Repository navigation
Guide release transitions and CI attribution during rebase - #183
Conversation
Teach live module and debug prompts to verify release pins together, audit adjacent upstream API contracts, and compare exact failure causes against recent main and branch builds. Keep parent-parity prompt goldens intact and document the operator handoff. Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 6df74d25d3
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| 1. Find the first causal traceback and every failed test in this job. Group | ||
| jobs by failure signature, not job label; one startup error can fail | ||
| several jobs, and one job can contain independent failures. |
There was a problem hiding this comment.
When a test job has multiple failures and an earlier traceback falls more than 200 lines before the end, the agent cannot reliably perform this instruction: both local and remote-CI callers pass only splitlines()[-200:] (rebase_v3.py:1410-1413 and rebase_v3.py:2131-2135), while the rendered prompt includes neither the source log path nor a truncation marker. The agent therefore sees the tail as the complete traceback and can group or fix the wrong signature; pass the log path/full failure summary or explicitly direct it to the complete artifact.
Useful? React with 👍 / 👎.
The previous commit made both debug-agent call sites prefer the adapter's debug_prompt_template_live and fall back to the parent-parity template, but left the rebase_engine and engine/steps/rebase_v3 spec pages unverified, so check_spec_freshness --strict failed CI. Record the selection rule in both pages and re-verify them. Signed-off-by: tzhouam <tzhouam@connect.ust.hk> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
1 participant
The vLLM 0.30.0 alignment exposed worker failures that import checks missed: one RoPE constructor mismatch broke six TTS/L4 jobs, while a changed sampling-mask contract broke two Simple Other jobs. A release rebase also needs to distinguish these failures from issues already present on main and from latency-only results.
This change updates Copilot's live rebase prompts to check release artifacts, dependencies, and source-install docs as one version set; audit adjacent constructor, dummy-run, execution-state, and async-output contracts after a startup fix; and compare exact failure signatures and build timing before assigning a regression to the rebase. It retains the parent-parity prompt and golden while routing module retries and CI debugging through the new live guidance. The README documents the operator handoff, including review comments and exact-commit CI reporting.
Validation: 78 adapter/CI tests passed; focused prompt tests passed again after the last edit;
git diff --checkpassed. Ruff reports 45 existing diagnostics in these files and none on changed lines.Source experience: vLLM-Omni rebase PR #7820, Buildkite #3065.