feat(skills): add agentation-fixer skill for watch & auto-fix loop

skills/agentation-fixer/SKILL.md

@@ -0,0 +1,79 @@
+---
+name: agentation-fixer
+description: >
+  Watch for browser annotations and auto-fix code. The "Fixer" half of the
+  two-session self-driving workflow — receives annotations created by a human or
+  by the agentation-self-driving Critic, locates the relevant source code, applies
+  minimal fixes, and resolves each annotation. Supports one-shot check and
+  continuous watch loop. Use when: "watch annotations", "watch mode", "fix annotations",
+  "agentation fixer", "agentation watch", or user wants an agent to continuously
+  process UI feedback from the browser.
+argument-hint: "[watch|check]"
+allowed-tools: "mcp__agentation__agentation_get_all_pending mcp__agentation__agentation_watch_annotations mcp__agentation__agentation_acknowledge mcp__agentation__agentation_resolve mcp__agentation__agentation_dismiss mcp__agentation__agentation_reply Edit Read Grep Glob"
+---
+
+# Agentation Fixer
+
+Receive browser annotations via MCP, locate source code, apply fixes, resolve.
+
+Designed as the **Fixer** in the [two-session workflow](../agentation-self-driving/references/two-session-workflow.md):
+- **Session 1 (Critic)**: `/agentation-self-driving` — scans page, adds annotations in a headed browser
+- **Session 2 (Fixer)**: This skill — watches for annotations, edits code to address each one
+
+Also works standalone when a human creates annotations manually.
+
+## Mode Selection
+
+- `/agentation-fixer` or `/agentation-fixer check` → **One-shot**: drain pending, fix, done
+- `/agentation-fixer watch` → **Watch loop**: continuous monitoring until stopped
+
+If no argument and zero pending, ask user: check only or enter watch mode?
+
+## One-shot
+
+1. `agentation_get_all_pending`
+2. Fix each via **Fix Cycle**
+3. Report summary (N fixed, N clarified, N dismissed)
+
+## Watch Loop
+
+1. Drain existing pending annotations first
+2. `agentation_watch_annotations` (timeoutSeconds: 300, batchWindowSeconds: 15)
+3. Annotations arrive → fix each → re-enter watch
+4. Timeout with no annotations → "Still watching..." → re-enter watch
+5. User says stop → print session summary, exit
+
+## Fix Cycle
+
+Per annotation:
+
+1. **Acknowledge**: `agentation_acknowledge(id)` — user sees you're working on it
+2. **Locate source**:
+   - Use `sourceFile` field if present (direct path)
+   - Otherwise use `reactComponents` to identify the component, then Grep/Glob to find the file
+   - Fallback: use `elementPath` or `cssClasses` to grep the codebase
+3. **Read** the relevant code section
+4. **Route by intent**:
+   - `fix` / `change` / (no intent) → apply minimal code change → `agentation_resolve(id, "summary")`
+   - `question` → `agentation_reply(id, "answer")`, leave open for human to confirm
+   - `approve` → `agentation_resolve(id, "Acknowledged, no changes needed")`
+5. **When unsure** → `agentation_reply(id, "clarifying question")`, leave open
+6. **When scope is too large** → `agentation_reply(id, "explanation of scope")`, leave open
+
+## Rules
+
+- **Minimal diffs** — fix exactly what's asked, nothing more
+- **No opportunistic refactoring** of surrounding code
+- **Severity ordering**: `blocking` first, then `important`, then `suggestion`
+- **Batch same-component annotations** into one edit when they touch the same file/section
+- **Never break the build** — if unsure about a change, reply instead of guessing
+
+## Install
+
+Symlink into your skills directory:
+
+```bash
+ln -s "$(pwd)/skills/agentation-fixer" ~/.claude/skills/agentation-fixer
+```
+
+Restart Claude Code after installing. Verify with `/agentation-fixer`.

skills/agentation-self-driving/references/two-session-workflow.md

@@ -23,12 +23,12 @@ This session opens the headed browser, scans the page, and adds design annotatio
 
 ```bash
 claude
-> Watch for agentation annotations and fix each one. Use agentation_watch_annotations
-> in a loop. For each annotation: read the target code, make the fix, then call
-> agentation_resolve with a summary of what you changed.
+> /agentation-fixer watch
 ```
 
-This session blocks on `agentation_watch_annotations`, receives each annotation as it's created by Terminal 1, and edits the codebase to address the feedback.
+This session blocks on `agentation_watch_annotations`, receives each annotation as it's created by Terminal 1, and edits the codebase to address the feedback. See the [`agentation-fixer` skill](../../agentation-fixer/SKILL.md) for full details.
+
+> **Without the skill**: You can also run the fixer with a manual prompt — `Watch for agentation annotations and fix each one. Use agentation_watch_annotations in a loop. For each annotation: read the target code, make the fix, then call agentation_resolve with a summary.`
 
 ## How It Connects