Repository navigation
fix(sessions): skip Codex replay a complete frontier covers - #4380
Conversation
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
|
I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".
|
|
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
There was a problem hiding this comment.
🔍 Devin Review: 1 flag
Not posted on this PR by your GitHub settings — view it in Devin Review. (Configure)
ScriptedAlchemy
left a comment
ScriptedAlchemy
left a comment
There was a problem hiding this comment.
Cross-referenced #4374. The replay shortcut is appropriately conditional on a complete durable frontier equal to the rebuilt index frontier, and the regression also checks a newly added rollout. No new concrete code defect found in this pass.
I changed the PR body's Fixes #4374 to Refs #4374: the issue also tracks the measured project-scope re-offer cost (approximately 28 ms/file), which this PR and #4374 (comment) explicitly leave open. The narrower restart/replay fix should leave that instrumentation/performance work tracked. Runtime tests were reviewed but not independently rerun here.
ScriptedAlchemy
left a comment
ScriptedAlchemy
left a comment
There was a problem hiding this comment.
Pushed 5e4184e after reviewing native job 119099357125 / run 39546581307 attempt 3. That run passed 167/169 targets; MCP had 667 passed and 6 failed, and runtime acceptance had 130 passed and 4 failed. The concurrent master merge at 8f3160a is preserved.
The branch now uses the canonical sized memory admission from #4393, which charges the actual text-build allocation atomically instead of adding the synthetic watermark margin. It includes the measured-headroom regression and the reviewed race-safe blocked-progress/recovery coverage: artifact refusals publish under their existing claim, and worker-entry refusals cannot overwrite a newer progress epoch or Ready.
The distinct admitted advisory-cycle timeout is still not localized by the available evidence. Independent source tracing found no demonstrated deadlock or lost owner to justify another production change. Its failure assertions now include at most 16 KiB of the existing daemon log captured before the invocation within the original phase deadline; the snapshot timing is labeled explicitly. Original timeout and retry rules remain, and admitted timeouts still fail.
Pinned rustfmt on all six changed Rust files, the whole-delta diff check and commitlint pass. Native compilation/execution is delegated to fresh Bazel CI. The memory-accounting defect is independently established, but these changes are not presented as proof that it caused every readiness or advisory failure.
…dex-replay-covered-consumer Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> # Conflicts: # crates/tracedecay-code-index-runtime/tests/code_index_scheduler/serving.rs
A restarted daemon skips replaying Codex rollouts to a late consumer when that consumer's durable complete frontier covers the replay index's exact corpus epoch. Acknowledgement returns the consumer to shared discovery. Incomplete frontiers and changed corpora still receive replay; newly added rollouts remain discoverable.
The regression restarts discovery with project and profile consumers, requires no redundant delivery for the complete profile frontier, then adds a rollout and requires delivery.
Validation and fleet repair:
Current head: f2a66c4. Full hosted CI on this head is required before merge.
Refs #4374.