[bugfix] post AI review inline comments instead of buffering them

[tiankongdeguiji]

Problem

The AI code review stopped posting inline comments. In run 34201597956 (PR #662) the reviewer made 11 create_inline_comment calls, every one succeeded, and none reached the PR (GET /pulls/662/comments returns 0). Only the top-level summary landed, because that goes through gh pr comment.

Each call came back:

{"success": true, "buffered": true,
 "message": "Comment buffered. It will be classified and posted after this session
             completes ... Set confirmed=true to post immediately."}

Root cause

claude-code-action's github-inline-comment-server.ts appends every unconfirmed call to /tmp/inline-comments-buffer.jsonl instead of posting it:

const CLASSIFY_ENABLED = process.env.CLASSIFY_INLINE_COMMENTS !== "false";
if (CLASSIFY_ENABLED && confirmed !== true) { appendFileSync(BUFFER_PATH, ...); return buffered; }

The buffer is drained by a separate entrypoint, src/entrypoints/post-buffered-inline-comments.ts, which the composite action runs as a post-step. Both jobs here drive the CLI directly with only --mcp-config / [mcp_servers], so that entrypoint never runs, the MCP server dies with the session, and the buffer stays on the runner.

The behavior has been upstream since March; the runner's git clone --branch v1 of the action was older than that until its last update. Run 34120554588 (2026-09-07 12:41) still posted 11 inline comments live; run 34201597956 (2026-09-08 07:53) buffered all of them. Nothing in this repo changed — .claude/commands/review-pr.md has never mentioned confirmed.

Fix

Set CLASSIFY_INLINE_COMMENTS=false in the MCP server's environment in both jobs, which restores immediate posting. The codex job needs its own copy of the variable: its [mcp_servers.github_inline_comment.env] block is the server's whole environment, not an inherited one, which is also why GITHUB_TOKEN and friends are listed there explicitly.

Two alternatives were rejected. Instructing the reviewers to pass confirmed: true depends on the model remembering the argument on every call and needs the same edit duplicated into the codex agent prompts. Adding the flush entrypoint as a post-step is worse still: post-buffered-inline-comments.ts reads the buffer but never truncates it, and on this persistent self-hosted runner the file has been accumulating since the regression — a flush without an rm -f first would spray one PR's comments onto the next.

Test Plan

• yaml.safe_load on the workflow: parses, and the RunCodeReview step env carries CLASSIFY_INLINE_COMMENTS: 'false' (string, so the server's !== "false" check sees it).
• Replayed the codex heredoc through tomllib: the generated config.toml parses with its preserved indentation and mcp_servers.github_inline_comment.env carries the new key alongside the existing four.
• pre-commit run --files .github/workflows/code_review.yml: all applicable hooks pass.
• End-to-end verification is the next claude-review / codex-review label on a PR — inline comments should appear during the run rather than being reported as buffered.

Unrelated follow-up worth doing on the runner: the setup instructions clone the action from a moving v1 branch, which is how a tool contract changed under CI unnoticed; pinning to a tag or SHA would prevent a repeat. The stale /tmp/inline-comments-buffer.jsonl there can also be deleted.

🤖 Generated with Claude Code

https://claude.ai/code/session_01BJg1zk3pFsjZdaUCWvjwBT