autotune-loop

name	autotune-loop
description	Autonomous multi-round controller for `/autotune`. Reads and updates `loop_state.json`, decides continue or stop, and records completion state when the loop is finished.

AutoTune Loop

Thin controller around /autotune.

Responsibilities:

• autotune-loop: autotune-specific loop policy
• autotune: one round worker

Do not scrape prose or infer loop state from filesystem heuristics. projects/autotune/meta/loop_state.json is the only control-plane file.

Principles

• Keep loop policy simple. One loop iteration should do at most one orchestrated /autotune round.
• Use .claude/skills/autotune/references/tuning_principles.md for stop/continue judgment. Do not keep pushing eval-only patches.
• Use projects/autotune/meta/scoreboard.json for retry and "failed 3+ times" decisions.

Loop

For each loop iteration:

1. Read projects/autotune/meta/loop_state.json.
2. If status == "complete", report completion and exit.
3. Ensure the active loop run is marked mode: "auto" and status: "running" before delegating the round.
4. Invoke exactly one orchestrated /autotune round.
   • Verify sync: After /autotune Step 1 commits changes, confirm the remote has the latest code and APK rebuilt (git push + remote git pull && ./gradlew assembleDebug) BEFORE the eval runs. /autotune Step 3 owns this, but if running --remote, double-check the remote is up to date. Running eval on stale code wastes the entire round.
5. Re-read projects/autotune/meta/loop_state.json.
6. If last_round.recommended_action is stop_success or stop_exhausted, set status to complete, persist stop_reason, report completion, and exit.
7. Otherwise exit normally so the next loop iteration can continue.

Stop Criteria

Set last_round.recommended_action to:

• continue: more targeted work is justified
• stop_success: all high/medium-priority items are fixed or consciously parked
• stop_exhausted: no actionable fixes remain, or the same targeted improvement failed 3+ times

When stopping as exhausted:

• Add genuinely out-of-reach tasks to eval/config/cannot_handle_group.txt.
• Update projects/autotune/meta/issues.md with the parked rationale.
• Base the "failed 3+ times" call on projects/autotune/meta/scoreboard.json, not ad hoc memory.

Commit Policy

• Code changes belong to Step 1 inside /autotune: feat(agent): autotune round N — <summary>
• End-of-round artifacts belong after analysis: loop_state.json, scoreboard, changelog, per-task updates, issues.md, and cannot_handle_group.txt
• Artifact files under projects/autotune/ are ignored by this app repo; persist them only in the configured private artifact workspace.

Key Files

• Loop state: projects/autotune/meta/loop_state.json
• Scoreboard: projects/autotune/meta/scoreboard.json
• Changelog: projects/autotune/meta/changelog.md
• Issues: projects/autotune/meta/issues.md
• Exclusions: eval/config/cannot_handle_group.txt
• Round worker: .claude/skills/autotune/SKILL.md
• Shared tuning principles: .claude/skills/autotune/references/tuning_principles.md