reflect

name	reflect
description	Analyze workflow runs and produce a daily reflection.

Reflect

You are running one phase of the reflect pipeline. The phase is specified after --- below as PHASE=<fetch|analyze-run|summarize>.

---

Phase: fetch

Download workflow run logs and artifacts for a date range.

Environment

• SINCE — start date (YYYY-MM-DD).
• UNTIL — end date inclusive (YYYY-MM-DD).
• OUTPUT_DIR — directory to write fetched data into.
• WORKFLOW — workflow name to filter by (optional).

Instructions

1. Run get_workflow_runs with:
   • repo set to whilp/working
   • since set to SINCE
   • until_ set to UNTIL
   • output_dir set to OUTPUT_DIR
   • workflow set to WORKFLOW (if provided)
2. The tool downloads logs and artifacts for each run into OUTPUT_DIR.
3. Verify the manifest file exists at OUTPUT_DIR/manifest.json.
4. Write OUTPUT_DIR/fetch-done containing the number of runs fetched.

Output

• OUTPUT_DIR/manifest.json — JSON array of run metadata with log/artifact paths.
• OUTPUT_DIR/<run-id>/log.txt — full log for each run.
• OUTPUT_DIR/<run-id>/artifacts/ — downloaded artifacts for each run.
• OUTPUT_DIR/fetch-done — completion marker.

---

Phase: analyze-run

Analyze a single workflow run. You are sandboxed with no network access.

Environment

• RUN_DIR — directory containing this run's log and artifacts.
• RUN_META — JSON string with run metadata (from manifest).
• OUTPUT_FILE — path to write the analysis.

Instructions

1. Parse RUN_META to get run metadata (workflowName, conclusion, displayTitle, etc.).
2. Analyze RUN_DIR/log.txt efficiently — do not read the full log. Logs are large (~100KB). Use targeted commands:
   • grep -n "==>" RUN_DIR/log.txt — phase transitions and timing
   • grep -n "done in\|verdict\|failed\|error\|exit" RUN_DIR/log.txt — outcomes and errors
   • grep -n "no_issues\|pr_limit\|issue.json" RUN_DIR/log.txt — pick results
   • head -30 RUN_DIR/log.txt and tail -30 RUN_DIR/log.txt — run boundaries
   • Use read with offset and limit to inspect specific line ranges when needed.
3. Check RUN_DIR/artifacts/ for structured data (these are small, safe to read fully):
   • */pick/issue.json — which issue was picked
   • */pick/reasoning.md — why it was picked
   • */plan/plan.md — what was planned
   • */do/do.md — what was done
   • */check/actions.json — verdict and actions
   • */do/feedback.md — if needs-fixes, what feedback
   • */act.json — final outcome
   • **/session*.db — session databases (note their presence/size)
4. Write a concise analysis to OUTPUT_FILE.

Output

Write OUTPUT_FILE in this format:

## <workflowName>: <displayTitle> (<conclusion>)

- **run**: <databaseId> | **event**: <event> | **branch**: <headBranch>
- **started**: <startedAt> | **conclusion**: <conclusion>

### Summary

1-3 sentence summary of what happened in this run.

### Issue

If a work run: issue number, title, verdict, PR link (if any).

### Failures

If failed: root cause, relevant log lines (keep brief).

### Notes

Anything notable: retries, feedback loops, unusual patterns.

Keep it concise — under 80 lines. Focus on what matters for the daily summary.

---

Phase: summarize

Combine per-run analyses into a final reflection. You are sandboxed with no network access.

Environment

• ANALYSIS_DIR — directory containing per-run analysis files.
• DATE — date string (YYYY-MM-DD).
• OUTPUT_FILE — path to write the reflection.

Instructions

1. Read all *.md files in ANALYSIS_DIR.
2. Synthesize across all runs:
   • success rate: how many runs passed vs failed vs cancelled, by workflow.
   • failure patterns: common failure modes grouped by root cause.
   • work loop outcomes: issues attempted, verdicts, PRs created.
   • patterns: trends, flakiness, duration observations.
   • recommendations: concrete, actionable improvements — each specific enough to become an issue.
3. Write OUTPUT_FILE.

Output

Write OUTPUT_FILE in this format:

# Reflection: YYYY-MM-DD

## Summary

Brief overview of the day.

## Runs

| workflow | total | pass | fail | cancel |
|---|---|---|---|---|

## Failure Analysis

Group failures by root cause. Include brief log excerpts where helpful.

## Work Loop Outcomes

Issues attempted, verdicts, PRs created.

## Patterns

Trends, flakiness, duration.

## Recommendations

Concrete improvements. Each should be specific enough to become an issue.