delegate

/delegate — Fable or Opus orchestrates, Opus implements

The division of labor: you (the main conversation) own architecture, decomposition, dispatch, integration, and everything the user sees. The implementer agent (Opus) turns written work-package specs into code — it is the only agent sanctioned to write files in this repo. The spec-reviewer agent (Opus, read-only) adversarially checks the result.

Phase 0 — Gate & ground

• Orchestrator gate. This workflow requires a Fable- or Opus-family orchestrator (claude-fable-5 or any claude-opus-* id) as the MAIN conversation model — check your own identity in your system prompt. If you are Sonnet, Haiku, or any smaller model: STOP. Tell the user /delegate is gated to Fable/Opus and offer to do the work inline instead. Never switch models to satisfy the gate; when working inline, apply the subagent-discipline skill's rituals to your own work.
• Opus-orchestrator addendum. If you are Opus (not Fable), you are a weaker orchestrator than this workflow was originally tuned for, so apply subagent-discipline to your OWN architecture, decomposition, and integration — not just the subagents' work. Read its references/orchestrator.md § "When the orchestrator is Opus or Sonnet" (you open that file for spec-shaping in Phase 1 regardless). Concretely: keep an epic-level constraint ledger in a scratchpad, and never accept a subagent's "done" as evidence — run the integration checks and read the load-bearing hunks yourself.
• Confirm this is an implementation session, not a grooming/planning session (if the user is researching or thinking out loud, produce plans instead — standing feedback).
• git --no-pager status — record the baseline. Never skip this; the Phase 5 footprint sweep depends on it.
• If the task shapes or polishes UI, run the impeccable skill BEFORE decomposing — its confirmed brief becomes part of the spec (standing feedback; an AskUserQuestion round is not a substitute).

Right-sizing — pick the mode before anything else

Review is never skipped; who reviews scales. Each avoided agent session is ~100K tokens, so this choice dominates cost:

• Lite mode — single package, few files: ONE implementer + your own orchestrator-review of the diff + one integration build. No Explore grounding when you already know the area; no separate reviewer.
• Full mode — genuinely multi-file, parallelizable work: packages + spec-reviewer(s), everything below.
• Threshold is conservative: orchestrator self-review loses the fresh-eyes property that catches idiom and spec drift — when in doubt, use the reviewer.

Phase 1 — Architect (in-loop; this is your job — don't delegate it)

Research with read-only fan-outs when needed: Explore for codebase questions, Plan for competing strategies, Workflow for bigger read-only sweeps (see Workflows below). Every research/review agent prompt must state: strictly read-only — no file writes, no git mutations, no test files. The implementer agent is the sole exception to the file-write rule, and only when executing a spec; the no-git-mutation rule has no exceptions, for any agent, ever.

Then write one work-package spec per package. Keep specs tight — they are replayed into both the implementer's and reviewer's contexts — but never under-specified: a well-specified first turn is what buys the turn-count savings. Before your first spec of a session, read the subagent-discipline skill's references/orchestrator.md — its spec-shaping checklist and Sonnet stage-prompt scaffolding compensate for the dispatched models' documented failure modes.

## Package: <name>
Objective: <one paragraph — what and why>
Files in scope: <explicit list; the implementer must not touch others>
Contracts: <types/interfaces/commands it must expose or consume>
Acceptance criteria: <checkable statements>
Verification: <SCOPED checks only — tsc + biome lint on the package's files;
  the whole-project build runs once, orchestrator-side, at integration>
Docs-sync: <README/site/help/changelog items, or "orchestrator handles">
Out of scope: <adjacent things it must leave alone>

Phase 2 — Decompose

• Package file scopes must be disjoint — that is what makes parallel dispatch safe. Shared contracts (types, src/lib/git/api.ts shapes, Tauri command signatures) go in a first, sequential package; dependents follow in parallel.
• Rust-side and frontend-side of one feature usually split cleanly at the IPC boundary — pin the exact command/payload shape in both specs.
• Fewer/larger packages beat many tiny ones (less integration risk, fewer agent sessions).

Phase 3 — Dispatch

• Spawn implementer (Agent tool, subagent_type: "implementer"), one package per agent, spec included verbatim plus any context the agent can't discover (branch intent, related in-flight packages, impeccable brief).
• Route effort per package (frontmatter high is the default when unrouted): effort: "low" for mechanical packages (docs-sync, changelog, renames); effort: "xhigh" for the single hardest package of a feature. Set via the Agent tool's effort param at dispatch.
• Parallel dispatches go in one message. Track which agent owns which files.
• Iterate with the SAME agent via SendMessage (it keeps its context) — don't spawn a fresh implementer to fix its own package.

