fix: responder uses gh api graphql to read review threads

.github/workflows/review-responder.md

@@ -47,26 +47,22 @@ This workflow runs when a review is submitted on a pull request.
 
 1. First, check if the PR has the `aw` label. If it does NOT have the `aw` label, stop immediately — this workflow only handles agent-created PRs.
 
-2. Check the review that triggered this workflow. If the review has no comments (e.g., a plain approval with no inline comments), check if there are any unresolved review threads on the PR. If the review has no comments AND there are no unresolved threads, stop — there is nothing to address. Do NOT add any labels when stopping here. If the review has no comments BUT there ARE unresolved threads from a previous review, continue to step 3 to address those threads.
+2. Check the review that triggered this workflow. If the review has no comments (e.g., a plain approval with no inline comments), stop — there is nothing to address.
 
-3. Read the unresolved review comment threads on the PR (not just the latest review — get all unresolved threads). If there are no unresolved threads, stop — there is nothing to address. Do NOT add any labels when stopping here.
+3. Check if the PR already has the label `review-response-attempted`. If it does, add a comment to the PR saying "Review response already attempted — stopping to prevent loops. Manual intervention needed." and stop.
 
-4. Check if the PR already has the label `review-response-attempted`. If it does, add a comment to the PR saying "Review response already attempted — stopping to prevent loops. Manual intervention needed." and stop.
+4. Add the label `review-response-attempted` to the PR.
 
-5. Add the label `review-response-attempted` to the PR. Only add this label HERE — after confirming there are actual threads to address.
+5. Read the unresolved review comment threads on the PR (not just the latest review — get all unresolved threads). If there are more than 10 unresolved threads, address the first 10 and leave a summary comment on the PR noting how many remain for manual follow-up.
 
 6. For each unresolved review comment thread (up to 10):
    a. Read the comment and understand what change is being requested
    b. Read the relevant file and surrounding code context
-   c. Make the requested fix in the code (edit the file locally — do NOT push yet)
+   c. Make the requested fix in the code
    d. Reply to the comment thread explaining what you changed
 
-7. If there are more than 10 unresolved threads, leave a summary comment on the PR noting how many remain for manual follow-up.
+7. After addressing all comments, run the CI checks locally to make sure your fixes don't break anything: `uv sync && uv run ruff check --fix . && uv run ruff format . && uv run pyright && uv run pytest --cov --cov-fail-under=80 -v`
 
-8. After addressing all comments, run the CI checks locally to make sure your fixes don't break anything: `uv sync && uv run ruff check --fix . && uv run ruff format . && uv run pyright && uv run pytest --cov --cov-fail-under=80 -v`
-
-9. Push all changes in a single commit with message "fix: address review comments". Reply to all threads BEFORE pushing — replies after a push will appear on outdated code.
+8. Push all changes in a single commit with message "fix: address review comments".
 
 If a review comment requests a change that would be architecturally significant or you're unsure about, reply to the thread explaining your concern rather than making the change blindly.
-
-NOTE: Thread resolution is handled separately by the pipeline orchestrator after this workflow completes. Your job is to fix the code and reply to threads — do NOT attempt to resolve threads.