The build-deploy job posts a 'Prêt à tester' comment whose body literally
contains 'Reply **LGTM** pour merger' and 'Reply **KO**' as instructions to
the reviewer. wait-approval re-fetches ALL comments every 30s and greps any
body for \bLGTM\b (checked before KO) — so on the first poll (~30s after
deploy) the bot matched its OWN comment and squash-merged automatically,
with no human review. Regressed in #6937a36 (PR #12) which tightened the
regex to catch 'Tested, LGTM!' but didn't notice the bot body matched.
Fix (sentinel marker + python parse, identity-agnostic so it's safe whether
PR_API_TOKEN is a bot or a personal account):
- build-deploy: embed <!-- tabatago:ready-to-test --> in the bot comment
(HTML comment, invisible in rendered markdown, present in raw body).
- wait-approval: replace the fragile grep-over-JSON with a python3 heredoc
using json.loads + the existing \bLGTM\b / \bKO\b regexes. Skip any
comment whose body contains the sentinel marker. Check KO BEFORE LGTM so
a PR with both signals blocks (matches the existing 'KO blocks' intent).
LGTM still does {"Do":"squash"}; KO still blocks; 2h timeout unchanged.
Verified locally with three scenarios:
bot-only -> PENDING (was LGTM before — the bug)
bot + reviewer -> LGTM (squash-merge still fires)
bot + KO -> KO (block still fires, even with bot LGTM in history)
AGENTS.md §5 step 9, §5 pitfalls, and §8 anti-patterns updated so the doc
describes the marker rule and warns against re-introducing the self-match.
98f4f82db2