/delegate: who does the work, and on which model

You are the main session. Your time and tokens are the scarce resource. Before doing work yourself, check whether a delegate below does it better or cheaper.

The delegates

Work	Delegate	How	Key
Code: implement a task or a fix	OpenCode GLM-5.3 Flash, then MiMo V2.6 Flash, then a Claude Sonnet agent	/run-task → tools/harness/implement.mjs	Z.AI's login (opencode auth login) and OpenCode Go's (opencode console login) to implement; OpenAI's (opencode auth login) to review
Review a PR	GPT-5.6 Luna on OpenAI, through OpenCode (default), then a Claude Opus agent; never the implementer's family	/run-task → tools/harness/review.mjs, or Agent	as above
Repeated decisions with known answers (triage, routing, gating, labelling)	Jev	/jev → tools/harness/jev.mjs	OPENROUTER_API_KEY
Images: sprites, icons, maps, mockups, illustrations	An image model on OpenRouter	pick one with models.mjs openrouter --output image	OPENROUTER_API_KEY
A second opinion from another model family, or reading audio or images	A model on OpenRouter	models.mjs openrouter --input image (or audio)	OPENROUTER_API_KEY
Speech, voices, sound effects, music	ElevenLabs	models.mjs elevenlabs	ELEVENLABS_API_KEY
Judgment that needs a written reason, design, planning	You, or a Claude agent	—	—

The session-start report says which keys are set. A delegate without its key is off; do the work another way, or tell the user what is missing (docs/environment.md).

Picking a model

1. Never from memory. Model ids and prices change faster than any document. List them live:

   • node tools/harness/models.mjs openrouter --output image;
   • node tools/harness/models.mjs elevenlabs;
   • opencode models opencode-go for implementers (OpenCode Go), opencode models openai for the reviewer.

   Do the same when you are unsure a model still exists. (L25) To change the implementer's or reviewer's model itself, use /switch-model.

2. Check the quota first (L50). Where quota-tracker runs (curl -sf localhost:8765/health), curl -s localhost:8765/avoid lists the providers out of quota and when each is usable again, and curl -s 'localhost:8765/recommend?tier=heavy' (or tier=light) ranks the models to use: pick is the best, ranking lists them with score, spare calls per day until the pool's reset (negative: it runs out before then at the current demand) (docs/environment.md). Never rank by headroom: pools differ in size and period, and Claude sessions draw on some. Skip a model whose provider is exhausted: name the next model of its chain that has quota (--model for implement.mjs, --reviewer for review.mjs) and say so in the run's report. Both scripts also skip such a model themselves and log why (L52); never pause a provider by editing harness.json for quota. When a task needs both an implementer and a reviewer, use /recommend's pair field (curl -s 'localhost:8765/recommend?tier=heavy' | jq .pair): pair.implementer and pair.reviewer, the reviewer from another family (claude, openai, glm, mimo, minimax; each free model its own). Add review_tier=light for a lighter reviewer; private=0 is for code that is not private (the free OpenRouter models may then be in the pair). With tier=heavy, the same response also offers free_reviewer — a free second opinion on the shared 1,000-requests-a-day allowance, lost at 00:00 UTC. Never use a free model for private or client code, secrets or NDA material. A null reviewer means no other family has quota: tell the owner. Project exclusions apply on top: if the suggested reviewer is one the project rules out (rule 3, the family rule), take the next row of ranking whose reviewer is from a different family than the implementer. Alibaba is never recommended and never appears in pair.

3. Choose on the evidence you have. In order of preference:

   • this project's own measurements (review rounds and tokens per task, from the merge comments);
   • a small trial;
   • the listed price and capabilities.

   Vendor benchmarks come last.

4. Record the choice where it is used. Implementer and reviewer models go in harness.json, and a task's own choice goes in its task file, with the date and the reason in the commit message. A generated asset gets a sidecar next to it, <asset>.json, with the provider, the model id, the prompt, the date and the cost. That records where it came from, and lets it be made again.

5. Say the cost first. Before a batch (many images, a long narration), tell the user the estimated cost from the listed prices, and ask when it is more than a few dollars.

Making media

Call the provider's API directly, with fetch from Node. Read the API reference for the endpoint before the first call in a project, because request shapes change. Then keep the working call in a small script under tools/, so the next asset is one command. Starting points (check them first):

• OpenRouter images: POST https://openrouter.ai/api/v1/chat/completions, with an image-output model and "modalities": ["image", "text"]. The image comes back in the assistant message.
• ElevenLabs: POST https://api.elevenlabs.io/v1/text-to-speech/<voice_id> for speech, and POST https://api.elevenlabs.io/v1/sound-generation for sound effects. Both use the xi-api-key header and return audio bytes.

Generated assets are project content: commit them with their sidecar, and review them the way the project reviews its UI (a person looks at the screen).

Never

• Name a model id you have not just seen in a live list.
• Send private repository content or personal data to a provider the user has not cleared for it.
• Spend on a batch without saying the cost first.

A rule's carve-out (process.md §10)

When a project's CLAUDE.md loosens a rule by carve-out, three changes merge before the first task that uses it: the known-good table in a finding, the gate in tools/ that checks every record against it, and the task file for the first run, which names the carve-out's lettered preconditions as Done-when lines. Delegate a model-driven task only once the gate accepts the implementer's run records; such a run corroborates the owner's runs, never replaces them, and its results amend the wording one claim at a time, never in bulk. (L63)