fix(agent): 注入目标文件完整原文 + 强令最小改动(修整文件重写噪声)

根因:agent 无本地读文件工具,只拿到 workspace 的截断采样,被要求"输出完整文件"
时只能凭残缺采样+自身知识重构整个文件 → 丢行/重排/改注释/版权头 → 巨噪声 diff,
要么 apply 不上,要么 apply 了破坏无关测试(SWE-bench Pro 两臂 0/10 的主因)。

修法:
- _load_exact_target_files():从子任务描述正则提取目标文件路径,注入其**完整未截断**
  的磁盘原文(上限6文件/各120KB,超限留给采样)。
- 执行 prompt 新增 "EXACT CURRENT CONTENT" 段 + 强令:返回内容必须是该原文逐字节保留、
  仅施加任务所需最小编辑,禁止 reformat/reorder/rename/clean-up 无关行。
- 强化原 full-file 输出说明。

目的:让 diff 只含必要行,救回 NodeBB 296/300、ansible 171/175 这类"差几个测试"的近失题。
契约测试全绿(runtime-contract/merge-smoke/workflow-e2e/contract-freeze)。

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

agent/task_executor.py

@@ -11,6 +11,7 @@ import asyncio
 import json
 import logging
 import os
+import re
 import time
 from pathlib import Path
 from typing import Callable, Optional
@@ -377,6 +378,7 @@ Return ONLY the JSON array, no other text."""
                         task_id, context.get("specialist_role", "general"), description)
             workspace_files = self._summarize_workspace()
             workspace_context = self._collect_workspace_context()
+            exact_target_files = self._load_exact_target_files(description)
             user_prompt = context.get("user_prompt") or context.get("run_goal") or context.get("root_task_description") or ""
             specialist_role = context.get("specialist_role", "general")
             dependency_artifacts = context.get("dependency_artifacts") or []
@@ -447,6 +449,16 @@ Current workspace files:
 {json.dumps(workspace_files, indent=2)}
 Relevant workspace file contents:
 {json.dumps(workspace_context, indent=2)}
+
+EXACT CURRENT CONTENT of the file(s) this task edits (verbatim from disk — authoritative):
+{json.dumps(exact_target_files, indent=2)}
+^ When a file appears above, your returned `content` for it MUST be that exact text with ONLY the
+  minimal edits the task requires. Copy every other line byte-for-byte: imports, comments, copyright
+  headers, blank lines, whitespace, and unrelated code. Do NOT reformat, reorder, rename, restyle,
+  "clean up", or re-wrap anything the task does not explicitly require. The resulting diff against the
+  base commit must contain ONLY the lines the task needs — unrelated changes break tests and fail the
+  patch. If a file you must edit is NOT shown above, read it conceptually from the listed workspace
+  contents and still change as little as possible.
 Peer specialist input:
 {json.dumps(peer_context, indent=2)}
 
@@ -457,7 +469,9 @@ Workspace & tool boundaries (IMPORTANT):
   it is not listed above: if the task references e.g. `qutebrowser/utils/log.py`, that file EXISTS
   on disk at {self.workspace_dir}/qutebrowser/utils/log.py — assume it is present and edit it.
 - Treat the local files as the single source of truth. Return the complete modified content of each
-  file you change in `files` (full file, not a diff).
+  file you change in `files` (full file, not a diff) — but it MUST be the file's exact current content
+  with only the minimal task-required edits applied (see "EXACT CURRENT CONTENT" above). Preserve all
+  unchanged lines verbatim; never regenerate a file from memory or reformat it.
 - The web search / read_url tools are for EXTERNAL knowledge ONLY (library/framework docs, language
   features, error-message lookups). NEVER use them to fetch THIS repository's own source — the web
   copy is a DIFFERENT version than your local base commit and will corrupt your patch.
@@ -610,9 +624,19 @@ Return ONLY the JSON, no other text."""
             rc = extra.get("reasoning_content") if isinstance(extra, dict) else None
         return (rc or "").strip()
 
+    @staticmethod
+    def _thinking_extra_body() -> Optional[dict]:
+        """qwen3.7-max defaults to THINKING mode: on heavy impl subtasks it spends the token
+        budget on `reasoning_content` and never emits the final `content` JSON, so json.loads("")
+        fails ("Expecting value: line 1 column 1"). Disable thinking for deterministic structured
+        execution. Set AGENT_ENABLE_THINKING=1 to opt back in (e.g. for non-structured analysis)."""
+        enabled = os.getenv("AGENT_ENABLE_THINKING", "false").strip().lower() in ("1", "true", "yes", "on")
+        return None if enabled else {"enable_thinking": False}
+
     async def _complete(self, prompt: str, max_tokens: int) -> str:
         """Call the LLM with optional Jina MCP tools; handles the tool-call loop."""
         extra_headers = self._model_attribution_headers()
+        extra_body = self._thinking_extra_body()
         tools = await self._load_jina_tools()
         messages = [{"role": "user", "content": prompt}]
 
@@ -623,6 +647,8 @@ Return ONLY the JSON, no other text."""
                 max_tokens=max_tokens,
                 extra_headers=extra_headers or None,
             )
+            if extra_body:
+                kwargs["extra_body"] = extra_body
             if tools:
                 kwargs["tools"] = tools
                 kwargs["tool_choice"] = "auto"
@@ -656,10 +682,13 @@ Return ONLY the JSON, no other text."""
 
