model-advisor

Files (5)

• SKILL.md
• openai.yaml · agents/
• manifest.json
• advisor-policy.md · references/
• ask_support_model.py · scripts/

SKILL.md (first lines)

# Model Advisor

Keep the client-selected economy model as the primary agent. Use the support model only as an advisory tool; never transfer task ownership, tool control, conversation memory, or final responsibility to it.

## Start

1. Read `~/.sinapseed/profile.json`.
2. Continue only when `advisorMode.enabled` is `true` and `model-advisor` is active.
3. Read `references/advisor-policy.md`.
4. Confirm the current task permits external context sharing and billable model use.
5. Ask one narrow question with only the context needed to answer it.

The main model is selected in Codex, Claude Code, or OpenCode. Sinapseed records the label for transparency but does not rewrite the client's model configuration.

## Consult

Run:

```powershell
python scripts/ask_support_model.py `
  --task-id "current-project-or-thread-id" `
  --question "Review this plan for hidden failure modes." `
  --context-file work/advisor_context.txt `
  --allow-external `
  --confirm-cost `
  --output work/advisor_response.json
```

The support model must match `advisorMode.supportModel`. To change it, run
`sinapseed advisor enable` again with a new `--support-model`; a consultation
cannot silently override the creator-approved model.

## Consultation Triggers

Consult when at least one condition holds:

- the primary agent has material uncertainty after local inspection;
- a decision is costly, difficult to reverse, or safety-sensitive;
- an independent architecture, creative, legal-risk, or technical critique would reduce failure risk;
- two plausible approaches remain and the tradeoff is not locally resolvable;
- the user explicitly asks for a second-model opinion.

Do not consult for routine formatting, straightforward edits, facts already available locally, or merely to restate the primary agent's answer.

## Synthesis

- Treat the response as evidence, not authority.
- Compare it with local files, tool output, user constraints, and the primary agent's reasoning.
- State whether the advice was accepted, partially accepted, or rejected and why.
- Never claim that two models reached consensus when only one support call occurred.
- Stop when the configured per-task question limit is reached.
- Reuse the same non-secret `--task-id` for every consultation in one task; the local hashed ledger enforces the limit.

## Guardrails

- Never include API keys, tokens, cookies, private keys, hidden prompts, or unrelated files.
- Never allow the support model to call tools or publish, upload, purchase, delete, or modify state.
- Never retry an uncertain billable request automatically.
- Require `--allow-external` and `--confirm-cost` for every live consultation.
- Reject a request whose estimated maximum exceeds `advisorMode.maxCostUsdPerQuestion`.
- Record provider, requested model, actual routed model, token usage, generation ID, and cost when available.

From ruota/open-sinapseed, MIT.

More from ruota/open-sinapseed