recap

Gathers all available context about work in progress and presents a clear "here's where you left off" summary so you can continue without losing momentum.

When to Use

• At the start of a new session
• After the session-load hook shows stale context
• When you want to review the full state of the project

What It Checks

1. Active work context — .claude/context/active-work.md (auto-saved by session-save hook)
2. Session summary — .claude/context/session-summary.md (generated by session-end hook)
3. Activity log — .claude/context/agent-activity.log (recent file changes and task completions)
4. In-progress features — All feature files in .claude/context/features/ with status in-progress
5. Git state — Current branch, uncommitted changes, recent commits
6. Pipeline state — .agent-pipeline/pipeline.json if an agent pipeline was running
7. Eval results — Latest eval report in .agent-eval-results/ if any

Workflow

Step 1: Read Session Context Files

Read the saved session context:

cat .claude/context/active-work.md
cat .claude/context/session-summary.md
tail -30 .claude/context/agent-activity.log

From these files, note:

• When the last session ended
• What files were edited
• What tasks were completed
• Current branch and uncommitted changes

Step 2: Scan Feature Files

Read all files in .claude/context/features/. For each file:

• Extract the title (first # heading)
• Extract the status line
• If status is in-progress, testing, or review, list the remaining unchecked tasks

Step 3: Check Git State

Run these commands to understand the current code state:

git branch --show-current
git status --short
git log --oneline -5

Note:

• What branch you're on
• Whether there are uncommitted changes
• What the recent commits were about

Step 4: Check Pipeline State

If .agent-pipeline/pipeline.json exists:

cat .agent-pipeline/pipeline.json

Report which stages completed, which failed, and which are pending.

Step 5: Check Latest Eval Results

Look for the most recent eval report:

ls -t .agent-eval-results/*.md 2>/dev/null | head -1

If found, read it and report the overall score and any NEEDS_FIXES items.

Output Format

Present the findings as a structured summary:

## Session Recap

### Last Session
- **Ended:** <timestamp from session-summary.md>
- **Branch:** <branch name>
- **Files edited:** <count>
- **Tasks completed:** <count>

### Recent Activity
<last 5-10 entries from agent-activity.log>

### In-Progress Features
1. **<Feature Title>** — <status>
   - Remaining: <count> tasks
   - Next task: <first unchecked item>

### Current Code State
- **Branch:** `<branch>`
- **Uncommitted changes:** <count> files modified
- **Recent commits:** <1-line summary of last 2-3 commits>

### Pipeline Status
- <stage>: <status> (or "No active pipeline")

### Suggested Next Steps
1. <most logical next action based on context>
2. <second action>

Key Rules

1. Read, don't guess. Only report information from actual files and git state.
2. Be concise. The user wants to quickly orient themselves, not read an essay.
3. Suggest next steps. Based on the state, recommend what to do next.
4. Don't modify anything. This skill is read-only. It never writes files or runs commands that change state.