weekly

name	weekly
description	Run every weekly research-orchestration task in parallel via subagents, each emitting its own redacted public gist. Manual trigger only — no cron. Returns one gist URL per task: weekly summary, workflow-optimization retrospective, code hygiene scan, mentor-prep agenda, mentor-prep slides (Marp deck), and SUMMARY gist. Add new weekly tasks by appending a row to the dispatch table and a corresponding subagent prompt section.
user_invocable	true

Weekly

End-of-week fan-out. Spawns N subagents in parallel (single message, multiple Agent tool calls), each running one self-contained task and returning a public gist URL. The orchestrator does no work itself — it only dispatches and collects URLs.

When to use

Manual trigger at the end of each work week. Manual trigger only — no cron. Use /schedule if you want to wire a cron later.

Dispatch table

Task	Subagent type	Returns
Weekly summary	general-purpose	gist URL — past-7d mentor update
Workflow optimization	retrospective	gist URL — CLAUDE.md / agent / skill / hook patches
Code hygiene	general-purpose	gist URL — dead code, refactor candidates, deps, .claude/ health, jscpd duplication, unmerged worktrees
Mentor-prep agenda	general-purpose	gist URL — clean-result TL;DRs assembled into a screen-shareable meeting doc
Mentor-prep slides	general-purpose	gist URL — Marp deck (deck.md + deck.pdf) for the actual mentor meeting; built via the mentor-update-slides skill
SUMMARY gist	general-purpose	gist URL — project SUMMARY (Motivation / Related work / Current results / Immediate next steps / Long-term goals / Glossary), regenerated from docs/SUMMARY.template.md + docs/claims.yaml + live issues

(Adding more is a 2-step change: append a row here, append a ## Subagent prompt: <name> section below.)

Procedure

Argument parsing. If "$ARGS" contains "--autonomous", set INTERACTIVE=false; else INTERACTIVE=true. When INTERACTIVE=false, every AskUserQuestion step (Step 0 below for free-form thoughts; Step 0.5 for self-reflection prompts in slice 6) is skipped; USER_NOTES="" and ANSWER_1..6="". The substitution rules in Step 2.5 then strip empty user-notes blocks and emit the static prompt list with "(unanswered)" placeholders inline.

0. Free-form thoughts (interactive). Determine INTERACTIVE per the rule above. If INTERACTIVE=true, call AskUserQuestion:

   "Any free-form thoughts you want to include in this week's gists? (Examples: surprises, questions, frustrations, decisions.) Reply with prose — empty answer skips the section."

   Capture into USER_NOTES. Empty / "skip" / "(none)" → USER_NOTES="". If INTERACTIVE=false, set USER_NOTES="" without asking.

0.5. Self-reflection (interactive). Determine INTERACTIVE per the rule above. If INTERACTIVE=true, call AskUserQuestion once per the 6 weekly prompts below, capturing each answer into ANSWER_1 … ANSWER_6. Empty answers → ANSWER_N="" (substituter prints (unanswered) inline). If INTERACTIVE=false, set every ANSWER_N="" without asking — the prompts are still emitted in the gist body, but each shows (unanswered).

The 6 weekly prompts (with sources cited inline; see "Self-reflection prompt sources" at the bottom of this file for the full URL list):

1. What is the ONE result this week I would put in a paper? (Nanda — How to Write ML Papers)

2. What did I actively try to falsify, and did it survive? (Nanda — Truth-Seeking)

3. If I had one more week before mentor meeting, what would I do? (Chua / Hughes)

4. Which of last week's plans turned out to be wrong, and what would I have predicted differently if I'd thought harder upfront? (Nanda — Research Taste)

5. What's the highest-information-gain experiment I could run next week, per GPU-hour? (Perez)

6. Am I in Explore, Understand, or Distill phase for each topic — and does my time allocation match? (Nanda — Explore/Understand/Distill)

7. Compute WEEK_TAG=$(date +%Y-W%V), TODAY=$(date +%Y-%m-%d), TS=$(date -Iseconds), and WEEK_AGO=$(date -d '7 days ago' +%Y-%m-%d).

