GeezerVison

Problem

AGENTS.md tells every agent to wait on CI with scripts/ci-watch.sh, but in Claude Code cloud sessions (where every routine in this fleet runs) GraphQL now returns HTTP 403: GitHub GraphQL is not available from Claude Code sessions. The script calls gh repo view (GraphQL) before anything else, so it exits 2 at startup there with no event line; gh pr view, gh pr checks and the gh api graphql poll are all GraphQL too. Verified in this session on 2026-10-17: gh repo view, gh pr checks 1590 and the queue GraphQL query all 403.

Agents then fall back to ad-hoc polling, which is exactly what AGENTS.md asks them not to do (and costs more of the shared API budget).

Root cause

The watcher is GraphQL-only; cloud sessions only allow REST.

Fix

• Repo comes from GH_REPO or the origin URL (falls back to gh repo view).
• Default PR: gh pr view, else GET pulls?head=<owner>:<branch>.
• On the first GraphQL is not available 403 the script switches to REST for good. Each poll reads pulls/N (state, merged, head, base), the head's check runs (filter=latest default) and commit statuses; the base branch's required checks (rules/branches/<base>, read once: ci-ok, clippy) must all have reported before it calls DONE, standing in for the rollup that counts expected checks.
• --merge over REST: REST has no merge-queue entry, so the queue lines follow the PR's unfinished merge_group runs (gh-readonly-queue/<base>/pr-N-*); left the queue needs two misses in a row so a rebuild isn't read as a dequeue. Position is shown as building.
• After a merge group's runs all pass, the watcher waits for MERGED (up to 10 min) instead of counting a dequeue; while the required-check list can't be read the rollup stays pending (Bugbot findings, fixed in b53f491).
• The GraphQL path is unchanged where GraphQL works (local terminals).

Budget: REST mode is 3 requests per poll (4 with --merge), at the default 120 s interval ~90-120/h per watcher.

Proof

Live, in this cloud session:

$ scripts/ci-watch.sh 1590 --interval 5
warn: GraphQL unavailable here; polling over REST
start: #1590 c38b16316, 55 pass 0 fail 0 pending
DONE checks pass (55 checks)
$ scripts/ci-watch.sh 1586 --merge --interval 5
warn: GraphQL unavailable here; polling over REST
start: #1586 bca69ddeb, 86 pass 0 fail 0 pending
DONE merged

Before this change both exit 2 immediately.

With a stubbed gh (pending→failure, required ci-ok arriving late, queue enter→merge, queue enter→dequeue after two misses):

start: #7 abc123def, 2 pass 0 fail 1 pending
fail: test
fail: ci-ok
DONE checks fail (2 of 4 failed)          # waited for ci-ok before DONE
---
queue: entered at building
DONE merged
---
queue: entered at building
queue: left the queue
DONE dequeued

bash -n clean; shellcheck isn't installed in this environment.

Where tests run

No tests or jobs change; this is an agent-side script only.

https://claude.ai/code/session_01AqtsyAYsehu1NAZhKbg3o5

Generated by Claude Code