xiaoheiandClaude Opus 4.8
435f720cfb
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>
965 lines
48 KiB
Python
965 lines
48 KiB
Python
"""Task execution engine with OpenAI-compatible LLM provider integration.
|
|
|
|
Merged from agent_swarm_v4:
|
|
- OpenAI-only provider configuration (OPENAI_*/MODEL_* env vars and custom base URLs)
|
|
- Peer-collaboration hook and specialist-role alignment prompting
|
|
- Preserves model invocation, handoff decision hooks, workspace summarization,
|
|
file-application behavior, and the Manager billing/audit usage attribution
|
|
(X-Agent/X-Agnet headers, usage payload, billing_source)
|
|
"""
|
|
import asyncio
|
|
import json
|
|
import logging
|
|
import os
|
|
import re
|
|
import time
|
|
from pathlib import Path
|
|
from typing import Callable, Optional
|
|
|
|
from openai import AsyncOpenAI
|
|
|
|
from .handoff_logic import should_handoff
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
class TaskExecutor:
|
|
"""Executes tasks using a configurable OpenAI-compatible API."""
|
|
|
|
def __init__(self, agent_id: str, workspace_dir: str):
|
|
self.agent_id = agent_id
|
|
self.workspace_dir = workspace_dir
|
|
|
|
api_key = (
|
|
os.getenv("OPENAI_API_KEY")
|
|
or os.getenv("MODEL_API_KEY")
|
|
)
|
|
if not api_key:
|
|
raise ValueError("OPENAI_API_KEY environment variable not set")
|
|
|
|
self.model = (
|
|
os.getenv("OPENAI_MODEL")
|
|
or os.getenv("MODEL_NAME")
|
|
or os.getenv("MODEL_ID")
|
|
or "gpt-4o-mini"
|
|
)
|
|
self.api_base = (
|
|
os.getenv("OPENAI_API_BASE")
|
|
or os.getenv("MODEL_API_BASE")
|
|
or "https://api.openai.com/v1"
|
|
)
|
|
self.api_mode = "openai"
|
|
self.client = AsyncOpenAI(api_key=api_key, base_url=self.api_base)
|
|
self.client_type = "openai"
|
|
|
|
self.usage = self._empty_usage()
|
|
self.current_context: dict = {}
|
|
|
|
# Jina MCP — loaded once at startup; empty list means no tools available
|
|
self.jina_api_key = os.getenv("JINA_API_KEY", "")
|
|
self._jina_tools: list[dict] = [] # OpenAI-format tool schemas
|
|
self._jina_tools_loaded = False
|
|
|
|
async def execute_task(
|
|
self,
|
|
task_id: str,
|
|
description: str,
|
|
context: dict,
|
|
handoff_callback: Optional[Callable] = None,
|
|
agent_capabilities: Optional[list[str]] = None,
|
|
peer_collaboration_callback: Optional[Callable] = None,
|
|
proposal_callback: Optional[Callable] = None,
|
|
) -> dict:
|
|
logger.info(f"Executing task {task_id}: {description}")
|
|
|
|
try:
|
|
started_at = time.time()
|
|
self.current_context = context or {}
|
|
self.usage = self._empty_usage(context)
|
|
|
|
# agent_swarm#7 — bottom-up decomposition (the swarm's canonical fan-out path): when this
|
|
# is a top-level seed, decompose it and PROPOSE the subtasks to the shared pool so peers
|
|
# self-select them, instead of doing the whole objective solo. Proposed/child tasks never
|
|
# re-propose (guarded in _maybe_propose_subtasks) → no loop. See autonomous-task-generation.md.
|
|
proposed = await self._maybe_propose_subtasks(task_id, description, context, proposal_callback)
|
|
if proposed:
|
|
return {
|
|
"success": True,
|
|
"task_id": task_id,
|
|
"summary": f"Decomposed the objective into {proposed} subtask(s) and proposed them to the swarm pool for peers to execute (agent_swarm#7).",
|
|
"subtasks": [],
|
|
"proposed_subtasks": proposed,
|
|
"agent_id": self.agent_id,
|
|
"usage": self._usage_payload(time.time() - started_at),
|
|
}
|
|
|
|
multi_agent_leaf_mode = self._multi_agent_leaf_mode(context)
|
|
allow_handoff = self._allow_dynamic_handoff(context)
|
|
|
|
if multi_agent_leaf_mode:
|
|
subtasks = [{
|
|
"description": description,
|
|
"complexity": "medium",
|
|
"estimated_time": 30,
|
|
"dependencies": [],
|
|
"required_capabilities": context.get("required_capabilities") or ["general"],
|
|
}]
|
|
elif self._subtask_handoff_enabled():
|
|
subtasks = await self._parse_task(description, context)
|
|
else:
|
|
subtasks = [{
|
|
"description": description,
|
|
"complexity": "medium",
|
|
"estimated_time": 30,
|
|
"dependencies": [],
|
|
"required_capabilities": ["general"],
|
|
}]
|
|
|
|
results = []
|
|
for subtask in subtasks:
|
|
if handoff_callback and self._subtask_handoff_enabled() and allow_handoff:
|
|
decision = should_handoff(
|
|
subtask,
|
|
self.agent_id,
|
|
agent_capabilities=agent_capabilities,
|
|
)
|
|
if decision.should_handoff:
|
|
await handoff_callback(
|
|
task_id=task_id,
|
|
subtask=subtask,
|
|
target_capabilities=decision.target_capabilities,
|
|
)
|
|
results.append({
|
|
"subtask": subtask,
|
|
"status": "handed_off",
|
|
"reason": decision.reason,
|
|
"target_capabilities": decision.target_capabilities,
|
|
})
|
|
continue
|
|
|
|
results.append(await self._execute_subtask(subtask, task_id, context, peer_collaboration_callback))
|
|
|
|
success = all(r["status"] in ["completed", "handed_off"] for r in results)
|
|
awaiting_handoff = any(r["status"] == "handed_off" for r in results)
|
|
payload = {
|
|
"success": success,
|
|
"task_id": task_id,
|
|
"subtasks": results,
|
|
"awaiting_handoff": awaiting_handoff,
|
|
"agent_id": self.agent_id,
|
|
"usage": self._usage_payload(time.time() - started_at),
|
|
}
|
|
# Hoist every subtask's generated files to a top-level `files` array so the produced
|
|
# artifacts travel in task.result and reach the orchestrator's aggregation node WITHOUT
|
|
# a git push of a work branch (the swarm's new convergence path; orchestrator
|
|
# quality.collect_generated_files consumes result['files']). Each entry keeps the frozen
|
|
# shape {path, content, action} with the COMPLETE file content; last writer wins per path
|
|
# so a later subtask that rewrites a file supersedes an earlier one. Deletes are carried
|
|
# through as {path, action:"delete"} (no content) for the aggregator to honor.
|
|
payload["files"] = self._aggregate_subtask_files(results)
|
|
if not success:
|
|
# Surface the model's OWN failure explanation (what it saw / what was missing) as a
|
|
# top-level `error`, so the orchestrator/Manager records WHY instead of a generic
|
|
# "Task failed" (agent/main.py uses this as the failure reason). Lead with the model's
|
|
# summary/error — NOT the task prompt — so the reason reads as the diagnosis, not the
|
|
# ask. summary is usually the fuller "saw X, missing Y" narrative; append a distinct
|
|
# error for the concise cause.
|
|
failed = [r for r in results if r.get("status") not in ("completed", "handed_off")]
|
|
explanations = []
|
|
for r in failed:
|
|
summary = str(r.get("summary") or "").strip()
|
|
error = str(r.get("error") or "").strip()
|
|
text = summary or error or "subtask failed without detail"
|
|
if summary and error and error not in summary:
|
|
text = f"{summary} ({error})"
|
|
explanations.append(text)
|
|
detail = "; ".join(explanations) or "subtask(s) failed without detail"
|
|
payload["error"] = detail
|
|
logger.error(f"Task {task_id} failed: {detail}")
|
|
return payload
|
|
except Exception as e:
|
|
logger.error(f"Error executing task {task_id}: {e}")
|
|
return {
|
|
"success": False,
|
|
"task_id": task_id,
|
|
"error": str(e),
|
|
"agent_id": self.agent_id,
|
|
"usage": self._usage_payload(time.time() - started_at if "started_at" in locals() else 0),
|
|
}
|
|
finally:
|
|
self.current_context = {}
|
|
|
|
@staticmethod
|
|
def _aggregate_subtask_files(results: list[dict]) -> list[dict]:
|
|
"""Flatten the per-subtask file specs into one ordered, de-duplicated `files` list.
