nanoclaw-check

name	nanoclaw-check
description	Monitor qwibitai/nanoclaw upstream for activity related to Chip's contributions. Classifies new issues/PRs/comments as FYI vs response-needed, drafts responses for review, delivers a summary to Telegram. Triggered by scheduled task; safe to run manually as `/nanoclaw-check`.

/nanoclaw-check — nanoclaw upstream monitor

Background skill. Runs on a schedule (typically 9am, 1pm, 5pm). Reads stable context from this skill folder; reads/writes mutable state from the group folder. Drafts responses but NEVER posts them autonomously — Chip approves before anything ships to a public thread.

Inputs

• Stable context (read-only): ./context.md in this skill's folder. Contains active threads, response policies, voice patterns, scope discipline, technical context.
• Mutable state (read/write): /workspace/group/nanoclaw-monitor/last-check.json (timestamp) and /workspace/group/nanoclaw-monitor/history.md (event log).
• Auth: gh CLI authenticated via GH_TOKEN env var (read-only public-repo access scope is enough).

Step 1 — Load context (required)

Read ./context.md in full before proceeding. Do not skip this step. The classification rules, response policies per thread, voice guide, and scope discipline live there. All downstream decisions depend on it.

If ./context.md is missing or empty, abort with a message to Chip via mcp__nanoclaw__send_message: "context.md missing in nanoclaw-check skill folder — skill cannot run safely". Do not improvise without context.

Step 2 — Read state

Read /workspace/group/nanoclaw-monitor/last-check.json. Expected shape:

{
  "lastCheckedIso": "2026-04-24T17:00:00Z",
  "lastDeliveryIso": "2026-04-24T17:05:00Z",
  "lastEventCount": 2
}

If missing, treat as first run: set lastCheckedIso to 7 days ago. Create the directory if needed.

Read /workspace/group/nanoclaw-monitor/history.md if present. Append-only event log; use it for recent-event memory (e.g., "we already flagged this PR yesterday — only mention if there's new activity since then").

Step 3 — Query upstream activity since last check

Goal: find any new issues, PRs, comments, or reviews on qwibitai/nanoclaw since lastCheckedIso.

Strategy:

1. Check tracked threads (numbers in context.md §Active threads — currently #1843, #1955, #1956, #1966, #1984, #1994, #1995). For each:

   gh issue view <N> --repo qwibitai/nanoclaw --json comments,state,closedAt
   gh pr view <N> --repo qwibitai/nanoclaw --json comments,reviews,state,mergedAt 2>/dev/null
   

   Filter comments/reviews by createdAt > lastCheckedIso.

2. Find new issues/PRs in the repo touching keywords from context.md (codex, provider, opencode, file-ops, latency, mcp, @chiptoe-svg):

   gh search issues "repo:qwibitai/nanoclaw updated:>=$LAST_CHECK codex" --json number,title,author,createdAt,state
   gh search issues "repo:qwibitai/nanoclaw updated:>=$LAST_CHECK provider"
   # ...repeat for other keywords
   

   Use --updated filter (catches issues with new comments without per-thread queries).

3. Watch for direct mentions of @chiptoe-svg anywhere in the repo:

   gh search issues "repo:qwibitai/nanoclaw mentions:chiptoe-svg updated:>=$LAST_CHECK"
   

Deduplicate results.

Step 4 — Classify each item

For each new event, classify per context.md §Scope discipline rules:

• FYI (no action) — informational, no response needed:

  • Comments by maintainer or others on threads that don't ask Chip anything
  • +1 reactions or simple acknowledgments
  • Activity on threads where context.md response policy says "Staying silent" (e.g., #1995)
  • Activity on threads we already commented on where conversation is quiet

• RESPONSE NEEDED — draft a reply per voice guide:

  • Direct @mention or quote of Chip
  • Maintainer responds to one of Chip's open issues (#1955, #1956) with design questions
  • Someone asks a question about Chip's #1843 code that needs context only Chip has
  • Reviewer feedback on Chip's posted comments asking for clarification

When in doubt, classify as FYI. Over-flagging "needs response" trains learned helplessness.

