewer': 'a11y-review',
   'performance-auditor': 'perf-report',
   'api-contract-reviewer': 'api-contract-review',
   'tech-writer': 'documentation',
+  'openspec-author': 'openspec-change',
+  'qa-tester': 'qa',
+  'unit-test-planner': 'unit-test-plan',
+  'unit-test-generator': 'unit-test-spec',
+  'unit-test-healer': 'unit-test-heal',
+  'integration-test-planner': 'integration-test-plan',
+  'integration-test-generator': 'integration-test-spec',
+  'integration-test-healer': 'integration-test-heal',
 };
 function artifactOutputIdFor(step, agent) {
   if (step && typeof step.artifactOutputId === 'string') return step.artifactOutputId;
@@ -242,7 +249,17 @@ function recordSkippedStep(step, stepIndex) {
 function writeStatus(next) {
   status = { ...status, ...next, slug, updatedAt: now };
   const temporary = `${statusPath}.${process.pid}.${Date.now()}.tmp`;
-  fs.writeFileSync(temporary, JSON.stringify(status, null, 2) + '\n');
+  // Open + write + fsync + rename so the durable file lands on disk before
+  // the orchestrator reads it. Closes the race window where a subagent that
+  // completed milliseconds before writeStatus() could otherwise invoke the
+  // next step on a stale status (in particular the human-gate write).
+  const fd = fs.openSync(temporary, 'w');
+  try {
+    fs.writeSync(fd, JSON.stringify(status, null, 2) + '\n');
+    try { fs.fsyncSync(fd); } catch {}
+  } finally {
+    fs.closeSync(fd);
+  }
   fs.renameSync(temporary, statusPath);
 }
 
@@ -337,6 +354,12 @@ const AGENT_HANDOFF = {
     `Follow tests-integration.mdc, integration-testing, and debug-investigation. Do not use browser e2e tooling. Upsert the owned terminal receipt in artifact-manifest.json; do not mutate status.json lifecycle fields.`,
   'tech-writer': (s) =>
     `Invoke tech-writer for slug "${s}". Write documentation.md from task artifacts. Upsert the owned terminal receipt in artifact-manifest.json; do not mutate status.json lifecycle fields.`,
+  'openspec-author': (s, scope) =>
+    `Invoke openspec-author for slug "${s}"${scope ? ` (scope: ${scope})` : ''}. ` +
+    `Read openspec/config.yaml schema + rules; if openspec/** is absent, emit blocked ("openspec/** not initialized"). ` +
+    `Author openspec/changes/<change-id>/{proposal.md, design.md, tasks.md, .openspec.yaml} and the capability spec delta per phase (explore|propose|apply|archive). ` +
+    `Run \`openspec validate <change-id> --type change --strict\` and \`openspec validate --specs --strict\` (Phase 3); surface FAIL output verbatim. ` +
+    `Upsert the owned terminal receipt (artifactOutputId \`openspec-change\`) in artifact-manifest.json with validator exit code + evidence paths; do not mutate status.json lifecycle fields. `,
 };
 
 const REVIEW_AGENTS = new Set([
@@ -510,9 +533,37 @@ if (!fs.existsSync(pipelinePath)) {
 }
 
 let pipeline;
+let pipelineParseError = null;
 try {
   pipeline = JSON.parse(fs.readFileSync(pipelinePath, 'utf8'));
-} catch {
+} catch (error) {
+  pipelineParseError = error;
+}
+// Modern pipelines carry an `intent` field. If pipeline.json is present but
+// unparseable, surface the error so the orchestrator can stop and report
+// instead of falling back to legacyFlow() — which silently drops modern
+// roles like `openspec-author`, `qa-tester`, and the test-planner trio.
+if (pipelineParseError) {
+  if (fs.existsSync(pipelinePath)) {
+    const raw = fs.readFileSync(pipelinePath, 'utf8');
+    const hasModernIntent = /"intent"\s*:/.test(raw);
+    appendMetric({
+      eventId: `${status.metricsRunId || slug}:pipeline_parse_error:${Date.now()}`,
+      type: 'pipeline_parse_error',
+      at: now,
+      source: 'hook',
+      attributionConfidence: 'high',
+      error: String(pipelineParseError.message || pipelineParseError),
+      hasModernIntent,
+    });
+    if (hasModernIntent) {
+      writeStatus({ state: 'blocked', phase: 'blocked', currentAgent: 'task-router' });
+      process.stdout.write(JSON.stringify({
+        followup_message: `pipeline.json is malformed for slug "${slug}" — JSON parse error: ${pipelineParseError.message}. Re-run the task-router to regenerate pipeline.json, then resume via /task-continue ${slug}.`,
+      }));
+      process.exit(0);
+    }
+  }
   legacyFlow();
 }
 

package/presets/claude/java/rules/README.md

@@ -30,6 +30,7 @@ Claude topic `.md` files are **derived** from Cursor SoT `presets/cursor/java/ru
 | `architecture-boundaries.mdc` | `architecture/boundaries.md` |
 | `anti-sycophancy-discipline.mdc` | `tooling-and-review/anti-sycophancy-discipline.md` |
 | `code-quality-and-refactoring.mdc` | `tooling-and-review/code-quality.md` |
+| `ai-local-overlay.mdc` | `tooling-and-review/ai-local-overlay.md` |
 | `design-guidance.mdc` | `tooling-and-review/design-guidance.md` |
 | `mcp-usage.mdc` | `tooling-and-review/mcp-usage.md` |
 | `code-review-mr.mdc` | `tooling-and-review/code-review.md` |

