Skip to content

promote: staging -> main (0012be9d) - #1208

Merged
dot-agi merged 4 commits into
mainfrom
staging
Aug 12, 2026
Merged

promote: staging -> main (0012be9d)#1208
dot-agi merged 4 commits into
mainfrom
staging

Conversation

@stateofkate

@stateofkate stateofkate commented Aug 12, 2026

Copy link
Copy Markdown
Collaborator

Caution

DO NOT USE THE MERGE BUTTON ON THIS PULL REQUEST.
THE BUTTON CREATES A NEW COMMIT AND BREAKS THE RELEASE MODEL.
COMMENT /promote TO COMPLETE THE PROMOTION.

Summary

Ships the cohort analysis work: the design pass, single-cohort analysis for tasks where every run failed or every run succeeded, per-model attribution, the QA verdict moving into the scrolling overview, and the trajectory deep-link fix.

Commits in this promotion

7cccd6c7 fix(trajectory): tag a pending deep-link step with the trial it arrived for (#1204)
2d6505de feat(cohort): analyse a single cohort, and attribute cohorts to models (#1209)
0012be9d fix(cohort): retitle, align the type scale, and land citations on their step (#1196)
a19340d0 Revert trial drawer header and shortened trial URLs (#1203)

Re-pinned from 0012be9d to 7cccd6c7. This PR was opened when staging held only the first two commits; #1209 and #1204 landed afterwards. Promoting the old pin would have shipped the citation links without the deep-link fix that makes them behave — the one combination worse than either alone.

Validation

Staging Deploy is green on 7cccd6c7, and on each commit beneath it. origin/main (09dafce3) is an ancestor, so the promotion fast-forwards. Bugbot has no unresolved threads on #1196, #1204 or #1209.

Exercised against real data on a local backend pointed at the prod database: rust-java-lsp-0a9c67d6-v20 (12 successful across 3 models, 133 failing across 14) produced a comparison that identified golden.jsonl cache-lookup reward hacking and named the four models doing it, with three citations per observation drawn from different trials. Citation validation dropped 3 unresolvable citations in that run, so the guard is doing work on live output.

Cost

SCHEMA_VERSION 1 → 3, so stored cohort comparisons regenerate on next view. Measured against analysis_costs: ~$0.28 each, single digits of stored rows, so a few dollars in total.

Known gaps

  • The trajectory summariser's coverage metric still divides by a trial's full step count while components come from drop_inert_steps, so it under-reports coverage for agents that pad with empty steps. The misleading warning built on it was removed rather than fixed; restoring it needs the summariser to persist its post-filter step count.
  • Single-cohort output has been generated for comparison-mode tasks but not yet eyeballed for an all-success task in the deployed environment.

@vercel

vercel Bot commented Aug 12, 2026

Copy link
Copy Markdown
Contributor

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated (UTC)
oddish-app Ready Ready Preview Aug 12, 2026 5:34am
oddish-app (staging) Ready Ready Preview Aug 12, 2026 5:34am

Comment thread frontend/src/components/trajectory-viewer.tsx Outdated

@cursor cursor Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Stale comment

Risk: medium. Left a non-blocking comment — Cursor Bugbot found 1 unresolved medium issue (stale step after trial switch) and the Bugbot check finished as skipped, so I am not approving. Assigned reviewers for human attention.

Open in Web View Automation 

Sent by Cursor Approval Agent: Pull Request Router and Approver

@cursor
cursor Bot requested review from charlesyhuang and kyle-compute August 12, 2026 04:13

@cursor cursor Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Risk: medium. Left a non-blocking comment — Cursor Bugbot did not finish on the current HEAD within the 8-minute wait, so I am not approving. Human reviewers are already requested; not assigning more.

Open in Web View Automation 

Sent by Cursor Approval Agent: Pull Request Router and Approver

@cursor cursor Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

Fix All in Cursor

❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.

Reviewed by Cursor Bugbot for commit 7cccd6c. Configure here.

take();
window.addEventListener("hashchange", take);
return () => window.removeEventListener("hashchange", take);
}, []);

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Deep link lost on viewer remount

Medium Severity

Citation deep links strip #step-N into pendingStep on mount, but that state lives only inside TrajectoryViewer. ActiveTabContent unmounts inactive tabs, so switching away before the trajectory loads (or any parent remount) drops the pending step after the fragment is already gone, and the cited step never scrolls into view.

Additional Locations (2)
Fix in Cursor Fix in Web

Reviewed by Cursor Bugbot for commit 7cccd6c. Configure here.

@stateofkate

Copy link
Copy Markdown
Collaborator Author

/promote 7cccd6c

@github-actions

Copy link
Copy Markdown
Contributor

Promoted. main now points at 7cccd6c7d7aa7756442d4c3d7e52ba4827a9c56f, the same commit as staging.

@dot-agi
dot-agi merged commit 7cccd6c into main Aug 12, 2026
20 checks passed
@railway-app
railway-app Bot temporarily deployed to pleasing-unity / production August 12, 2026 05:40 Inactive
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants