pi-crew 0.9.68 → 0.10.2

package/CHANGELOG.md

@@ -2,6 +2,228 @@
 
 > **Note:** `atomic-write-v2.ts` / `AtomicWriter` mentioned in historical entries below was consolidated into `atomic-write.ts` as of v0.9.42. This changelog is preserved as historical record — the migration was completed (the v2 class was never adopted; v1 won on simplicity + symlink-safety + link+unlink atomicity). See `docs/migration/atomic-write-v2-migration.md` for the decision rationale.
 
+## [0.10.2] — UI rewrite + adaptive default team (2026-08-24)
+
+40 commits after `v0.10.1` (≈3,500 LOC, 54 files). Headline: the UI surface
+goes from "two modal overlays + status widget" to a proper in-document
+panel — task list above the editor, dock at the very bottom, and an inline
+agent transcript that opens with `↓` from the empty prompt. The default
+team also moves from a fixed 4-step chain to a single adaptive `assess` →
+parallel-execute → verify DAG.
+
+### UI: task list above the editor (`src/ui/widget/task-list.ts`)
+
+- **Pi-tasks style plan rows** (commits `3d930579`, `7434afd7`, `f37619c0`,
+  `54e638ef`). One numbered row per task painted in plan order: `#1`, `#2`,
+  … — no more `01_explore` style technical ids leaking into the user view.
+  Completed rows dim and strike through; the running row carries a spinner
+  frame with elapsed time and live token counts; queued rows name the open
+  dependencies they wait on (`› blocked by #2`).
+- **Header is Claude-Code style**: `● 4 tasks (1 done, 1 in progress, 2 open)` —
+  plain counts, no agent/role identity (those belong to the dock).
+- **Role/agent/model visibility removed from the row** — they live in the dock
+  below the editor. The list above the editor is the *plan*, not a worker
+  report.
+- **10-row cap with `… and N more`** — unfinished work never falls out of the
+  cap; finished rows collapse behind the overflow marker.
+- **Templated titles / descriptions** (commit `f51f7095`) — the task packet
+  emits `title` + `description` fields, `{goal}` substitution lands in the
+  persisted task record (not the run goal), so the list reflects what was
+  actually scoped per phase.
+
+### UI: inline agent panel (`src/ui/inline-panel/`)
+
+A new module (1,761 LOC across 10 files) that replaces the modal-only agent
+surface with an in-document path. Design rationale: `docs/design/2026-08-20-inline-agent-panel.md`.
+Attribution: `NOTICE.md` §"Inline agent panel" — adapted from `pi-subtask`
+v0.7.4 (Victor Mustar, MIT); no code copied verbatim; steering rides
+pi-crew's existing `team steer` channel rather than pi-subtask's stdin pipe.
+
+- **`↓` from the empty prompt opens the dock** (commits `cfcad621`, `cf4895e8`,
+  `bff74c4f`). Static-icon flat rows, footer usage meter, `belowEditor`
+  placement. Enter on a row drops you into the worker's transcript.
+- **Worker view is an in-document pane, never a session switch** (commits
+  `11c6649a`, `21ff211d`, `87497bd5`, `de3d44a0`, `50f41f30`, `5526c279`,
+  `402bb2a5`, `3dd68cc8`, `2b3e899d`, `9ee89866`, `ec8efd0a`, `e1708b20`,
+  `27c5c402`, `56279b00`, `c16cb332`). A sequence of fixes that landed
+  together: a) the `/crew-view` dispatch was routed through pi's immediate
+  command path (not a custom overlay), b) the agent view is a *byte-copy*