8. In a single assistant message, issue one Agent tool call per row in the dispatch table. For the retrospective row use subagent_type: "retrospective"; the rest use general-purpose. Each subagent gets the corresponding "Subagent prompt" section verbatim as its prompt.

   2.5. String-substitute USER_NOTES + ANSWER_1..6 into the subagent prompt template. For each subagent prompt that contains a ## User notes\n{{USER_NOTES}} block: - If USER_NOTES is non-empty, replace {{USER_NOTES}} with the value. - If USER_NOTES is empty, REMOVE the entire ## User notes\n{{USER_NOTES}}\n\n block (do NOT leave a stub (none)).

    For each `{{ANSWER_N}}` placeholder in the
    `## Self-reflection (canonical)` block (slice 6):
    - If the answer is non-empty, substitute it directly.
    - If empty, substitute the literal string `(unanswered)`.
    Pass the substituted prompt to the Agent tool.
   

9. Wait for all subagents to complete (they run concurrently).

10. Classify each subagent's return value into one of three statuses:

    • success — returned a https://gist.github.com/... URL
    • skipped — returned a literal (skipped — <reason>) style string (today: only mentor-prep does this, when zero clean-results in the window)
    • failed — crashed, errored, or returned nothing parseable

11. Log each success and skipped outcome to both files via Bash — one batch append per file, not four parallel writes. failed outcomes are NOT logged. For skipped rows, put the bracketed skip message in the url column (markdown) or set "status":"skipped" with no url field (JSONL).

    # Ensure the markdown log exists with a header (idempotent).
    if [ ! -f docs/update_log.md ]; then
      mkdir -p docs
      printf '# Update log\n\nGist URLs from /daily and /weekly. Newest at top.\n\n| date | scope | task | url-or-status |\n|---|---|---|---|\n' > docs/update_log.md
    fi
    
    # Ensure the JSONL log directory exists.
    mkdir -p .claude/cache
    
    # For each task, append one row to each file. Use the URL when
    # status=success; use the literal "(skipped — <reason>)" when
    # status=skipped. Skip the row entirely when status=failed.
    # Example with all four tasks success:
    {
      echo "| ${TODAY} | weekly (${WEEK_TAG}) | summary | <url-or-skip> |"
      echo "| ${TODAY} | weekly (${WEEK_TAG}) | workflow-optimization | <url-or-skip> |"
      echo "| ${TODAY} | weekly (${WEEK_TAG}) | code-hygiene | <url-or-skip> |"
      echo "| ${TODAY} | weekly (${WEEK_TAG}) | mentor-prep | <url-or-skip> |"
      echo "| ${TODAY} | weekly (${WEEK_TAG}) | mentor-slides | <url-or-skip> |"
    } >> docs/update_log.md
    
    # JSONL: success rows have "status":"success","url":"...";
    #        skipped rows have "status":"skipped","reason":"...".
    {
      echo '{"date":"'"${TODAY}"'","ts":"'"${TS}"'","scope":"weekly","week":"'"${WEEK_TAG}"'","task":"summary","status":"success","url":"<url>"}'
      echo '{"date":"'"${TODAY}"'","ts":"'"${TS}"'","scope":"weekly","week":"'"${WEEK_TAG}"'","task":"workflow-optimization","status":"success","url":"<url>"}'
      echo '{"date":"'"${TODAY}"'","ts":"'"${TS}"'","scope":"weekly","week":"'"${WEEK_TAG}"'","task":"code-hygiene","status":"success","url":"<url>"}'
      # Example skipped row (mentor-prep agenda or mentor-slides skip when no clean-results in window):
      # echo '{"date":"'"${TODAY}"'","ts":"'"${TS}"'","scope":"weekly","week":"'"${WEEK_TAG}"'","task":"mentor-prep","status":"skipped","reason":"no clean-results this week"}'
      # echo '{"date":"'"${TODAY}"'","ts":"'"${TS}"'","scope":"weekly","week":"'"${WEEK_TAG}"'","task":"mentor-slides","status":"skipped","reason":"no clean-results this week"}'
      echo '{"date":"'"${TODAY}"'","ts":"'"${TS}"'","scope":"weekly","week":"'"${WEEK_TAG}"'","task":"mentor-prep","status":"success","url":"<url>"}'
      echo '{"date":"'"${TODAY}"'","ts":"'"${TS}"'","scope":"weekly","week":"'"${WEEK_TAG}"'","task":"mentor-slides","status":"success","url":"<url>"}'
    } >> .claude/cache/update_log.jsonl
    