package/presets/claude/java/rules/tooling-and-review/agent-team-orchestrator.md

@@ -23,8 +23,6 @@ Work request without `/task` — see **agent-team-intake** (when to suggest the
 |---------|------|
 | **`/task <desc>`** | **Preferred** — router → dynamic pipeline → first agent |
 | `/task-continue <slug>` | After human gate or pause |
-| `/feature-start <desc>` | Legacy: analyst-only start (no router) |
-| `/feature-continue <slug>` | Alias of task-continue |
 | `/spec-start [desc]` | Generate/bootstrap `docs/specs` — router intent `spec-gen` → analyst author mode → human gate |
 | `/technical-retro [slug]` | Retro with agent team block |
 
@@ -163,7 +161,7 @@ After each agent completes, verify its terminal manifest receipt, then let the o
 
 ## Legacy mode
 
-If `pipeline.json` is missing (old `/feature-start` tasks), fall back to fixed phases: analysis → development → review → testing. Hook supports both.
+If `pipeline.json` is missing, fall back to fixed phases: analysis → development → review → testing. Hook supports both.
 
 ## Skip pipeline when
 
@@ -208,4 +206,4 @@ When the work edits `packages/ai-rules/presets/**` (or dogfood `.cursor/` sync o
 - Chain hook handoffs: `integration-test-*`, `post-change-test`, `java-tooling` — never `lint:js`, Playwright, XCUITest, or `xcodebuild`.
 - Skills: `feature-delivery`, `unit-testing`, `integration-testing`, `ci-investigation`, `debug-investigation`, `code-review`, `technical-retro`, `write-adr`.
 
-**Product specs:** When `/task`, `/feature-start`, or `/spec-start` touches pages/features/`docs/specs`: if INDEX exists and no generate intent → instruct next agent to load requestable **`product-specs`** (consume; soft-fail; never invent). If generate/bootstrap/`spec-gen` → load **`product-specs-authoring`** + skill **`spec-authoring`** inside **task-analyst** (not a separate pipeline step); pass created spec paths + INDEX diff in handoff; human gate must approve before feature-developer. No-op if INDEX absent and no author intent.
+**Product specs:** When `/task` or `/spec-start` touches pages/features/`docs/specs`: if INDEX exists and no generate intent → instruct next agent to load requestable **`product-specs`** (consume; soft-fail; never invent). If generate/bootstrap/`spec-gen` → load **`product-specs-authoring`** + skill **`spec-authoring`** inside **task-analyst** (not a separate pipeline step); pass created spec paths + INDEX diff in handoff; human gate must approve before feature-developer. No-op if INDEX absent and no author intent.

package/presets/claude/java/rules/tooling-and-review/ai-local-overlay.md

@@ -0,0 +1,48 @@
+---
+description: Local overlay from .ai.local/ (consumer-owned rules with priority over preset rules) — MANDATORY
+paths:
+  - "**/.ai.local/**"
+  - "**/AGENTS.md"
+  - "**/CLAUDE.md"
+alwaysApply: false
+impact: CRITICAL
+---
+
+# Local overlay (`.ai.local/`) — MANDATORY
+
+This rule is mandatory. **You MUST** obey it on every turn, without exception.
+
+## On every turn
+
+1. **Probe** — before answering, run `ls -la .ai.local` from the repository root.
+2. **Walk** — if `.ai.local/` exists at multiple levels (monorepo), read every level from root to leaf. **Deeper wins** in practice: more specific rules appear later in the assembled context and take precedence.
+3. **Apply** — treat every `.ai.local/rules/*.md` and `.ai.local/conventions.md` as a higher-priority rule for this repo. The block below ("Authoring `.ai.local/rules/*.md`") defines the conventions.
+4. **Conflict** — if a local rule conflicts with a preset rule, **the local rule wins**. If a local rule body contains `override: kill <preset-rule-name>`, drop that preset rule from your context window entirely.
+5. **Acknowledge** — when answering a user question, briefly state which `.ai.local/` rules apply to the current task (one line is enough). Format: `Using .ai.local: <rule-name>(, <rule-name>)*`. If no `.ai.local/` exists, do NOT mention this rule.
+
+If `.ai.local/` does not exist → **no-op**, do not error. Do not produce any log line about its absence.
+
+## Authoring `.ai.local/rules/*.md` (consumer-side reference)
+
+Each `.ai.local/rules/<name>.md` is a plain-markdown rule body. The model parses it as:
+
+- Lines before `## Body` are scope metadata. Only `glob:` / `globs:` / `override:` lines are recognised.
+- Everything from `## Body` onward is the rule body.
+- `glob: always` → always-on. Counts against your context budget.
+- `globs: <comma-separated globs>` → on-demand; applies only when matching files are in focus.
+- `override: kill <preset-rule-name>` → drop that preset rule from context.
+
+Keep files ≤ 120 lines. Match the style of preset rules (short sections, bullets, code examples).
+
+## Limitations (read this)
+
+- Cursor/Claude do **not** natively read `.ai.local/`. This rule is a **prompt-only** mechanism: it asks you, the model, to read and apply the files. Reliability is high but not 100%.
+- If a user adds a file to `.ai.local/` mid-session, **restart the session** to be sure it is picked up.
+- "Priority" is **contextual**, not enforced. Always re-read the local rule before answering questions that touch it.
+- This rule is **not** session-start (Claude convention); it loads on touch of `.ai.local/**` or `AGENTS.md`/`CLAUDE.md`. The "every turn" mandate is enforced by the body text, not by Cursor/Claude loading semantics.
+
+## Pointers
+
+- Authoring guide: `docs/LOCAL-OVERLAY.md` in the `@bonesofspring/ai-rules` package.
+- Living gotchas alternative: `.claude/team/conventions.md`.
+- Root `AGENTS.md` reinforces this rule with a `MANDATORY` section.