|
|
|
|
Each subtask result carries the model's `files` (the same {path, action, content} specs the
|
|
executor applied to the workspace). We re-emit them at the top level so the artifacts travel
|
|
in task.result instead of a pushed git branch. Last writer wins per path (a later subtask
|
|
rewriting/deleting a file supersedes an earlier write); ordering follows last occurrence.
|
|
Only writes with non-None content and deletes are kept; malformed entries are skipped.
|
|
"""
|
|
by_path: dict[str, dict] = {}
|
|
for r in results:
|
|
if not isinstance(r, dict):
|
|
continue
|
|
for f in (r.get("files") or []):
|
|
if not isinstance(f, dict):
|
|
continue
|
|
path = f.get("path")
|
|
if not path:
|
|
continue
|
|
action = f.get("action", "write")
|
|
if action == "delete":
|
|
by_path[path] = {"path": path, "action": "delete"}
|
|
elif action == "write":
|
|
content = f.get("content")
|
|
if content is None:
|
|
continue # an incomplete write (no content) is not a usable artifact
|
|
by_path[path] = {"path": path, "content": content, "action": "write"}
|
|
# unknown actions are ignored (the orchestrator only consumes write/delete)
|
|
return list(by_path.values())
|
|
|
|
async def _maybe_propose_subtasks(self, task_id: str, description: str, context: dict,
|
|
proposal_callback: Optional[Callable]) -> int:
|
|
"""agent_swarm#7: decompose a top-level seed and propose each subtask to the shared pool.
|
|
|
|
Returns the number of subtasks proposed (0 = not a seed / disabled / undecomposable, caller
|
|
then executes normally). Guards against loops: only top-level seeds (no parent, source not
|
|
agent_proposed/dynamic_handoff) propose; the proposed children won't re-propose."""
|
|
if not proposal_callback:
|
|
return 0
|
|
ctx = context or {}
|
|
if ctx.get("parent_task_id") is not None:
|
|
return 0
|
|
if ctx.get("source") in ("agent_proposed", "dynamic_handoff"):
|
|
return 0
|
|
if os.getenv("ENABLE_AUTONOMOUS_PROPOSALS", "true").lower() not in {"1", "true", "yes"}:
|
|
return 0
|
|
try:
|
|
subtasks = await self._parse_task(description, ctx)
|
|
except Exception as e:
|
|
logger.warning(f"Seed decomposition failed: {e}")
|
|
return 0
|
|
if not subtasks or len(subtasks) < 2:
|
|
return 0 # nothing to fan out — execute as a single task
|
|
proposed = 0
|
|
for st in subtasks:
|
|
desc = (st.get("description") or "").strip()
|
|
if not desc:
|
|
continue
|
|
caps = st.get("required_capabilities") or []
|
|
try:
