orchestrate

Files (1)

• SKILL.md

SKILL.md (first lines)

# orchestrate

Combined orchestration: get work done while spending as little paid Claude usage as possible.
The orchestrator plans, briefs, routes, and reviews. It never writes implementation code
itself.

## The model

- **Orchestrator** = Claude Opus in this session. The only component that spends Claude
  tokens. Holds the plan, writes each brief, routes each task, adjudicates reviews.
- **Tier B** = the delegation pipeline (`~/.claude/bin/delegate`, `free` / `deepseek` /
  `nvidia` / `openrouter`). Zero Claude tokens. **Default tier for both implementation and
  review.** `free` (OmniRoute) is the only $0-unlimited local tier; `nvidia` / `deepseek` /
  `openrouter` are remote (credits, or free-with-daily-limits).
- **Tier A** = Claude subagents (Agent tool: `haiku` / `sonnet` / `opus`). Costs Claude
  tokens. Escape hatch only.

## Routing (per task, three axes)

Any single axis landing in the escalate column sends that task to Tier A; otherwise Tier B.

- **Complexity:** anything specifiable and diff-verifiable, from boilerplate up to a whole
  module or a real refactor -> Tier B (do not pick a backend, see below); needs broad
  codebase judgment -> Tier A `sonnet` / `opus`.
- **Iteration depth:** verifiable in ~one shot -> Tier B; long autonomous run-fail-edit loop,
  or needs a live service / Docker / the app running -> Tier A subagent.
- **Sensitivity:** ordinary code -> Tier B; security-sensitive or full-context debugging ->
  Tier A or the orchestrator itself.

### Tier B backend: let it default

Once a task is Tier B, **do not pick a backend**: run `delegate "<task>"` and let the CLI
default to `free` and escalate to `deepseek` itself. Measured on `bench/tasks.json`
(2026-10-01), reviewing a `free` diff costs the same as reviewing a paid one, so `free` is
the less reliable tier, not the weaker one, and hand-routing by difficulty is what left it
unused. Name a backend only for the specific jobs below:

- **`free` (OmniRoute, $0 local, unlimited):** the default, applied for you. Slow; the pool
  drains under sustained load, which the automatic escalation now covers. Name it explicitly
  only to force $0 and fail rather than fall back.
- **`openrouter` ($0 `:free`, rate-limited):** two jobs. (1) **free-remote fallback** when
  OmniRoute is dry, preferred over `nvidia` so nvidia's finite credits stay in reserve.
  (2) the **pin lever**: a specific fast/quality free model, one not on the nvidia account, or
  a **vision** model (`--model ...-vl:free --image <path|url>`). `:free` ids are rate-limited
  (~50/day at $0, ~1000/day with ~$10 credit) and slow (~3 min/trivial), so not for high
  volume or when you are blocking on the result.
- **`nvidia` (free trial credits, finite):** a **stronger free model** than a `:free` id gives,
  when openrouter's free models are too weak but you still want it free. Second to openrouter
  for ordinary free work because the credits are exhaustible. `--list-models` first, the
  catalog churns.
- **`deepseek` (cheap, fast, reliable, ~Sonnet-class):** default for **substantial
  well-specified work** (a whole module, a real refactor, a non-trivial first draft) and for
  **reviews**. Go straight here, do not burn time on weak/slow free models, when the logic is
  real or the result must be dependable.
- **`kimi` (premium):** only when the user explicitly asks.

One line: mechanical -> `free` -> (dry) `openrouter:free` -> `nvidia`; substantial or
reliability-critical -> `deepseek`; need a pinned or vision model -> `openrouter --model`
(`--image` for vision); reviews -> `deepseek`. Confirm any pick with
`delegate <backend> --list-models`; the delegation-pipeline repo's `MODELS.md` has the
current shortlist.

**Cost gate (your call).** Delegation spends *your* tokens on the brief, the review, and any
re-run. Before routing to Tier B, weigh that against doing the task inline; if the brief plus
review would cost as much as or more than the edit itself, do it inline. The pipeline only
saves tokens when the work is bulkier than its description, so one-liners stay in-session.

## The loop (per task)

1. **Brief.** Tight and self-contained: exact files, the change, a pattern to mirror, exact
   values. Save to `brief.md` so review can cite it. The worker has no conversation context.
2. **Route and run** with verify + commit, so you get back a tested, committed result:
   ```
   ~/.claude/bin/delegate --verify "<test/typecheck cmd>" --commit "<msg>, per brief" "<brief>"
   ```
   Name no backend: it defaults to `free` and falls back to `deepseek` by itself.
   `--verify` runs after editing; a non-zero exit prints output and exits 2 without
   committing. **Exit 3 means the worker edited nothing**, so verify was not run at all:
   that is a brief it could not act on, not a test failure. `--commit` fires only on green.
   On Windows `--verify` runs through `cmd.exe`, so pass one command, not a `&&` / `;` chain.
   Allow minutes: the free tier is slow and it is the default.
3. **On red** (exit 2, no commit): read the printed output and either re-delegate with the
   failure folded in as a sharper spec, or escalate that task to a Tier A subagent.
4. **Review (Tier B):** `git show HEAD > review.diff`, then
   `delegate deepseek "Review the diff in review.diff against the brief in brief.md; report spec compliance and quality, most severe first. Do not edit."`
   Adjudicate yourself: accept, fix inline, or re-delegate. Clean up scratch files after.
5. **Escalate to Tier A** only when the loop cannot converge or the task hits an axis above.

## Guardrails

- Never write implementation code yourself when a tier can do it.
- The worker **model** never runs shell or git. Only the caller's `--verify` / `--commit`
  flags run commands, so the no-shell / no-git trust boundary stays intact.
- Delegating sends repo content to external providers. Keep sensitive repos, or a session the
  user told you not to delegate, in-session on Tier A or the orchestrator.

For a single self-contained task, the `delegate` skill is enough. Reach for this when there is
a plan to route across tiers.

From aayushpokhrel1/delegation-pipeline, MIT.

More from aayushpokhrel1/delegation-pipeline