package/presets/claude/java/rules/tooling-and-review/preset-pr-checklist.md

@@ -19,11 +19,15 @@ Checklist for PRs that touch `packages/ai-rules/presets/**` (and dogfood `.curso
 - [ ] **Core:** if `_shared/core/**` changed — all `<!-- shared-core: … -->` consumers updated (or documented intentional fork)
 - [ ] **Twin:** Cursor change → Claude twin **same PR** (**FAIL** if missing on next, nuxt, svelte, ios-swift, android-kotlin, go, java, mcp-ts, php-hexagonal, and php-laravel — see `preset-twin-sync`)
 - [ ] **Mapping:** `claude/<stack>/rules/README.md` mapping table is current
-- [ ] **Token budget:** alwaysApply / session-start ≤ 3; meta rules not always-on
+- [ ] **Token budget:** Cursor `alwaysApply` ≤ 4 (4th slot reserved for `ai-local-overlay`); Claude session-start exactly 3; meta rules not always-on
+- [ ] **Local overlay coverage:** every installable stack ships `ai-local-overlay.mdc` (Cursor) and `ai-local-overlay.md` (Claude) with `MANDATORY` + acknowledgement format + `docs/LOCAL-OVERLAY.md` pointer; Claude twin listed in `claude/<stack>/rules/README.md` mapping table. Run `packages/ai-rules/scripts/check-local-overlay.sh`.
 - [ ] **Leakage:** no foreign stack agent stems / handoff strings (`preset-no-cross-stack-leakage`)
+- [ ] **Hook-only lifecycle:** `status.json` / `metrics.json` writes happen only inside `hooks/`; agents read but never rewrite lifecycle fields. Run `packages/ai-rules/scripts/check-hook-only-lifecycle.sh`.
+- [ ] **Chain fork:** the byte-wise MD5 uniqueness of `chain-team-phases.sh` across the 10 stacks × 2 platforms (20 hooks) holds. Run `packages/ai-rules/scripts/check-chain-hook-coverage.sh` and `packages/ai-rules/scripts/test-chain-team-phases-coverage.mjs`.
 - [ ] **Twin depth:** domain ≥15 lines; thin aliases documented
 - [ ] **README:** stack `rules/README.md` core→stack mapping / catalog updated if stems added
 - [ ] **CHANGELOG / package README:** note when public preset paths or install surface change
+- [ ] **i18n anchors:** kept-marked Cyrillic fragments in `tests-unit.md(c)` stay consistent with their English counterparts. Run `packages/ai-rules/scripts/check-preset-i18n.sh`.
 - [ ] **Dogfood:** run `packages/ai-rules/scripts/sync-dogfood-cursor.sh` (or document intentional skip) when maintaining this monorepo
 - [ ] **shared-core drift:** for embed-mode stems, run `packages/ai-rules/scripts/check-shared-core-drift.sh` (or document intentional fork)
 

package/presets/claude/java/rules/tooling-and-review/preset-token-budget.md

@@ -17,12 +17,16 @@ Invariants for rule loading in an **installed** preset (consumer app).
 
 | Platform | Always-on | Cap |
 |----------|-----------|-----|
-| Cursor | `alwaysApply: true` | **≤ 3** |
+| Cursor | `alwaysApply: true` | **≤ 4** |
 | Claude | rules **without** `paths:` | **exactly 3** |
 
+**Cursor 4th slot is reserved for `ai-local-overlay`** (consumer-owned `.ai.local/` rules). If you do not need local overlay, demote that rule to `globs: ["**/.ai.local/**"]` and revert to ≤3.
+
 **Claude Option A (README exclusion):** when counting Claude rules without `paths:` / session-start, **exclude `**/README.md`** (topic indexes are not session-start rules). The intended session-start trio remains **exactly 3**; body budget sums **only those 3** rules.
 
-Session-start body budget (where documented): ≤140 lines total (stretch ≤120).
+Session-start body budget:
+- Claude trio: ≤140 lines total (stretch ≤120).
+- Cursor 4-set: ≤170 lines total (4 rules including `ai-local-overlay`).
 