+  of the worker's own session log via a derived sessions ROOT, c) live
+  refresh polls the worker transcript and re-renders, d) `Enter` always
+  opens a real Pi session (no broken overlay), e) the worker-143 kill on
+  switches is root-caused (stale `ctx` was being passed around) and the
+  `AbortError` post-exit crash is fixed (`9db66cda`).
+- **Dock survives until the run is done** (commit `67fa202a`) — rows no
+  longer disappear at "completed", the 3-row scroll window keeps the
+  in-progress task visible, and each row remembers its model.
+- **Full-screen overlay still works for end-of-run review** (commit `c16cb332`)
+  — transcripts are complete; `Escape` returns to the dock.
+
+### Workflow: default is now adaptive (BREAKING for templates)
+
+`workflows/default.workflow.md` was rewritten (commit `fff4b98d`) — old
+sequential `explore → plan → execute → verify` was too rigid for the
+common case of "small goal, fast turnaround". The new default is a
+single-`assess` adaptive plan:
+
+- Topology: `complex-dag` (was `sequential`).
+- Frontmatter: `adaptive: true`.
+- The `assess` step (planner role) inspects the repo and emits an
+  `ADAPTIVE_PLAN_JSON` block with the concrete phases + tasks. Independent
+  tasks go into the SAME phase so they run in parallel; the verifier still
+  closes the loop at the end of any implementation plan.
+- Hard cap: 12 tasks per phase. A simple goal may produce 1-2 phases with
+  1-2 tasks; broader goals fan out.
+- Workflow file is **runtime data** (no rebuild needed) — the change is live
+  the next time a worker loads it.
+- Migration: anyone pinning the old `explore/plan/execute/verify` chain
+  should switch to `workflow='plan-execute'` (still ships the fixed DAG).
+
+### Misc
+
+- **Model-routing transparency**: `console.warn` per-model passthrough
+  muted (`273c4f83`) + test pins the muted behaviour (`b019e645`).
+- **Biome 2.5.3 lint + format sync** (commit `69a0460e`, no logic changes).
+- 5 `docs/real-test/reports/real-test-2026-08-2{0..4}-*.md` reports landed
+  alongside the B5 real-test battery (the subagent-v2 release was
+  end-to-end verified before this UI batch).
+- Real-test re-run on the released bundle (`b0cf3a3a` onward): test:critical
+  102/102, typecheck clean, 9a 10/10, 9b 5/5, chain 306.8s observation;
+  opportunistic finding #5 (SPEC-EVIDENCE marker format mismatch `===` vs
+  `:`) filed separately.
+
+### See also
+- `NOTICE.md` §"Inline agent panel" — pi-subtask attribution.
+- `NOTICE.md` §"Source inspiration" — pi-tasks attribution for task list rendering.
+- `docs/design/2026-08-20-inline-agent-panel.md` — inline panel design rationale.
+- `workflows/default.workflow.md` — adaptive default workflow (runtime data).
+- `docs/real-test/reports/real-test-2026-08-11-full-battery.md` — full 9-tier
+  battery on the post-v0.10.1 HEAD (`54e638ef`).
+
+## [0.10.1] — subagent v2: governed delegation, plan objects, spec system, transparency (2026-08-20)
+
+The full subagent-v2 effort (design `docs/design/subagent-v2-design.md`, plan
+`docs/design/subagent-v2-implementation-plan.md`) shipped as ONE release.
+Trains T1-T5 merged as internal phases (PRs #47-#51); batteries B1-B5 in