|
|
await proposal_callback(
|
|
description=desc,
|
|
reason="bottom-up decomposition of the seed objective (agent_swarm#7)",
|
|
confidence=0.8,
|
|
origin_task_id=task_id,
|
|
trigger_event="seed_decomposition",
|
|
shared_state_snapshot={"origin_task": task_id},
|
|
# Capabilities are a HINT (agent_role), not a dispatch gate: leave
|
|
# required_capabilities empty so ANY idle agent can self-select the subtask
|
|
# (like the seed). LLM-specific caps (e.g. flask/sqlite) aren't a subset of the
|
|
# fixed pool caps → can_agent_run_task would reject → task stuck PENDING (the
|
|
# dispatch gap we hit). Self-selection + τ handle routing instead.
|
|
agent_role=",".join(caps) or "general",
|
|
required_capabilities=[],
|
|
)
|
|
proposed += 1
|
|
except Exception as e:
|
|
logger.warning(f"propose_task failed for a subtask: {e}")
|
|
if proposed:
|
|
logger.info(f"Proposed {proposed} subtask(s) to the shared pool (agent_swarm#7)")
|
|
return proposed
|
|
|
|
async def _parse_task(self, description: str, context: dict) -> list[dict]:
|
|
# agent_swarm#75 (LOCAL fix — not for upstream as-is): the old impl capped output at
|
|
# max_tokens=2000 + strict json.loads + silently fell back to a SINGLE task on any error,
|
|
# so large tasks (their breakdown JSON exceeds 2000 → truncated → parse fail) never split →
|
|
# the swarm degraded to one agent. Fix: bigger configurable budget, lenient parsing
|
|
# (accept bare array or {"subtasks":[...]}), one retry, and explicit decompose/fallback logs.
|
|
# Decomposition output budget (env-tunable). qwen3.7-max max output = 65536 tokens
|
|
# (gateway rejects >65536 with HTTP 400 InvalidParameter), so default to the full 65536
|
|
# so large-task breakdown JSON never truncates. It's a cap, not a target.
|
|
max_tokens = int(os.getenv("AGENT_PLAN_MAX_TOKENS", "65536") or 65536)
|
|
prompt = f"""You are decomposing a software change for a DECENTRALIZED SWARM of autonomous coding agents.
|
|
How the swarm executes (this CONSTRAINS what a valid subtask is — read carefully):
|
|
|
|
- The target repository is ALREADY checked out locally at the correct base commit. Each agent can
|
|
READ any file in the repo and its ONLY output is the COMPLETE new content of the file(s) it owns.
|
|
- Subtasks go into a shared pool and are picked up by DIFFERENT agents IN PARALLEL. Agents do NOT
|
|
share a filesystem and there is NO guaranteed order — every subtask must be independently doable
|
|
straight from the base commit, without seeing any other agent's output.
|
|
- All agents' file outputs are merged BY PATH (one file = one owner) and diffed against the base
|
|
commit to produce the final patch.
|
|
|
|
So every subtask MUST be a concrete CODE CHANGE scoped to specific file(s) that yields complete file
|
|
content. A subtask must NOT be a process/meta step.
|
|
|
|
FORBIDDEN subtasks (an agent literally cannot do these — they will fail and waste the run):
|
|
exploring/surveying the repo, running or executing tests, moving/renaming/deleting files as an
|
|
\"action\", reviewing or verifying other agents' work, setting up the environment, installing
|
|
dependencies, committing/pushing, \"ensure/confirm/check that ...\".
|
|
ALLOWED subtask shape: \"Implement/modify <specific change> in <file(s)>, outputting the complete
|
|
updated file.\" Writing a NEW source or test FILE is allowed (it produces content); \"run the
|
|
tests\" is not.
|
|
|
|
Rules:
|
|
1. Partition by FILE OWNERSHIP: each file that must change is owned by EXACTLY ONE subtask. Never let
|
|
two subtasks edit the same file (parallel agents would clobber each other).
|
|
2. One subtask may own several files only if they must change together; keep ownership disjoint.
|
|
3. Make the SMALLEST change that satisfies the task. If it fits in one file, return EXACTLY ONE
|
|
subtask — do not invent extra steps to look thorough.
|
|
4. Ground every subtask in the ACTUAL files present at the base commit, not assumptions.
|
|
5. dependencies = real content dependencies only (e.g. a test file depends on the intended interface
|
|
of an impl file); each subtask must still be writable independently from the base commit.
|
|
|
|
Task: {description}
|
|
|
|
Context: {json.dumps(context, indent=2)}
|
|
|
|
Return ONLY a JSON array. Each element:
|
|
- description: imperative and file-scoped — WHAT to change, in WHICH file(s), and that it must output
|
|
the complete final file content.
|
|
- target_files: array of repo-relative paths this subtask OWNS (disjoint across subtasks).
|
|
- complexity: \"low\" | \"medium\" | \"high\"
|
|
- dependencies: array of subtask indices (content dependencies only; empty if none)
|
|
- required_capabilities: e.g. [\"python\"], [\"python\",\"testing\"]
|
|
|
|
Return ONLY the JSON array, no other text."""
|
|
last_err = None
|
|
for attempt in range(2): # one retry before degrading
|
|
try:
|
|
content = await self._complete(prompt, max_tokens=max_tokens)
|
|
subtasks = self._coerce_subtasks(content)
|
|
if subtasks:
|
|
logger.info(f"Decomposed task into {len(subtasks)} subtask(s)")
|
|
return subtasks
|
|
last_err = "no subtasks parsed"
|
|
except Exception as e:
|
|
last_err = e
|
|
logger.warning(f"Task decomposition parse failed (attempt {attempt + 1}/2): {e}")
|
|
logger.warning(f"Task decomposition fell back to a single task (reason: {last_err}); the run will not fan out")
|
|
return [{
|
|
"description": description,
|
|
"complexity": "medium",
|
|
"estimated_time": 30,
|
|
"dependencies": [],
|
|
"required_capabilities": ["general"],
|
|
}]
|
|
|
|
def _coerce_subtasks(self, content: str) -> list[dict]:
|
|
"""Lenient parse of the planner output into a subtask list (agent_swarm#75).
|
|
|
|
Accepts a bare JSON array, a ```-fenced array, or an object like {"subtasks": [...]}.
|
|
Returns [] when nothing usable is found (caller retries / degrades)."""
|
|
text = self._strip_json_fence(content or "")
|
|
obj = None
|
|
try:
|
|
obj = json.loads(text)
|
|
except Exception:
|
|
start, end = text.find("["), text.rfind("]")
|
|
if start != -1 and end > start:
|
|
try:
|
|
obj = json.loads(text[start:end + 1])
|
|
except Exception:
|
|
obj = None
|
|
if isinstance(obj, dict):
|
|
obj = obj.get("subtasks") or obj.get("tasks")
|
|
if not isinstance(obj, list):
|
|
return []
|
|
return [s for s in obj if isinstance(s, dict) and s.get("description")]
|
|
|
|
async def _execute_subtask(self, subtask: dict, task_id: str, context: dict, peer_collaboration_callback: Optional[Callable]) -> dict:
|
|
content = ""
|
|
try:
|
|
description = subtask["description"]
|
|
logger.info("════════ [任务开始] task=%s role=%s\n 描述: %s",
|
|
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 []
|
|
implementation_artifacts = [
|
|
artifact for artifact in dependency_artifacts
|
|
if "implementation" in (artifact.get("task_id") or "")
|
|
]
|
|
testing_artifacts = [
|
|
artifact for artifact in dependency_artifacts
|
|
if "testing" in (artifact.get("task_id") or "")
|
|
]
|
|
peer_context = []
|
|
peer_agents = context.get("peer_agents") or []
|
|
should_consult_peers = specialist_role in {"testing", "documentation"}
|
|
if peer_collaboration_callback and peer_agents and should_consult_peers:
|
|
max_peer_consults = int((context or {}).get("max_peer_consults", 2) or 2)
|
|
preferred_roles = []
|
|
if specialist_role in {"testing", "documentation"}:
|
|
preferred_roles = ["implementation"]
|
|
|
|
ordered_peers = sorted(
|
|
peer_agents,
|
|
key=lambda peer: 0 if peer.get("role") in preferred_roles else 1,
|
|
)
|
|
|
|
for peer in ordered_peers[:max_peer_consults]:
|
|
try:
|
|
reply = await peer_collaboration_callback(
|
|
task_id=task_id,
|
|
target_agent_id=peer.get("agent_id"),
|
|
content=(
|
|
f"Specialist role: {specialist_role}. "
|
|
f"Please provide guidance and confirm behavior for: {description}. "
|
|
f"Original user request: {user_prompt}"
|
|
),
|
|
timeout_seconds=float((context or {}).get("peer_timeout_seconds", 10.0) or 10.0),
|
|
)
|
|
peer_context.append(
|
|
{
|
|
"agent_id": peer.get("agent_id"),
|
|
"role": peer.get("role"),
|
|
"content": reply.get("content", ""),
|
|
}
|
|
)
|
|
except Exception as exc:
|
|
logger.warning(f"Peer collaboration failed with {peer.get('agent_id')}: {exc}")
|
|
prompt = f"""You are a programming assistant working in a collaborative agent system.
