Address PR Comments

[ableinc]

Add the feature to monitor comments in PRs. If you're tagged in the comment via "@coding-agent", then react to the comment with the 👀 emoji to acknowledge that you're addressing the comment. React with the ✅ (or GH equivalent) when the comment has been addressed.

[ableinc]

The coding agent could not complete this issue (attempt 1).

claude run failed: claude exited with error: exit status 1 (stderr: )

It will try again after Sat, 22 Aug 2026 10:57:42 UTC (the wait doubles with each consecutive failure). Remove the agent-ready label to stop it retrying.

_{coding-agent-loop run e81482e8-92f9-4076-9859-255ddbec7933}

[ableinc]

Plan

Add PR review-comment handling: acknowledge @coding-agent mentions with 👀, address them on the PR branch, then confirm with 👍

(Note: I could not write this to /home/node1/.claude/plans/… — that path is on a read-only filesystem. The plan is below in full.)

Context

Today the daemon's only input is a labelled GitHub issue: discover → plan → approve → implement → draft PR (internal/orchestrator/loop.go). Once the draft PR exists, the loop is done with it. A reviewer who leaves feedback on that PR has no way to hand it back to the agent — they must re-label the issue, and the agent would start from a fresh branch.

Issue #4 asks for the missing half of that conversation: watch comments on pull requests, and when one mentions @coding-agent, react 👀 to show it has been picked up, do the work, then react with the "done" equivalent.

Scope decision (the issue does not spell it out): "addressed" means the agent actually makes the requested change on the PR's existing branch and pushes it — reacting alone would be theatre. If the comment is a question rather than a change request, the agent answers it in a reply comment and still marks it done.

The one prior automated attempt failed before producing anything (claude exited with error: exit status 1), so there is no partial work to build on.

---

Design decisions taken here

1. Which PRs are watched: open PRs authored by the daemon's own GitHub account whose head branch starts with workspace.branch_prefix. This is a hard safety rule, not a config knob: the daemon only ever pushes to branches it created itself. A mention on any other PR is ignored (documented in the README).
2. Which comments can trigger: PR conversation comments (/issues/{n}/comments) and inline review comments (/pulls/{n}/comments). Both are reactable through the REST reactions API. Review summary bodies are read as context but cannot trigger — REST has no reactions endpoint for a PullRequestReview, so the 👀/👍 contract can't be honoured there. See Open decisions.
3. Reactions: ack = eyes; done = +1 (GitHub has no ✅ reaction; +1 is the conventional "done", rocket the alternative). Both configurable.
4. Who may command the agent: only comment authors whose author_association is OWNER, MEMBER, or COLLABORATOR, or who appear in an explicit login allowlist. Without this, any drive-by commenter on a public repo can trigger a bypassPermissions Claude run.
5. Reuse of the existing concurrency model: GitHub numbers PRs in the same sequence as issues, so claims(repo, issue) can hold a PR number with no schema change. Store.TryClaim/RepoBusy/RenewClaim/ReleaseClaim, Orchestrator.reserveRepo/releaseRepo and renewLease are reused verbatim — one unit of work per repo at a time still holds across both work kinds.
6. Priority: the PR-comment pass runs before issue discovery in each tick and consumes the same capacity budget. A reviewer waiting on a reply outranks starting a new issue.

---

Changes by file

internal/config/config.go (+ config.example.json, README)

Add a nested block to GitHubConfig:

// PRComments controls responding to @-mentions on the daemon's own pull requests.
PRComments PRCommentsConfig `json:"pr_comments"`

type PRCommentsConfig struct {
 Enabled bool `json:"enabled"`
 Mention string `json:"mention"` // "@coding-agent"
 SearchLimit int `json:"search_limit"` // 30
 MaxAge Duration `json:"max_age"` // "168h"; 0 = no limit
 AckReaction string `json:"ack_reaction"` // "eyes"
 DoneReaction string `json:"done_reaction"` // "+1"
 AllowedAuthors []string `json:"allowed_authors"` // empty = fall back to associations
 AllowedAssociations []string `json:"allowed_associations"` // OWNER, MEMBER, COLLABORATOR
}

