Sol Luna Router

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 metadata.

Every invocation also appends one privacy-safe record to $CODEX_HOME/state/sol-luna-router/runs.jsonl (normally under ~/.codex). It captures the commander session ID, Luna thread ID, token usage, duration, profile, warnings or failure class, and a prompt fingerprint, but not prompt text, final response text, or raw errors. The current Codex session is detected from CODEX_THREAD_ID; use --parent-session-id only when an explicit override is required. Use --run-log to select another absolute ledger or --no-run-log for an intentional one-off opt-out. A ledger write failure is reported without discarding a successful worker result. Read references/run-log.md before analyzing or exporting the ledger.

Use --allow-non-git only when the user explicitly wants work outside a Git repository. Raw events are off by default because they can contain task and answer content. Use --events-file /absolute/path/events.jsonl only when a durable raw trace is explicitly needed. The path must be absolute and new; the runner writes mode-0600 JSONL there while Luna runs and emits a heartbeat to stderr every 30 seconds.

4. Verify and review

1. Inspect the actual diff and changed-file list; reject out-of-ownership edits.
2. For read-only work, inspect command side effects and use no-write settings, external caches, or a disposable copy.
3. Run required builds, type checks, and focused tests in the current session.
4. Run broader checks when Luna returns requires_commander_verification; never weaken tests.
5. Compare Git status before and after read-only verification; generated artifacts fail the no-mutation check.
6. Review correctness, security, data integrity, error handling, and missing coverage.
7. After verification, append the outcome using the run_id returned under telemetry:

python3 <skill-dir>/scripts/run_luna_worker.py annotate \
  --run-id <run_id> \
  --outcome verified \
  --checks-passed <count> \
  --checks-failed 0

Use needs_correction, blocked, or rejected instead when that is the evidence-backed result. verified requires at least one fresh passing check and no failed checks. The annotation is append-only and contains no free-text notes. Summarize accumulated reliability, quality, and cost:

python3 <skill-dir>/scripts/analyze_run_log.py --format json

Treat the report as observational evidence. Its token totals cover Luna only, not the Sol commander. Use comparable task cohorts or controlled A/B benchmarks that include both agents before claiming that routing caused an efficiency improvement.

When evaluating router effectiveness, read the 2026-08-12 transport-warning benchmark and its machine-checkable record for measured scope, arithmetic, and claim limits.

Optional historical credit estimate

Credit estimation is opt-in. Without --rate-card, the analyzer does not estimate credits and keeps the existing token-only report behavior. The bundled card is a dated benchmark assumption, not current pricing:

python3 <skill-dir>/scripts/analyze_run_log.py \
  --run-log /absolute/private/runs.jsonl \
  --rate-card <skill-dir>/references/rate-card-2026-08-05.json \
  --format json

The card calculates Luna worker credits from input_tokens, cached_input_tokens, and output_tokens; input_tokens includes cached input, so uncached input is the difference. Every run, including failed runs, is costed when it has exact valid usage. Missing usage is unresolved rather than zero; malformed, negative, or inconsistent supplied usage is excluded and reported as unresolved. Worker-only normalized metrics are null when worker usage coverage is incomplete. gpt-5.6-luna remains fixed at max reasoning; the estimate does not change routing or reasoning.

Joining Sol commander usage is a separate explicit opt-in and reads only parent IDs already present in the ledger:

python3 <skill-dir>/scripts/analyze_run_log.py \
  --run-log /absolute/private/runs.jsonl \
  --rate-card <skill-dir>/references/rate-card-2026-08-05.json \
  --codex-sessions-root /absolute/private/.codex/sessions \
  --format json

The join reads only session metadata, token-count event timestamps/types, and cumulative token usage. For the union of each parent’s merged run windows, it subtracts the last snapshot at or before each window start from the first snapshot at or after its end. It never copies prompt, response, or raw event text into the report or ledger. Sol preflight before the first run start and work after the last run completion are outside this attribution window. Shared parent windows are charged once; missing baselines/endpoints, counter resets, malformed data, missing sessions, and ambiguous files remain visible as unresolved coverage. Resolved partial commander components may be shown, but commander-plus-worker totals and total-scope normalized metrics are null until every required parent window resolves. A complete total additionally requires every ledger run to have valid worker usage; commander-window coverage alone is insufficient. The runner preserves exact usage on failed Codex exits, failed turns, or timeouts when Codex emitted it, but absent usage remains unresolved. The normalized credit metrics are observational cost-per-outcome measures; controlled A/B remains the causal total-cost proof.

5. Request a correction

When verification or Sol review finds an actionable defect, write a new temporary prompt containing the exact failure evidence. Annotate the original run as needs_correction, then resume the same worker thread:

python3 <skill-dir>/scripts/run_luna_worker.py resume \
  --cwd /absolute/path/to/repo \
  --thread-id <thread_id> \
  --prompt-file /absolute/path/to/correction.md \
  --profile <original-profile>

The resume command reasserts Luna, Max reasoning, disabled native agents, and the selected sandbox; it does not rely on the historical thread configuration. Repeat verification after every correction. Do not open a new worker thread unless the previous thread is unavailable or the task has materially changed.

Gotchas and failure handling

• If the runner reports an incompatible or unavailable model, stop and report the exact error.
• If Luna requests broader ownership, network, or permissions, return it to the user or revise the plan; do not grant it silently.
• On timeout, retain the partial thread_id, events path, and any emitted usage; resume with a smaller prompt, or treat a run without a thread ID as unrecoverable.
• Treat capacity_exhausted as infrastructure capacity, not task quality; never lower Luna effort automatically.
• Treat malformed JSONL, nonzero exit, turn.failed, missing completion, or missing final response as failure; recovered transport errors remain warnings.
• Do not claim the Skill is effective from invocation count, worker success, or token totals alone; require commander evaluations with fresh check evidence and report the sample size.
• If unrelated user changes block safe verification, report the boundary instead of reverting them.
• If bounded-review requests or starts an unbudgeted full suite, stop the run and tighten the task instead of increasing its timeout.