 ## Meta / author rules
 
@@ -38,4 +42,4 @@ Orchestrator, feature-delivery, code-review, post-change gates, domain globs —
 
 ## Verifier
 
-build-verifier **FAIL** if alwaysApply / session-start count exceeds 3 (Claude count **excludes `**/README.md`**), or if a `preset-*` meta rule is marked always-on.
+build-verifier **FAIL** if Cursor `alwaysApply` count exceeds 4, Claude session-start count exceeds 3 (Claude count **excludes `**/README.md`**), or if a `preset-*` meta rule is marked always-on.

package/presets/claude/java/rules/tooling-and-review/preset-twin-sync.md

@@ -39,6 +39,8 @@ Cursor `.mdc` — **SoT per stack**. Claude topic `.md` — derived twin.
 
 - **Domain** rules: ≥15 body lines (FAIL if thin without documented alias reason).
 - **Thin aliases** (`agent-team-intake`, `technical-retro`): may be shorter (soft) if they only point to a command/skill.
+- **Slim-companion rules** (UI-edit bundle, slim app-cores, slim tooling): may be shorter than 15 lines **if and only if** the body is a pointer into shared-core / stack-prose (no standalone prose). Examples live in stack `rules/README.md` (e.g. `cursor/next/rules/README.md` "UI-edit bundle" intent).
+  - **Mechanical enforcement is pending.** Today build-verifier excludes only `agent-team-intake` / `technical-retro` / `preset-layering`; slim-companion rules currently pass under the same human-judgement path as thin aliases. Follow-up: add `allow_slim_stems` to `packages/ai-rules/scripts/check-preset-token-budget.sh` analogous to `trio_stems_allowlist`.
 
 ## Severity (build-verifier)
 

package/presets/claude/java/skills/code-review/SKILL.md

@@ -20,6 +20,12 @@ description: Reviews pull requests and branch diffs for Java hexagonal services
 
 Never require iOS/Next-only gates (XCUITest, Playwright, `lint:js`).
 
+## Receiving review feedback
+
+<!-- shared-core: review/receiving-code-review.md -->
+
+When **implementing** review comments (not when issuing them): verify each item against the codebase before changing code; do not perform agreement (`You're absolutely right!`). Unclear items → ask, do not guess.
+
 ## Design guidance
 
 - Load `design-guidance` (rule stem) when assessing structure, smells, or pattern fit.

package/presets/claude/java/skills/debug-investigation/SKILL.md

@@ -3,21 +3,49 @@ name: debug-investigation
 description: Investigates bugs, failing tests, and unexpected behavior in Java services with runtime evidence before applying minimal fixes. Use for bugfix tasks, debugger agent runs, or regressions.
 ---
 
-# Debug investigation (Java)
+<!-- shared-core: quality/debug-investigation.md -->
+
+# Debug investigation
+
+Stack-agnostic root-cause loop. Consumed by skill `debug-investigation` and agent `debugger`. Platform paths, MCP, and stack commands belong in adapter **Stack notes**.
+
+## When
+
+Any bug, test failure, build failure, or unexpected behavior — **before** proposing a fix.
+
+## Phases (in order)
+
+1. **Evidence** — reproduce, or collect the closest evidence (error output, failing test, logs, user steps).
+2. **Isolate** — expected vs actual in one sentence; smallest relevant slice of the codebase.
+3. **Root cause** — name the cause before editing. One hypothesis at a time; do not stack speculative fixes.
+4. **Minimal fix + verify** — smallest change that preserves architecture and public contracts. Add a regression test when the behavior is user-facing or likely to recur. Re-run the **failing** check first, then the relevant lint/type/test slice.
+
+## 3-fix architecture gate
+
+If **three** fix attempts fail, or the same symptom returns after two supposed fixes: **STOP**. Do not attempt fix #4. Escalate — the design is likely wrong (human partner or `solution-architect`). Chasing symptoms is the failure mode.
+
+## Do not
+
+- Broad refactor during a bugfix.
+- Fix only where the error surfaces if invalid data originates elsewhere.
+- Claim `fixApplied: true` without re-running the failing check (see `quality/verification-before-completion.md`).
+
+## Stack notes
+
+### Stack process
 
 1. Capture failing `mvn test` / `./gradlew test` / logs / stack traces.
 2. Hypothesize layer (domain vs adapter vs wiring).
 3. Minimal fix + regression test when appropriate.
 4. Document in `debug-report.md`.
 
-
-## MCP usage
+### MCP usage
 
 Load requestable rule `mcp-usage` when verifying third-party library APIs (Context7) or when a local overlay MCP is present and the task fits. Discover tools at runtime; do not invent schemas. Overlay absence is not a blocker unless the user required that MCP as source of truth.
 
 If a Sentry/Datadog (or similar) MCP is ready and the task is a production error, event, or trace, use it. Do not invent telemetry.
 
 
-## Design guidance
+### Design guidance
 
 - Load `design-guidance` (rule stem) when assessing structure, smells, or pattern fit.

package/presets/claude/mcp-ts/agents/build-verifier.md

@@ -18,6 +18,12 @@ You are the build verifier for **TypeScript MCP servers** (and for **ai-rules pr
 
 **Never** require Playwright, XCUITest, `lint:js`/`lint:css`, `xcodebuild`/`SwiftLint`, or Go tools (`go test` / `go vet` / `golangci-lint`) as the MCP gate.
 
+## Verification before completion
+
+<!-- shared-core: quality/verification-before-completion.md -->
+
+Before writing `validation-report.md` or claiming PASS: IDENTIFY the commands, RUN them, READ exit codes, VERIFY they match the verdict, THEN claim. Do not mark PASS from memory of a prior developer run.
+
 ## Preset-structure validation (when editing `packages/ai-rules/presets/**`)
 
 From `packages/ai-rules/`:

package/presets/claude/mcp-ts/agents/feature-developer.md

@@ -43,6 +43,13 @@ Implement MCP server features following **mcp-ts** boundaries.
 ## Design guidance / anti-sycophancy
 
 - Load `design-guidance` and `anti-sycophancy-discipline` when writing or changing code.
+- When acting on review feedback (`changes_requested` / human comments): verify each item against the codebase before implementing; no performative agreement.
+
+## Verification before completion
+
+<!-- shared-core: quality/verification-before-completion.md -->
+
+Before handoff, IDENTIFY→RUN→READ→VERIFY the scoped checks for this WP, THEN claim. Do not say the work is done without that evidence in the same turn. Full-project lint is `build-verifier`.
 
 
 ## Product specs (docs/specs)

package/presets/claude/mcp-ts/agents/task-analyst.md

@@ -1,11 +1,21 @@
 ---
 name: task-analyst
-description: Task analysis specialist. Clarifies requirements, validates acceptance criteria, and decomposes work into concrete tasks. Use for /feature-start, task intake, or when the user asks to analyze or decompose a feature before coding. Writes only team artifacts; never writes production code.
+description: Task analysis specialist. Clarifies requirements, validates acceptance criteria, and decomposes work into concrete tasks. Use for task intake, or when the user asks to analyze or decompose a feature before coding. Writes only team artifacts; never writes production code.
 readonly: false
 ---
 
 Analyze MCP server work requests for the **mcp-ts** preset.
 
+## Classify first
+
+<!-- shared-core: agent-team/task-classification.md -->
+
+Before decomposing, name the class: **spike** / **bounded** / **architectural**. Maps to existing pipeline profiles — do not invent a second orchestrator.
+
+- **Spike** — research DoR; no implement tasks.
+- **Bounded** — few tasks, only layers actually touched; do not pad types/api/store/ui.
+- **Architectural** — full layer table; architect when boundaries are unclear.
+
 ## Deliverables
 
 1. `brief.md` — goal, AC, out of scope, assumptions, risks.

package/presets/claude/mcp-ts/agents/task-router.md

@@ -46,6 +46,20 @@ Append **`security-reviewer` after `build-verifier`** for `feature` / `bugfix` /
 
 **Never skip `build-verifier`** after developer when TypeScript sources changed.
 
+## Task class (spike / bounded / architectural)
+
+<!-- shared-core: agent-team/task-classification.md -->
+
+Name the class in `routingReasons`. Maps to profiles already in this file — do not add a parallel step table.
+
+| Class | Use existing |
+|-------|----------------|
+| Spike | intent `spike` |
+| Bounded | `profile: light` (or skip analyst when AC is complete) |
+| Architectural | `profile: full` + `solution-architect` when boundaries unclear |
+
+SoT: `agent-team/task-classification` (shared-core marker above; do not invent a second pipeline).
+
 ## Routing decision rubric
 
 Keep top-level `profile` and emit this top-level object in every new `pipeline.json`:

package/presets/claude/mcp-ts/commands/README.md

@@ -6,6 +6,4 @@ Slash commands are copied to `.cursor/commands/` (Cursor) or `.claude/commands/`
 |---------|---------|
 | `/task` | Router + pipeline entry |
 | `/task-continue` | Resume after human gate |
-| `/feature-start` | Legacy analyst-only start |
-| `/feature-continue` | Alias of task-continue |
 | `/technical-retro` | Retro |

package/presets/claude/mcp-ts/commands/task-continue.md

@@ -5,7 +5,7 @@ description: Continue an agent-team task after a human gate, pause, or review ch
 
 # Task continue — after human gate or pause
 
-Resume the pipeline from `pipeline.json`. Semantic alias: `/feature-continue`.
+Resume the pipeline from `pipeline.json`.
 
 ## Arguments
 

package/presets/claude/mcp-ts/commands/task.md

@@ -33,7 +33,7 @@ Examples:
    - **do not** write lifecycle fields in `status.json` / `metrics.json` yourself — orchestration runtime/hook owns them (see **agent-artifact-contracts**). If `status.json` is missing, only minimal bootstrap is allowed (`slug`, `intent`, `pipelineIndex: 0`, `currentAgent: steps[0]`, `state: in_progress`) — **do not** increment `pipelineIndex` manually;
    - invoke subagent for `pipeline.steps[0].agent` (slug, artifact paths, scope).
 6. **After each agent** — wait for hook handoff / reconciliation terminal receipt → `status.state`; do not start the next step until `status.json` reflects handoff (or until hook returns `followup_message`).
-7. **Human gate** — if there is a gate in `humanGates` after the current step (`after:<agent>`), **stop** after the agent finishes and ask for `/task-continue <slug>` (or `/feature-continue <slug>`).
+7. **Human gate** — if there is a gate in `humanGates` after the current step (`after:<agent>`), **stop** after the agent finishes and ask for `/task-continue <slug>`.
 
 ## Do not
 
@@ -46,9 +46,6 @@ Examples:
 - `awaitingHumanGate` / `state: awaiting_approval` → offer artifact review or `/task-continue <slug>`.
 - `in_progress` → ask: continue or new slug.
 
