Automl

/automl v6.0

Calibrator subagent → alignment dialogue → round loop → audit → terminal lifecycle, with 6 always-on gates (iteration / wall-time / repeat-loop / budget / context / quota) via FIXED ORDER tick_gate orchestrator + RED_TEAM opt-in dispatch + multi-session safety (active_session lock, cross-session quota coordination, worktree advisory, orphan recovery) + calibrator self-improvement telemetry (run_summary.md per terminal run, weekly aggregator alerts on Discord).

Spec: docs/superpowers/specs/2026-05-06-automl-v6-design.md Algorithm: ALGORITHM.md Changelog: CHANGELOG.md

Shared quota gate

/automl has an always-on quota gate. Use the shared helper first so this skill and Codex/Codex CLI share the same semantics:

/Users/fredchu/bin/agent-orch quota check --provider claude --threshold 85 --on-error fail-open --json

• exit 0 / decision=="allow"：continue the normal tick_gate loop.
• exit 2 / decision=="wait"：transition lifecycle to quota-wait/paused as the run state already supports, call ScheduleWakeup for JSON resume_at, and resume from .automl/<run_id>/state.json after waking. Re-run quota check before continuing.
• decision=="probe_failed"：retry after retry_at/retry_after_seconds; do not treat probe failure as a 5h reset.
• If agent-orch is unavailable, fall back to the legacy scripts/quota_check.py path for compatibility.

Commands

/automl <freeform goal description>     start a new run (calibrator + alignment + round loop)
/automl --spec <path/to/goal.md>        start from existing goal.md (skip alignment)
/automl pause [run_id]                  pause an active run (release session lock)
/automl resume [run_id]                 resume a paused run (claim session lock)
/automl status [run_id]                 multi-section run status (spec §8.3 format)
/automl clear <run_id>                  delete a terminal run's directory
/automl list                            scan cwd .automl/ + per-run summary
/automl history                         calibrator vs actual telemetry across runs

Flags

--no-budget                  disable calibrated budget cap
--budget <int>               override calibrator's budget estimate
--depth shallow|normal|deep  override calibrator's dialogue depth
--red-team                   force RED_TEAM dispatch
--no-red-team                force skip RED_TEAM
--no-codex                   route Codex tasks to sonnet/opus fallback
--max-iter <int>             ultimate safety cap (default 10000)
--max-wall <hours>           ultimate safety cap (default 72)
--force-fallback <dep>       debug: force a soft dep to fallback path
--cli claude|codex|gemini    debug: force adapter (v6.0 only ships claude)
--allow-cwd-conflict         override worktree advisory for concurrent same-cwd
--autonomous                 silent orphan takeover, no interactive prompts

--red-team + --no-red-team are mutually exclusive.

Invocation routing (main session — read this before dispatching)

Use scripts/cli_parser.py::parse_args to parse the user's /automl ... arguments into a ParsedCommand, then route as follows:

For --spec bypass specifically: the run inherits its run_id from the goal.md frontmatter. run_dir is cwd/.automl/{run_id}/. Because no calibrator was dispatched, no calibrator.json is written; run_summary.py::write_run_summary is null-safe and will produce a slimmer telemetry block (estimated_tokens=0, diff_pct=null, etc.) at the terminal hook.

Migration from v5.10

/automl-legacy runs frozen v5.10 during the 4-6 week migration window. Removed flags: --cap (calibrator handles), --max-ticks (renamed --max-iter). --goal replaced by --no-budget for clearer semantics.

Adapter status

• v6.0: claude-code (this skill)
• v6.1: Codex /goal adapter (≥4 weeks after v6.0 stable)
• v6.2: Gemini adapter (≥8 weeks after v6.0)