[dolfim-ibm]

Adds the AI PR Triage workflow. It runs on pull_request_target for every non-draft PR, before a maintainer approves CI. It gives an advisory first look with Bob Shell. No job checks out or runs code from the PR.

What it does

• Duplicate PRs. It collects candidates with deterministic rules: other PRs that cite the same issue (issue timeline cross-references), and open PRs with overlapping diff hunks in the same files. Bob Shell then rates each candidate as duplicate, related, or unrelated. The model can rate only PR numbers from this candidate list.

• CI safety. Path rules always force ai:ci-needs-care: CI files, build and dependency files, conftest.py, shell scripts, binaries, and agent instruction files (AGENTS.md, .agents/, .bob/, …). Pattern hints in added lines (process execution, eval, network, encoded payloads, secret access) go to the model as context. The model can only make the verdict stricter, never less strict.

• Reference data summary. It sorts changed files under tests/data/**/groundtruth/ into formatting-only, coordinates-only, text, table, and structure changes, with per-file details (for example text: 386 → 385, table 0: shape 3x4 → 3x5, reading order changed). The model says if the changes are expected from the PR intent.

• Topic labels. The triage adds labels from the existing repository set (bug, enhancement, docx, markdown, ocr, table structure, …). Rules map the conventional-commit title (fix → bug, scope (docx) → docx) and the changed source paths to labels, and the model can add up to 3 more from the same allowlist (.github/scripts/pr_topic_labels.py). They are added only on the first triage of a PR, labels that do not exist are never created, and topic labels are never removed. On the last 40 PRs, the rules gave every label that maintainers set, except xml on an epub PR and tests on a dependency PR.

Trust model

Job	Token	LLM	PR code
prepare	read-only	no	read through the API and blob-less git fetch, never checked out
analyze	contents: read, no write	Bob Shell, read-only custom mode, --disable-tool-groups edit,execute,browser,mcp,skill,artifact,subagent,mode, MCP and subagents off	data files only
publish	pull-requests/issues: write	no	validates the JSON again, label allowlist, sanitizes text (no mentions, no HTML)

Bob Shell run pre-approves all tools of a mode. For this reason the mode has only the read group. The prompt and the mode come from the base branch. The answer is discarded if it contains the API key.

The labels are advisory: ai:ci-safe, ai:ci-needs-care, ai:possible-duplicate. They never add tests:full. A new push removes them.

Never a blocking check

The workflow is a review aid, not a gate. Every job and every step uses continue-on-error. Long steps have a timeout below the job timeout. An error (missing secret, API outage, invalid model answer) shows as a warning annotation on the run, and the PR check stays green. The comment then shows only the parts that finished.

Setup needed before merge

• Secret BOB_API_KEY (API key with the Inference scope).
• Optional variables: BOB_TEAM_ID (for general keys) and BOB_TRIAGE_MAX_COST (default 2).

Checklist:

• Documentation has been updated, if necessary. (.github/CI_LABELS.md)
• Examples have been added, if necessary.
• Tests have been added, if necessary.

Review evidence:

• uv run pytest tests/test_ai_pr_triage.py tests/test_groundtruth_diff.py tests/test_pr_topic_labels.py: 22 passed.
• make validate: passed.
• I ran prepare locally against real PRs:
  • fix(docx): ignore XML comments and processing instructions #4627: 5 candidates. fix(docx): keep XML comments in NOTES, drop processing instructions #4584 is linked through the same issue DOCX conversion fails or drops table text when document.xml contains an XML comment #4583, and fix(docx): skip non-element XML nodes #4564/fix(docx): read the active branch of a body-level mc:AlternateContent #4615 overlap in msword_backend.py.
  • fix: preserve paragraphs across overlapping text predictions #4571: the summary shows structure for the two PDF JSON files (text: 386 → 385) and separates the Markdown paragraph-break changes from whitespace-only changes.
• Bob Shell 2.0.5, local run of the analyze step on the fix(docx): ignore XML comments and processing instructions #4627 data. It ended with status: success in 31 s, with 8 tool calls and a cost of 0.167 (limit 2). It wrote no files. extract accepted the answer.
  • fix(docx): skip non-element XML nodes #4564 and fix(docx): keep XML comments in NOTES, drop processing instructions #4584 were rated duplicate. fix(docx): read the active branch of a body-level mc:AlternateContent #4615, fix(docx): keep w:sym characters and map Symbol font text to Unicode #4622 and fix(docx): keep the text of wrapped runs in plain table cells #4455 were rated related. CI safety was safe.
  • The labels were ai:possible-duplicate and ai:ci-safe.
  • Duplicates show in the comment, related PRs are collapsed in <details>, and unrelated PRs are hidden.
• Not run on GitHub: the workflow runs only from main, because it uses pull_request_target. The npm install --ignore-scripts install step is not tested on a runner yet.

[github-actions]

✅ DCO Check Passed

Thanks @dolfim-ibm, all your commits are properly signed off. 🎉

[mergify]

Merge Protections

🟢 Merge protection satisfied — ready to merge.

Show 1 satisfied protection

🟢 Enforce conventional commit

Make sure that we follow https://www.conventionalcommits.org/en/v1.0.0/

• title ~= ^(fix|feat|docs|style|refactor|perf|test|build|ci|chore|revert)(?:\(.+\))?(!)?:

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!