+`docs/real-test/reports/`. ADRs: governed-nesting, plan-object,
+waiting-producer-ask, spec-system (+ round-1 errata).
+
+### T1 — waiting/ask producer + session surface
+- Ask tool (worker→leader blocking questions; parked `waiting` tasks,
+  structured events `ask.requested/answered/timedout`), follow-up
+  notifications, session-ownership map.
+
+### T2 — plan object (ADR-4)
+- PlanRecord revisions (append-only, stable item ids, carried linkage),
+  `plans approve/reject` dual-write, scheduler linkage, adaptive re-plan
+  (revision switch, dropped-item soft-cancel advisory).
+
+### T3 — governed nesting (ADR-5)
+- delegate tool for executor-class roles at depth 1 (dormant flag;
+  `nesting.enabled` USER-only sensitive), nested-slot budget (fail-fast,
+  max(1,⌊sem/2⌋)), depth-capped broker credentials (grandchild tokens
+  subId-scoped + shadow task records — unbounded chains terminate),
+  read-only grandchild roles, usage roll-up, fenced results via mailbox.
+
+### T4 — spec system (ADR-6 + erratum §11)
+- SpecRecord store: workspace `state/specs/` (generated-only) + USER store
+  `~/.pi/agent/specs/` with DIGEST-BOUND trust sidecar — workers can never
+  mint re-executable specs (provenance v2; content-swap + TOCTOU closed).
+- SPEC-EVIDENCE footer contract + mechanical coverage gate (default:
+  `unverified` badge on gaps, never blocks); strict mode = coverage AND
+  machine-check in a hardened sandbox (`unshare -rn`, ulimit, env scrub,
+  digest-only outcomes, fail-closed everywhere); verifier-role reject-start.
+- Freeze at dispatch: snapshots immutable; later spec edits never rewrite
+  what a running task was held to. Mint CLI: `scripts/spec-import.mjs`.
+
+### T5 — P2 transparency + hygiene
+- R7 Plan UI (`PI_CREW_PLAN_UI=1`): plans snapshot slice, dashboard pane 7
+  (tree, progress, depth badges, approval keys, revision diff), powerbar
+  plan phases.
+- R8 model-routing transparency: pre-run budget summary (chain +
+  worst-case spawns/task), deduped unvalidated-passthrough warnings,
+  per-attempt models in the output pane.
+- R9 worker self-reporting channel: bounded `worker.*` append channel
+  (schema tag, sliding-window rate limit, FIFO cap, partial-line
+  quarantine); `PI_CREW_EVENTS_PATH` unconditional (read-only roles too).
+- R10 docs hygiene (phantom commands removed, real ones documented,
+  widget TTL corrected).
+
+### Fixes surfaced by the batteries
+- bug-029: symlinked-TMPDIR repo-root escape (macOS `/var`↔`/private/var`)
+  — deterministic `getCacheStats` CI failures; canonicalized boundaries.
+- dwf-setresult determinism (ambient `PI_CREW_TRUST_PROJECT_DWF` leak).
+- B2/B3 live catches: hang-class timeout group-kill gaps, delegate
+  deadline precedence, double-save clobber on plan revision switch.
+
+**Full suites at release tree: 6977 unit + 188 integration, 0 fail; CI 3-OS green on every train PR.**
+
+## [0.10.0] — maintainability refactor: full remediation + reliability (2026-08-16)
+
+Branch `refactor/maintainability`: 16-round audit → 5-phase refactor →
+finding-by-finding sweep → full remediation of every open finding (Waves
+1A/1B/2A/2B/B/C). Full execution log with commit map:
+`docs/refactor-plan.md`. Audit inventory: `docs/refactor-plan.review.md`.
+
+### Security (config tiering)
