space/gitea-codex
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

improvements: Big Reviewer Update #4

Merged
space merged 7 commits from improv/big-reviewer-update into main 2026-05-23 14:22:33 +02:00
Conversation 13 Commits 7 Files Changed 19 +766 -663
space commented 2026-05-23 13:37:23 +02:00
Owner
Copy Link
No description provided.
space added 1 commit 2026-05-23 13:37:24 +02:00
feat: Enhance review prompt with detailed instructions and placeholders for empty sections
All checks were successful
ci / test (pull_request) Successful in 34s
Details
ci / publish (pull_request) Has been skipped
Details
d6a9397914
space commented 2026-05-23 13:53:06 +02:00
Author
Owner
Copy Link

@codex review

@codex review
gitea-codex commented 2026-05-23 13:53:08 +02:00
First-time contributor
Copy Link

👀 Codex review queued for commit d6a9397.

👀 Codex review queued for commit `d6a9397`.
space added 1 commit 2026-05-23 13:53:30 +02:00
Resolve Runner issue, runner refused to launch
All checks were successful
ci / test (pull_request) Successful in 35s
Details
ci / publish (pull_request) Has been skipped
Details
c3925f37e1
gitea-codex commented 2026-05-23 13:54:18 +02:00
First-time contributor
Copy Link

Review Result

This PR introduces a checkout robustness regression that can cause valid PR reviews to fail.

Findings

1) PR checkout assumes both refs exist on one remote

In src/gitea_codex_bot/workers/container_runner.py (around the new fetch step), the runner does:

  • clone one clone_url as origin
  • git fetch --no-tags origin <base_ref> <head_ref>

For fork and other cross-repo topologies, base_ref and head_ref are not guaranteed to both exist on that same remote. If either ref is missing, fetch fails and review aborts before checkout.

This is a correctness issue because it can reject reviewable PRs purely due to ref layout.

Suggested Fix

Use SHA-based fetch (or independent tolerant fetches) rather than requiring both branch names from one remote. The runner already has base_sha and head_sha, which are stable identifiers for this purpose.

Test Gap

I did not see a test covering fork/cross-remote PR checkout where one of base_ref/head_ref is absent on origin. Adding that case would prevent regressions here.

Note: model gpt-5.3-codex, input 26991, output 2212 tokens used.

<!-- codex-review:head_sha=d6a9397914696fcb0ded6e081b1e44fd71b0c755 --> ## Review Result This PR introduces a checkout robustness regression that can cause valid PR reviews to fail. ## Findings ### 1) PR checkout assumes both refs exist on one remote In `src/gitea_codex_bot/workers/container_runner.py` (around the new fetch step), the runner does: - clone one `clone_url` as `origin` - `git fetch --no-tags origin <base_ref> <head_ref>` For fork and other cross-repo topologies, `base_ref` and `head_ref` are not guaranteed to both exist on that same remote. If either ref is missing, fetch fails and review aborts before checkout. This is a correctness issue because it can reject reviewable PRs purely due to ref layout. ## Suggested Fix Use SHA-based fetch (or independent tolerant fetches) rather than requiring both branch names from one remote. The runner already has `base_sha` and `head_sha`, which are stable identifiers for this purpose. ## Test Gap I did not see a test covering fork/cross-remote PR checkout where one of `base_ref`/`head_ref` is absent on `origin`. Adding that case would prevent regressions here. _Note: model `gpt-5.3-codex`, input `26991`, output `2212` tokens used._
space added 1 commit 2026-05-23 13:57:58 +02:00
[fix]. Harden fork PR fetch + config marker
All checks were successful
ci / test (pull_request) Successful in 33s
Details
ci / publish (pull_request) Has been skipped
Details
01e10abc71
space added 1 commit 2026-05-23 14:03:25 +02:00
[refactor]. Remove legacy review prompt path
All checks were successful
ci / test (pull_request) Successful in 34s
Details
ci / publish (pull_request) Has been skipped
Details
30aa737516
space changed title from feat: Enhance review prompt with detailed instructions and placeholders for empty sections to improvements: Big Reviewer Update 2026-05-23 14:04:16 +02:00
space added 1 commit 2026-05-23 14:07:54 +02:00
[feat]. Add @codex -h help summary command
All checks were successful
ci / test (pull_request) Successful in 38s
Details
ci / publish (pull_request) Has been skipped
Details
9392591429
space commented 2026-05-23 14:08:53 +02:00
Author
Owner
Copy Link

@codex -h

@codex -h
gitea-codex commented 2026-05-23 14:08:53 +02:00
First-time contributor
Copy Link

Codex Help

Supported commands:

  • @codex review [security|performance|tests] [--full]
  • @codex rerun
  • @codex explain
  • @codex fix [--branch ...]
  • @codex ignore
  • @codex -h / @codex --help / @codex help

Status note:

  • Pending jobs on this PR: 1
  • Fix command enabled: false
  • Latest review command: review status succeeded. Summary: The PR introduces a likely functional regression in PR checkout for some pull request topologies (especially forks) by fetching both base and head refs from a single origin remot

Discussion summary (4 comments, human 2, bot 2):

  • @space: @codex review
  • @gitea-codex: 👀 Codex review queued for commit d6a9397.
  • @gitea-codex: ## Review Result This PR introduces a checkout robustness regression that can cause valid PR reviews to fail...
  • @space: @codex -h
## Codex Help Supported commands: - `@codex review [security|performance|tests] [--full]` - `@codex rerun` - `@codex explain` - `@codex fix [--branch ...]` - `@codex ignore` - `@codex -h` / `@codex --help` / `@codex help` Status note: - Pending jobs on this PR: `1` - Fix command enabled: `false` - Latest review command: `review` status `succeeded`. Summary: The PR introduces a likely functional regression in PR checkout for some pull request topologies (especially forks) by fetching both base and head refs from a single `origin` remot Discussion summary (4 comments, human `2`, bot `2`): - @space: @codex review - @gitea-codex: 👀 Codex review queued for commit `d6a9397`. - @gitea-codex: <!-- codex-review:head_sha=d6a9397914696fcb0ded6e081b1e44fd71b0c755 --> ## Review Result This PR introduces a checkout robustness regression that can cause valid PR reviews to fail... - @space: @codex -h
space commented 2026-05-23 14:10:16 +02:00
Author
Owner
Copy Link

@codex review

@codex review
gitea-codex commented 2026-05-23 14:10:18 +02:00
First-time contributor
Copy Link

👀 Codex review queued for commit 9392591.

👀 Codex review queued for commit `9392591`.
gitea-codex commented 2026-05-23 14:11:43 +02:00
First-time contributor
Copy Link

Review Result

I found functional regressions introduced by this PR that affect review correctness and behavior.

Findings

1) PR-scoped context removed from model input (high)

The new ephemeral runner path no longer passes explicit PR review context (changed files, unified diff, base/head scope, etc.) and now sends only a short free-form prompt (review: ...).

This can cause reviews to drift from "issues introduced by this PR" and produce incorrect findings.

2) .codex-review.yml behavior regression (medium)

The code still loads repo config, but key config-driven behaviors previously applied in prompt construction are no longer enforced in the new path (notably default_mode fallback for @codex review without explicit mode, and other scope-shaping behavior).

Repos depending on these settings will observe changed review behavior.

3) Output marker parsing can fail on collisions (medium)

Result extraction uses global first-occurrence marker matching. If marker text appears inside generated JSON content, extraction can truncate and fail.

Use end-marker search anchored after the located start marker, and ideally randomized per-run markers.

Suggested Follow-up

  1. Restore explicit PR scope/context in runner prompting or use a native base/head review target.

  2. Reintroduce config semantics (default_mode and other scope controls) in the new container path.

  3. Harden marker parsing and add tests for marker-collision scenarios.

Note: model gpt-5.3-codex, input 68257, output 3309 tokens used.

