Workflow Engine

The reusable execution body behind /craft:orch:workflow. The command owns args (--dry-run, --resume, --refine); this skill owns the work.

Unlike drive-engine (improvises "what next" each turn) the control flow here is fixed in the definition — only the count of agents in a parallel stage flexes to upstream data. That determinism is real only because the mechanical steps are delegated to scripts/workflow_parse.py, never improvised. You orchestrate and judge the advisory semantic layer; the Python core decides everything that must be reproducible.

Hard rule — delegate the deterministic mechanics

Never eyeball these. Always shell out to the core:

Agent Prompt Composition (Lever B — prompt-trim)

Every subagent prompt MUST contain exactly three parts — nothing more:

1. Spec slice — only the portion of the WORKFLOW definition scoped to THIS agent's files/phase. Do NOT pass the entire WORKFLOW-*.yaml to every agent. Extract and pass only the stage block(s) the agent needs to execute.
2. Summarized prior outputs — a brief structured summary of earlier stage results (~200 tokens max per stage). Do NOT include full transcripts or raw prior-agent outputs. Derive the summary from the structured return values cached in .craft/workflow-runs/<run-id>/.

3. Structured-return instruction — end every prompt with exactly:

   Return a concise structured summary:
   { files_changed: [...], tests_run: bool, tests_passed: bool, issues: [...] }
   Do NOT include full file contents or verbose explanations in your response.

Anti-patterns (explicitly forbidden)

• Passing full WORKFLOW-*.yaml to every subagent — use spec slice instead. A 300-line workflow YAML passed to 8 parallel agents = 2400 lines of wasted context; the agent needs only its own stage block.
• Passing raw prior-agent transcripts — use structured summary instead. Transcripts can be 10–50× larger than the information they carry; always compress to the structured-return schema before forwarding.
• Asking for long narrative responses — use the structured-return instruction instead. Narrative responses are unstructured, hard to gate, and inflate downstream context for every subsequent agent in the wave.

Responsibilities

1. Compile — run the parser on the WORKFLOW-*.yaml (or shape-DSL) file to get the canonical wave plan. The YAML and shape-DSL forms compile to an identical plan given identical upstream outputs (D1).
2. Per wave, resolve fan-out — for a parallel stage, call resolve_fanout against the real upstream outputs. An empty bound array hard-aborts the run naming the upstream stage (D6/FR8) — never silently skip it.
3. Dispatch under the semaphore — for each bound item, sem_acquire against the run-wide ceiling before launching a file-scoped subagent; refuse (queue) when at the ceiling; sem_release on completion. The counter is the file semaphore.count, not context state, so it survives compression (D5).
4. Gate every output (hybrid, D2) — gate_output returns {ok, structural_errors, semantic_warning}. ok=False (a structural miss) fails just that branch with a structured marker (failure isolation); the run continues and synthesize receives partials + error markers. Then judge the output's semantic plausibility yourself and pass it as semantic_warning — it is surfaced but never blocks (blocking on LLM judgment would reintroduce the non-determinism D2 designed out).
5. Cache / replay (D4) — write each agent output under .craft/workflow-runs/<run-id>/ keyed by cache_key. On --resume, recompute keys; reuse unchanged stages, and cascade_invalidate the downstream of any changed stage (derive the dependency graph from the plan with build_deps(plan) — never hand-trace it). Replay reuses cached outputs — it does not re-invoke the model for unchanged stages (a fresh agent run is not byte-identical; the spec never claims otherwise).
6. Reconcile at each wave boundary — before opening the next wave, count the actually-live agents and sem_reconcile the counter to it. This heals a leaked count from a mid-run crash between increment and dispatch (the accepted D5 residual risk).

The verify gate (D8 — first class)

A verify stage runs the project's actual verification command and treats its exit status as the authoritative pass/fail. A green transcript is NOT sufficient — the command must really run. This is the drive-engine gate lifted into the engine so drive stays expressible as a workflow definition; keep the semantics a strict superset.

See ../references/verify-gate-detection.md for the full detection table and how to use it.

Run substrate

.craft/workflow-runs/<run-id>/ holds per-agent output JSON, a human-readable manifest.json (wave-by-wave progress + any semantic warnings), and semaphore.count. All are plain files — inspectable mid-run.

Outputs

A structured run result: { run_id, waves: [...], warnings: [...], verify: {command, exit_code, passed} | null }. The skill never opens a PR; on a passing verify it reports verified-green and hands back to the command.

Cache Strategy and Model Routing (Lever C)

Cache Strategy (5-min TTL)

Keep the tool set and system prompt byte-stable across all agents in a single run. Claude's prompt cache has a 5-minute TTL; any change to the tool list or system prompt between agent calls busts the cache and forces a full re-tokenization of every prior turn.

Rules:

• Determine the tool list once at run start and hold it constant for the entire run. Do NOT add, remove, or reorder tools between agent dispatches.
• If two sequential agents share the same system context, batch them in the same turn rather than separate sessions. Separate sessions reset the cache clock even when the content is identical.
• Target cache_hit_ratio > 0.5 for runs with more than 3 agents. This is measurable via scripts/orchestrate-token-report.py after the run completes.

What busts the cache (avoid these mid-run):

• Adding a tool that was not in the initial tool list
• Removing a tool from the initial tool list
• Reordering the tool list
• Changing the system prompt text (even whitespace)
• Starting a new session instead of a new turn

Model Routing (Haiku for cheap stages)

Not every stage needs Sonnet. Route by stage cost:

Cheap stages → model: claude-haiku-4-5-20251001 (10–20× cheaper):

• Grep a file for a pattern
• Count lines or tokens in a file
• Read and summarize a short section (< 200 lines)
• Emit structured JSON from a file under 200 lines

These stages require no reasoning — Haiku handles them correctly and at dramatically lower cost.

Reasoning-required stages → session-default model (do not specify model):

• Architecture decisions
• Multi-file synthesis
• Verify gate semantic judgment
• Any stage requiring cross-file reasoning or prose explanation

Heuristic: a stage is "cheap" if:

1. The prompt fits in under 500 tokens, AND
2. The output is simple structured JSON with no prose reasoning required.

When dispatching a cheap agent, add model: "claude-haiku-4-5-20251001" to the agent call. When dispatching a reasoning-required agent, omit model entirely and inherit the session default.

Routing responsibility: The workflow-engine skill is responsible for model routing decisions. The WORKFLOW yaml does NOT specify models — routing is a runtime concern, not a definition concern.