+- Schema-driven project-config sanitize: `sensitive:true` drop-list +
+  CONDITIONAL drops — project config may only TIGHTEN (guards `off`/
+  loosening dropped). goalWrap subtree + magicKeywords user-only.
+  ADR: `docs/decisions/2026-08-15-schema-driven-sanitize.md`.
+- P1 fix: project-tier `.pi/crew-settings.json` no longer bypasses sanitize
+  (tiers API; legacy flat fns JSDoc'd user-tier-only).
+- scheduledJobs user-tier opt-in gate (both `schedulingEnabled:true` AND
+  `allowProjectScheduledJobs:true` required for project jobs).
+
+### Performance
+- R10-1 O(B²) batch closeout → per-run `ResultArtifactReadCache`
+  (bench b10: −50% fs reads, byte-identical outputs; b11 extends to
+  dependency-context path).
+- R10-2 path-scoped flush; R10-3 broker mtime-gated manifest parse cache
+  (N waiters ≤2 parses); R10-4 sync/async snapshot parity; R10-5
+  child-executor fs batching (≤250ms documented window).
+
+### Hygiene / dead code
+- FIFO caps on all unbounded collections (10 sites); named+removed abort
+  listeners; dead GroupJoinManager + 7 dead fns removed; security telemetry
+  wired into doctor; parser/schema bounds aligned (schema wins).
+
+### Runner reliability (bug-026)
+- Sub-A: stderr-only result artifacts now FAIL the task
+  (`empty-or-stderr-only-result`) instead of persisting noise as completed.
+- Sub-B: fatal-fs errno classifier — ENOSPC/EDQUOT/EMFILE surfaced as
+  `failed (disk full)`-style failureCause in task/run/status/doctor.
+- Sub-C: terminal runs evicted from in-memory preloaded manifests +
+  heartbeat aggregator terminal-task defense (kills phantom dead-worker
+  alerts on long-lived sessions).
+
+### Runner reliability (bug-027 / bug-028 — post-Wave-D, found via real-test + CI)
+- bug-027 (`f6e605ef`): `finalizeRun` heals stale non-terminal in-memory
+  snapshots from disk (in-flight guard) — a run whose disk manifest is
+  terminal but whose preloaded snapshot says "running" no longer blocks
+  finalize with a misleading in-flight error.
+- bug-028 (`e9586aca`): fan-out stale-fallback resurrection —
+  `persistSingleTaskUpdate` previously wrote the dispatch-time
+  `fallbackTasks` array (the O(1) "F4" shortcut), which could resurrect
+  disk-terminal sibling tasks back to "running". It now always re-reads
+  the latest manifest from disk inside the run lock. Deterministic
+  regression test:
+  `test/unit/runtime/task-runner/state-helpers-bug028-stale-fallback.test.ts`
+  (verified fail-before-fix via git stash).
+
+### Verification (real-test 9-tier, v0.10.0 premerge)
+- Full 9-tier real-test battery incl. the live mid-run battery: steer into
+  a running task, live cancel (cross-session ownership guard + force),
+  plan-approval waiting flow, and respond handler — run via a second pi
+  session in tmux. Report:
+  `docs/real-test/reports/real-test-2026-08-16-v0.10.0-premerge.md`.
+- macOS CI teardown hardening (`8197f054`): retry recursive rmSync against
+  the worker-write/rmdir ENOTEMPTY race in `resume-cancel.test.ts`.
+
+### Repo hygiene
+- bench suite b1–b11 tracked; per-run perf reports + bench results
+  gitignored; package.json-referenced scripts committed (fresh-clone fix).
+
 ## [0.9.68] — RLM fixes after deep-review verification (2026-08-12)
 
