## Problem
A second `@claude` request on a PR while a Claude run is already in
progress gets silently cancelled. Concrete case on #21443: [this
comment](https://github.com/twentyhq/twenty/pull/21443#discussion_r3404037004)
("@claude please investigate and report", 14:24:28) produced [a
run](https://github.com/twentyhq/twenty/actions/runs/27421867558) that
was cancelled 3 seconds after creation, while [the run for an earlier
comment](https://github.com/twentyhq/twenty/actions/runs/27421856742)
was still in progress. Claude never responded.
## Root cause
The `concurrency` block is declared at the **workflow** level, keyed on
the PR number. Two GitHub Actions behaviors combine badly here:
1. A concurrency group holds at most one running + **one pending** run;
every new run entering the group cancels the previously pending one
(`cancel-in-progress: false` only protects the *running* run).
2. Workflow-level concurrency is acquired **before** job-level `if`
conditions are evaluated — so every `issue_comment` /
`pull_request_review_comment` / `pull_request_review` event on the PR
enters the group, even ones with no `@claude` that end up skipped.
So while a Claude run is in progress, any PR activity evicts the queued
`@claude` run. It's even self-defeating: a review-thread reply fires
*two* webhook events (`pull_request_review_comment` + a companion
`pull_request_review` with an empty body), so the companion event
cancels the queued comment run ~2s later. Claude's own "finished" reply
also fires events that kill whatever is queued.
## Fix
Move concurrency to the **job** level. A job whose `if` evaluates false
is skipped before it ever requests the concurrency slot, so only genuine
`@claude` jobs enter the queue. Real Claude runs on the same PR are
still serialized (no parallel pushes to the same branch).
Also gives `claude-cross-repo` its own group keyed on source repo +
issue number — previously a cross-repo dispatch for issue N shared a
group with PR N in this repo and they could needlessly queue behind each
other.
## Remaining limitation
GitHub keeps only one pending job per group: posting three `@claude`
requests while the first is still running will still cancel the second
when the third arrives. Zero-loss queueing would require a unique group
per comment, which would allow concurrent runs pushing to the same PR
branch — not worth it.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
<!-- This is an auto-generated description by cubic. -->
<a
href="https://cubic.dev/pr/twentyhq/twenty/pull/21498?utm_source=github"
target="_blank" rel="noopener noreferrer"
data-no-image-dialog="true"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://www.cubic.dev/buttons/review-in-cubic-dark.svg"><source
media="(prefers-color-scheme: light)"
srcset="https://www.cubic.dev/buttons/review-in-cubic-light.svg"><img
alt="Review in cubic"
src="https://www.cubic.dev/buttons/review-in-cubic-dark.svg"></picture></a>
<!-- End of auto-generated description by cubic. -->
Félix Malfait
79f7c96939
