agent: implement issue #5 - Feat: 实现 agent 与 human 在 PR 中的交互

README.md

@@ -2,7 +2,7 @@
 
 [中文说明](README_CN.md)
 
-Local, auditable CLI service that scans configured Gitea repositories for eligible issues, runs an implementer agent in an isolated workspace, opens a pull request, runs a reviewer agent, and posts a standardized human review summary.
+Local, auditable CLI service that scans configured Gitea repositories for eligible issues, runs an implementer agent in an isolated workspace, opens a pull request, runs a reviewer agent, and posts a standardized human review summary. After the PR is opened, it also watches PR comments from humans and can continue the same task on the existing branch to push follow-up commits.
 
 The MVP never merges pull requests.
 
@@ -18,12 +18,13 @@ agent-gitea --config config.yaml run-once
 ```
 
 `sync-repos` discovers repositories owned by the authenticated Gitea user from `/user/repos`; repositories are not listed in the config file.
-`worker` continuously syncs repositories, scans issues, and processes eligible tasks.
+`worker` continuously syncs repositories, scans issues, scans open agent PRs for new human comments, and processes eligible tasks.
 
 ## Commands
 
 - `agent-gitea sync-repos`
 - `agent-gitea scan-issues`
+- `agent-gitea scan-pr-feedback`
 - `agent-gitea run-once`
 - `agent-gitea worker`
 - `agent-gitea show-task <task_id>`

README_CN.md

@@ -2,7 +2,7 @@
 
 [English](README.md)
 
-一个本地运行、可审计的 CLI 服务，用于扫描已配置 Gitea 仓库中的合格 issue，在隔离工作区中运行实现 agent，创建 pull request，运行评审 agent，并发布标准化的人工评审摘要。
+一个本地运行、可审计的 CLI 服务，用于扫描已配置 Gitea 仓库中的合格 issue，在隔离工作区中运行实现 agent，创建 pull request，运行评审 agent，并发布标准化的人工评审摘要。PR 创建后，它还会监控来自 human 的 PR 评论，并在同一个任务、已有分支和 workspace 上继续处理，推送后续 commit。
 
 MVP 版本不会合并 pull request。
 
@@ -18,12 +18,13 @@ agent-gitea --config config.yaml run-once
 ```
 
 `sync-repos` 会通过 `/user/repos` 发现已认证 Gitea 用户拥有的仓库；仓库不会列在配置文件中。
-`worker` 会持续同步仓库、扫描 issue，并处理符合条件的任务。
+`worker` 会持续同步仓库、扫描 issue、扫描 agent 已创建 PR 中的新 human 评论，并处理符合条件的任务。
 
 ## 命令
 
 - `agent-gitea sync-repos`
 - `agent-gitea scan-issues`
+- `agent-gitea scan-pr-feedback`
 - `agent-gitea run-once`
 - `agent-gitea worker`
 - `agent-gitea show-task <task_id>`

src/agent_gitea/cli.py

@@ -9,7 +9,12 @@ import typer
 from .config import AppConfig, load_config
 from .db import Database
 from .gitea import GiteaClient
-from .service import TaskRunner, scan_issues as scan_issues_service, sync_repositories
+from .service import (
+    TaskRunner,
+    scan_issues as scan_issues_service,
+    scan_pull_request_feedback,
+    sync_repositories,
+)
 
 
 app = typer.Typer(no_args_is_help=True, help="Agentic Gitea issue-to-PR manager.")
@@ -74,6 +79,17 @@ def run_once(ctx: typer.Context) -> None:
         typer.echo(f"task {task.id} -> {task.state.value}")
 
 
+@app.command("scan-pr-feedback")
+def scan_pr_feedback(ctx: typer.Context) -> None:
+    cli: CliContext = ctx.obj
+    client = cli.gitea()
+    try:
+        queued = scan_pull_request_feedback(cli.db, client)
+    finally:
+        client.close()
+    typer.echo(f"queued {queued} tasks")
+
+
 @app.command("worker")
 def worker(ctx: typer.Context) -> None:
     cli: CliContext = ctx.obj

src/agent_gitea/db.py

@@ -83,6 +83,13 @@ CREATE TABLE IF NOT EXISTS agent_runs (
     report TEXT,
     created_at TEXT NOT NULL
 );
+
+CREATE TABLE IF NOT EXISTS pull_request_feedback (
+    task_id INTEGER PRIMARY KEY REFERENCES tasks(id),
+    last_seen_comment_id INTEGER NOT NULL DEFAULT 0,
+    pending INTEGER NOT NULL DEFAULT 0,
+    updated_at TEXT NOT NULL
+);
 """
 