-## Legacy
-
-`/feature-start` and `/feature-continue` remain compatible; for new tasks prefer `/task`.
 
 ## Metrics
 

package/presets/claude/mcp-ts/commands/technical-retro.md

@@ -10,7 +10,7 @@ The agent acts as a **neutral facilitator of a technical retrospective**, not as
 ## When to use
 
 - User asks: "retro", "technical retro", "sprint/iteration review", "action items after release".
-- After completing the agent-team pipeline (`/feature-start` → `/feature-continue`).
+- After completing the agent-team pipeline.
 - Context available: period, topic (release, incident), task slug, notes.
 
 ## Inputs (gather; do not invent)
@@ -53,7 +53,7 @@ Do not attribute facts not present in artifacts, diff, or the user's message.
 
 ## "Agent work" block (required when slug present)
 
-If retro is tied to a `/feature-start` pipeline:
+If retro is tied to an agent-team pipeline:
 
 | Question | What to capture |
 |----------|-----------------|

package/presets/claude/mcp-ts/hooks/chain-team-phases.sh

@@ -128,17 +128,10 @@ const DEFAULT_ARTIFACT_OUTPUT_IDS = {
   'task-router': 'routing',
   'task-analyst': 'analysis',
   'solution-architect': 'architecture',
-  'migration-specialist': 'migration-plan',
   debugger: 'debug',
-  'ci-investigator': 'ci',
   'feature-developer': 'implementation',
   'build-verifier': 'validation',
-  'code-reviewer': 'review',
   'security-reviewer': 'security-review',
-  'accessibility-reviewer': 'a11y-review',
-  'performance-auditor': 'perf-report',
-  'api-contract-reviewer': 'api-contract-review',
-  'tech-writer': 'documentation',
 };
 function artifactOutputIdFor(step, agent) {
   if (step && typeof step.artifactOutputId === 'string') return step.artifactOutputId;
@@ -243,7 +236,17 @@ function recordSkippedStep(step, stepIndex) {
 function writeStatus(next) {
   status = { ...status, ...next, slug, updatedAt: now };
   const temporary = `${statusPath}.${process.pid}.${Date.now()}.tmp`;
-  fs.writeFileSync(temporary, JSON.stringify(status, null, 2) + '\n');
+  // Open + write + fsync + rename so the durable file lands on disk before
+  // the orchestrator reads it. Closes the race window where a subagent that
+  // completed milliseconds before writeStatus() could otherwise invoke the
+  // next step on a stale status (in particular the human-gate write).
+  const fd = fs.openSync(temporary, 'w');
+  try {
+    fs.writeSync(fd, JSON.stringify(status, null, 2) + '\n');
+    try { fs.fsyncSync(fd); } catch {}
+  } finally {
+    fs.closeSync(fd);
+  }
   fs.renameSync(temporary, statusPath);
 }
 
@@ -475,9 +478,37 @@ if (!fs.existsSync(pipelinePath)) {
 }
 
 let pipeline;
+let pipelineParseError = null;
 try {
   pipeline = JSON.parse(fs.readFileSync(pipelinePath, 'utf8'));
-} catch {
+} catch (error) {
+  pipelineParseError = error;
+}
+// Modern pipelines carry an `intent` field. If pipeline.json is present but
+// unparseable, surface the error so the orchestrator can stop and report
+// instead of falling back to legacyFlow() — which silently drops modern
+// roles like `openspec-author`, `qa-tester`, and the test-planner trio.
+if (pipelineParseError) {
+  if (fs.existsSync(pipelinePath)) {
+    const raw = fs.readFileSync(pipelinePath, 'utf8');
+    const hasModernIntent = /"intent"\s*:/.test(raw);
+    appendMetric({
+      eventId: `${status.metricsRunId || slug}:pipeline_parse_error:${Date.now()}`,
+      type: 'pipeline_parse_error',
+      at: now,
+      source: 'hook',
+      attributionConfidence: 'high',
+      error: String(pipelineParseError.message || pipelineParseError),
+      hasModernIntent,
+    });
+    if (hasModernIntent) {
+      writeStatus({ state: 'blocked', phase: 'blocked', currentAgent: 'task-router' });
+      process.stdout.write(JSON.stringify({
+        followup_message: `pipeline.json is malformed for slug "${slug}" — JSON parse error: ${pipelineParseError.message}. Re-run the task-router to regenerate pipeline.json, then resume via /task-continue ${slug}.`,
+      }));
+      process.exit(0);
+    }
+  }
   legacyFlow();
 }
 

package/presets/claude/mcp-ts/rules/README.md

@@ -24,6 +24,7 @@ Claude topic `.md` files are **derived** from Cursor SoT `presets/cursor/mcp-ts/
 | `agent-team-orchestrator.mdc` | `tooling-and-review/agent-team-orchestrator.md` |
 | `anti-sycophancy-discipline.mdc` | `tooling-and-review/anti-sycophancy-discipline.md` |
 | `code-quality-and-refactoring.mdc` | `tooling-and-review/code-quality.md` |
+| `ai-local-overlay.mdc` | `tooling-and-review/ai-local-overlay.md` |
 | `code-review-mr.mdc` | `tooling-and-review/code-review.md` |
 | `design-guidance.mdc` | `tooling-and-review/design-guidance.md` |
 | `mcp-usage.mdc` | `tooling-and-review/mcp-usage.md` |

package/presets/claude/mcp-ts/rules/tooling-and-review/agent-team-orchestrator.md