• Populate all of it in config.Default() (defaults above, Enabled: true).
• Extend Config.Validate(): when enabled, mention must be non-empty and start with @; the two reaction values must be one of the eight GitHub contents (+1 -1 laugh confused heart hooray rocket eyes); search_limit ≥ 1.
• Every field must be additive and default-safe — config.Load uses DisallowUnknownFields, so an existing deployment's config.json that omits the block must still load.
• Mirror the block into config.example.json and the README config table. config.json is gitignored and generated by make config, so nothing else to touch.

internal/gh/gh.go — new API surface

New type, in the same "typed struct per gh JSON payload" style as Issue/Comment:

// PRComment is one reactable comment on a pull request.
type PRComment struct {
 ID int64 // REST numeric id — what the reactions endpoint keys on
 Kind string // CommentKindIssue | CommentKindReview
 Author string
 Association string // author_association
 Body string
 URL string // html_url
 CreatedAt time.Time
 // review-comment only, used to give Claude the anchor:
 Path, DiffHunk string
 Line int
}

New methods:

Method	Underlying command	Notes
CurrentLogin(ctx)	gh api user --jq .login	cached by the caller; used for --author and to skip our own comments
SearchPRs(ctx, author, owners, limit)	gh search prs --state open --author <login> --owner … --json number,repository,title,url,updatedAt	mirrors SearchIssues, including the defensive ownedBy() filter already in this file
ViewPR(ctx, repo, number)	gh pr view <n> --repo <r> --json number,title,body,url,state,isDraft,headRefName,baseRefName	extend the existing PullRequest struct with BaseRefName, IsDraft
PRComments(ctx, repo, number)	gh api --paginate repos/<repo>/issues/<n>/comments + …/pulls/<n>/comments	one merged []PRComment sorted by CreatedAt; gh api --paginate merges JSON arrays
PRReviewBodies(ctx, repo, number)	gh api --paginate repos/<repo>/pulls/<n>/reviews	context only, never a trigger
React(ctx, repo, c PRComment, content)	gh api --method POST repos/<repo>/{issues|pulls}/comments/<id>/reactions -f content=<content>	path segment chosen by c.Kind; must honour DryRun via logf, exactly like Comment and CreatePR
CommentOnPR(ctx, repo, number, body)	same endpoint as the existing Comment (a PR is an issue here)	may just call Comment; keep the alias for readability

All go through the existing run/runJSON helpers, so CmdError context and dry-run behaviour come for free.

internal/store/store.go — migration 5

Append one migration string to migrations (the existing migrate() loop handles versioning):

ALTER TABLE runs ADD COLUMN kind TEXT NOT NULL DEFAULT 'issue';

CREATE TABLE IF NOT EXISTS pr_comment_tasks (
 repo TEXT NOT NULL,
 pr INTEGER NOT NULL,
 comment_kind TEXT NOT NULL, -- 'issue' | 'review'
 comment_id INTEGER NOT NULL,
 status TEXT NOT NULL, -- 'acked' | 'done' | 'failed'
 attempts INTEGER NOT NULL DEFAULT 0,
 last_attempt_at INTEGER NOT NULL DEFAULT 0,
 run_id TEXT NOT NULL DEFAULT '',
 created_at INTEGER NOT NULL,
 updated_at INTEGER NOT NULL,
 PRIMARY KEY (comment_kind, comment_id)
);
CREATE INDEX IF NOT EXISTS pr_comment_tasks_pr ON pr_comment_tasks(repo, pr);

• Run.Kind field + runColumns + scanRun + CreateRun updated (RunKindIssue/RunKindPRComment constants beside the status constants). Add AND kind = 'issue' to IssueHistory's aggregate so PR runs can never colour an issue's back-off.
• New status StatusAddressed = "addressed", added to IsTerminal.
• Four small methods in the existing ExecContext + wrapped-error style: PRCommentTasks(ctx, repo, pr), MarkPRCommentAcked, MarkPRCommentDone, MarkPRCommentFailed (the last bumps attempts, stamps last_attempt_at).

Why a table rather than reading reactions back: it is the only thing that survives a crash between the 👀 and the run, it gives per-comment attempt counts for the back-off, and it keeps dedup independent of who else may have reacted.

internal/orchestrator/prcomments.go (new)

Pure, unit-testable helpers first:

// mentionsAgent reports whether body addresses the agent by handle. Match is
// case-insensitive, must sit on a word boundary (so "@coding-agent-loop" is not
// a match for "@coding-agent"), and quoted lines (leading ">") are ignored so a
// quoted mention cannot re-trigger the agent.
func mentionsAgent(body, handle string) bool

// pendingMentions filters a PR's comments to the ones this run should act on:
// mention present, author permitted, not older than maxAge, not authored by the
// daemon, not one of the daemon's own marker comments, and either unseen or a
// failed task whose back-off has elapsed (reusing retryDelay from loop.go).
func pendingMentions(...) []gh.PRComment

Then the pass, mirroring tick/work/execute:

• tickPRComments(ctx, capacity) int — called from tick right after the gate/capacity checks and before the issue loop; returns remaining capacity. For each SearchPRs hit: Owned / Excluded / in-memory-busy / Store.RepoBusy guards (the same ones eligible applies), then ViewPR, then the head-branch-prefix check, then pendingMentions. Nothing pending → next PR.
• workPRComments(ctx, cand, comments) — TryClaim on (repo, prNumber), defer ReleaseClaim, register the cancel func in o.cancels, renewLease, CreateRun{Kind: pr_comment, Issue: prNumber, Branch: headRef}, Discord.RunClaimed.
  1. Ack immediately — GH.React(ack) on every pending comment and MarkPRCommentAcked, before any cloning. This is the user-visible promise; it must not wait on a clone.
  2. Workspace — repoMetadata, Git.EnsureRepo, Git.AssertRemote, then Git.AddWorktree(ctx, repoPath, worktree, headRef, headRef). Passing the PR head branch as both branch and base makes the existing helper run worktree add -B <headRef> <path> origin/<headRef> — a clean checkout of the PR head, and a retry after a human pushed to the branch picks their commits up naturally. No new git code. Worktree path from Git.WorktreePath(repo, prNumber) unchanged.
  3. Model — Store.CooledDownModels + Registry.Ladder(models.RoleImplement, cooled) + models.Head, as in the implement phase, demoting by the task's attempt count.
  4. Runner.Run with prCommentSystemPrompt / prCommentTaskPrompt and cfg.Claude.PermissionMode; the same OnEvent session capture, RecordUsage/recordSession bookkeeping, and gate.DetectLimit → errRetryable handling as execute, so a usage limit defers rather than fails.
  5. Git.HasWork(worktree, headRef). No work is not a failure here — a question answered in prose is a legitimate outcome. With work: Git.CommitAll (message Address review feedback on #<pr> + run id), Verify.Run, Git.Push(worktree, headRef, repo). The push is the existing non-force --set-upstream origin b:b; a rejected push (someone pushed meanwhile) is a normal failure and retries after the back-off from a fresh checkout.
  6. GH.CommentOnPR with the marker-tagged summary, then GH.React(done) + MarkPRCommentDone per comment, SetRunStatus(StatusAddressed), events, Discord, cleanup.
  7. Failure path — MarkPRCommentFailed per comment (back-off is per comment), a marker-tagged failure comment on the PR, Store.FailRun, Discord.RunFailed. The 👀 stays: the comment was seen, and the next pass retries it.

internal/orchestrator/phase.go

Add markerPRComment = markerPrefix + "pr-comment -->" beside the existing markers, so the harness's own PR replies are recognised by isAgentComment and can never be read back as a mention.

internal/orchestrator/prompt.go

• prCommentSystemPrompt(repo, branch, worktree) — same harness contract as systemPrompt (no push, no PR/branch creation, no history rewriting, edits confined to the worktree, nobody is watching), reframed: an open pull request already exists on this branch, you are addressing reviewer feedback on it, keep the change to what the feedback asks for.
• prCommentTaskPrompt(repo, pr, comments, reviews) — PR title/body, then each triggering comment as author + body (+ path, line, and fenced diff_hunk for inline comments), then review summary bodies as context. Reuse the existing truncate, maxCommentChars, maxBodyChars. Instruct: address every comment listed; if one is a question rather than a change request, answer it in the final summary; the final summary is posted back to the PR.

internal/orchestrator/report.go

• prCommentComment(handled, summary, runID, v verify.Result, pushed bool) — marker, one line per handled comment linking its html_url, the agent's summary, verify status (reusing issueComment's store.VerifyPassed/Failed switch), run-id footer.
• prCommentFailureComment(...) — the equivalent of failureComment for this path.

internal/discord/notifier.go

One method, PRCommentsAddressed(r RunRef, handled int, res *claude.Result, v verify.Result, elapsed time.Duration), shaped like PROpened. Claim/fail/defer reuse the existing RunClaimed/RunFailed/RunDeferred with RunRef{Issue: prNumber, Title: pr.Title, URL: pr.URL}.

internal/server/server.go

No handler changes needed; Run.Kind flows through /runs and /runs/:id once it is on the struct. One README line noting kind distinguishes the two work types.

README.md

New section "Responding to PR comments" after "Lifecycle of one issue"; the config block in the config reference; a row in Safety boundaries (author allowlist, agent-branch-only rule, never force-pushes, never merges); the pr_comments knobs in the table; and the top-of-file ASCII flow updated to show the second entry point.

---

Safety boundaries this adds (please review)

• Only PRs authored by the daemon's own account and on a workspace.branch_prefix branch.
• Only comment authors with OWNER/MEMBER/COLLABORATOR association, or an explicit login allowlist — otherwise an arbitrary GitHub user can drive a bypassPermissions Claude run on your machine.
• Own comments and marker-tagged comments are skipped; quoted (>) lines don't count as mentions — no self-triggering loop.
• max_age (default 7 days) stops a first poll from stampeding through a year of comment history.
• Push stays non-force and on the PR's own branch; AssertRemote still runs immediately before it.
• --dry-run suppresses reactions as well as comments and pushes.

Tests

• internal/orchestrator/prcomments_test.go: table test for mentionsAgent (plain, mid-sentence, @Coding-Agent, @coding-agent-loop non-match, inside a > quote, inside a fenced code block, absent); pendingMentions filtering (own login, marker body, disallowed association, too old, already done, failed inside vs. outside the back-off).
• prompt_test.go / report_test.go: the prompt includes every comment body and diff hunk; the reply comment carries markerPRComment and each handled comment's URL.
• internal/gh/gh_test.go: reuse the existing stubGH helper to assert exact argv for React (both kinds), SearchPRs (owner scoping, no cross-owner leakage), and PRComments decoding; assert React is a no-op under DryRun.
• internal/store/store_test.go: migration applies to an existing DB; ack → done and ack → failed transitions; attempts increments; IssueHistory ignores kind='pr_comment' rows.
• internal/config/config_test.go: defaults populated, invalid reaction rejected, a config without the pr_comments block still loads.

Verification

1. make test (go test -race ./...).
2. go build ./... and --check with the new config block, then again with an old config.json lacking it (must still boot).
3. --once --dry-run --log-level debug against a repo with a live agent PR carrying an @coding-agent comment: logs show the PR discovered, the mention matched, and dry-run: would … for the reaction, the push, and the reply — nothing mutated on GitHub.
4. End-to-end on a scratch repo: label an issue, let the loop open a draft PR, comment @coding-agent please rename X to Y, and confirm — 👀 within one poll, a commit on the PR branch, a marker-tagged reply, 👍, and a second poll that does not re-process the comment. Then restart the daemon mid-run and confirm the comment is retried rather than stranded.

Open decisions for the reviewer

1. Done reaction: +1 (recommended) vs rocket. Configurable either way; this only sets the default.
2. Review summary bodies. Excluded from triggering because REST cannot react to a PullRequestReview. Including them means moving the read + react path to GraphQL (addReaction on any Reactable, which would also give viewerHasReacted as a free cross-restart dedup marker). That is a materially larger change and one big embedded query string; this plan deliberately takes the smaller REST path first. Say if review bodies should be in v1.
3. Polling cost. Each poll costs one search plus ~2 API calls per open agent PR. Negligible at a 5m interval with a handful of PRs; if a repo routinely has dozens, the follow-up is to skip PRs whose updatedAt hasn't moved (one more column on pr_comment_tasks). Not included now.
4. runs.kind column. Recommended for legible /runs output and clean issue back-offs, but it does touch scanRun/runColumns/CreateRun. Say if you'd rather PR runs be indistinguishable in the run table and skip the ALTER TABLE.

---

Reply with exactly implement to approve this plan and start the change. Reply with anything else and the plan will be revised to address it.

_{coding-agent-loop run 1fdb4ab7-042a-4e44-a841-1d6e125e5c7a, model claude-opus-5, cost $2.3082}

[ableinc]

implement