Phase 4 — Verify (never skip; never trust the report alone)

1. Full mode: one whole-feature spec-reviewer given ALL the specs + touched-file lists is the DEFAULT for ≤2 coupled packages (ask for per-package verdicts; raise its maxTurns per-dispatch for large reviews — a capped-out reviewer returns a truncated review); use per-package reviewers only for larger fan-outs. Lite mode: you review the diff yourself against the spec and gd-conventions.
2. Run the integration checks yourself in the main loop: pnpm build, the check-only lint, and the cargo suite when Rust changed — package-level green doesn't guarantee the merged working tree is green.
3. UI work: dogfood live via the drive-ui skill — edge cases (first/last/ empty, boundaries) and visual polish, before declaring done.
4. Batch fix rounds: collect ALL findings for a package into ONE SendMessage to the owning implementer (every extra round replays the agent's whole context — ~100k tokens even for a one-line diff). Re-review = the delta + prior findings only, not the full protocol again.
5. Trivial-fix exception (user-authorized 2026-07-02): the orchestrator may apply a trivial fix directly instead of a fix round — ≤ ~3 lines, confirmed by the spec-reviewer or by live validation, and changing no contract, API, or design. Disclose every such fix in the report and re-run the package's verification. Anything beyond trivial still round-trips to the owning implementer; this never loosens the no-git-mutation rule.

Phase 5 — Close out

• Docs-sync complete (README, site, help content, CHANGELOG — per CLAUDE.md).
• Footprint sweep: git --no-pager status against the Phase 0 baseline — every difference is accounted for by a package or the user's own parallel work; nothing stray, nothing staged, no probe/scratch files left behind.
• Report to the user: outcome first, then per-package summary, verification results verbatim, and open concerns. Never commit — the user commits.

When things go wrong

• Package blocked (spec wrong, dependency missing): resolve it yourself — rewrite the spec or implement that piece inline — don't ask the agent to improvise around it.
• Reviewer verdict "rework": the approach is wrong, so rewrite the SPEC first; then redispatch — to the same agent if the correction is communicable, to a fresh implementer if the old approach would anchor it.
• Scope collision (two packages need the same file): serialize them; re-issue the later spec against the post-merge state of that file.
• An agent stops without a usable report: check its file scope with git --no-pager status, then SendMessage "produce your report now"; if resume fails, spawn fresh with the spec + a summary of the prior state.
• An agent mutated git state despite the rules: stop all dispatches, do not attempt cleanup, report exactly what happened to the user and wait for direction.

Workflows (the Workflow tool) — read-only stages only

• Read-only fan-outs via agent(prompt, {agentType: 'spec-reviewer'}) are supported — the same agent registry resolves, with per-call model / effort overrides available. Good for review panels and research sweeps.
• Never dispatch implementer via Workflow. Workflow agents run in the background where no permission prompt can surface — the run burns a full read-and-plan session and dies at its first Edit.
• Background agents get silent DENIALS for anything off the allowlist — tell Workflow-spawned reviewers to treat tool denials as environment limits (note and continue), never as findings or reasons to stall.
• SendMessage-resume is validated only for foreground Agent-tool dispatches; follow-ups to Workflow-spawned agents go via fresh dispatch.

Environment notes (hard-won; trust these)

• Agent definitions normally register at session start (observed hot-loading mid-session once — don't count on it). If implementer / spec-reviewer aren't in the available-agents list, restart the session or use /agents. Skills hot-load reliably.
• First dispatch of a session doubles as a smoke test. Start with the smallest package. If the harness denies the implementer's Write, don't fight it: implement that package inline from the same spec and note the environment limit to the user.
• Subagents receive CLAUDE.md and the session-start memory snapshot automatically; they do NOT see mid-session memory edits or your conversation. Everything a package needs must be in its spec.
• Worktree runs (validated 2026-07-02): fresh worktrees check out CRLF (core.autocrlf=true), so tree-wide biome check fails on formatting noise — the reliable worktree gates are tsc -b, biome lint, and pnpm build. Deps need pnpm install first. If the main checkout advances mid-run, the final patch may not apply clean — re-apply divergent hunks by hand against current HEAD. Clean up with git worktree remove + prune (Windows: node_modules may need cmd /c rd /s /q afterwards).