[Sync] llm stuff

common/.config/claude/skills/pr-loop/SKILL.md

@@ -2,7 +2,7 @@
 description: Babysit the current branch's PR — address review comments that don't need my input, re-check every ~10 min, and stop when there are no open comments left or the automated reviewers (Codex, Claude) hit their limit. Prep-only — stages fixes and drafts the commit message; never commits, pushes, or merges.
 argument-hint: [pr number or url — optional; defaults to the current branch's PR]
 disable-model-invocation: true
-allowed-tools: Bash(gh:*), Bash(git status:*), Bash(git diff:*), Bash(git log:*), Bash(git rev-parse:*), Bash(git fetch:*), Bash(git add:*), Bash(wl-copy:*), Read, Edit, Write, Grep, Glob, ScheduleWakeup
+allowed-tools: Bash(gh:*), Bash(git status:*), Bash(git diff:*), Bash(git log:*), Bash(git rev-parse:*), Bash(git fetch:*), Bash(git add:*), Bash(wl-copy:*), Bash(~/.config/git/claude-commit-msg:*), Read, Edit, Write, Grep, Glob, ScheduleWakeup
 ---
 
 # Watch this PR and resolve review comments (prep-only)
@@ -36,11 +36,12 @@ comments left or the automated reviewers (Codex/Claude) hit a usage limit.
   session stays open**. **To stop it: press `Esc` while I'm idle between cycles** (that clears the
   queued wake-up); closing the session also stops it. A plain message does **not** cancel a pending
   wake-up — use `Esc`.
-- **Every token counts.** Each wake-up re-reads the whole conversation with a cold prompt cache
-  (the 10-min cadence outlives the 5-min cache TTL), so cost compounds with every cycle and every
-  extra line in context. Therefore: fetch **filtered** data only (use the `--jq` filters below —
-  never let raw unfiltered JSON into the conversation), summarize deltas rather than restating
-  unchanged state, and keep idle cycles to **one cheap API call and ≤2 lines of output**.
+- **Every token counts.** Each wake-up re-sends the whole conversation. On a Claude subscription
+  the prompt cache holds for **1 hour**, so the 10-min cadence stays warm; on an API key it drops to
+  5 min and every cycle is a cold read. Either way every extra line in context rides along on every
+  later cycle. Therefore: fetch **filtered** data only (use the `--jq` filters below — never let raw
+  unfiltered JSON into the conversation), summarize deltas rather than restating unchanged state,
+  and keep idle cycles to **one cheap API call and ≤2 lines of output**.
 
 Current state (auth + the current branch's PR):
 
@@ -233,10 +234,11 @@ Current state (auth + the current branch's PR):
 9. **Draft the commit message — exactly like `/commit-msg`:** from the staged diff and the repo's
    recent `git log` style (this repo uses `[Scope] summary`), write a message that matches. Then:
    - **(a) Print it** in your reply as a fenced ` ```text ` block — the durable copy.
-   - **(b) Write it** with the Write tool to the path from `git rev-parse --git-path CLAUDE_COMMIT_MSG`
-     (**not** `COMMIT_EDITMSG` — git overwrites that on `git commit`; the `prepare-commit-msg` hook
-     prefills the editor from `CLAUDE_COMMIT_MSG`).
-   - **(c) Copy it:** `wl-copy < <that path>`.
+   - **(b) Hand it off:** `~/.config/git/claude-commit-msg save <<'EOF'` … `EOF` (the message as a
+     quoted heredoc, so quotes and backticks survive) — it is stored with the paths you staged this cycle, so the `prepare-commit-msg` hook prefills it only
+     for the commit that contains them (other sessions' handoffs are not overwritten). Never write
+     `CLAUDE_COMMIT_MSG` / `COMMIT_EDITMSG` directly.
+   - **(c) Copy it:** `wl-copy < <file path printed by save>`.
    - Tell me all three locations. **Never** run `git commit`.
    (Skip this step when nothing new was staged this cycle.)
 
@@ -252,9 +254,10 @@ Current state (auth + the current branch's PR):
     needs-input questions are posted (reactions don't bump `updatedAt`, comments do — so your own
     activity doesn't defeat the next cheap check), then save state
     to `<git-dir>/pr-loop-state.json`. Then schedule:
-    - **normal cycle:** `ScheduleWakeup(delaySeconds=600, prompt="/pr-loop", reason="recheck PR
-      #<num> review comments")`;
-    - **idle cycle** (step 3 short-circuited): back off — `delaySeconds=1200`, I'm clearly away;
+    - **normal cycle:** `ScheduleWakeup(delaySeconds=600, noop=false, prompt="/pr-loop",
+      reason="recheck PR #<num> review comments")`;
+    - **idle cycle** (step 3 short-circuited): back off — `delaySeconds=1200, noop=true` (idle
+      ticks then collapse into one line in my terminal instead of stacking), I'm clearly away;
       re-running `/pr-loop` checks immediately.
 
     End the turn with a sign-off that states when you'll run again and how to stop, e.g.:
@@ -268,7 +271,9 @@ Current state (auth + the current branch's PR):
 
 - **Never** `git commit`, `git push`, `gh pr merge`, `gh pr create`, `gh pr close`, or `gh pr edit`
   — I do all of those. You stage, react, resolve, and draft; nothing more. The only comment you may
-  post is a needs-input decision question.
+  post is a needs-input decision question. This is enforced, not just asked: `settings.json` denies
+  those commands and the `git-guard.py` PreToolUse hook blocks any spelling of them, so a slip is
+  refused with a reason rather than executed.
 - Act only on the target PR. Don't touch unrelated files or other PRs.
 - **Keep context lean:** no raw JSON dumps in replies, no re-listing threads that haven't changed,
   idle cycles ≤2 lines. Every line you emit is re-read (uncached) on every later cycle.