 
@@ -271,6 +278,72 @@ class Database:
         ).fetchall()
         return [self._task(row) for row in rows]
 
+    def list_tasks_awaiting_pr_feedback(self) -> list[TaskRecord]:
+        rows = self.conn.execute(
+            """
+            SELECT *
+            FROM tasks
+            WHERE state = ?
+              AND pr_number IS NOT NULL
+            ORDER BY id
+            """,
+            (TaskState.HUMAN_REVIEW_READY.value,),
+        ).fetchall()
+        return [self._task(row) for row in rows]
+
+    def get_pr_feedback_cursor(self, task_id: int) -> int:
+        row = self.conn.execute(
+            "SELECT last_seen_comment_id FROM pull_request_feedback WHERE task_id = ?",
+            (task_id,),
+        ).fetchone()
+        return int(row["last_seen_comment_id"]) if row else 0
+
+    def upsert_pr_feedback_state(
+        self,
+        task_id: int,
+        *,
+        last_seen_comment_id: int | None = None,
+        pending: bool | None = None,
+    ) -> None:
+        row = self.conn.execute(
+            "SELECT last_seen_comment_id, pending FROM pull_request_feedback WHERE task_id = ?",
+            (task_id,),
+        ).fetchone()
+        now = dt_to_db(utcnow())
+        if row is None:
+            self.conn.execute(
+                """
+                INSERT INTO pull_request_feedback (task_id, last_seen_comment_id, pending, updated_at)
+                VALUES (?, ?, ?, ?)
+                """,
+                (task_id, last_seen_comment_id or 0, int(bool(pending)), now),
+            )
+        else:
+            next_cursor = int(row["last_seen_comment_id"]) if last_seen_comment_id is None else last_seen_comment_id
+            next_pending = bool(row["pending"]) if pending is None else pending
+            self.conn.execute(
+                """
+                UPDATE pull_request_feedback
+                SET last_seen_comment_id = ?, pending = ?, updated_at = ?
+                WHERE task_id = ?
+                """,
+                (next_cursor, int(next_pending), now, task_id),
+            )
+        self.conn.commit()
+
+    def mark_pr_feedback_pending(self, task_id: int) -> None:
+        self.upsert_pr_feedback_state(task_id, pending=True)
+
+    def clear_pr_feedback_pending(self, task_id: int, *, last_seen_comment_id: int) -> None:
+        self.upsert_pr_feedback_state(task_id, last_seen_comment_id=last_seen_comment_id, pending=False)
+
+    def has_pending_pr_feedback(self, task_id: int) -> bool:
+        row = self.conn.execute(
+            "SELECT pending FROM pull_request_feedback WHERE task_id = ?",
+            (task_id,),
+        ).fetchone()
+        return bool(row and row["pending"])
+
     def claim_next_task(self, worker_id: str, lease_seconds: int) -> TaskRecord | None:
         now = utcnow()
         expires = now + timedelta(seconds=lease_seconds)

src/agent_gitea/gitea.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 from dataclasses import dataclass
+from datetime import datetime
 from typing import Any
 
 import httpx
@@ -27,6 +28,16 @@ class GiteaPullRequest:
     merged: bool
 
 
+@dataclass(frozen=True)
+class GiteaComment:
+    id: int
+    body: str
+    author: str
+    html_url: str
+    created_at: datetime | None
+    updated_at: datetime | None
+
+
 @dataclass(frozen=True)
 class GiteaRepository:
     owner: str
@@ -140,12 +151,32 @@ class GiteaClient:
         response.raise_for_status()
         return pull_request_from_payload(response.json())
 
