fix(cloud-agent): serialize concurrent codex prompts to fix usage accounting

[tatoalo]

Problem

codex-acp and codex-rs serialize submissions at the conversation level, so concurrent prompts are safe from codex's perspective. But our adapter accumulates per-turn usage into sessionState.accumulatedUsage via the codex-client sessionUpdate handler, and each prompt() starts with resetUsage(). If two prompts run concurrently on the JS side — e.g. two concurrent /command POSTs (a Slack follow-up racing a web-UI follow-up, or a retry after an HTTP timeout) — the second prompt's resetUsage wipes the first's in-flight counters and both TURN_COMPLETE notifications report garbled totals.

Changes

Chain CodexAcpAgent.prompt calls through a FIFO promise mutex so the usage accumulator has a single owner per turn. Errors on one prompt do not block subsequent prompts (.catch(() => {}) on the previous tail).

How did you test this?

• New unit test: two concurrent prompt() calls serialize — the second does not invoke codexConnection.prompt until the first resolves.
• New unit test: a failing prompt does not block the next one.
• Full @posthog/agent suite green (263 tests).

[tatoalo]

• fix(cloud-agent): serialize concurrent codex prompts to fix usage accounting #1721 👈 (View in Graphite)
• fix(cloud-agent): propagate taskRunId through codex loadSession #1720
• fix(cloud-agent): inject prContext into codex prompts #1719
• fix(cloud-agent): broadcast user prompt on codex cloud runs #1718
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[tatoalo]

Merge activity

• Apr 20, 9:55 AM UTC: @tatoalo merged this pull request with Graphite.