Let Claude follow up on reviews and open PRs from issues

[ddaspit]

Quick summary

The code review follows up on its earlier findings instead of posting them again. A comment with @claude on an issue can now open a pull request. No code under machine/ or tests/ is changed.

Where to look

• The review matches its earlier findings by reading the thread. Nothing pins this. The action restores .claude/ from the base branch, so the review of this PR still runs the skill from main.
• The review resolves its own threads with resolveReviewThread. I did not run the mutation. It uses the Claude app token that already posts the reviews.
• Bash(poetry run:*) lets Claude run any command. The author_association checks in claude.yml only let it run when the issue or PR and the comment come from an owner, member or collaborator.
• I do not know whether Reviewable picks up a thread resolved on GitHub.

Deliberately not included

• Allowing only the specific check commands instead of poetry run:*. The tests would still run the branch's code.
• Filtering outsider comments with include_comments_by_actor. It needs a list of users to maintain.

Validation

• ./local_check.sh --agent-strict -- comment hygiene found no in-scope files. black, flake8 and isort passed. pyright and pytest failed on missing optional dependencies (clearml, datasets, thot.alignment, sentencepiece) because my local environment does not have the extras. This PR changes no Python.
• git diff --check -- no output.
• yaml.safe_load on both workflow files -- parsed.
• The review thread query in the skill, run against Migrate to transformers 5 #363 -- returned its threads. This showed that GraphQL names the bot claude rather than claude[bot].

Issue / porting context

The @claude run on #369 could not read sillsdev/machine#472 or #382 and did not open a pull request.

---

Reading this a year from now

The review runs on every push. Before this change, each run reviewed the PR from scratch and posted whatever it found, so the same findings came back on every push. The skill now settles the open findings first. It replies when a finding was addressed, withdrawn or accepted, and resolves the thread when it was addressed or withdrawn. It drops any new finding that an open or resolved thread already raises. New findings are limited to code changed since the last reviewed commit, unless they are Critical. The summary records Reviewed at <head-sha> so the next run knows where to start.

The @claude action only allows a small set of tools by default. Anything else is denied because the run is headless. The action also ends an issue run with a "Create a PR" link rather than a PR. claude.yml now allows the gh and git commands the port-pr skill needs, installs the dependencies, and appends a prompt that has Claude open the PR itself.

The skill change and the claude.yml change only take effect after this is merged to main. The concurrency group in claude-code-review.yml applies to this PR's own runs.

🤖 Generated with Claude Code

---

This change is 

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 92.13%. Comparing base (615cd59) to head (e634127).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #384   +/-   ##
=======================================
  Coverage   92.13%   92.13%           
=======================================
  Files         390      390           
  Lines       24644    24644           
=======================================
  Hits        22705    22705           
  Misses       1939     1939           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Enkidu93]

@Enkidu93 reviewed 3 files and all commit messages, and made 1 comment.
Reviewable status: complete! all files reviewed, all discussions resolved (waiting on ddaspit).