@@ -19,8 +19,6 @@ Work request without `/task` — see **agent-team-intake** (when to suggest the
 |---------|------|
 | **`/task <desc>`** | **Preferred** — router → dynamic pipeline → first agent |
 | `/task-continue <slug>` | After human gate or pause |
-| `/feature-start <desc>` | Legacy: analyst-only start (no router) |
-| `/feature-continue <slug>` | Alias of task-continue |
 | `/spec-start [desc]` | Generate/bootstrap `docs/specs` — router intent `spec-gen` → analyst author mode → human gate |
 | `/technical-retro [slug]` | Retro with agent team block |
 
@@ -159,7 +157,7 @@ After each agent completes, verify its terminal manifest receipt, then let the o
 
 ## Legacy mode
 
-If `pipeline.json` is missing (old `/feature-start` tasks), fall back to fixed phases: analysis → development → review → testing. Hook supports both.
+If `pipeline.json` is missing, fall back to fixed phases: analysis → development → review → testing. Hook supports both.
 
 ## Skip pipeline when
 
@@ -190,4 +188,4 @@ When the work edits `packages/ai-rules/presets/**` (or dogfood `.cursor/` sync o
 - **Review-only:** pipeline step = `security-reviewer` for threat/capability diffs; parent applies skill `code-review` for schema/contract/style (no `code-reviewer` agent in Phase 1).
 - After TS changes, pipeline includes **build-verifier** with repository typecheck/test via `post-change-test` / `mcp-ts-tooling`.
 
-**Product specs:** When `/task`, `/feature-start`, or `/spec-start` touches pages/features/`docs/specs`: if INDEX exists and no generate intent → instruct next agent to load requestable **`product-specs`** (consume; soft-fail; never invent). If generate/bootstrap/`spec-gen` → load **`product-specs-authoring`** + skill **`spec-authoring`** inside **task-analyst** (not a separate pipeline step); pass created spec paths + INDEX diff in handoff; human gate must approve before feature-developer. No-op if INDEX absent and no author intent.
+**Product specs:** When `/task` or `/spec-start` touches pages/features/`docs/specs`: if INDEX exists and no generate intent → instruct next agent to load requestable **`product-specs`** (consume; soft-fail; never invent). If generate/bootstrap/`spec-gen` → load **`product-specs-authoring`** + skill **`spec-authoring`** inside **task-analyst** (not a separate pipeline step); pass created spec paths + INDEX diff in handoff; human gate must approve before feature-developer. No-op if INDEX absent and no author intent.

package/presets/claude/mcp-ts/rules/tooling-and-review/ai-local-overlay.md

@@ -0,0 +1,48 @@
+---
+description: Local overlay from .ai.local/ (consumer-owned rules with priority over preset rules) — MANDATORY
+paths:
+  - "**/.ai.local/**"
+  - "**/AGENTS.md"
+  - "**/CLAUDE.md"
+alwaysApply: false
+impact: CRITICAL
+---
+
+# Local overlay (`.ai.local/`) — MANDATORY
+
+This rule is mandatory. **You MUST** obey it on every turn, without exception.
+
+## On every turn
+
+1. **Probe** — before answering, run `ls -la .ai.local` from the repository root.
+2. **Walk** — if `.ai.local/` exists at multiple levels (monorepo), read every level from root to leaf. **Deeper wins** in practice: more specific rules appear later in the assembled context and take precedence.
+3. **Apply** — treat every `.ai.local/rules/*.md` and `.ai.local/conventions.md` as a higher-priority rule for this repo. The block below ("Authoring `.ai.local/rules/*.md`") defines the conventions.
+4. **Conflict** — if a local rule conflicts with a preset rule, **the local rule wins**. If a local rule body contains `override: kill <preset-rule-name>`, drop that preset rule from your context window entirely.
+5. **Acknowledge** — when answering a user question, briefly state which `.ai.local/` rules apply to the current task (one line is enough). Format: `Using .ai.local: <rule-name>(, <rule-name>)*`. If no `.ai.local/` exists, do NOT mention this rule.
+
+If `.ai.local/` does not exist → **no-op**, do not error. Do not produce any log line about its absence.
+
+## Authoring `.ai.local/rules/*.md` (consumer-side reference)
+
+Each `.ai.local/rules/<name>.md` is a plain-markdown rule body. The model parses it as:
+
+- Lines before `## Body` are scope metadata. Only `glob:` / `globs:` / `override:` lines are recognised.
+- Everything from `## Body` onward is the rule body.
+- `glob: always` → always-on. Counts against your context budget.
+- `globs: <comma-separated globs>` → on-demand; applies only when matching files are in focus.
+- `override: kill <preset-rule-name>` → drop that preset rule from context.
+
+Keep files ≤ 120 lines. Match the style of preset rules (short sections, bullets, code examples).
+
+## Limitations (read this)
+
+- Cursor/Claude do **not** natively read `.ai.local/`. This rule is a **prompt-only** mechanism: it asks you, the model, to read and apply the files. Reliability is high but not 100%.
+- If a user adds a file to `.ai.local/` mid-session, **restart the session** to be sure it is picked up.
+- "Priority" is **contextual**, not enforced. Always re-read the local rule before answering questions that touch it.
+- This rule is **not** session-start (Claude convention); it loads on touch of `.ai.local/**` or `AGENTS.md`/`CLAUDE.md`. The "every turn" mandate is enforced by the body text, not by Cursor/Claude loading semantics.
+
+## Pointers
+
+- Authoring guide: `docs/LOCAL-OVERLAY.md` in the `@bonesofspring/ai-rules` package.
+- Living gotchas alternative: `.claude/team/conventions.md`.
+- Root `AGENTS.md` reinforces this rule with a `MANDATORY` section.