-    def post_issue_comment(self, *, owner: str, name: str, issue_number: int, body: str) -> None:
+    def list_issue_comments(self, *, owner: str, name: str, issue_number: int) -> list[GiteaComment]:
+        comments: list[GiteaComment] = []
+        page = 1
+        limit = 50
+        while True:
+            response = self.client.get(
+                f"/repos/{owner}/{name}/issues/{issue_number}/comments",
+                params={"page": page, "limit": limit},
+            )
+            response.raise_for_status()
+            payload = response.json()
+            if not payload:
+                break
+            comments.extend(comment_from_payload(item) for item in payload)
+            if len(payload) < limit:
+                break
+            page += 1
+        return comments
+
+    def post_issue_comment(self, *, owner: str, name: str, issue_number: int, body: str) -> GiteaComment:
         response = self.client.post(
             f"/repos/{owner}/{name}/issues/{issue_number}/comments",
             json={"body": body},
         )
         response.raise_for_status()
+        return comment_from_payload(response.json())
 
     def close_issue(self, *, owner: str, name: str, issue_number: int) -> None:
         response = self.client.patch(
@@ -192,3 +223,24 @@ def pull_request_from_payload(payload: dict[str, Any]) -> GiteaPullRequest:
         state=payload.get("state") or "",
         merged=merged,
     )
+
+
+def comment_from_payload(payload: dict[str, Any]) -> GiteaComment:
+    user_payload = payload.get("user") or payload.get("poster") or {}
+    author = user_payload.get("login") or user_payload.get("username") or ""
+    return GiteaComment(
+        id=int(payload["id"]),
+        body=payload.get("body") or "",
+        author=str(author),
+        html_url=payload.get("html_url") or payload.get("url") or "",
+        created_at=parse_gitea_dt(payload.get("created_at")),
+        updated_at=parse_gitea_dt(payload.get("updated_at")),
+    )
+
+
+def parse_gitea_dt(value: str | None) -> datetime | None:
+    if not value:
+        return None
+    if value.endswith("Z"):
+        value = value[:-1] + "+00:00"
+    return datetime.fromisoformat(value)

src/agent_gitea/rendering.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 import re
 from dataclasses import dataclass
 
+from .gitea import GiteaComment
 from .models import IssueRecord, RepositoryRecord
 
 
@@ -46,6 +47,48 @@ Keep the section headings exactly as written below, but write the section conten
 """
 
 
+def render_pr_feedback_prompt(
+    repo: RepositoryRecord,
+    issue: IssueRecord,
+    pr_number: int,
+    branch_name: str,
+    comments: list[GiteaComment],
+) -> str:
+    return f"""# Agent PR Feedback Task
+
+Repository: {repo.full_name}
+Base branch: {repo.default_branch}
+Working branch: {branch_name}
+Issue: #{issue.issue_number} {issue.title}
+Issue URL: {issue.html_url}
+Pull request: #{pr_number}
+
+## Issue Body
+
+{issue.body or "(no issue body)"}
+
+## Human PR Comments To Address
+
+{render_comment_thread(comments)}
+
+## Instructions
+
+Continue work in the current workspace and on the existing PR branch.
+Address the human PR comments above with the smallest scoped changes that satisfy the review.
+Create code changes only when needed, run the relevant tests before finishing, and leave the branch ready for a new commit.
+
+Write `.agent-output/AGENT_IMPLEMENTATION_REPORT.md` using this exact section contract.
+Keep the section headings exactly as written below, but write the section content in Chinese:
+
+- Summary
+- Files changed
+- Test commands run
+- Test results
+- Known risks
+- Follow-up suggestions
+"""
+
+
 def render_reviewer_prompt(repo: RepositoryRecord, issue: IssueRecord, pr_number: int) -> str:
     return f"""# Agent Review Task
 
@@ -87,6 +130,20 @@ def render_pr_body(issue: IssueRecord, implementation_report: str) -> str:
 """
 
 
+def render_comment_thread(comments: list[GiteaComment]) -> str:
+    if not comments:
+        return "(no new human PR comments found)"
+    blocks: list[str] = []
+    for comment in comments:
+        created = comment.created_at.isoformat() if comment.created_at else "unknown time"
+        header = f"- Comment #{comment.id} by {comment.author or 'unknown'} at {created}"
+        if comment.html_url:
+            header += f"\n  URL: {comment.html_url}"
+        body = "\n".join(f"  {line}" for line in comment.body.strip().splitlines())
+        blocks.append(f"{header}\n\n{body or '  (empty comment)'}")
+    return "\n\n".join(blocks)
+
+
 def parse_review_report(raw: str) -> ReviewReport:
     verdict_match = re.search(r"(?im)^\s*(?:[-*]\s*)?(?:##\s*)?Verdict\s*:?\s*`?([A-Z_]+)`?", raw)
     verdict = verdict_match.group(1) if verdict_match else "NEEDS_HUMAN_DECISION"
@@ -123,5 +180,5 @@ def render_human_review_summary(review: ReviewReport) -> str:
 