12. Report to the user as a bulleted list:

    Weekly updates posted (week <WEEK_TAG>):
    - Summary: <url>
    - Workflow optimization: <url>
    - Code hygiene: <url>
    - Mentor-prep agenda: <url-or-skip-message>
    - Mentor-prep slides: <url-or-skip-message>
    (logged to docs/update_log.md + .claude/cache/update_log.jsonl)
    

13. If any subagent failed, list it with ❌ <task> — <reason> and continue; one task failing must not suppress the others. Failed tasks are NOT logged (skipped tasks ARE logged — see step 5).

---

Subagent prompt: Weekly summary

You are generating this week's research mentor update for the
explore-persona-space project. Lead with the result, not the process.
Reading-time target: under 7 minutes. The 7-day window is "past 7 days
from now".

# Data sources (gather in parallel via Bash; read-only)

WEEK_AGO=$(date -d '7 days ago' +%Y-%m-%d)
WEEK_TAG=$(date +%Y-W%V)

1. Git history past 7 days:
   git log --since="7 days ago" --no-merges --oneline --stat
   git diff --stat HEAD~$(git log --since="7 days ago" --oneline | wc -l)..HEAD 2>/dev/null

2. Clean-result issues created or updated this week:
   gh issue list --label clean-results --state all \
     --search "created:>=${WEEK_AGO}" --json number,title,body,createdAt,labels
   gh issue list --label clean-results --state all \
     --search "updated:>=${WEEK_AGO}" --json number,title,body,updatedAt,labels
   For each: extract TL;DR + confidence + hero figure URL.

3. Done experiment + done impl issues this week:
   gh issue list --search "is:issue updated:>=${WEEK_AGO} (label:status:done-experiment OR label:status:done-impl)" \
     --json number,title,labels,updatedAt

4. Recent figures:
   find figures -type f \( -name "*.png" -o -name "*.pdf" \) -mtime -7
   READ each .png with the Read tool before captioning.

5. Eval results past 7 days:
   find eval_results -name "*.json" -mtime -7 -type f

6. Pending / blocked items:
   gh issue list --state open --label status:blocked
   gh issue list --state open --label status:proposed
   gh issue list --state open --label status:running

# Output structure

# Weekly Update — week <WEEK_TAG>

## TL;DR
[2-3 sentences. Single most important thing learned this week + what it means.]

## User notes
{{USER_NOTES}}

## Self-reflection (canonical)
1. **What is the ONE result this week I would put in a paper?**
   {{ANSWER_1}}
2. **What did I actively try to falsify, and did it survive?**
   {{ANSWER_2}}
3. **If I had one more week before mentor meeting, what would I do?**
   {{ANSWER_3}}
4. **Which of last week's plans turned out to be wrong, and what would I have predicted differently if I'd thought harder upfront?**
   {{ANSWER_4}}
5. **What's the highest-information-gain experiment I could run next week, per GPU-hour?**
   {{ANSWER_5}}
6. **Am I in Explore, Understand, or Distill phase for each topic — and does my time allocation match?**
   {{ANSWER_6}}

## Headline Findings (clean-results from this week)
[For each clean-result issue created or updated in past 7 days:
title + confidence + 1-line TL;DR + hero figure URL.]

## Done This Week
[Group by type:experiment / type:infra / type:analysis. 1-2 lines each.]

## Running experiments
[Currently on pods, with expected completion.]

## Blockers
[Open `status:blocked`. "None" is valid.]

## Next Week (top 3 priorities)
[Ordered by information gain per GPU-hour. Action + expected cost + issue link.]

# Writing rules

- Lead with result, not process. No legacy taxonomy / jargon.
- Quantify everything (N, p, effect, CI). Be honest about negatives.
- Bold key numbers; structured bullets only.
- Read figures before captioning. Max 5 figures.
- Never fabricate. Cross-reference with clean-result bodies.

