retrospective（retrospective 分析）

Agent Retrospective

Run a structured self-review covering the last 7 days from the run date. Produce honest, specific analysis — not generic platitudes.

Process

1. Gather Context

• Read memory/ files for the review period (daily logs)
• Read MEMORY.md for long-term context (pending tasks, blocked items, decisions)
• Check cron job history: cron(action="list") — note failures and consecutive errors
• Check session status for usage patterns

2. Analyse: What Went Well

List concrete wins with evidence:

• Tasks completed successfully
• Automations that ran reliably
• Problems solved creatively
• Time/effort saved for the user

Keep it brief. 3-7 bullet points max.

3. Analyse: What Didn't Go Well

Be brutally honest. For each struggle:

• What: The specific task or goal
• Why it failed: Root cause (not symptoms)
• Time wasted: Rough estimate of turns/attempts spent
• Pattern: Is this a recurring struggle or one-off?

Categories to check:

• Bot detection / scraping failures
• Browser automation reliability
• API quota / auth issues
• Tasks attempted manually that a skill could handle
• Blocked items that stayed blocked too long
• Over-engineering (built when a simpler approach existed)
• Responsiveness issues (slow replies, timeouts, retries)
• Communication issues (user had to repeat themselves)

4. Skill Gap Analysis

Search for skills that address identified struggles. For each promising skill:

• Note name, version, relevance score
• Assess whether it genuinely solves the problem vs. being a low-quality wrapper
• Recommend install only if it clearly saves future effort

Also check: are any installed skills underused or misconfigured?

5. Config & Infrastructure Review

Check for improvements to:

• Cron jobs: Any consistently failing? Should any be added/removed?
• Auth/credentials: Any expired, rotated, or missing?
• Memory/context: Is MEMORY.md getting stale? Are daily logs being written?
• Tool config: Any env vars, API keys, or integrations that need attention?
• Workspace hygiene: Temp files, stale scripts, orphaned projects?

6. Recommendations

Produce a prioritised action list:

1. Quick wins (< 5 min each) — config fixes, skill installs, cleanup
2. Medium effort (< 1 hour) — new automations, script fixes, unblocking items
3. Larger projects — new integrations, architectural changes

Each recommendation must be specific and actionable (not "improve scraping" but "install X to bypass Cloudflare on Y").

Output Format

Write the retro to memory/retro-YYYY-MM-DD.md:

# Retrospective — [date range]

## Wins
- [concrete win with evidence]

## Struggles
| Task | Root Cause | Time Spent | Recurring? |
|------|-----------|------------|------------|
| ... | ... | ... | yes/no |

## Skill Gaps
| Problem | Skill | Action |
|---------|-------|--------|
| ... | skill-name v1.0 | install / skip / evaluate |

## Config Issues
- [specific issue + fix]

## Action Items
### Quick Wins
1. [action]

### Medium Effort
1. [action]

### Larger Projects
1. [action]

Scheduling

Set up a weekly cron job on first use (check cron(action="list") first — only create once):

cron(action="add", job={
  "name": "Weekly Retrospective",
  "schedule": {"kind": "cron", "expr": "0 16 * * 5", "tz": "local"},
  "sessionTarget": "isolated",
  "payload": {
    "kind": "agentTurn",
    "message": "Run a full retrospective for the last 7 days following the retrospective skill. Read memory files, check cron status, search for skills addressing struggles, write retro to memory/retro-YYYY-MM-DD.md, update MEMORY.md with key findings, deliver a concise summary."
  },
  "delivery": {"mode": "announce"}
})

Guidelines

• No fluff. Every line should contain information.
• Be self-critical. The point is improvement, not a highlight reel.
• Cite evidence. Reference specific memory files, dates, error messages.
• Don't recommend skills blindly. Only recommend if genuinely useful.
• Cap action items at 10. Prioritise ruthlessly.
• Update MEMORY.md with key findings after the retro.