diff --git a/AGENTS.md b/AGENTS.md
index a99f9b616..91a5a4abb 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -11,3 +11,51 @@ details of a change in its commit messages and PR description instead.
 If a PR you work on already changes `CHANGELOG.md`, restore that file to its
 merge-base version. The only exceptions are release PRs: a `release/v*`
 branch, or the release train's `release-sync` PR.
+
+## Wait on CI with a monitor, not a polling loop
+
+To wait on a PR's checks or its trip through the merge queue, run
+`scripts/ci-watch.sh` under the Monitor tool instead of looping on `sleep` and
+`gh pr checks`, or polling with `/loop`. The script stays silent while
+nothing changes and prints one line per event: a start summary, each failed
+check by name, progress, queue position, and a final `DONE ...` line when it
+exits.
+
+```
+scripts/ci-watch.sh [PR]            # until every check settles (exit 0 pass, 1 fail)
+scripts/ci-watch.sh [PR] --merge    # through the merge queue (exit 0 merged, 1 closed/dequeued)
+```
+
+PR defaults to the current branch's. Give Monitor the maximum `timeout_ms`
+and re-arm it when it expires before `DONE`. Act on a `fail:` line as soon as
+it arrives; there's no need to wait for the remaining checks. If a session
+has no Monitor tool, run the script with Bash `run_in_background` and read
+its output when it exits.
+
+## Run /code-review before a PR leaves draft
+
+Before you mark a PR ready for review or add it to the merge queue, run
+`/code-review high` on its diff. Fix every confirmed finding
+(`/code-review high --fix` applies them) and push the fixes before going on.
+For a finding you decide not to fix, say why in the PR description. When you
+review someone else's PR, use `/code-review <PR> --comment` to post the
+findings as inline comments.
+
+## Share the GitHub API budget
+
+Every agent and routine here calls GitHub as the same account, so all of
+them share one budget of 5,000 REST requests and 5,000 GraphQL points
+per hour. When it runs out, every agent stalls until the hourly reset.
+
+- Before a read of more than ~20 requests, check what is left with
+  `gh api -i repos/SocketDev/socket-patch | grep -i '^x-ratelimit-remaining'`.
+  Trust that header over `gh api rate_limit`. Below 1,000, do only your
+  most important write and finish early.
+- Never `--paginate` over Actions runs, jobs or issues without a date
+  filter and a page cap, and never run more than 2 GitHub requests in
+  parallel.
+- On a 403 or 429 that mentions a rate limit, stop calling GitHub and
+  report what you have. Don't retry in a loop, and don't schedule a
+  reminder just to retry after the reset.
+- Keep at most one pending re-check reminder per PR, at least 30 minutes
+  out.
diff --git a/scripts/ci-watch.sh b/scripts/ci-watch.sh
new file mode 100755
index 000000000..9a3ca2370
--- /dev/null
+++ b/scripts/ci-watch.sh
@@ -0,0 +1,120 @@
+#!/usr/bin/env bash
+# Event stream of a PR's CI, for Claude Code's Monitor tool (or a terminal).
+#
+#   scripts/ci-watch.sh [PR] [--merge] [--interval SECONDS]
+#
+# PR defaults to the current branch's PR. Prints one line per event and
+# nothing while nothing changes, so it stays quiet enough for Monitor:
+#   start: #<pr> <sha>, <p> pass <f> fail <q> pending
+#   fail: <check name>                            -- one line per failed check
+#   progress: <settled>/<total> settled, <f> fail -- at most every 5th poll
+#   queue: entered at <n> | position <n> <state> | left the queue
+#   DONE checks pass|fail (<p> pass, <f> fail)   -- exits here without --merge
+#   DONE merged|closed|dequeued                  -- exits here with --merge
+#   warn: ...                                    -- a failed poll; keeps going
+#
+# Every agent on this account shares one GitHub API budget, so each poll
+# is one small GraphQL query; the full check list (several pages for a PR
+# here) is fetched only when the rollup state changes or every 5th poll.
+# A rate-limited poll pauses for 10 minutes instead of retrying.
+# Exit status: 0 on pass/merged, 1 on fail/closed/dequeued, 2 on bad usage.