Step 5 — Draft responses (only for RESPONSE NEEDED items)

For each item classified RESPONSE NEEDED:

1. Read the full thread (the relevant comment + immediate prior context). Use gh pr view <N> --comments or gh issue view <N> --comments.
2. Match Chip's voice per context.md §Voice guide:
   • Acknowledge the work first
   • Bold-label structure for multi-point responses
   • Honest about scope; avoid hedges ("really", "actually", "just")
   • End with scope clarity
3. Generate draft. Keep it tight — drafts that exceed ~250 words probably overstuffed.
4. Provide reasoning for why you suggested this response (1–2 sentences). Helps Chip evaluate.

If drafting feels uncertain (e.g., novel design question, unclear voice fit), don't draft — flag as "needs Chip's eyes" with no draft. Better than a bad draft.

Step 6 — Deliver summary via Telegram

Use mcp__nanoclaw__send_message to send a summary to the current group (Telegram main).

When to send vs. stay silent

• Send: any RESPONSE NEEDED items, OR any FYI items the agent thinks Chip wants to know about (significant merges, new contributors engaging Chip's threads, etc.)
• Stay silent: all events were trivial FYI (minor non-Chip comments on quiet threads, etc.) AND less than 7 days since last delivery
• Heartbeat: if 7+ days since last delivery, send a brief "all quiet — last 7 days had N small events, none requiring action" so Chip knows the skill is still running

Format

📬 Nanoclaw update — N items since <last check>

## No action
- [#NNNN](url) — 1-line summary
- [#NNNN](url) — 1-line summary

## Worth responding to
### [#NNNN](url) — 1-line context
**Suggested response:**

``` **Why:** <1-2 sentence reasoning>

---

(or if quiet:)

📬 Nanoclaw quiet check — no new activity since

Use Markdown that Telegram renders. Truncate any URL with > 80 chars.

## Step 7 — Update mutable state

Atomic-ish write to `/workspace/group/nanoclaw-monitor/last-check.json`:

```json
{
  "lastCheckedIso": "<now ISO>",
  "lastDeliveryIso": "<now ISO if delivered, else preserve prior>",
  "lastEventCount": <N>
}

Append to /workspace/group/nanoclaw-monitor/history.md:

## 2026-04-25 13:00 EST — N events
- FYI: #NNNN — 1-line summary
- RESPONSE NEEDED: #NNNN — 1-line summary (draft delivered)
- HEARTBEAT (no events) — when applicable

Keep history.md append-only. Trimming/archiving is Chip's call.

Step 8 — Flag stale context (occasional)

If during analysis the agent notices context.md looks stale (e.g., a thread is closed and no longer needs monitoring, a contributor hasn't been active in 30+ days), include a one-liner at the bottom of the delivery message:

🔧 Context note: thread #1955 closed last week — consider removing from active threads list on next rebuild.

Don't edit context.md directly. Flag it for Chip.

Failure modes

• gh auth fails → send error message to Chip, don't crash silently
• context.md missing → abort per Step 1
• Network/rate-limit error → log to history.md, retry on next scheduled run
• Drafting can't decide between FYI and RESPONSE → classify as RESPONSE without draft, let Chip judge
• State file corruption → recreate with lastCheckedIso = 24 hours ago (conservative restart)

Manual invocation

This skill is safe to run manually any time as /nanoclaw-check. Will use the same state file, so manual + scheduled runs interleave cleanly.

What this skill does NOT do

• Post comments to GitHub (autonomously or otherwise)
• Edit context.md (only flag for Chip's manual update)
• Notify Chip of events Chip already commented on within the last 24h (unless someone responds to him)
• Engage threads explicitly listed as "Staying silent" in context.md
• Make architectural recommendations to Chip beyond what's in context.md voice guide