         # Fallback: ask for a final answer without tools
         messages.append({"role": "user", "content": "Please provide your final answer now."})
-        response = await self.client.chat.completions.create(
+        fallback_kwargs: dict = dict(
             model=self.model, messages=messages, max_tokens=max_tokens,
             extra_headers=extra_headers or None,
         )
+        if extra_body:
+            fallback_kwargs["extra_body"] = extra_body
+        response = await self.client.chat.completions.create(**fallback_kwargs)
         self._record_openai_usage(response)
         return self._message_text(response.choices[0].message)
 
@@ -837,6 +866,48 @@ Return ONLY JSON:
                 logger.warning(f"Failed to read workspace file {relative}: {e}")
         return context
 
+    #: Path-like tokens in a subtask description, e.g. `lib/ansible/module_utils/urls.py`.
+    _PATH_RE = re.compile(
+        r"[A-Za-z0-9_][A-Za-z0-9_./\-]*\.(?:py|js|jsx|ts|tsx|go|rb|java|c|h|cc|cpp|hpp|"
+        r"rs|php|cs|kt|swift|scala|vue|json|ya?ml|toml|cfg|ini|sh)\b"
+    )
+
+    def _load_exact_target_files(self, description: str, max_files: int = 6,
+                                 max_bytes: int = 120000) -> dict[str, str]:
+        """Load the FULL, UNTRUNCATED on-disk content of the file(s) a subtask names.
+
+        Root cause of noisy/non-applying patches: the agent was only shown a truncated SAMPLE of the
+        repo, so it reconstructed whole files from partial input → dropped/reordered/reformatted lines
+        → huge diffs that break tests. Here we read the exact current content of the files referenced
+        in the subtask description so the model can preserve every unchanged line verbatim and edit
+        only what the task requires. Paths that don't resolve inside the workspace are skipped."""
+        if not description:
+            return {}
+        out: dict[str, str] = {}
+        # Longest paths first: prefer `a/b/c.py` over a bare `c.py` mention of the same file.
+        cands = sorted(set(m.group(0) for m in self._PATH_RE.finditer(description)),
+                       key=len, reverse=True)
+        for cand in cands:
+            if len(out) >= max_files:
+                break
+            try:
+                p = self._resolve_workspace_path(cand)
+            except ValueError:
+                continue
+            if not p.is_file():
+                continue
+            try:
+                data = p.read_bytes()
+            except Exception:
+                continue
+            if len(data) > max_bytes:
+                continue  # too large to inject in full; leave to the sampled context
+            try:
+                out[cand] = data.decode("utf-8")
+            except UnicodeDecodeError:
+                continue
+        return out
+
     def _resolve_workspace_path(self, file_path: str) -> Path:
         if not file_path or os.path.isabs(file_path):
             raise ValueError(f"Invalid relative path: {file_path}")

orchestrator/main.py

@@ -776,13 +776,29 @@ async def refresh_swarm_run_status(run):
                 if collect_generated_files([t]).get("impl"):
                     if await task_queue.reopen_task(t.task_id):
                         reopened += 1
-            run.metadata["review_cycles"] = agg_cycles + 1
-            run.metadata["rework_reopens"] = int(run.metadata.get("rework_reopens", 0) or 0) + reopened
-            await swarm_runtime.save_run(run)
-            logger.info("aggregation: bounced run %s — reopened %d for rework (cycle %d, pass_rate=%s)",
-                        run.swarm_id, reopened, agg_cycles + 1,
-                        (agg.get("validation") or {}).get("pass_rate"))
-            return
+            # Deadlock guard: when the impl subtask(s) FAILED (0 impl files) and only test tasks
+            # produced output, the loop above finds nothing to reopen → reopens=0 → the run would
+            # spin 'running' forever. Reopen the FAILED tasks so the next cycle retries the impl.
+            if reopened == 0:
+                for t in tasks:
+                    if t.status == TaskStatus.FAILED and await task_queue.reopen_task(t.task_id):
+                        reopened += 1
+            if reopened > 0:
+                run.metadata["review_cycles"] = agg_cycles + 1
+                run.metadata["rework_reopens"] = int(run.metadata.get("rework_reopens", 0) or 0) + reopened
+                await swarm_runtime.save_run(run)
+                logger.info("aggregation: bounced run %s — reopened %d for rework (cycle %d, pass_rate=%s)",
+                            run.swarm_id, reopened, agg_cycles + 1,
+                            (agg.get("validation") or {}).get("pass_rate"))
+                return
+            # Nothing reopenable (no impl files and no failed tasks to retry): bouncing here would
+            # deadlock the run in 'running'. Stop bouncing and accept the existing artifacts so the
+            # run terminates via the normal completion/convergence path below.
+            agg["decision"] = "accept_no_rework"
+            run.metadata["aggregation"] = agg
+            logger.warning("aggregation: bounce on run %s but nothing reopenable (reopens=0) — "
+                           "accepting existing artifacts to avoid deadlock (cycle %d)",
+                           run.swarm_id, agg_cycles + 1)
         await swarm_runtime.save_run(run)
 
     if run.status == next_status: