Superpower Daily
GE Appliances Deploys 800 AI Agents, Reports 25% Fewer Back Ord...GPT-6 Astra Hits 99.9% on ARC-AGI-3, but Scores 62.7% in a Shar...Crusoe Signs Reported $13 Billion, Five-Year AI Cloud Deal With...Microsoft Releases MAI-Transcribe-2 at 10 Cents an HourxAI Opens Grok Bot to Enterprises With Controls for Autonomous...Figure Signs $3.5B Nscale Deal for Up to 100,000 GPUs for Human...GE Appliances Deploys 800 AI Agents, Reports 25% Fewer Back Ord...GPT-6 Astra Hits 99.9% on ARC-AGI-3, but Scores 62.7% in a Shar...Crusoe Signs Reported $13 Billion, Five-Year AI Cloud Deal With...Microsoft Releases MAI-Transcribe-2 at 10 Cents an HourxAI Opens Grok Bot to Enterprises With Controls for Autonomous...Figure Signs $3.5B Nscale Deal for Up to 100,000 GPUs for Human...
Productspublished

OpenAI’s Codex Desktop Contains a 10-Minute Pull-Request Repair Loop

Recovered instructions describe Codex revisiting GitHub, repairing PR-caused failures and optionally merging the branch. The packaged code has not been tested on a live pull request, so availability and reliability remain unconfirmed.

By 2 min read
OpenAI’s Codex Desktop Contains a 10-Minute Pull-Request Repair Loop
OpenAI’s Codex Desktop Contains a 10-Minute Pull-Request Repair Loop

Listen to this story

The audio brief

About 1:35
0:001:35
Read transcript
Codex appears to be getting a persistent pull-request repair loop: every ten minutes, it can revisit a GitHub pull request, diagnose new failures, make a fix, and potentially merge the branch. The capability was recovered from OpenAI’s latest desktop package, where it’s called Watch and fix. For each watched pull request, Codex creates a scheduled automation named Auto-fix PR, followed by the pull-request number. On every pass, it reads the current GitHub state again instead of relying on an earlier diagnosis. It checks whether the request is mergeable, reviews CI logs and annotations, and limits its work to failures caused by that pull request or to conflicts with the base branch. That boundary matters: it is not supposed to take responsibility for unrelated repository breakage. If code needs changing, Codex works in an isolated Git worktree, then can verify the repair, commit it, push it to the pull-request branch, and clean up when safe. If credentials or developer judgment are required, it asks a concise question and pauses. There’s also an Auto-merge when ready setting. Once required checks pass, Codex can merge; a failed merge attempt sends it back through diagnosis and branch updates, with retries intended to continue until the request is merged or closed. But this is package evidence, not a live test. The logic appeared in version 26.901.20858, not the earlier 26.820.60940, while real availability and reliability remain unconfirmed.

Story brief

3 key points

OpenAI’s Codex desktop package appears to add an autonomous pull-request maintenance workflow: every 10 minutes, it re-reads GitHub state, diagnoses PR-caused CI failures or base-branch conflicts, and can edit in an isolated worktree, commit, push, and optionally merge. The discovery points to a shift from one-shot coding assistance toward persistent repository operations. However, the evidence comes from version...

  1. 01

    The recovered heartbeat is named “Auto-fix PR” plus the pull-request number and pauses when credentials or developer judgment are required.

  2. 02

    Codex is constrained to failures attributable to the PR or base-branch conflicts, rather than unrelated repository breakage.

  3. 03

    Auto-merge retries diagnosis and branch updates after failed merge attempts, continuing until merge or closure.

A reverse-engineering analysis of OpenAI’s latest Codex desktop package recovered a Watch and fix mode built to keep tending a pull request after the initial task. The packaged implementation creates a scheduled agent that checks live GitHub state every 10 minutes, then is instructed to repair CI failures caused by the pull request and conflicts with its base branch.

The differentiator is the recurring task. For each watched pull request, the implementation creates a heartbeat automation called Auto-fix PR followed by the pull-request number. Rather than trust an earlier diagnosis, its instructions tell Codex to read the current GitHub state again on each run.

A repair loop built around changing repository state

The recovered instructions direct the agent to inspect mergeability and current checks before editing, using CI logs and annotations to diagnose a failure. They limit its target to failures caused by the pull request and merge conflicts with the base branch—an important boundary when other changes may also be landing in a repository.

  • For code changes, Codex is instructed to use an isolated Git worktree rather than the developer’s active checkout.
  • After a repair, it is told to verify the change, commit it, push it to the pull-request branch, and clean up the temporary worktree when safe.
  • If a developer decision or credential is needed, the instructions say to ask a concise question and pause the heartbeat.

An optional path from repair to merge

The package also contains an Auto-merge when ready preference. When enabled, the recovered instructions tell Codex to merge once required checks pass and the pull request is mergeable. A failed merge attempt sends it back to diagnosis and branch updates, with retries intended to continue until the pull request is merged or closed.

Package evidence is not a live deployment

The controls were found in Codex desktop version 26.901.20858, build 7658. Package searches did not find the same configuration keys and automatic-fix markers in the earlier analyzed version, 26.820.60940, indicating the capability was added between those examined builds.

That establishes recovered product logic and intended behavior, not a working customer workflow. The analysis did not run Watch and fix end to end on a live GitHub pull request, leaving account-level availability and its ability to handle real CI failures, conflicts, and merges unconfirmed.

Related coverage

Continue reading