Do

Phase 3: ENHANCE

Stack on signals.

Review overlap: real-diff row wins; fallback only without diff.

Explicit "interview me" or "grill me" opts into exhaustive material coverage until shared understanding or a user stop; it has no arbitrary question or round cap. An implicit interview must earn its interruption cost through likely avoided rework and remains bounded. "Just build it," "skip questions," and equivalents use recommended defaults and continue.

An interview is not terminal when it suspends an active delivery objective. Compile the decisions and automatically resume the originating build, fix, deploy, validation, or other execution flow; never report the decision artifact as completion of the original objective. Stop after compilation only when the user explicitly requested only an interview artifact or excluded implementation.

Ask the current independent frontier in one logical round. In Markdown, include the full frontier with a recommendation for each and wait once. A harness native structured question UI may chunk the frontier only at its capacity per call; do not recompute between chunks unless an answer invalidates a pending question. Single question turns are reserved for true dependency branches. Interviews initiated by the router cap at five questions and three decision rounds, plus at most one concise confirmation response; explicit grills construct and exhaust the material decision tree, with the kinds and number of questions determined by what shared understanding requires. Do not add ceremonial questions. Answering the last frontier does not authorize execution by itself: ask one concise shared understanding confirmation. Skip that extra confirmation only when the same response explicitly says "proceed", "build it", "looks right, continue", or equivalent. Resume nested execution only after explicit proceed or confirmation; a request for only an interview stops at the artifact.

Check pairs_with before stacking. Skills with built-in verification gates may suffice.

Step G: GATHER (Simple+) — fill the Task Spec before the Gate.

1. Keep request_verbatim unchanged. State this worker's intent, constraints (including authority), files, ownership, and acceptance.
2. Include decisions, prior results, and gaps when they affect this assignment. Summarize findings and link durable evidence; quote exact text only when its wording matters. Do not copy the whole investigation into every handoff.
3. Verify named paths. Add excerpts only when they explain a decision or let the worker act; the builder validates paths even when gathering is off.
4. Use context_mode: "summary" for current git status, diff stat, and recent commits. Use files when initial excerpts help, or none when the worker already has valid context. The legacy default is files. Mode none retains a notice that no fresh state was gathered, so the handoff envelope stays complete. Legacy --no-gather removes that envelope too; use none for Medium+ handoffs. Context modes never omit the supplied Task Spec or required injections.

Reuse reads only while their source and task context remain unchanged; a fresh worker needs access to the relevant content or references. For worker transitions, use process. Verification evidence follows testing; repeat checks when their inputs or relevant environment change, not just because a new phase starts.

Gate: Enhancements applied, Task Spec filled. Phase 4.

Phase 4: EXECUTE

Step 0: Creation — ADR at adr/{name}.md, adr-query.py register, plan.

Step 1: Plan (Simple+) — task_plan.md; skip Trivial.

Step 1b: Quality-loop (Medium+ code mod) — references/quality-loop.md 14 phases. P2 agent=implementation. Force-route in loop. Skip non-code/Trivial/Simple.

Step 1c: Workflow — Pipeline pick or Complex no pick or explicit → ${CLAUDE_SKILL_DIR}/references/workflow-dispatch.md. Both 1b+1c → quality-loop OUTER, workflow in IMPLEMENT.

Step 2: Invoke agent

build-dispatch.py — source for [do-route], exact Skill-tool calls, thinking, budget, Task Spec, injections, worktree/local-only. Do not hand-assemble it.

python3 "$SDIR/build-dispatch.py" --json '{
  "agent": "<agent>", "skill": "<skill; omit when agent-only>",
  "pipeline": "<pipeline; omit when Phase 2 returned null>",
  "complexity": "<trivial|simple|medium|complex>",
  "model": "inherit",
  "context_mode": "summary",
  "provider": "<anthropic|openai|other>",
  "manual_model_override": false,
  "health": "-",
  "fallback_reason": "<REQUIRED when agent=general-purpose; omit otherwise>",
  "stack": ["s1","s2"],
  "task_spec": {"request_verbatim": "<user message, unchanged>", "intent": "...",
                "constraints": "<applicable rules, limits, and authorization>",
                "decisions": "...", "prior_results": "...", "gaps": "...",
                "acceptance": "<command> → <expected>",
                "files": "<owned paths; optional line ranges>", "ownership": "<worker scope>",
                "operator_context": "..."},
  "flags": {"worktree": false, "local_only": false, "thinking_override": null},
  "token_remaining": 480000
}'

agent/skill/complexity: Phase 2 (null→-). pipeline: the Phase 2 pick, passed so the marker carries it; omit when null. The builder validates each name against its index, then emits this exact action contract once per callable skill, primary first with ordered stack de-duplication: Call the Skill tool with \skill-name`.Shared-pattern stack entries remain prompt injections. Agents and pipelines stay out of Skill-tool calls. Fan-out: one call per agent, sameskill/pipeline. model: **required Medium+**; use inheritby default. Explicit overrides followreferences/model-selection.md. providerdescribes the active harness (anthropic|openai|other), not an installed directory.health: -(in-context weights read retired —docs/route-loop-validation.md). fallback_reason: **required when agent=general-purpose** — the one-line reason from the Agent-greediness gate, any prose; build-dispatch.pyslugifies it and appendsfallback=<slug>to the marker so every fallback is countable. Dispatch fails without it.stack: Phase 3. task_spec: mandatory Simple+ (Phase 3 Step G); the script rejects an empty spec at Medium+; creation+"match ADR". thinking_override`: slow=security/arch/5+files; fast=lookups.

[do-route] = SOLE signal for routing-decision-recorder. Sub-agents excluded.

Fallback: [do-route] agent={a} skill={s|-} complexity={c}[ pipeline={p}] health=- model={m|-}, Task Spec inline, dispatch.

Model selection. Default to model: "inherit". Omit model_policy, model_effort, and tool-level model/effort overrides. Do not pass the word inherit to an agent tool as a model name. This uses the current session model when the harness supports inheritance. If it cannot, report the limitation rather than silently selecting another model.

The marker records the requested selection, not an observed worker model. The actual model remains unknown unless the harness reports it. Do not infer session identity from installed script directories or historical model tables.

Medium+ must provide model: "inherit", a supported explicit model, or a policy. For a deliberate override, load references/model-selection.md; existing provider policies and explicit choices remain supported. Use scripts for deterministic work. Change model or effort only for a concrete task need or a missed acceptance criterion. Session configuration stays under the user's control. Codex prompts stay read-only and public unless the task requires otherwise.

Complex (3+ sources):

Simple/Medium: direct. Feature-branch; mods commit. isolation:"worktree"→flags.worktree. Non-org: 3 reviews→fix→PR. Org: confirm git.

Step 3: Multi-part / fan-out — deps sequential; independent parallel (max 10). Phase 2 agents → ONE build-dispatch.py call and ONE Agent dispatch per agent: N agents = N calls = N markers, one marker each. Emit the parallel Agent calls in a single message. Each agent gets its own files and scope. Sequential stages pass relevant prior results and evidence locations; synthesis receives the findings and access to evidence from every required stage. Packing several markers into one Bash/Workflow script keeps them recorded but forfeits route-fit scoring, which reads a lone marker per event.

Step 4: Auto-Pipeline Fallback (no match, Simple+) — auto-pipeline. None → closest+workflow. Never empty skill.

Lazy-completion check. "Done" on enumerable → compare scope; short → reject, re-dispatch (references/lazy-completion-detector.md). Re-dispatch → route failure.

Gate: Agent invoked, results delivered.

Routing telemetry (automatic)

Hooks record routing. On an observed route failure, or to re-derive a quoted telemetry figure → load ${CLAUDE_SKILL_DIR}/references/routing-telemetry.md (hooks table, outcome fidelity, route-failure protocol, figure queries).

Error Handling

On any routing error → load ${CLAUDE_SKILL_DIR}/references/error-handling.md.

References

• ${CLAUDE_SKILL_DIR}/references/progressive-depth.md
• agents/INDEX.json, skills/INDEX.json
• skills/workflow/SKILL.md, skills/workflow/references/pipeline-index.json
• scripts/routing-manifest.py