<!-- codex-review:head_sha=93925914290f32b8ff25b598a4969de7b4d2c811 --> ## Review Result I found functional regressions introduced by this PR that affect review correctness and behavior. ## Findings ### 1) PR-scoped context removed from model input (high) The new ephemeral runner path no longer passes explicit PR review context (changed files, unified diff, base/head scope, etc.) and now sends only a short free-form prompt (`review: ...`). This can cause reviews to drift from "issues introduced by this PR" and produce incorrect findings. ### 2) `.codex-review.yml` behavior regression (medium) The code still loads repo config, but key config-driven behaviors previously applied in prompt construction are no longer enforced in the new path (notably `default_mode` fallback for `@codex review` without explicit mode, and other scope-shaping behavior). Repos depending on these settings will observe changed review behavior. ### 3) Output marker parsing can fail on collisions (medium) Result extraction uses global first-occurrence marker matching. If marker text appears inside generated JSON content, extraction can truncate and fail. Use end-marker search anchored after the located start marker, and ideally randomized per-run markers. ## Suggested Follow-up 1. Restore explicit PR scope/context in runner prompting or use a native base/head review target. 2. Reintroduce config semantics (`default_mode` and other scope controls) in the new container path. 3. Harden marker parsing and add tests for marker-collision scenarios. _Note: model `gpt-5.3-codex`, input `68257`, output `3309` tokens used._
space added 1 commit 2026-05-23 14:15:07 +02:00
[fix]. Restore PR-scoped review + remove fix cmd
All checks were successful
ci / test (pull_request) Successful in 32s
Details
ci / publish (pull_request) Has been skipped
Details
08075cb3c4
space commented 2026-05-23 14:16:12 +02:00
Author
Owner
Copy Link

@codex review "give me a merge score"

@codex review "give me a merge score"
gitea-codex commented 2026-05-23 14:16:14 +02:00
First-time contributor
Copy Link

👀 Codex review queued for commit 08075cb.

👀 Codex review queued for commit `08075cb`.
gitea-codex commented 2026-05-23 14:17:34 +02:00
First-time contributor
Copy Link

Review Result

I found one behavior regression introduced by this PR.

Findings

1) @codex fix now gets silently ignored

src/gitea_codex_bot/services/commands.py removes fix from SUPPORTED_COMMANDS, so parse_command() returns None for @codex fix ....

Because of that, no job is queued for new fix comments, and the dispatcher's new deprecation path ("@codex fix is no longer supported") is not reached for normal webhook traffic.

This changes behavior from explicit handling to silent no-op for users.

Suggested Fix

Either:

  • Keep fix parseable as a deprecated command and let dispatcher post the deprecation message, or
  • Add explicit webhook fallback handling for @codex fix parse failures that posts the same deprecation notice.

Merge Score

Not ready to merge yet due to the user-visible command regression above.

Note: model gpt-5.3-codex, input 40143, output 2475 tokens used.

<!-- codex-review:head_sha=08075cb3c49ec6c2a000f633e2407033beadb7e3 --> ## Review Result I found one behavior regression introduced by this PR. ## Findings ### 1) `@codex fix` now gets silently ignored `src/gitea_codex_bot/services/commands.py` removes `fix` from `SUPPORTED_COMMANDS`, so `parse_command()` returns `None` for `@codex fix ...`. Because of that, no job is queued for new fix comments, and the dispatcher's new deprecation path (`"@codex fix is no longer supported"`) is not reached for normal webhook traffic. This changes behavior from explicit handling to silent no-op for users. ## Suggested Fix Either: - Keep `fix` parseable as a deprecated command and let dispatcher post the deprecation message, or - Add explicit webhook fallback handling for `@codex fix` parse failures that posts the same deprecation notice. ## Merge Score Not ready to merge yet due to the user-visible command regression above. _Note: model `gpt-5.3-codex`, input `40143`, output `2475` tokens used._
space added 1 commit 2026-05-23 14:20:43 +02:00
[fix]. Reply on unsupported @codex commands
All checks were successful
ci / test (pull_request) Successful in 36s
Details
ci / publish (pull_request) Has been skipped
Details
d1ca1052f4
space commented 2026-05-23 14:20:55 +02:00
Author
Owner
Copy Link

@codex penis

@codex penis
gitea-codex commented 2026-05-23 14:20:56 +02:00
First-time contributor
Copy Link

⚠️ Command @codex penis is not supported. Try @codex -h.

⚠️ Command `@codex penis` is not supported. Try `@codex -h`.
space merged commit cf95343e06 into main 2026-05-23 14:22:33 +02:00
space deleted branch improv/big-reviewer-update 2026-05-23 14:22:36 +02:00
Sign in to join this conversation.
Reviewers
No Reviewers
Labels
No items
No Label
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
luna space
No Assignees
2 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: space/gitea-codex#4
Delete Branch "improv/big-reviewer-update"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?