 ## 需要人工处理
 
-请人工审核该 PR。agent-manager 不会自动合并、关闭 PR 或提交变更请求。
+请人工审核该 PR。agent-manager 不会自动合并或关闭 PR；如果继续在该 PR 中留言，agent-manager 会检测评论并尝试提交后续修复。
 """

src/agent_gitea/service.py

@@ -8,12 +8,13 @@ from pathlib import Path
 from .agents import CommandRunner, read_report, render_command, write_prompt
 from .config import AppConfig
 from .db import Database
-from .gitea import GiteaClient
+from .gitea import GiteaClient, GiteaComment
 from .models import IssueRecord, RepositoryRecord, TaskRecord, TaskState
 from .rendering import (
     parse_review_report,
     render_human_review_summary,
     render_implementer_prompt,
+    render_pr_feedback_prompt,
     render_pr_body,
     render_reviewer_prompt,
 )
@@ -93,6 +94,41 @@ def close_issues_for_merged_pull_requests(db: Database, client: GiteaClient) ->
     return closed
 
 
+def scan_pull_request_feedback(db: Database, client: GiteaClient) -> int:
+    tasks = db.list_tasks_awaiting_pr_feedback()
+    if not tasks:
+        return 0
+    username = client.get_authenticated_username()
+    queued = 0
+    for task in tasks:
+        repo, _issue = load_task_context(db, task)
+        assert task.pr_number is not None
+        pull_request = client.get_pull_request(owner=repo.owner, name=repo.name, pr_number=task.pr_number)
+        if pull_request.merged or pull_request.state != "open":
+            continue
+        comments = client.list_issue_comments(owner=repo.owner, name=repo.name, issue_number=task.pr_number)
+        last_seen_comment_id = db.get_pr_feedback_cursor(task.id)
+        newest_comment_id = max((comment.id for comment in comments), default=last_seen_comment_id)
+        human_comments = [
+            comment
+            for comment in comments
+            if comment.id > last_seen_comment_id and comment.author.casefold() != username.casefold()
+        ]
+        if not human_comments:
+            if newest_comment_id > last_seen_comment_id:
+                db.upsert_pr_feedback_state(task.id, last_seen_comment_id=newest_comment_id)
+            continue
+        db.mark_pr_feedback_pending(task.id)
+        db.transition(
+            task.id,
+            TaskState.DISCOVERED,
+            message=f"queued PR feedback from {len(human_comments)} human comment(s)",
+            clear_lease=True,
+        )
+        queued += 1
+    return queued
+
+
 class TaskRunner:
     def __init__(
         self,
@@ -124,6 +160,8 @@ class TaskRunner:
                 logger.info("synced %d repositories", len(repos))
                 closed = close_issues_for_merged_pull_requests(self.db, self.gitea)
                 logger.info("closed %d issues for merged pull requests", closed)
+                feedback_tasks = scan_pull_request_feedback(self.db, self.gitea)
+                logger.info("queued %d tasks from pull request feedback", feedback_tasks)
                 task_ids = scan_issues(self.db, self.config, self.gitea)
                 logger.info("created %d tasks from issue scan", len(task_ids))
                 task = self.run_once()
@@ -138,6 +176,19 @@ class TaskRunner:
 
