メニュー
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.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
One eval-tune round. Applies approved fixes, selects tasks, runs eval, analyzes results, and updates loop state. Standalone use stops for human review; `/autotune-loop` can orchestrate repeated rounds.
Fix Gradle/Kotlin build errors incrementally. No refactoring - just get the build green. Use when build verification fails.
Analyze ClosePaw agent cognition using debug-run traces/replay artifacts and eval results, then propose and implement improvements to prompts, tool definitions, context packing (todo/scratchpad/history), and multi-agent coordination. Use when a debug run feels wrong, when eval metrics regress, when tuning context engineering for generalizable gains, or when reviewing LLM input/output and tool usage; produce both a report and code/doc changes.
Apply prompt, tool description, and app skill changes based on cog-tune diagnosis. Enforces the three-layer ownership model (core prompt → tool descriptions → app skills) to keep changes in the right place.
Run end-to-end Android UX QA from a user perspective through ADB-driven interaction and visual evidence capture. Use when iterating app UX, validating interaction regressions after UI changes, reproducing reported UX bugs, or checking flows in different app states (including Main app and Smart Capsule states such as Takeover/Supplement and text input paths). Prefer this skill when the goal is UX behavior validation rather than internal agent reasoning/debug logs.
Debug failed tool executions from eval runs or debug-runs by isolating the action execution layer. Use when /cog-tune classifies the root cause as "Execution" (action failed or hit wrong target), when actions report success but UI doesn't change (false success), or when you need to compare adb baseline vs. accessibility service execution paths. Designed to complement /cog-tune — this skill focuses on action reliability, not LLM reasoning.
| 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. |
Thin controller around /autotune.
Responsibilities:
autotune-loop: autotune-specific loop policyautotune: one round workerDo not scrape prose or infer loop state from filesystem heuristics. projects/autotune/meta/loop_state.json is the only control-plane file.
/autotune round..claude/skills/autotune/references/tuning_principles.md for stop/continue judgment. Do not keep pushing eval-only patches.projects/autotune/meta/scoreboard.json for retry and "failed 3+ times" decisions.For each loop iteration:
projects/autotune/meta/loop_state.json.status == "complete", report completion and exit.mode: "auto" and status: "running" before delegating the round./autotune round.
/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.projects/autotune/meta/loop_state.json.last_round.recommended_action is stop_success or stop_exhausted, set status to complete, persist stop_reason, report completion, and exit.Set last_round.recommended_action to:
continue: more targeted work is justifiedstop_success: all high/medium-priority items are fixed or consciously parkedstop_exhausted: no actionable fixes remain, or the same targeted improvement failed 3+ timesWhen stopping as exhausted:
eval/config/cannot_handle_group.txt.projects/autotune/meta/issues.md with the parked rationale.projects/autotune/meta/scoreboard.json, not ad hoc memory./autotune: feat(agent): autotune round N — <summary>loop_state.json, scoreboard, changelog, per-task updates, issues.md, and cannot_handle_group.txtprojects/autotune/ are ignored by this app repo; persist them only in the configured private artifact workspace.projects/autotune/meta/loop_state.jsonprojects/autotune/meta/scoreboard.jsonprojects/autotune/meta/changelog.mdprojects/autotune/meta/issues.mdeval/config/cannot_handle_group.txt.claude/skills/autotune/SKILL.md.claude/skills/autotune/references/tuning_principles.md