delegate-kimi-k3

Files (4)

• SKILL.md
• openai.yaml · agents/
• stable-policy.md · references/
• worker-cli.mjs · scripts/

SKILL.md (first lines)

# Delegate Kimi K3

Resolve `WORKER_CLI` to the absolute `scripts/worker-cli.mjs` path inside this installed skill directory. Run every worker command as `node "$WORKER_CLI" ...`. If the file is missing, stop and report a broken plugin installation; never install npm packages dynamically or fall back to an unpinned remote command.

Keep Codex responsible for design, review, acceptance, and integration. Never present Kimi as superior; measure both patch quality and delegation quality.

1. Read `references/stable-policy.md` and the project-local `.external-worker/capability-profile.yaml`. Treat missing or invalid generated data as safe defaults.
2. Run `node "$WORKER_CLI" route "<task>"`. Do not delegate prohibited ownership. For sensitive narrow implementation, require a Sol-defined design, invariants, file scope, executable checks, security review, and correctness review.
3. Write and validate a version `1.0` task-contract JSON. Use argument arrays for every command. Require human review for UI.
4. Require a dedicated non-production Kimi home. Tell the user to authenticate it themselves with `env HOME=<dir> kimi login`, then run `node "$WORKER_CLI" preflight --kimi-home <dir>`. Never install Kimi, handle credentials, or resume a session.
5. Run `node "$WORKER_CLI" run --contract <file> --repo <repo> --worktree-root <outside-repo-path> --kimi-home <dir> --keep-worktree`.
6. Inspect the evidence bundle. Reject contract mutation, scope violation, failed blocking verification, missing required review, or incomplete UI review.
7. Dispatch installed read-only reviewers in parallel where independent. If unavailable, spawn built-in Codex agents with the exact category instructions in `../review-external-worker-run/references/inline-reviewers.md`. Nesting depth is one. Do not give initial reviewers Kimi's summary.
8. Adjudicate evidence: merge duplicates, retain disagreements, reject unsupported claims, identify pre-existing issues, and choose accept, one repair pass, Sol takeover, or discard. Never count findings as votes.
9. For UI, capture evidence, serve the loopback report, and wait for recorded human review. Rendered changes after repair require fresh review.
10. Append the final outcome to the ledger and regenerate the local profile. Never rewrite this skill from evidence.

Allow at most one automatic repair pass. Its contract contains exact failed commands and accepted blocking findings, keeps path scope unchanged, and prohibits unrelated redesign.

From MAbbara/codex-kimi-worker, Apache-2.0.

More from MAbbara/codex-kimi-worker