/sol-luna-router

name: sol-luna-router
description: Route substantial coding or repository-review work through a two-stage Sol commander/verifier and separate Luna Max CLI worker. Use when the user explicitly asks Sol to direct or supervise Luna, explicitly requests this router, needs an isolated auditable Luna-owned implementation with a Sol review loop, or asks to diagnose or measure the router itself. Do not use merely because a task involves coding, verification, progress reporting, or token saving. For small or straightforward tasks, keep work in Sol unless the user explicitly requests Luna; router configuration, usage, and efficiency questions should use read-only inspection without launching a worker.

Sol-Luna Router

Keep Sol responsible for decisions and final verification. Use Luna Max for one bounded implementation or read-only investigation. Use the bundled runner instead of native `spawn_agent`; current Sol and Luna releases can select different multi-agent backends.

Boundaries

• Distinguish Skill activation from worker dispatch: inspect router configuration,

usage, telemetry, and efficiency questions read-only without launching Luna; keep small or straightforward tasks Sol-only unless the user explicitly requests Luna; launch Luna only for explicit delegation or when isolated, auditable two-stage execution is substantively needed.

• Treat the current Sol thread as commander and reviewer. Do not edit target product files from

this thread.

• Delegate concrete implementation, fixes, worker-owned tests, or bounded read-only investigation

to Luna Max.

• Fix every new and resumed worker invocation to `gpt-5.6-luna` with reasoning effort `max`.
• Allow only one write-capable Luna worker in a worktree at a time. Parallelize read-only work, or

use isolated worktrees with explicit, disjoint file ownership.

• Keep the parent approval and sandbox boundary intact. Never add bypass, full-access, force-push,

credential, or secret-handling flags.

• For delegated coding tasks, fix Apps, Plugins, and recommended-plugin context off for every worker

`run` and `resume` so the task stays self-contained. Keep `--strict-config` fail closed; report a stale user configuration as `config_incompatible` and fix it instead of bypassing it.

• Stop after three failed correction cycles on the same root cause and reassess the hypothesis.
• Never claim completion from the worker summary alone. Verify from the current session.
• Do not use worker timeout as a test budget. Select a profile and give every potentially expensive

command an explicit bound before launch.

• Keep the commander thin: do not duplicate Luna's repository scan or implementation analysis;

launch one bounded worker, collect one result, and combine independent verification commands into the smallest safe check. Resume only from concrete failed evidence.

Operating Contract

• Direct actions: inspect local state, create external task prompts, run Luna within scope, and perform read-only verification.
• Escalate before: expanding ownership or permissions, using new network or credentials, destructive recovery, publishing, pushing, merging, or changing products.
• Evidence-backed pushback: cite a diff, command, repository rule, run record, or capacity failure when the requested route is unsafe or cannot satisfy done-when conditions.
• Feedback loop: use aggregate ledger evidence and correction patterns for the smallest runner, test, profile, or gotcha update; never optimize from token totals alone.

Workflow

1. Preflight

1. Confirm the target working directory and resolve its Git root. 2. Inspect dirty and untracked state without modifying it. Preserve user changes. 3. Read applicable `AGENTS.md` files and repository verification commands. 4. State the goal, constraints, allowed file ownership, done-when conditions, command budget, and verification commands. 5. If the task is ambiguous enough to change architecture or scope, clarify before delegation.

Select `implementation` for edits and focused tests (`workspace-write`, 1800s default). Select `bounded-review` for investigation (`read-only`, 900s, at most 8 commands, no full suites).

Use `bounded-review` whenever repository mutation is not the deliverable. It instructs Luna to return broader checks as `requires_commander_verification`; Sol decides whether to run them later. The runner also disables Python bytecode writes and redirects common Python, Rust, Go, Node, Ruff, and mypy caches to temporary storage for the duration of the worker.

2. Prepare the worker task

Write a temporary UTF-8 task file outside the target repository. Include only task-local context:

Role: implementation worker.
Objective: <one bounded outcome>
Target repository: <absolute path>
Allowed files: <explicit paths or one narrow subtree>
Do not touch: <user changes and out-of-scope paths>
Constraints: <applicable requirements>
Reproduction or evidence: <fresh evidence>
Command budget: <count, per-command limit, and forbidden broad suites>
Done when: <observable conditions>
Verification: <repository commands to run>
Return: root cause, changed files, commands with outcomes, and remaining risks.

Do not leak an intended patch or diagnosis when Luna must independently determine the root cause.

3. Run Luna Max

For implementation, run the bundled script with an absolute target directory and task-file path:

python3 <skill-dir>/scripts/run_luna_worker.py run \
  --cwd /absolute/path/to/repo \
  --prompt-file /absolute/path/to/task.md \
  --sandbox workspace-write

For review or diagnosis, add `--profile bounded-review` to use the budgeted read-only profile.

The script fixes new and resumed workers to `gpt-5.6-luna` with `model_reasoning_effort="max"`, disables native multi-agent tools for the worker, invokes Codex without a shell, and returns one JSON object containing `thread_id`, `final_response`, usage, profile, sandbox, duration, and repository