     def run_claimed(self, task: TaskRecord) -> TaskRecord:
         try:
+            if self.db.has_pending_pr_feedback(task.id):
+                return self._run_pr_feedback(task)
+            return self._run_initial_implementation(task)
+        except Exception as exc:
+            return self.db.transition(
+                task.id,
+                TaskState.FAILED,
+                message="task failed",
+                error_message=str(exc),
+                clear_lease=True,
+            )
+
+    def _run_initial_implementation(self, task: TaskRecord) -> TaskRecord:
         repo, issue = self._load_context(task)
         validate_transition(task.state, TaskState.PLANNING)
         task = self.db.transition(task.id, TaskState.PLANNING, message="rendering implementation prompt")
@@ -183,33 +234,108 @@ class TaskRunner:
         task = self.db.transition(task.id, TaskState.REVIEWING, message="running reviewer agent")
         review_report_raw = self._run_reviewer(task, repo, issue, pr.number, workspace)
         review = parse_review_report(review_report_raw)
-            self.gitea.post_issue_comment(
+        reviewer_comment = self.gitea.post_issue_comment(
             owner=repo.owner,
             name=repo.name,
             issue_number=pr.number,
             body=review_report_raw.strip() or "Reviewer did not produce a report.",
         )
-            self.gitea.post_issue_comment(
+        summary_comment = self.gitea.post_issue_comment(
             owner=repo.owner,
             name=repo.name,
             issue_number=pr.number,
             body=render_human_review_summary(review),
         )
+        self.db.clear_pr_feedback_pending(
+            task.id,
+            last_seen_comment_id=max(reviewer_comment.id, summary_comment.id),
+        )
         final_task = self.db.transition(
             task.id,
             TaskState.HUMAN_REVIEW_READY,
             message=f"human review summary posted with verdict {review.verdict}",
             clear_lease=True,
         )
-            if self.config.workspace.cleanup_on_success and final_task.workspace_path:
-                self.workspace_manager.cleanup(final_task.workspace_path)
         return final_task
-        except Exception as exc:
+
+    def _run_pr_feedback(self, task: TaskRecord) -> TaskRecord:
+        repo, issue = self._load_context(task)
+        if task.pr_number is None:
+            raise ValueError(f"task {task.id} has pending PR feedback without a PR number")
+        if not task.branch_name:
+            raise ValueError(f"task {task.id} has pending PR feedback without a branch")
+        validate_transition(task.state, TaskState.PLANNING)
+        task = self.db.transition(task.id, TaskState.PLANNING, message="rendering PR feedback prompt")
+        workspace = self.workspace_manager.prepare_existing_branch(repo, issue, task.branch_name, task.workspace_path)
+        task = self.db.transition(
+            task.id,
+            TaskState.IMPLEMENTING,
+            message=f"running implementer agent for PR #{task.pr_number} feedback",
+            workspace_path=workspace,
+        )
+        last_seen_comment_id = self.db.get_pr_feedback_cursor(task.id)
+        comments = self.gitea.list_issue_comments(owner=repo.owner, name=repo.name, issue_number=task.pr_number)
+        username = self.gitea.get_authenticated_username()
+        human_comments = [
+            comment
+            for comment in comments
+            if comment.id > last_seen_comment_id and comment.author.casefold() != username.casefold()
+        ]
+        if not human_comments:
+            newest_comment_id = max((comment.id for comment in comments), default=last_seen_comment_id)
+            self.db.clear_pr_feedback_pending(task.id, last_seen_comment_id=newest_comment_id)
             return self.db.transition(
                 task.id,
-                TaskState.FAILED,
+                TaskState.HUMAN_REVIEW_READY,
-                message="task failed",
+                message="PR feedback queue was stale; no new human comments found",
-                error_message=str(exc),
+                clear_lease=True,
+            )
+        implementation_report = self._run_pr_feedback_implementer(
+            task,
+            repo,
+            issue,
+            task.branch_name,
+            task.pr_number,
+            human_comments,
+            workspace,
+        )
+        task = self.db.transition(task.id, TaskState.TESTING, message="checking PR feedback diff")
+        commit_message = f"agent: address PR #{task.pr_number} feedback for issue #{issue.issue_number}"
+        commit_id = self.workspace_manager.commit_changes(workspace, commit_message)
+        self.db.add_event(task.id, TaskState.TESTING, TaskState.TESTING, f"committed PR feedback update {commit_id}")
+        self.workspace_manager.push_branch(workspace, task.branch_name)
+        self.gitea.post_issue_comment(
+            owner=repo.owner,
+            name=repo.name,
+            issue_number=task.pr_number,
+            body=f"## 代理已处理 PR 反馈\n\n{implementation_report.strip()}",
+        )
+        task = self.db.transition(task.id, TaskState.PR_OPENED, message=f"updated PR #{task.pr_number}")
+        validate_transition(task.state, TaskState.REVIEWING)
+        task = self.db.transition(task.id, TaskState.REVIEWING, message="running reviewer agent after PR feedback")
+        review_report_raw = self._run_reviewer(task, repo, issue, task.pr_number, workspace)
+        review = parse_review_report(review_report_raw)
+        reviewer_comment = self.gitea.post_issue_comment(
+            owner=repo.owner,
+            name=repo.name,
+            issue_number=task.pr_number,
+            body=review_report_raw.strip() or "Reviewer did not produce a report.",
+        )
+        summary_comment = self.gitea.post_issue_comment(
+            owner=repo.owner,
+            name=repo.name,
+            issue_number=task.pr_number,
+            body=render_human_review_summary(review),
+        )
+        newest_comment_id = max(
+            [comment.id for comment in comments] + [reviewer_comment.id, summary_comment.id],
+            default=max(reviewer_comment.id, summary_comment.id),
+        )
+        self.db.clear_pr_feedback_pending(task.id, last_seen_comment_id=newest_comment_id)
+        return self.db.transition(
+            task.id,
+            TaskState.HUMAN_REVIEW_READY,
+            message=f"PR feedback handled; human review summary posted with verdict {review.verdict}",
             clear_lease=True,
         )
 