|
|
|
|
Original user request:
|
|
{user_prompt}
|
|
|
|
Your specialist role:
|
|
{specialist_role}
|
|
|
|
Task: {description}
|
|
|
|
Dependency artifacts from other specialists:
|
|
{json.dumps(dependency_artifacts, indent=2)}
|
|
|
|
Implementation artifacts relevant to alignment:
|
|
{json.dumps(implementation_artifacts, indent=2)}
|
|
|
|
Testing artifacts relevant to alignment:
|
|
{json.dumps(testing_artifacts, indent=2)}
|
|
|
|
Workspace directory: {self.workspace_dir}
|
|
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)}
|
|
|
|
Workspace & tool boundaries (IMPORTANT):
|
|
- The COMPLETE target repository is ALREADY cloned on local disk at: {self.workspace_dir} (checked
|
|
out at the correct base commit). The file list and contents shown above are only a PARTIAL SAMPLE
|
|
of that repo — NOT the whole tree. So NEVER conclude a file or directory is "missing" just because
|
|
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) — 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.
|
|
- Do not pass local paths or file:// URLs to read_url; it only fetches public web URLs.
|
|
|
|
Alignment requirements:
|
|
- If your role is testing, align your tests with the implementation artifacts and their stated error semantics.
|
|
- If your role is documentation, align your docs with both implementation and testing artifacts.
|
|
- Do not invent behavior that conflicts with dependency artifacts unless you explicitly surface an error.
|
|
- Treat implementation artifacts as the source of truth for API behavior and exception semantics.
|
|
- If peer specialist input conflicts with implementation artifacts, prefer implementation semantics and explain the correction in your changes summary.
|
|
- If your role is documentation or testing, update only your specialist outputs to converge on implementation behavior unless the implementation artifact is clearly missing or contradictory.
|
|
|
|
Return your response as JSON with this structure:
|
|
{{
|
|
\"status\": \"completed\" or \"failed\",
|
|
\"summary\": \"Brief description of what was done\",
|
|
\"files\": [{{\"path\": \"relative/path.py\", \"action\": \"write\", \"content\": \"complete file content\"}}],
|
|
\"changes\": \"Detailed description of changes\",
|
|
\"error\": \"Error message if failed, null otherwise\"
|
|
}}
|
|
|
|
Return ONLY the JSON, no other text."""
|
|
# Full-file rewrites (the `files` JSON carries complete file contents) easily exceed a few
|
|
# thousand tokens → 4000 truncated the JSON mid-string → parse failure → task failed.
|
|
# Raise to the model's output ceiling (qwen3.7-max = 65536); env-tunable for other gateways.
|
|
exec_max_tokens = int(os.getenv("AGENT_EXEC_MAX_TOKENS", "65535") or 65535)
|
|
content = await self._complete(prompt, max_tokens=exec_max_tokens)
|
|
result = self._parse_json_response(content)
|
|
logger.info("──────── [任务产出] task=%s status=%s 文件=%s\n 说明: %s",
|
|
task_id, result.get("status"),
|
|
[f.get("path") for f in (result.get("files") or [])],
|
|
(result.get("changes") or result.get("summary") or "")[:600])
|
|
if result.get("status") == "completed":
|
|
apply_result = await self._apply_file_changes(result.get("files", []))
|
|
result["files_modified"] = apply_result["files_modified"]
|
|
result["files_deleted"] = apply_result["files_deleted"]
|
|
if apply_result["errors"]:
|
|
result["status"] = "failed"
|
|
result["error"] = "; ".join(apply_result["errors"])
|
|
logger.warning(
|
|
f"Subtask for task {task_id} failed applying file changes: {result['error']}"
|
|
)
|
|
else:
