Merge pull request 'fix(agent): 最小改动产 patch — 注入目标文件完整原文,消除整文件重写噪声' (#27) from fix/minimal-diff-execution into main

Reviewed-on: #27

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.
@@ -852,6 +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}")