@@ -250,6 +376,46 @@ class TaskRunner:
             raise RuntimeError(f"implementer failed with exit code {result.exit_code}")
         return report or "(implementer did not produce AGENT_IMPLEMENTATION_REPORT.md)"
 
+    def _run_pr_feedback_implementer(
+        self,
+        task: TaskRecord,
+        repo: RepositoryRecord,
+        issue: IssueRecord,
+        branch_name: str,
+        pr_number: int,
+        comments: list[GiteaComment],
+        workspace: Path,
+    ) -> str:
+        prompt = render_pr_feedback_prompt(repo, issue, pr_number, branch_name, comments)
+        output_dir = workspace / ".agent-output"
+        output_dir.mkdir(exist_ok=True)
+        prompt_path = output_dir / "AGENT_PR_FEEDBACK_PROMPT.md"
+        write_prompt(prompt_path, prompt)
+        command = render_command(
+            self.config.agents.implementer.command,
+            workspace_path=workspace.resolve(),
+            prompt_path=prompt_path.resolve(),
+            issue_number=issue.issue_number,
+            issue_title=issue.title,
+            pr_number=pr_number,
+            branch_name=branch_name,
+        )
+        result = self.command_runner.run(command, workspace, stdin=prompt)
+        report = read_report(output_dir / "AGENT_IMPLEMENTATION_REPORT.md")
+        self.db.add_agent_run(
+            task_id=task.id,
+            role="implementer_pr_feedback",
+            command=command,
+            prompt=prompt,
+            stdout=result.stdout,
+            stderr=result.stderr,
+            exit_code=result.exit_code,
+            report=report,
+        )
+        if not result.ok:
+            raise RuntimeError(f"implementer failed with exit code {result.exit_code}")
+        return report or "(implementer did not produce AGENT_IMPLEMENTATION_REPORT.md)"
+
     def _run_reviewer(
         self,
         task: TaskRecord,

src/agent_gitea/workspace.py

@@ -35,11 +35,33 @@ class WorkspaceManager:
         self.exclude_runtime_artifacts(path)
         return path
 
+    def prepare_existing_branch(
+        self,
+        repo: RepositoryRecord,
+        issue: IssueRecord,
+        branch_name: str,
+        workspace_path: str | Path | None,
+    ) -> Path:
+        path = Path(workspace_path) if workspace_path else self.task_workspace(repo, issue)
+        if not path.exists():
+            path.parent.mkdir(parents=True, exist_ok=True)
+            self._git(["clone", repo.clone_url, str(path)], Path.cwd())
+        self.exclude_runtime_artifacts(path)
+        self._git(["fetch", "origin"], path)
+        checkout = self._git(["checkout", branch_name], path, check=False)
+        if checkout.returncode != 0:
+            self._git(["checkout", "-B", branch_name, f"origin/{branch_name}"], path)
+        pull = self._git(["pull", "--ff-only", "origin", branch_name], path, check=False)
+        if pull.returncode != 0:
+            raise RuntimeError(f"git pull --ff-only origin {branch_name} failed: {pull.stderr.strip()}")
+        return path
+
     def exclude_runtime_artifacts(self, workspace: str | Path) -> None:
         exclude_path = Path(workspace) / ".git" / "info" / "exclude"
         entries = {
             ".agent-output/",
             "AGENT_IMPLEMENTER_PROMPT.md",
+            "AGENT_PR_FEEDBACK_PROMPT.md",
             "AGENT_REVIEWER_PROMPT.md",
         }
         existing = exclude_path.read_text(encoding="utf-8") if exclude_path.exists() else ""