 Fixes from `docs/rlm-deep-review-2026-08-12.md` (verified) +

package/NOTICE.md

@@ -6,6 +6,7 @@
 
 - Primary design and Pi-extension implementation inspiration: `pi-subagents` by Nico Bailon, MIT license.
 - Team orchestration, state, and worktree contract inspiration: `oh-my-claudecode` by Yeachan Heo, MIT license.
+- Task-list-above-the-editor rendering (numbered plan rows, dependency hints, one-line-per-task) draws on `pi-tasks` (tintinweb, MIT) and Claude Code's plan header convention; no code copied — the pi-crew implementation lives in `src/ui/widget/task-list.ts` and is original work informed by those surfaces.
 - Conceptual inspiration only: `oh-my-openagent` / `oh-my-opencode`, SUL-1.0. No source code from this project should be copied into `pi-crew` unless explicitly reviewed for license compatibility and documented here.
 - Built-in skill topics are original pi-crew guidance informed by common agent-skill patterns in `Source/awesome-agent-skills`, `Source/oh-my-claudecode`, and related local references; no verbatim skill text was copied.
 
@@ -13,6 +14,26 @@
 
 When code is copied or substantially adapted from an MIT source, add the source path and license note here.
 
+### Inline agent panel (2026-08-20)
+
+The inline agent panel (`src/ui/inline-panel/`) adapts the display architecture
+of `pi-subtask` v0.7.4 by Victor Mustar (MIT,
+https://github.com/gary149/pi-subtask):
+
+- width-budgeted single-line agent rows (compact widget row style),
+- identity-tracked panel cursor (`panel-selection.ts` matches its
+  `panelSelId` state machine),
+- in-document `aboveEditor` transcript pane reusing pi's native transcript
+  components instead of a viewport overlay,
+- `CustomEditor` wrapper with `↓`/`↑`/`enter`/`x`/`escape` panel navigation
+  and the `@agent` editor-border label, and the
+  `!ctx.ui.getEditorComponent()` yield rule for editor ownership.
+
+No code was copied verbatim; the pi-crew implementation is a fresh port onto
+pi-crew's own widget/store/event architecture (steering rides pi-crew's
+existing `team steer` steering-file channel rather than pi-subtask's stdin
+pipe).
+
 Current scaffold status: no substantial source files have been copied verbatim; implementation is a fresh scaffold based on documented design lessons.
 
 ## crew-vibes font assets

package/README.md

@@ -53,7 +53,9 @@ repo: https://github.com/baphuongna/pi-crew
 - **Durable state** — manifest, tasks, events, artifacts all persisted to disk
 - **Async/background runs** — detached runs survive session switches with completion notifications
 - **Worktree isolation** — opt-in git worktrees per task for safe parallel edits
-- **Rich UI** — live widget, dashboard, progress tracking, model/token display
+- **Rich UI** — task list above the editor (pi-tasks style: numbered plan rows, dependency hints, `… and N more` overflow), dock at the very bottom (static icons, per-row model, footer usage), and an inline agent panel (open a worker's transcript with `↓` from the empty prompt — never a session switch). Live widget, dashboard, and progress tracking unchanged.
+- **Inline agent panel** (`src/ui/inline-panel/`) — status rows at the bottom, transcript in-document with pi's native components, `CustomEditor` overlay for steering (rides the existing `team steer` channel). Adapted from `pi-subtask` v0.7.4 (MIT) — attribution `NOTICE.md` §"Inline agent panel"; no code copied verbatim.
+- **Adaptive default team** (`workflows/default.workflow.md`) — the built-in default is now a single `assess` → adaptive-DAG step instead of a fixed `explore → plan → execute → verify` chain. The planner inspects the repo and emits a JSON plan; independent tasks run in parallel; the verifier closes the loop. Workflow files are runtime data — the change is live without a rebuild. Pin to `workflow='plan-execute'` if you need the old fixed DAG.
 - **Observability** — metrics registry, Prometheus/OTLP exporters, heartbeat watching, deadletter queue
 - **Resource management** — create/update/delete agents, teams, workflows with validation
 - **Import/export** — portable run bundles for sharing and archiving
@@ -212,7 +214,7 @@ When unsure which team/workflow fits:
 
 | Team | Workflow | Purpose |
 |------|----------|----------|
-| `default` | explore → plan → execute → verify | Balanced, general-purpose |
+| `default` | adaptive: planner derives concrete tasks from the goal, parallel phases, verify | Balanced, general-purpose |
 | `fast-fix` | explore → execute → verify | Quick bug fixes |
 | `implementation` | Adaptive planner decides fanout | Multi-file implementation |
 | `review` | explore → code-review → security-review → verify | Code review + security audit |
@@ -289,6 +291,46 @@ The advisory is **informational only** — there is no `force:true` flag needed
 