|
|
# LLM returned 2xx but declared the subtask not completed: log its reason and a
|
|
# bounded snippet of the model output so the failure is diagnosable from agent logs.
|
|
logger.warning(
|
|
"Subtask for task %s reported status=%r (error=%s summary=%s); llm_response[:500]=%s",
|
|
task_id, result.get("status"), result.get("error"),
|
|
result.get("summary"), (content or "").strip()[:500],
|
|
)
|
|
result["subtask"] = subtask
|
|
return result
|
|
except Exception as e:
|
|
# Parse failures / API errors land here; include a bounded model-output snippet
|
|
# (model-generated text, no secrets) to explain why parsing/execution failed.
|
|
logger.error(
|
|
f"Error executing subtask for task {task_id}: {e}; "
|
|
f"llm_response[:500]={(content or '').strip()[:500]}"
|
|
)
|
|
return {
|
|
"subtask": subtask,
|
|
"status": "failed",
|
|
"error": str(e),
|
|
"summary": f"Failed to execute: {e}",
|
|
}
|
|
|
|
# Jina MCP endpoint (StreamableHTTP). Read-only web tools (search_web/read_url/…).
|
|
_JINA_MCP_URL = "https://mcp.jina.ai/v1"
|
|
|
|
def _jina_mcp_headers(self) -> dict:
|
|
return {"Authorization": f"Bearer {self.jina_api_key}"}
|
|
|
|
async def _load_jina_tools(self) -> list[dict]:
|
|
"""Fetch tool schemas from Jina MCP via the standard `mcp` SDK (StreamableHTTP transport).
|
|
Cached after first call. Returns OpenAI function-calling tool specs. The SDK handles the
|
|
MCP handshake, SSE framing, and session — no hand-rolled JSON-RPC/SSE parsing."""
|
|
if self._jina_tools_loaded:
|
|
return self._jina_tools
|
|
self._jina_tools_loaded = True
|
|
if not self.jina_api_key:
|
|
return []
|
|
try:
|
|
from mcp import ClientSession
|
|
from mcp.client.streamable_http import streamablehttp_client
|
|
|
|
async with streamablehttp_client(
|
|
self._JINA_MCP_URL, headers=self._jina_mcp_headers()
|
|
) as (read, write, _):
|
|
async with ClientSession(read, write) as session:
|
|
await session.initialize()
|
|
tools = (await session.list_tools()).tools
|
|
# Constrain the toolset: the agent only needs web SEARCH + fetch a page's full
|
|
# CONTENT. Exposing all ~21 Jina tools bloats the prompt and tempts the model to
|
|
# misuse them (e.g. read_url on local file:// paths, or fetching the repo's own
|
|
# source from the web = WRONG version vs the local base_commit checkout).
|
|
allow = {t.strip() for t in os.getenv(
|
|
"JINA_TOOL_ALLOWLIST", "search_web,read_url").split(",") if t.strip()}
|
|
self._jina_tools = [
|
|
{
|
|
"type": "function",
|
|
"function": {
|
|
"name": t.name,
|
|
"description": t.description or "",
|
|
"parameters": t.inputSchema or {"type": "object", "properties": {}},
|
|
},
|
|
}
|
|
for t in tools
|
|
if not allow or t.name in allow
|
|
]
|
|
logger.info("Loaded %d Jina MCP tools (allowlist=%s, of %d offered)",
|
|
len(self._jina_tools), sorted(allow) or "*", len(tools))
|
|
except Exception as exc:
|
|
logger.warning("Failed to load Jina MCP tools: %s", exc)
|
|
self._jina_tools = []
|
|
return self._jina_tools
|
|
|
|
async def _call_jina_tool(self, tool_name: str, arguments: dict) -> str:
|
|
"""Invoke a single Jina MCP tool via the standard `mcp` SDK and return its text result."""