tests/test_gitea_service.py

@@ -8,7 +8,13 @@ import httpx
 from agent_gitea.config import AppConfig
 from agent_gitea.gitea import GiteaClient
 from agent_gitea.models import AgentResult, TaskState
-from agent_gitea.service import TaskRunner, close_issues_for_merged_pull_requests, scan_issues, sync_repositories
+from agent_gitea.service import (
+    TaskRunner,
+    close_issues_for_merged_pull_requests,
+    scan_issues,
+    scan_pull_request_feedback,
+    sync_repositories,
+)
 
 
 def make_config(tmp_path: Path, **overrides: object) -> AppConfig:
@@ -167,10 +173,17 @@ class FakeWorkspaceManager:
         self.root = root
         self.diff = diff
         self.pushed: list[str] = []
+        self.resumed: list[tuple[str, Path | None]] = []
 
     def prepare(self, repo, issue, branch_name):
         path = self.root / branch_name.replace("/", "_")
-        path.mkdir(parents=True)
+        path.mkdir(parents=True, exist_ok=True)
+        return path
+
+    def prepare_existing_branch(self, repo, issue, branch_name, workspace_path):
+        self.resumed.append((branch_name, workspace_path))
+        path = Path(workspace_path) if workspace_path else self.root / branch_name.replace("/", "_")
+        path.mkdir(parents=True, exist_ok=True)
         return path
 
     def has_diff(self, workspace, base_ref="origin/HEAD"):
@@ -233,6 +246,16 @@ def seed_task(db):
     return db.create_task(repo.id, 1)
 
 
+def transition_to_human_review_ready(db, task_id: int, *, pr_number: int = 5, branch_name: str | None = None):
+    db.transition(task_id, TaskState.CLAIMED)
+    db.transition(task_id, TaskState.PLANNING)
+    db.transition(task_id, TaskState.IMPLEMENTING, branch_name=branch_name)
+    db.transition(task_id, TaskState.TESTING)
+    db.transition(task_id, TaskState.PR_OPENED, pr_number=pr_number)
+    db.transition(task_id, TaskState.REVIEWING)
+    return db.transition(task_id, TaskState.HUMAN_REVIEW_READY, clear_lease=True)
+
+
 def test_run_task_success_posts_review_comments(db, tmp_path):
     config = make_config(tmp_path)
     seed_task(db)
@@ -272,6 +295,115 @@ def test_run_task_success_posts_review_comments(db, tmp_path):
     assert [path for _, path, _ in requests].count("/api/v1/repos/acme/service/issues/5/comments") == 2
 