package/presets/claude/mcp-ts/rules/tooling-and-review/preset-pr-checklist.md

@@ -19,11 +19,15 @@ Checklist for PRs that touch `packages/ai-rules/presets/**` (and dogfood `.curso
 - [ ] **Core:** if `_shared/core/**` changed — all `<!-- shared-core: … -->` consumers updated (or documented intentional fork)
 - [ ] **Twin:** Cursor change → Claude twin **same PR** (**FAIL** if missing on next, nuxt, svelte, ios-swift, android-kotlin, go, java, mcp-ts, php-hexagonal, and php-laravel — see `preset-twin-sync`)
 - [ ] **Mapping:** `claude/<stack>/rules/README.md` mapping table is current
-- [ ] **Token budget:** alwaysApply / session-start ≤ 3; meta rules not always-on
+- [ ] **Token budget:** Cursor `alwaysApply` ≤ 4 (4th slot reserved for `ai-local-overlay`); Claude session-start exactly 3; meta rules not always-on
+- [ ] **Local overlay coverage:** every installable stack ships `ai-local-overlay.mdc` (Cursor) and `ai-local-overlay.md` (Claude) with `MANDATORY` + acknowledgement format + `docs/LOCAL-OVERLAY.md` pointer; Claude twin listed in `claude/<stack>/rules/README.md` mapping table. Run `packages/ai-rules/scripts/check-local-overlay.sh`.
 - [ ] **Leakage:** no foreign stack agent stems / handoff strings (`preset-no-cross-stack-leakage`)
+- [ ] **Hook-only lifecycle:** `status.json` / `metrics.json` writes happen only inside `hooks/`; agents read but never rewrite lifecycle fields. Run `packages/ai-rules/scripts/check-hook-only-lifecycle.sh`.
+- [ ] **Chain fork:** the byte-wise MD5 uniqueness of `chain-team-phases.sh` across the 10 stacks × 2 platforms (20 hooks) holds. Run `packages/ai-rules/scripts/check-chain-hook-coverage.sh` and `packages/ai-rules/scripts/test-chain-team-phases-coverage.mjs`.
 - [ ] **Twin depth:** domain ≥15 lines; thin aliases documented
 - [ ] **README:** stack `rules/README.md` core→stack mapping / catalog updated if stems added
 - [ ] **CHANGELOG / package README:** note when public preset paths or install surface change
+- [ ] **i18n anchors:** kept-marked Cyrillic fragments in `tests-unit.md(c)` stay consistent with their English counterparts. Run `packages/ai-rules/scripts/check-preset-i18n.sh`.
 - [ ] **Dogfood:** run `packages/ai-rules/scripts/sync-dogfood-cursor.sh` (or document intentional skip) when maintaining this monorepo
 - [ ] **shared-core drift:** for embed-mode stems, run `packages/ai-rules/scripts/check-shared-core-drift.sh` (or document intentional fork)
 

package/presets/claude/mcp-ts/rules/tooling-and-review/preset-token-budget.md

@@ -17,12 +17,16 @@ Invariants for rule loading in an **installed** preset (consumer app).
 
 | Platform | Always-on | Cap |
 |----------|-----------|-----|
-| Cursor | `alwaysApply: true` | **≤ 3** |
+| Cursor | `alwaysApply: true` | **≤ 4** |
 | Claude | rules **without** `paths:` | **exactly 3** |
 
+**Cursor 4th slot is reserved for `ai-local-overlay`** (consumer-owned `.ai.local/` rules). If you do not need local overlay, demote that rule to `globs: ["**/.ai.local/**"]` and revert to ≤3.
+
 **Claude Option A (README exclusion):** when counting Claude rules without `paths:` / session-start, **exclude `**/README.md`** (topic indexes are not session-start rules). The intended session-start trio remains **exactly 3**; body budget sums **only those 3** rules.
 
-Session-start body budget (where documented): ≤140 lines total (stretch ≤120).
+Session-start body budget:
+- Claude trio: ≤140 lines total (stretch ≤120).
+- Cursor 4-set: ≤170 lines total (4 rules including `ai-local-overlay`).
 
 ## Meta / author rules
 
@@ -38,4 +42,4 @@ Orchestrator, feature-delivery, code-review, post-change gates, domain globs —
 
 ## Verifier
 
-build-verifier **FAIL** if alwaysApply / session-start count exceeds 3 (Claude count **excludes `**/README.md`**), or if a `preset-*` meta rule is marked always-on.
+build-verifier **FAIL** if Cursor `alwaysApply` count exceeds 4, Claude session-start count exceeds 3 (Claude count **excludes `**/README.md`**), or if a `preset-*` meta rule is marked always-on.