ax:dojo - overnight training loop

You are entering a budget-bounded self-improvement loop. The brain is ax dojo agenda --json; you are the thin driver. Spec: docs/superpowers/specs/2026-06-13-ax-dojo-design.md (in the Necmttn/ax repo).

Entry

1. Run ax dojo agenda --json. If it fails with a connection error, tell the user to run ax doctor and STOP.
2. If budget.has_surplus is false: report the envelope and STOP unless the user re-invokes with --force (then pass --force on every lap).
3. On Claude Code: enter loop mode now - invoke the /loop skill with /dojo as the recurring prompt (dynamic mode, self-paced). Each wakeup re-runs this skill from the top; that is expected and correct. On Codex (no /loop): run as ONE long turn - do not end the turn until a stop condition below is met.

The lap

1. ax dojo agenda --json -> agenda.
2. STOP conditions (write the report, then stop):
   • budget.has_surplus is false
   • now >= budget.deadline
   • items is empty
3. Otherwise: take items[0], follow its playbook below, then go to 1. Completed work self-clears: the item vanishes from the next agenda because the underlying system recorded it (verdict locked, brief consumed, proposal created). If the same item survives 2 laps untouched, skip it and note why in the report.

Playbooks by kind

• verdict_pending - ax improve verdict <id> to see the suggested verdict + checkpoint evidence; confirm with --set <verdict> only when the evidence supports it. Distinguish "pattern resolved" from "artifact never fired" before locking no_longer_needed.
• brief_unfilled - open the .ax/tasks/*.md brief, do what it says in the target files, then run the reconciler it names (ax skills lint / ax improve lint).
• routing_backtest - judgment-flagged routing classes: backtest the pattern against dispatch history (ax dispatches --candidates), check false-positive risk, then ax routing tune --apply=<ids> --days=<window> or reject with a written rationale in the report.
• proposal_mint - ax improve recommend; accept the grounded ones (ax improve accept <id>) so briefs exist for the next lap.
• experiment - heavy item. Work ONLY in a fresh worktree (git worktree add .claude/worktrees/dojo-<slug> -b dojo/<slug>). Reproduce the churn pattern, attempt the fix/hook/skill, capture evidence. If it will not finish inside this budget: package it as a goal file (objective + checkpoint index + gates) under docs/superpowers/goals/ so the NEXT dojo session resumes it. Output = an improve proposal; merging the proposal is what activates anything. NEVER merge, never touch main.
• New hooks specifically - author via @ax/hooks-sdk, then run BOTH validators and embed their output in the proposal:
  1. ax hooks backtest <file> --json → cases caught (benefit side): would-block/ would-warn rates, false-positive count, cases with evidence.
  2. ax hooks bench <file> --json → per-fire p50/p95 from real bun spawns, est fires/day from tool_call history, installed-chain budget vs --budget-ms default 250 (cost side). Reject the hook when daily cost (fires/day × p95) or an installed-chain budget overrun outweighs the benefit shown by backtest. Both ledgers must appear in the proposal; neither alone is sufficient.
• spar - only present when invoked with --spar and spendable >= 30%. One task, one delta, scored. Concrete flow:
  1. Pick a landed task: ax sessions here --days=30 - note its commit sha from ax sessions near <sha> or git log.
  2. ax dojo spar-plan <sha> - captures the baseline (prompt + cost/turns/churn) and writes ~/.ax/dojo/spar/<id>.md; the command prints the exact git worktree add command to run next.
  3. Read the brief at ~/.ax/dojo/spar/<id>.md; run the printed git worktree add .claude/worktrees/dojo-spar-<id> -b dojo/spar-<id> <parentSha> command to pin the worktree at the parent SHA.
  4. Apply exactly ONE delta in the delta section (skill on/off, hook on/off, prompt change, thinking level, or model override) - no compound changes.
  5. Do the task in that worktree; let it finish naturally.
  6. ax dojo spar-score <id> - auto-discovers the variant session from the worktree cwd; or pass --variant-session=<id> if there are multiple sessions. Writes the receipt to ~/.ax/dojo/spar/<id>-report.md.
  7. Append the receipt to the dojo report. Track multi-run campaigns as goal files under docs/superpowers/goals/ so the next session can resume.
• explore - free investigation, retro-meta style: follow a hunch through ax recall / ax sessions churn, and convert anything real into a proposal or outbox draft.
• Upstream findings (any lap) - an ax bug or improvement found while training (items of kind upstream_draft are handled by this same rule): run ax dojo draft --title=<title> --kind=bug|improvement to stage it to ~/.ax/dojo/outbox/<slug>.md (complete issue draft: title, body, repro, session refs written by the command). NEVER publish from the dojo - the user reviews and publishes in the morning (ax-repo skill / gh).

Exit - the morning report

Run ax dojo report --since=<loop-start-iso> --notes-file=<lap-notes-path> to write ~/.ax/dojo/reports/<YYYY-MM-DD>.md. The command collects the budget envelope, per-lap item log (from the lap notes file), proposals created, verdicts locked, and outbox drafts awaiting review - pass it the ISO timestamp you recorded when the loop started and the scratch file you appended notes to. Then tell the user the report path and the top 3 things awaiting their review.

For upstream findings (ax bugs or improvements discovered during training), stage them with ax dojo draft --title=<title> --kind=bug|improvement before the report step - never publish directly. The draft lands in ~/.ax/dojo/outbox/<slug>.md; the user reviews and publishes via ax-repo skill / gh in the morning.

Hard rails

• worktrees only; never write on main; never merge anything
• proposals are the only activation path
• outbox only; nothing leaves the machine
• respect the deadline even mid-item: checkpoint, report, stop