 
+def test_scan_pull_request_feedback_queues_task_for_new_human_comment(db):
+    task = seed_task(db)
+    transition_to_human_review_ready(db, task.id, pr_number=5, branch_name="agent/issue-1-ready-issue")
+    db.clear_pr_feedback_pending(task.id, last_seen_comment_id=2)
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        if request.url.path == "/api/v1/user":
+            return httpx.Response(200, json={"login": "agent-bot"})
+        if request.url.path == "/api/v1/repos/acme/service/pulls/5":
+            return httpx.Response(200, json={"number": 5, "state": "open", "merged": False})
+        if request.url.path == "/api/v1/repos/acme/service/issues/5/comments":
+            return httpx.Response(
+                200,
+                json=[
+                    {"id": 1, "body": "review report", "user": {"login": "agent-bot"}},
+                    {"id": 2, "body": "summary", "user": {"login": "agent-bot"}},
+                    {"id": 3, "body": "Please add a regression test.", "user": {"login": "alice"}},
+                    {"id": 4, "body": "internal note", "user": {"login": "agent-bot"}},
+                ],
+            )
+        return httpx.Response(404)
+
+    queued = scan_pull_request_feedback(db, make_client(handler))
+    task_after_scan = db.get_task(task.id)
+
+    assert queued == 1
+    assert task_after_scan is not None
+    assert task_after_scan.state == TaskState.DISCOVERED
+    assert db.has_pending_pr_feedback(task.id)
+    assert db.get_pr_feedback_cursor(task.id) == 2
+
+
+def test_scan_pull_request_feedback_advances_cursor_without_human_comments(db):
+    task = seed_task(db)
+    transition_to_human_review_ready(db, task.id, pr_number=5, branch_name="agent/issue-1-ready-issue")
+    db.clear_pr_feedback_pending(task.id, last_seen_comment_id=2)
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        if request.url.path == "/api/v1/user":
+            return httpx.Response(200, json={"login": "agent-bot"})
+        if request.url.path == "/api/v1/repos/acme/service/pulls/5":
+            return httpx.Response(200, json={"number": 5, "state": "open", "merged": False})
+        if request.url.path == "/api/v1/repos/acme/service/issues/5/comments":
+            return httpx.Response(200, json=[{"id": 3, "body": "summary", "user": {"login": "agent-bot"}}])
+        return httpx.Response(404)
+
+    queued = scan_pull_request_feedback(db, make_client(handler))
+
+    assert queued == 0
+    assert not db.has_pending_pr_feedback(task.id)
+    assert db.get_pr_feedback_cursor(task.id) == 3
+    assert db.get_task(task.id).state == TaskState.HUMAN_REVIEW_READY  # type: ignore[union-attr]
+
+
+def test_run_task_with_pending_pr_feedback_updates_existing_pr(db, tmp_path):
+    config = make_config(tmp_path)
+    task = seed_task(db)
+    workspace_path = tmp_path / "work" / "existing"
+    transition_to_human_review_ready(db, task.id, pr_number=5, branch_name="agent/issue-1-ready-issue")
+    db.transition(task.id, TaskState.DISCOVERED, clear_lease=True)
+    db.clear_pr_feedback_pending(task.id, last_seen_comment_id=2)
+    db.mark_pr_feedback_pending(task.id)
+    db.conn.execute("UPDATE tasks SET workspace_path = ? WHERE id = ?", (str(workspace_path), task.id))
+    db.conn.commit()
+    requests: list[tuple[str, str, dict]] = []
+    next_comment_id = 3
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        nonlocal next_comment_id
+        payload = json.loads(request.content.decode() or "{}")
+        requests.append((request.method, request.url.path, payload))
+        if request.url.path == "/api/v1/user":
+            return httpx.Response(200, json={"login": "agent-bot"})
+        if request.url.path == "/api/v1/repos/acme/service/issues/5/comments" and request.method == "GET":
+            return httpx.Response(
+                200,
+                json=[
+                    {"id": 1, "body": "review report", "user": {"login": "agent-bot"}},
+                    {"id": 2, "body": "summary", "user": {"login": "agent-bot"}},
+                    {"id": 3, "body": "Please add a regression test.", "user": {"login": "alice"}},
+                ],
+            )
+        if request.url.path == "/api/v1/repos/acme/service/issues/5/comments" and request.method == "POST":
+            next_comment_id += 1
+            return httpx.Response(201, json={"id": next_comment_id, "body": payload["body"], "user": {"login": "agent-bot"}})
+        return httpx.Response(404)
+
+    workspace_manager = FakeWorkspaceManager(tmp_path / "work")
+    finished = TaskRunner(
+        db=db,
+        config=config,
+        gitea=make_client(handler),
+        workspace_manager=workspace_manager,
+        command_runner=FakeRunner(),
+        worker_id="worker",
+    ).run_once()
+
+    assert finished is not None
+    assert finished.state == TaskState.HUMAN_REVIEW_READY
+    assert finished.pr_number == 5
+    assert workspace_manager.resumed == [("agent/issue-1-ready-issue", workspace_path)]
+    assert workspace_manager.pushed == ["agent/issue-1-ready-issue"]
+    assert workspace_manager.commit_message == "agent: address PR #5 feedback for issue #1"
+    assert not any(path == "/api/v1/repos/acme/service/pulls" for _, path, _ in requests)
+    assert [run["role"] for run in db.list_agent_runs(task.id)] == ["implementer_pr_feedback", "reviewer"]
+    assert db.get_pr_feedback_cursor(task.id) == 6
+    assert not db.has_pending_pr_feedback(task.id)
+
+
 def test_close_issues_for_merged_pull_requests_closes_linked_issue(db):
     repo = db.upsert_repository(
         owner="acme",