divecode-pr-watch

name	divecode-pr-watch
description	Watch a PR — auto-respond to comments and CI failures. Uses bin/divecode-pr-watch which polls via gh and emits a routing decision (status + route). 6-status / 7-route routing per agent-flow: green→merge, has_comments→fix-loop, ci_failed→fix-loop, pending→block, closed→block, merged→cleanup, error→block. Use after /divecode-push-pr, before /divecode-merge.
triggers	["divecode pr-watch","watch pr","poll pr"]
allowed-tools	["Bash","Read","Write"]

divecode-pr-watch — status-driven routing

You are running divecode-pr-watch. Read the PR's live status, decide the next route per agent-flow's 6-status contract.

Iron Laws

1. Status is computed, not assumed. Always actually fetch the PR.
2. One status → one route. No "maybe try X then Y" — the routing table is deterministic.
3. Block ≠ fail. Pending CI is not a failure; just don't merge yet.
4. Closed (not merged) requires user input. Never auto-cleanup a closed-without-merge PR.

Preamble

PROJ_DIR="${DIVECODE_PROJECT_DIR:-$PWD}"
DIVECODE_HOME="${DIVECODE_HOME:-$HOME/.divecode}"
[ -x "$DIVECODE_HOME/bin/divecode-pr-watch" ] || DIVECODE_HOME="$HOME/Trappist/divecode"

# Find PR number from push-pr.md
PR=$(grep -E '^url:' "$PROJ_DIR/divecode/push-pr.md" 2>/dev/null | head -1 | grep -oE '/pull/[0-9]+' | grep -oE '[0-9]+')
[ -z "$PR" ] && { echo "could not find PR in divecode/push-pr.md"; exit 1; }
echo "PR: $PR"

Workflow

Step 1 — Fetch + route

bash "$DIVECODE_HOME/bin/divecode-pr-watch" "$PR" > "$PROJ_DIR/divecode/pr-watch.md"
cat "$PROJ_DIR/divecode/pr-watch.md"

Routing table (matches agent-flow exactly):

status	route	action
green	merge	invoke /divecode-merge
has_comments	fix-loop	invoke /divecode-fix-loop with PR comments converted into must-fix items
ci_failed	fix-loop	invoke /divecode-fix-loop; reproduce the failure locally, fix, push
pending	block	sleep / wait — re-invoke pr-watch after ~5min
closed (not merged)	block	ask user before proceeding
merged	cleanup	invoke /divecode-cleanup
error	block	surface error; do not retry until user input

Step 2 — Branch per route

Per the table. For pending, suggest the user re-invoke later or set a timer.

Step 3 — Append round entry to pr-watch.md

Each poll/action is recorded:

## Round <N> — <date>
status: <s>
route:  <r>
action: <what was done>

Done criteria

• pr-watch.md shows latest status
• Next skill invoked (or user prompted for pending/closed)