# Publish

  uv run python scripts/redact_for_gist.py --in /tmp/weekly-body.md --out /tmp/weekly-body.redacted.md
  gh gist create --public \
    --filename "weekly-${WEEK_TAG}.md" \
    --desc "Weekly research update — ${WEEK_TAG}" \
    /tmp/weekly-body.redacted.md

RETURN the gist URL as the SOLE output. No commentary.

---

Subagent prompt: Workflow optimization

This task uses the existing retrospective agent, which already knows the playbook. Spawn it with subagent_type: "retrospective" and the prompt below. It does the JSONL transcript review + GitHub-side activity scan itself.

End-of-week workflow retrospective for the explore-persona-space project.
Run with --lookback-days 7. Read every JSONL transcript modified in the
past 7 days from
~/.claude/projects/-home-thomasjiralerspong-explore-persona-space/, plus
GitHub-side activity over the same window:

  WEEK_AGO=$(date -d '7 days ago' +%Y-%m-%d)
  gh issue list --search "updated:>=${WEEK_AGO}" --state all \
    --json number,title,labels,state,updatedAt --limit 200
  gh pr list --search "updated:>=${WEEK_AGO}" --state all \
    --json number,title,state,updatedAt --limit 200

Aggregate findings into a numbered list of patches to propose against:
- CLAUDE.md
- .claude/agents/*.md
- .claude/skills/*/SKILL.md
- .claude/settings.json (hooks)

Output structure:

# Weekly Workflow Optimization — week <WEEK_TAG>

## Summary
- N sessions reviewed
- M user corrections / friction events
- K successful agent dispatches

## Top friction patterns (proposed patches)
1. <pattern> — proposed change
2. ...

## What worked (reinforce these)
[1-3 bullets — focus stays on friction.]

## Metrics
- Time spent debugging vs research: <estimate>
- Most-spawned agent: <name>, <count>
- Agents never spawned this week: <list>

## Proposed CLAUDE.md / agent / skill diffs
[Numbered unified-diff blocks the user can apply with patch(1).]

# Publish

  WEEK_TAG=$(date +%Y-W%V)
  # Body already in /tmp/weekly-workflow-body.md after you build it.
  uv run python scripts/redact_for_gist.py --in /tmp/weekly-workflow-body.md --out /tmp/weekly-workflow-body.redacted.md
  gh gist create --public \
    --filename "weekly-workflow-${WEEK_TAG}.md" \
    --desc "Weekly Claude Code workflow optimization — ${WEEK_TAG}" \
    /tmp/weekly-workflow-body.redacted.md

This skill is READ-ONLY for the project — never modify CLAUDE.md or any
agent / skill / hook directly. Every proposed change is a diff in the gist.

RETURN the gist URL as the SOLE output. No commentary.

---

Subagent prompt: Code hygiene

End-of-week code hygiene scan for the explore-persona-space project.
Combines repo-wide dead-code analysis, refactor candidates, dependency
freshness, .claude/ health audit, code duplication (jscpd), and unmerged
worktree branches into one report. READ-ONLY for the project — never
auto-refactor.

# Hard requirement

Node v18+ + jscpd accessible via `npx jscpd`. If unavailable, abort and
return the message:

  "install Node v18+ and re-run; jscpd is required for duplication
   detection. \`npm install -g jscpd\` or use the ephemeral \`npx jscpd\`."

# Procedure

WEEK_TAG=$(date +%Y-W%V)

1. Lint sweep (safe auto-fix):
   uv run ruff check --fix .
   uv run ruff format .
   Capture: N files reformatted, M lint fixes applied.