 ## Recent changes
 
+### v0.10.2: UI rewrite + adaptive default team (2026-08-24)
+
+40 commits after `v0.10.1` (≈3,500 LOC, 54 files). Headline: the UI surface
+goes from "two modal overlays + status widget" to a proper in-document
+panel — task list above the editor, dock at the very bottom, and an inline
+agent transcript that opens with `↓` from the empty prompt. The default
+team also moves from a fixed 4-step chain to a single adaptive `assess` →
+parallel-execute → verify DAG.
+
+- **Task list above the editor** — pi-tasks style numbered plan rows
+  (`#1`, `#2`, …) instead of `01_explore` ids; completed rows dim and
+  strike through; running row shows spinner + elapsed time + token
+  counts; queued rows name dependencies (`› blocked by #2`). Header
+  is Claude-Code style (`● 4 tasks (1 done, 1 in progress, 2 open)`).
+  Role/agent/model identity moved to the dock — the list is the *plan*,
+  not a worker report. 10-row cap with `… and N more` overflow.
+- **Inline agent panel** (`src/ui/inline-panel/`, ~1,761 LOC, 10 files) —
+  `↓` from the empty prompt opens the dock; Enter drops into the worker's
+  full transcript rendered in-document with pi's native components;
+  steering rides the existing `team steer` channel (no stdin pipe).
+  Worker view is a *byte-copy* of the worker's own session log, polled
+  for live refresh, and Enter on the dock row always opens a real Pi
+  session (the worker-143 kill on switches and the `AbortError`
+  post-exit crash are both fixed). Full-screen overlay preserved for
+  end-of-run review.
+- **Dock survives until the run is done** — rows no longer disappear at
+  "completed"; 3-row scroll window keeps the in-progress task visible;
+  per-row model display.
+- **Adaptive default team** (`workflows/default.workflow.md`, runtime
+  data — no rebuild needed) — old fixed `explore → plan → execute →
+  verify` replaced by a single `assess` step (planner role) that emits
+  an `ADAPTIVE_PLAN_JSON` block. Independent tasks go into the SAME
+  phase so they run in parallel; verifier still closes the loop.
+  **Pin to `workflow='plan-execute'` if you need the old fixed DAG.**
+- **Model-routing passthrough muted** + Biome 2.5.3 lint/format sync.
+- Real-test re-run on the released v0.10.1 bundle: test:critical 102/102,
+  typecheck clean, 9a 10/10, 9b 5/5, chain 306.8s observation.
+  See `CHANGELOG.md` §Unreleased for full notes and `NOTICE.md` for
+  pi-subtask / pi-tasks attributions.
+
 ### v0.9.65: team-tool schema empty-string guard + effectiveness empty-result guard (2026-08-10)
 
 - **`budgetTotal` empty-string unset marker accepted**: `budgetTotal` was the only numeric `TeamToolParams` field missing the `Literal("")` union branch its siblings had. Calling models that emit every schema key with defaults were rejected by pi-ai's pre-handler validation → `Validation failed for tool "team"` on every action. The `MISCONFIGURATION GUARD` (rejects 1-999) is preserved. Caught by the Tier 9 feature battery — Tiers 1-8 stayed green while the team tool was broken for emitting models.

package/agents/analyst.md

@@ -5,7 +5,7 @@ model: false
 systemPromptMode: replace
 inheritProjectContext: true
 inheritSkills: false
-tools: read, grep, find, ls
+tools: read, grep, find, ls, ask
 ---
 
 You are a requirements analyst. Identify what is known, unknown, risky, ambiguous, or underspecified. Produce clarifying assumptions and acceptance criteria.

package/agents/cold-verifier.md

@@ -5,7 +5,7 @@ model: false
 systemPromptMode: replace
 inheritProjectContext: true
 inheritSkills: false
-tools: read, grep, find, ls, bash
+tools: read, grep, find, ls, bash, ask
 maxTurns: 12
 ---
 
