Sol-Luna-Terra Router

概述

Sol Luna Router

Use the parent task as the coordination and acceptance layer. For the intended role split, the parent should run on gpt-5.6-sol; do not use Luna or Terra as the final decision-maker. The skill must remain usable without preinstalled custom-agent files: select worker models explicitly when spawning them.

Route the task

1. Inspect enough of the repository to determine whether the request is clear and bounded.
2. Keep requirements clarification, design choices, architecture, complex diagnosis, code review, scope decisions, and final acceptance in the parent task.
3. For a bounded implementation, testing, local bug fix, or mechanical refactor, create a task card and spawn an implementation worker with model gpt-5.6-luna and reasoning effort medium. If the host already provides a compatible luna-executor custom agent, it may be used instead. Use no more than three workers at once, and only when their write sets do not overlap.
4. Wait for every delegated worker before integrating results. Do not delegate a task merely to avoid making an architectural or product decision.

Task card

Every Luna or Terra delegation must state:

• Objective: the desired behavior or concrete change.
• Allowed scope: files or modules it may alter.
• Non-goals: behavior and files that must remain unchanged.
• Acceptance criteria: observable behavior and required checks.
• Validation: exact test, build, lint, or reproduction commands when known.
• Return format: status, changed files, validation result, failure evidence, and risks.

Tell a Luna worker to make only the assigned change, run the specified validation, and perform at most one focused self-repair. Require this result schema:

Status (complete | incomplete | blocked); Failure category (implementation | transient | requirements | architecture | permissions | environment, if not complete); Changed files; Validation commands and results; Failure evidence; Risks or blockers.

Luna → Terra escalation

Treat a Luna result as needing recovery when it reports incomplete, reports blocked with failure category implementation, or when its required validation still fails after one self-repair attempt. Also escalate when parent verification finds a reproducible defect, missing required coverage, or a task-card boundary violation.

• A transient tool, network, or package-fetch error gets one Luna retry before escalation.
• Requirements ambiguity, architecture conflicts, missing permissions, and environment failures use blocked with their matching failure category. They are parent decisions, not Terra escalation triggers.
• For an escalation, spawn a recovery worker with model gpt-5.6-terra and reasoning effort high. If the host already provides a compatible terra-recovery custom agent, it may be used instead. Give it the original task card, Luna's report, changed files, validation output, and current workspace state. Ask it to continue from the existing work and make the smallest defensible repair; it must not discard correct Luna work or restart the task without cause.
• Require the Terra result schema: Status (complete | incomplete | blocked); Escalation reason; Failure category; Recovery changes; Changed files; Validation commands and results; Failure evidence; Remaining risks or blockers.
• After Terra returns, the parent inspects the diff and reruns relevant validation. If it still cannot meet acceptance, the parent diagnoses, replans, or reports the concrete blocker to the user.

Completion standard

The parent task owns the final response. Report which worker handled the task, any escalation reason, changed files, validation evidence, and remaining risk. Preserve all normal approval, permission, and safety boundaries.

For optional persistent Codex defaults and named-agent installation, read [references/codex-setup.md](references/codex-setup.md) only when the user asks to install, configure, migrate, or publish this skill.