2. Repo-wide dead-code analysis (no auto-fix — re-exports may be flagged):
   uv run ruff check . --select F401,F811,F841 --no-fix

3. Refactoring candidates (ranked by severity, do NOT auto-refactor):
   - Python files > 500 lines (excluding tests, generated code)
   - Functions > 60 lines
   - Functions with > 4 levels of indentation

4. Dependency audit:
   uv pip list --outdated 2>/dev/null || echo "uv pip list not available"
   Flag packages > 2 major versions behind; note known security advisories.

5. .claude/ health audit:
   - .claude/agents/*.md: grep for file paths or function names that no
     longer exist in the codebase
   - .claude/skills/*/SKILL.md: check for broken refs to other skills/agents
   - .claude/plans/: flag plans > 30 days old that reference issues now
     in `status:done-*`
   - .claude/settings.json: check allow rules for tools that no longer exist

6. Unmerged worktree branches:
   git worktree list --porcelain | awk '/^worktree/ {print $2}' | grep '\.claude/worktrees/'
   For each, count `git log main..<branch> --oneline | wc -l`.

7. Code duplication via jscpd (min 10 lines / 50 tokens):
   npx jscpd --min-lines 10 --min-tokens 50 --reporters json \
     --output /tmp/jscpd src/ scripts/ 2>/tmp/jscpd-stderr.log || true
   Top 10 duplicates from /tmp/jscpd/jscpd-report.json.

8. Skill / agent description-overlap (Jaccard on description bigrams,
   threshold 0.4):
   Inline Python over .claude/skills/*/SKILL.md + .claude/agents/*.md
   frontmatter `description:` blocks.

# Output structure

# Weekly Code Hygiene — week <WEEK_TAG>

## Lint + Format
- N files reformatted, M lint fixes applied

## Dead Code (repo-wide)
- N unused imports across M files
- [top 10]

## Refactoring Candidates
- N files > 500 lines, M functions > 60 lines
- [top 5 by severity]

## Dependencies
- N packages outdated; [list if any]; security advisories [if any]

## .claude/ Health
- N stale references found; [list]

## Unmerged Worktree Branches
- <wt path>: branch <name>, <N> unmerged commits
- (or "(none)")

## Code Duplication (jscpd)
- [top 10 duplicate pairs with file:line ranges]

## Skill / Agent Description Overlap (jaccard > 0.4)
- [list pairs] (or "no overlapping pairs above 0.4")

## Recommended Actions
1. <highest-priority>
2. ...

# Publish

  uv run python scripts/redact_for_gist.py --in /tmp/weekly-hygiene-body.md --out /tmp/weekly-hygiene-body.redacted.md
  gh gist create --public \
    --filename "weekly-hygiene-${WEEK_TAG}.md" \
    --desc "Weekly code hygiene + refactor candidates — ${WEEK_TAG}" \
    /tmp/weekly-hygiene-body.redacted.md

RETURN the gist URL as the SOLE output. No commentary.

---

Subagent prompt: Mentor-prep agenda

Assemble recent clean-result GitHub issues into a single
screen-shareable mentor-meeting agenda. Read-only — never edit source
issues; never paraphrase TL;DRs (copy verbatim). Default window: past 7
days from now (label `clean-results`).

# Step 1: Fetch issues

WEEK_AGO=$(date -d '7 days ago' +%Y-%m-%d)
WEEK_TAG=$(date +%Y-W%V)

gh issue list --label clean-results --state all \
  --search "updated:>=${WEEK_AGO}" \
  --json number,title,body,labels,updatedAt --limit 100

For each issue, fetch the full body:
  gh issue view <N> --json number,title,body,labels,updatedAt

# Step 2: Parse each clean-result body

Extract for each:
- TL;DR block: everything between `## TL;DR` and the horizontal rule
  preceding `# Detailed report`.
- Subsections to verify present: ### Background, ### Methodology,
  ### Results, ### Next steps. Warn (not fail) if any missing.
- Hero figure URL: first `raw.githubusercontent.com` markdown image under
  ### Results.
- Confidence tag: parse the `**Confidence: HIGH | MODERATE | LOW**` line
  in ### Results.
- Status icon:
  - ✅ if confidence = HIGH
  - ⚠️  if confidence = MODERATE
  - ❌ if confidence = LOW or any CRITICAL caveat present
- Next steps subsection — used for the aggregated "Open questions / asks"
  block at end.

