Codeck

Codeck

Use Codeck as the execution bridge while the current coding agent remains the host and moderator. Treat a host product's model selector as unrelated: switching the active chat model is not the same as consulting external models through Codeck.

Choose the workflow

• Use a single ask for a read-only second opinion from a named executor.
• Use a council for independent opinions from two or more genuinely different underlying model providers.
• Use delegate only when the user explicitly asks an external executor to implement changes.
• Use pick before auto when routing is uncertain.

When a routed tool result has status: "running" or status: "pending", treat it as an in-progress job rather than a model answer. Call wait_run in bounded intervals (no more than 45 seconds per call) until the run is terminal, then use its output. A short MCP call can be cut off by the host while Agy or another executor continues in the background; use the returned runId to recover the final answer.

Never describe one model playing several roles as multi-model consensus. Label it single-model-multi-role.

Preflight

1. Run command -v codeck.
2. If unavailable, stop and explain that the Codeck CLI is required. Do not install it without user authorization.
3. Run codeck doctor and codeck list in the target project.
4. Use only healthy, configured executors whose declared mode and permissions fit the task.
5. If the project is not initialized, ask before running codeck init because it creates .codeck/ files.

Read references/providers.md when configuring executors, choosing providers for a regional network, or explaining why a requested model is unavailable.

Prepare the handoff

Write a compact Markdown brief for non-trivial work. Include the objective, constraints, questions, acceptance criteria, and only the files needed by the external model. Prefer a file attachment over embedding a long brief in the command:

codeck ask <executor> "Read the attached task brief and return an evidence-backed review." -f task.md

Do not include secrets, credentials, unrelated source files, or private user data. Tell the user when task content will leave the host product and be sent to another provider.

Read references/artifacts.md for the input brief and final artifact schemas.

Run a multi-model council

1. Keep the current coding agent as moderator.
2. Select at least two healthy executors backed by distinct providers. Prefer capability fit and availability over brand order.
3. Give every first-round expert the same facts and core question, plus a distinct review role.
4. Collect first-round answers independently before exposing one expert's answer to another.
5. Extract claims, evidence, assumptions, risks, and disagreements.
6. Ask targeted follow-ups only about material disagreements. Attach the relevant prior outputs as Markdown files.
7. Stop when the acceptance criteria are met, remaining disagreements are explicitly documented, or the configured round limit is reached.
8. Produce report.md, summary.md, and decisions.json when writing artifacts is in scope; otherwise return the same structure in the response.

Use codeck compare <executor-a>,<executor-b> "<task>" -f task.md when all experts should receive the same prompt. Use separate codeck ask calls when roles or follow-up questions differ. Do not claim parallel execution unless the observed run actually ran concurrently.

Read references/moderation.md for role prompts, disagreement handling, and convergence rules.

Route safely

• Default to ask, which is intended for read-only consultation.
• Do not add -y or call delegate unless the user authorized external implementation.
• Treat permission fields as declarations unless the underlying CLI also enforces a sandbox.
• Do not silently substitute a provider. Report the requested executor, actual executor/provider, and fallback reason.
• Do not hardcode current model IDs. Discover or read them from local configuration because provider model names change.
• If fewer than two distinct providers are available, continue only as a single-model consultation and state the limitation.
• Preserve raw external outputs or Codeck run IDs so the synthesis remains auditable.
• Codeck archives each routed request inside the project. Use list_runs or search_runs to find prior work, and get_run(includeContent=true) only when the redacted prompt/context is actually needed. Do not inject archive history automatically into a new handoff.
• When a run is genuinely reusable knowledge, call curate_run with focused tags and a note. Curation is explicit; do not mark every historical run as knowledge.
• Use replay_run only when the user wants a new model call. Historical-snapshot replay is the default; currentContext=true is an explicit alternative. Every replay creates a new linked run.

Return the result

Lead with the moderator's conclusion. Then state:

• which executors and underlying providers actually ran;
• where they agreed and disagreed;
• what evidence supports the decision;
• unresolved risks and recommended verification;
• any fallback, failure, estimated usage, or regional limitation.

Keep the short summary decision-oriented. Keep detailed reasoning and source attribution in the full report.