Coordinator Planning: Grill + Auto-DoD

The coordinator follows a graduated planning motion before executing work:

grill (gradient) → generate DoD → task DAG (linked to DoD) → critic → execute → verify DoD

Not every prompt needs full ceremony. The coordinator self-classifies incoming work into one of three tiers and applies the appropriate level of rigour.

---

Trigger Gradient

Tier	Signal	Behaviour
Trivial	Single job, no ambiguity	Skip grill + DoD entirely — proceed directly to task/execute.
Mechanical	Multi-part but well-defined (bugfix, refactor, chore)	Silent self-grill — the coordinator internally answers 3 inquisitor angles (UX, technical, edge-cases) without prompting the user — then auto-generates a DoD.
Novel / ambiguous	Feature-shaped, open-ended, or unclear scope	Bounded interactive grill — ≤1 round, ≤5 questions via ask_user — followed by a user-confirmed DoD.

Unattended / no ask_user fallback

Under --unattended or when the ask_user tool is unavailable, the interactive tier degrades to the self-answered path. Any assumptions the coordinator makes are prefixed with ⚠️ ASSUMPTION. The planning motion never blocks — it always respects the "stay responsive" rule.

---

Definition of Done (DoD)

After grilling, the coordinator calls dod_generate to write structured DoD lines into .plans/<slug>/06-definition-of-done.md:

DoD-1: All endpoints return 200 on valid input — VERIFY: integration test passes
DoD-2: Error responses include machine-readable code — VERIFY: grep response schema
DoD-3: No regressions in existing test suite — VERIFY: `npm test` green

Each line follows the format DoD-N: <requirement> — VERIFY: <method>.

The coordinator auto-creates the .plans/<slug>/ directory for ad-hoc DAG planning so DoD and grill artifacts always have a home.

---

Task ↔ DoD Linking

When the coordinator creates tasks via task_create, it can include dod_ids: string[] to link each task to the DoD items it satisfies.

For already-created tasks, task_update also accepts dod_ids for in-place remapping (replace-on-provided; omit = unchanged; empty array = clears all links). This allows the self-healing gate to remap existing unmapped tasks without recreating them.

Coverage tracking:

• dod_coverage reports which tasks lack a DoD mapping and which DoD items have no implementing task.
• The coordinator uses this feedback to fill gaps before spawning workers.

Ordering note

dod_generate should be called before task_create so that DoD IDs exist when tasks are created. Tasks created before activePlanSlug was set may not appear in coverage reports.

---

Coordinator Tools

Two new coordinator tools power the DoD workflow:

Tool	Purpose
dod_generate	Writes DoD-N lines to .plans/<slug>/06-definition-of-done.md. Deterministic — no LLM call.
dod_coverage	Returns a coverage report: unmapped tasks, uncovered DoD items, and an allow/warn signal.

These tools will appear in the auto-generated Coordinator Tools Reference after the next npm run gen:reference run.

---

DoD Gate — AGENTS_FLEET_DOD_GATE

An environment variable controls enforcement at execute-wave time. The gate is self-healing (remediate-first): on a coverage gap the coordinator auto-remediates before deciding whether to proceed or escalate.

Remediation behaviour

When dod_coverage detects a gap (unmapped tasks or uncovered DoD items), the coordinator automatically:

1. Calls dod_generate to author missing DoD-N: <requirement> — VERIFY: <method> items.
2. Maps unmapped tasks — new tasks via task_create dod_ids, existing unmapped tasks via task_update dod_ids (in-place remapping).
3. Re-runs dod_coverage to verify the gap is closed.

Remediation is capped at one pass — the coordinator never loops. After the single remediation attempt, the gate mode determines what happens next.

Gate modes

Value	Behaviour
off	No gating — coverage is informational only.
warn (default)	Auto-remediates, then proceeds regardless; only emits a chip/warning if a gap genuinely could not be closed.
enforce	Auto-remediates, then holds the execute wave (escalates to the user) if a gap remains. Under --unattended or when ask_user is unavailable, proceeds after the remediation pass.

Safety guarantees

• The gate can never wedge the coordinator: if no DoD file exists, no DoD items are present, coverage is fully mapped, or the gate value is off, execution proceeds unconditionally.
• An unrecognised env value is treated as warn.
• Remediation is bounded to one pass — no infinite loop risk.

---

Known Cosmetic Edges

These are non-blocking and documented for future hardening:

1. Tasks created before activePlanSlug was set may not appear in dod_coverage reports — but they can be remapped in-place via task_update dod_ids once the plan slug is active (the self-healing gate handles this automatically).
2. Zero tasks + existing DoD items yields fullyMapped: false with an allow signal (cosmetic inconsistency).
3. No .max() bound on the dod_generate items array — callers could theoretically pass an unbounded list.
4. O(n) max-DoD regex scan on append — negligible in practice.