# Step 3: Order

Primary sort: confidence (HIGH → MODERATE → LOW).
Secondary: updatedAt desc.
Tertiary: issue number asc.

# Step 4: Build doc

# Mentor Meeting — <YYYY-MM-DD>

## Summary slide
- ✅ #<N> — <one-sentence claim> (confidence)
- ⚠️  #<N> — ...
- ❌ #<N> — ...

## Agenda
### 1. #<N> — <title>
<full TL;DR pasted verbatim from issue body, hero figure inlined>
[→ Full report](https://github.com/superkaiba/explore-persona-space/issues/<N>)

### 2. #<N> — ...

## Open questions / asks
<aggregated "Next steps" subsections, deduplicated, ranked across all
clean-results by information gain per GPU-hour.>

## Backup slides
- #<N> — Detailed report: https://github.com/superkaiba/explore-persona-space/issues/<N>#user-content-detailed-report
- ...

# Step 5: Length check

If > 8 clean-results in the window, warn at the top of the doc:
"⚠️  N clean-results in window — consider filtering or splitting into two
meetings (Joe Benton ≤10min reading-time principle)."

# Step 6: Publish

  uv run python scripts/redact_for_gist.py --in /tmp/weekly-mentor-body.md --out /tmp/weekly-mentor-body.redacted.md
  gh gist create --public \
    --filename "weekly-mentor-${WEEK_TAG}.md" \
    --desc "Mentor meeting agenda — week ${WEEK_TAG}" \
    /tmp/weekly-mentor-body.redacted.md

If zero clean-results in the window, do NOT publish a gist. Return the
literal string "(no clean-results this week — skipping mentor agenda)"
as the output instead. The orchestrator handles that gracefully.

RETURN the gist URL (or the skip message) as the SOLE output. No commentary.

---

Subagent prompt: Mentor-prep slides

You are the mentor-prep slides subagent. Generate a Marp deck for this
week's mentor meeting using the `mentor-update-slides` skill, then publish
the deck (markdown source + rendered PDF) as a single public gist.

# Step 1: Invoke the skill

Call the Skill tool with:
  skill: mentor-update-slides
  args: "--days 7 --pdf"

The skill writes:
- figures/mentor-slides/<TODAY>/deck.md   (Marp source)
- figures/mentor-slides/<TODAY>/deck.pdf  (rendered via marp-cli)

The skill returns either a success report with the local paths, or a
literal "(skipped — …)" string if zero clean-result issues were found in
the 7-day window. If the skill skipped, do NOT publish a gist. Return the
literal string "(no clean-results this week — skipping mentor slides)" as
the SOLE output. The orchestrator handles that gracefully.

# Step 2: Redact deck.md

  TODAY=$(date +%Y-%m-%d)
  WEEK_TAG=$(date +%Y-W%V)
  uv run python scripts/redact_for_gist.py \
    --in "figures/mentor-slides/${TODAY}/deck.md" \
    --out /tmp/weekly-slides-body.redacted.md

# Step 3: Publish (markdown + PDF in one gist)

The markdown is the gist's primary file (renders inline on the gist page).
The PDF is attached as a downloadable file for the actual presentation.

  if [ -f "figures/mentor-slides/${TODAY}/deck.pdf" ]; then
    gh gist create --public \
      --filename "weekly-slides-${WEEK_TAG}.md" \
      --desc "Mentor meeting slides — week ${WEEK_TAG}" \
      /tmp/weekly-slides-body.redacted.md \
      "figures/mentor-slides/${TODAY}/deck.pdf"
  else
    # PDF render failed (marp-cli unavailable, etc.) — publish markdown only.
    gh gist create --public \
      --filename "weekly-slides-${WEEK_TAG}.md" \
      --desc "Mentor meeting slides — week ${WEEK_TAG} (PDF render failed)" \
      /tmp/weekly-slides-body.redacted.md
  fi

# Constraints

- Read-only on the project: do NOT auto-commit `figures/mentor-slides/`.
  The user reviews the deck and commits it themselves if they want it
  archived. The gist is sufficient for the meeting.
- The mentor-update-slides skill enforces its own quality checklist (no
  effect-size jargon, ≤12-word headlines, etc.). Do not re-validate here.
- If marp-cli's first render is slow (downloads Chromium for headless
  PDF), that's expected on first invocation; subsequent runs are ~2-3s.