@@ -64,3 +64,5 @@ EVIDENCE: file:line references + your own test output
 ```
 
 If you cannot refute a claim after honest effort, that is itself evidence the claim is solid — say so explicitly rather than inventing doubt.
+
+Intentionally unreferenced by builtin teams (CLI/custom use).

package/agents/critic.md

@@ -5,7 +5,7 @@ model: false
 systemPromptMode: replace
 inheritProjectContext: true
 inheritSkills: false
-tools: read, grep, find, ls, glob
+tools: read, grep, find, ls, glob, ask
 ---
 
 You are a critical reviewer. Find flaws, missing steps, unsafe assumptions, overengineering, underengineering, and verification gaps. Return concrete fixes to the plan.

package/agents/executor.md

@@ -5,7 +5,7 @@ model: false
 systemPromptMode: replace
 inheritProjectContext: true
 inheritSkills: false
-tools: read, grep, find, ls, bash, edit, write, scratchpad
+tools: read, grep, find, ls, glob, bash, edit, write, scratchpad, ask, delegate
 ---
 
 You are an implementation specialist. Follow the provided plan, make targeted changes, keep edits minimal, and report changed files plus validation status. Do not broaden scope without explaining why.

package/agents/explorer.md

@@ -5,7 +5,7 @@ model: false
 systemPromptMode: replace
 inheritProjectContext: true
 inheritSkills: false
-tools: read, grep, find, ls, glob, bash
+tools: read, grep, find, ls, glob, bash, ask
 ---
 
 You are a fast codebase explorer. Map relevant files, symbols, data flow, and constraints. Do not modify files. Return concise findings with paths and evidence.

package/agents/planner.md

@@ -5,7 +5,7 @@ model: false
 systemPromptMode: replace
 inheritProjectContext: true
 inheritSkills: false
-tools: read, grep, find, ls, glob
+tools: read, grep, find, ls, glob, ask
 ---
 
 You are a planning specialist. Convert the goal and discovery notes into a concrete, ordered plan. Identify dependencies, risks, validation steps, and handoff instructions for implementers.

package/agents/reviewer.md

@@ -5,7 +5,7 @@ model: false
 systemPromptMode: replace
 inheritProjectContext: true
 inheritSkills: false
-tools: read, grep, find, ls, glob, bash
+tools: read, grep, find, ls, glob, bash, ask
 ---
 
 You are a code reviewer. Review the implementation for bugs, regressions, maintainability issues, missing tests, and project-rule violations. Return prioritized findings with evidence.

package/agents/security-reviewer.md

@@ -5,7 +5,7 @@ model: false
 systemPromptMode: replace
 inheritProjectContext: true
 inheritSkills: false
-tools: read, grep, find
+tools: read, grep, find, ask
 ---
 
 You are a security reviewer. Look for injection, authn/authz flaws, insecure defaults, secret exposure, unsafe filesystem/network behavior, and dependency risks. Return severity and remediation.

package/agents/test-engineer.md

@@ -5,7 +5,7 @@ model: false
 systemPromptMode: replace
 inheritProjectContext: true
 inheritSkills: false
-tools: read, edit, write, bash, ls, scratchpad
+tools: read, edit, write, bash, ls, glob, grep, find, scratchpad, ask, delegate
 ---
 
 You are a test engineer. Identify the right test level, add or adjust tests when asked, detect flaky assumptions, and report exact validation commands and results.

package/agents/verifier.md

@@ -5,7 +5,7 @@ model: false
 systemPromptMode: replace
 inheritProjectContext: true
 inheritSkills: false
-tools: read, grep, find, ls, bash, scratchpad
+tools: read, grep, find, ls, bash, scratchpad, ask
 maxTurns: 15
 ---
 

package/agents/writer.md

@@ -5,7 +5,7 @@ model: false
 systemPromptMode: replace
 inheritProjectContext: true
 inheritSkills: false
-tools: read, edit, write, ls
+tools: read, edit, write, ls, ask
 ---
 
 You are a documentation specialist. Produce clear, concise, maintainable docs and summaries. Preserve technical accuracy and avoid marketing fluff.