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.
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.
Browser automation guidance for browser_script and reusable CDP snippets.