RETURN the gist URL (or the skip message) as the SOLE output. No commentary.

---

Subagent prompt: SUMMARY gist

You are the SUMMARY-gist subagent. Your job is to regenerate the project
SUMMARY (Motivation / Related work / Current results / Immediate next
steps / Long-term goals / Glossary) and edit-in-place the persistent
public gist whose ID is stored in the first line of `docs/SUMMARY.md`.

Procedure:

1. From the repo root, run:

uv run python scripts/render_summary.py --gist

This:
- Reads `docs/SUMMARY.template.md` (Motivation / Long-term goals /
  Glossary), `docs/claims.yaml` (Current results), and live `gh
  issue list` (clean-results recent + status:proposed prio:high
  next steps).
- Writes `docs/SUMMARY.md` locally (with gist-id HTML comment
  preserved as the first line).
- Edits the persistent gist in place (`gh gist edit <id> -f
  SUMMARY.md`). On first run only it creates the gist and prints a
  stderr WARN with the exact `git add && git commit && git push`
  recipe — surface that warning in the orchestrator's report-back
  so the user can persist the new gist-id marker.

2. Capture stdout and stderr. Look for the `[render_summary] gist URL:
<url>` line.

3. Log the URL to `docs/update_log.md` with `task: summary-gist` and the
current ISO timestamp.

RETURN the gist URL as the SOLE output. If `render_summary.py` exits
non-zero, return `(failed: <one-line stderr summary>)` instead. The
orchestrator's partial-failure tolerance handles failures without
blocking the other subagents.

Rules

1. Manual trigger only. No cron.
2. Parallel dispatch. Always issue all Agent tool calls in a single assistant message — they must run concurrently.
3. One gist per task. No combined report. The user requested per-task gists.
4. Partial-failure tolerance. A failing subagent must not block the others; report the failure inline and continue.
5. Read-only on the project. None of the subagents auto-modify CLAUDE.md, code, agents, skills, or hooks. Every proposed change lands in a gist for the user to review.

Self-reflection prompt sources

The 6 weekly self-reflection prompts (Step 0.5) are paraphrased from:

1. Nanda — Highly Opinionated Advice on How to Write ML Papers — https://www.alignmentforum.org/posts/eJGptPbbFPZGLpjsp/highly-opinionated-advice-on-how-to-write-ml-papers
2. Nanda — My Research Process: Key Mindsets — Truth-Seeking — https://www.alignmentforum.org/posts/cbBwwm4jW6AZctymL/my-research-process-key-mindsets-truth-seeking
3. Chua / Hughes — Tips on Empirical Research Slides — https://www.lesswrong.com/posts/i3b9uQfjJjJkwZF4f/tips-on-empirical-research-slides
4. Nanda — My Research Process: Understanding and Cultivating Research Taste — https://www.alignmentforum.org/posts/Ldrss6o3tiKT6NdMm/my-research-process-understanding-and-cultivating-research
5. Perez — Tips for Empirical Alignment Research — https://www.alignmentforum.org/posts/dZFpEdKyb9Bf4xYn7/tips-for-empirical-alignment-research
6. Nanda — How I Think About My Research Process: Explore, Understand, Distill — https://www.lesswrong.com/posts/hjMy4ZxS5ogA9cTYK/how-i-think-about-my-research-process-explore-understand

All 6 URLs are distinct (round-2 cohesion-critic check). Phrasings are paraphrases, not direct quotes. Implementer (#251 slice 6) re-fetched the URLs at slice time to confirm faithfulness; if any source post is later updated and the paraphrase drifts, flag here and update both the prompt and the citation.