+set -uo pipefail
+
+pr="" merge=0 interval=120
+while [ $# -gt 0 ]; do
+  case "$1" in
+    --merge) merge=1 ;;
+    --interval) interval="$2"; shift ;;
+    -h|--help) sed -n '2,15p' "$0"; exit 0 ;;
+    [0-9]*) pr="$1" ;;
+    *) echo "usage: $0 [PR] [--merge] [--interval SECONDS]" >&2; exit 2 ;;
+  esac
+  shift
+done
+
+if [ -z "$pr" ]; then
+  pr=$(gh pr view --json number -q .number 2>/dev/null) || { echo "no PR for the current branch" >&2; exit 2; }
+fi
+repo=$(gh repo view --json nameWithOwner -q .nameWithOwner) || exit 2
+owner=${repo%/*} name=${repo#*/}
+
+# shellcheck disable=SC2016 # GraphQL variables, not shell
+query='query($o:String!,$n:String!,$p:Int!){repository(owner:$o,name:$n){pullRequest(number:$p){
+  state headRefOid mergeQueueEntry{position state}
+  commits(last:1){nodes{commit{statusCheckRollup{state}}}}}}}'
+
+errf=$(mktemp) || exit 2
+trap 'rm -f "$errf"' EXIT
+seen="" head="" queued="" was_in_queue=0 settled_before=-1 rollup_before="" polls=0
+total=0 pending=1 fail=0
+while true; do
+  if ! json=$(gh api graphql -f query="$query" -F o="$owner" -F n="$name" -F p="$pr" 2>&1); then
+    if grep -qiE 'rate limit|HTTP 403|HTTP 429' <<<"$json"; then
+      echo "warn: GitHub rate limit hit; pausing 10 minutes"; sleep 600; continue
+    fi
+    echo "warn: gh ${json%%$'\n'*}"; sleep "$interval"; continue
+  fi
+  state=$(jq -r '.data.repository.pullRequest.state' <<<"$json")
+  sha=$(jq -r '.data.repository.pullRequest.headRefOid' <<<"$json")
+  if [ "$sha" != "$head" ]; then
+    [ -n "$head" ] && echo "push: head now ${sha:0:9}, checks restart"
+    head=$sha seen="" settled_before=-1
+  fi
+
+  rollup=$(jq -r '.data.repository.pullRequest.commits.nodes[0].commit.statusCheckRollup.state // "NONE"' <<<"$json")
+  polls=$((polls + 1))
+  if [ "$settled_before" = -1 ] || [ "$rollup" != "$rollup_before" ] || [ $((polls % 5)) = 0 ]; then
+    # gh pr checks pages through every check; its exit status is 8 while any
+    # is pending and 1 when one failed, so only its output is trusted.
+    checks=$(gh pr checks "$pr" -R "$repo" --json name,bucket 2>"$errf")
+    if ! jq -e 'type == "array"' <<<"$checks" >/dev/null 2>&1; then
+      # Keep the last counts and retry this fetch on the next poll; the
+      # paginated call is the expensive one, so back off on a rate limit.
+      rollup_before=""
+      if grep -qiE 'rate limit|HTTP 403|HTTP 429' "$errf"; then
+        echo "warn: GitHub rate limit hit; pausing 10 minutes"; sleep 600; continue
+      fi
+      echo "warn: gh pr checks returned no JSON"; sleep "$interval"; continue
+    fi
+    total=$(jq 'length' <<<"$checks")
+    pending=$(jq '[.[] | select(.bucket == "pending")] | length' <<<"$checks")
+    fail=$(jq '[.[] | select(.bucket == "fail" or .bucket == "cancel")] | length' <<<"$checks")
+    failed=$(jq -r '.[] | select(.bucket == "fail" or .bucket == "cancel") | .name' <<<"$checks" | sort -u)
+    settled_now=$((total - pending))
+    if [ "$settled_before" = -1 ]; then
+      echo "start: #$pr ${sha:0:9}, $((settled_now - fail)) pass $fail fail $pending pending"
+      printf '%s\n' "$failed" | grep -v '^$' | sed 's/^/fail: /'
+    else
+      comm -13 <(printf '%s\n' "$seen") <(printf '%s\n' "$failed") | grep -v '^$' | sed 's/^/fail: /'
+      [ "$settled_now" != "$settled_before" ] && [ "$pending" != 0 ] \
+        && echo "progress: $settled_now/$total settled, $fail fail"
+    fi
+    seen=$failed settled_before=$settled_now
+  fi
+  rollup_before=$rollup
+
+  entry=$(jq -r '.data.repository.pullRequest.mergeQueueEntry | if . then "\(.position) \(.state | ascii_downcase)" else "" end' <<<"$json")
+  if [ -n "$entry" ] && [ "$entry" != "$queued" ]; then
+    [ "$was_in_queue" = 1 ] && echo "queue: position $entry" || echo "queue: entered at $entry"
+    queued=$entry was_in_queue=1
+  elif [ -z "$entry" ] && [ "$was_in_queue" = 1 ] && [ "$state" = OPEN ]; then
+    echo "queue: left the queue"; queued="" was_in_queue=0
+    [ "$merge" = 1 ] && { echo "DONE dequeued"; exit 1; }
+  fi
+
+  case "$state" in
+    MERGED) echo "DONE merged"; exit 0 ;;
+    CLOSED) echo "DONE closed"; exit 1 ;;
+  esac
+
+  # gh pr checks lists only check runs that exist; the rollup also counts
+  # expected and re-queued ones, so both must agree before calling it done.
+  if [ "$merge" = 0 ]; then
+    if [ "$total" -gt 0 ] && [ "$pending" = 0 ] \
+      && { [ "$rollup" = SUCCESS ] || [ "$rollup" = FAILURE ] || [ "$rollup" = ERROR ]; }; then
+      if [ "$fail" = 0 ]; then echo "DONE checks pass ($total checks)"; exit 0; fi
+      echo "DONE checks fail ($fail of $total failed)"; exit 1
+    fi
+  fi
+  sleep "$interval"
+done