|
|
try:
|
|
from mcp import ClientSession
|
|
from mcp.client.streamable_http import streamablehttp_client
|
|
|
|
async with streamablehttp_client(
|
|
self._JINA_MCP_URL, headers=self._jina_mcp_headers()
|
|
) as (read, write, _):
|
|
async with ClientSession(read, write) as session:
|
|
await session.initialize()
|
|
result = await session.call_tool(tool_name, arguments)
|
|
parts = [
|
|
c.text for c in result.content
|
|
if getattr(c, "type", None) == "text"
|
|
]
|
|
return "\n".join(parts) or str(result.content)
|
|
except Exception as exc:
|
|
return f"[tool error: {exc}]"
|
|
|
|
@staticmethod
|
|
def _message_text(msg) -> str:
|
|
"""Robustly extract the assistant's textual answer. qwen3.7-max (a THINKING model) can return
|
|
an empty `content` with the real answer in `reasoning_content`; reading only `content` then
|
|
yields "" → `json.loads("")` → "Expecting value: line 1 column 1 (char 0)" → task fails.
|
|
Fallback chain: content → reasoning_content (attribute or OpenAI-SDK model_extra)."""
|
|
text = (getattr(msg, "content", None) or "").strip()
|
|
if text:
|
|
return text
|
|
rc = getattr(msg, "reasoning_content", None)
|
|
if not rc:
|
|
extra = getattr(msg, "model_extra", None) or {}
|
|
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}]
|
|
|
|
for round_i in range(8): # max 8 tool-call rounds
|
|
kwargs: dict = dict(
|
|
model=self.model,
|
|
messages=messages,
|
|
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"
|
|
response = await self.client.chat.completions.create(**kwargs)
|
|
self._record_openai_usage(response)
|
|
msg = response.choices[0].message
|
|
|
|
# 记录这一轮 LLM 的"思考/回复"(可观测 agent 怎么想的)
|
|
if msg.content:
|
|
logger.info("[LLM·第%d轮] 思考/回复:\n%s", round_i + 1, msg.content.strip()[:1200])
|
|
|
|
if not msg.tool_calls:
|
|
return self._message_text(msg)
|
|
|
|
# Execute each tool call and feed results back
|
|
messages.append(msg.model_dump(exclude_unset=True))
|
|
for tc in msg.tool_calls:
|
|
args = json.loads(tc.function.arguments or "{}")
|
|
# 记录调用了哪个 tool、query 是什么
|
|
logger.info("[TOOL·调用] %s 参数=%s", tc.function.name,
|
|
json.dumps(args, ensure_ascii=False)[:400])
|
|
result = await self._call_jina_tool(tc.function.name, args)
|
|
# 记录 tool 返回了什么(搜索结果内容,截断)
|
|
logger.info("[TOOL·返回] %s (%d字):\n%s", tc.function.name, len(result),
|
|
(result or "").strip()[:1000])
|
|
messages.append({
|
|
"role": "tool",
|
|
"tool_call_id": tc.id,
|
|
"content": result,
|
|
})
|
|
|
|
# Fallback: ask for a final answer without tools
|
|
messages.append({"role": "user", "content": "Please provide your final answer now."})
|
|
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)
|
|
|
|
async def peer_reply(self, *, query: str, capabilities: list[str],
|
|
last_summary: Optional[str], max_tokens: Optional[int] = None) -> dict:
|
|
"""Compose a substantive, grounded reply to a peer agent's query (one bounded LLM call).
|
|
|
|
Returns {stance, content, evidence, refs}. `content` is what the requesting agent reads.
|
|
Bounded by PEER_CONSULT_MAX_TOKENS (default 500).
|
|
"""
|
|
if max_tokens is None:
|
|
max_tokens = int(os.getenv("PEER_CONSULT_MAX_TOKENS", "500"))
|
|
workspace_files = self._summarize_workspace()
|
|
workspace_context = self._collect_workspace_context(max_files=8)
|
|
prompt = f"""You are a specialist agent being consulted by a peer in a collaborative swarm.
|
|
Your capabilities: {', '.join(capabilities)}
|
|
Your latest completed work (summary): {last_summary or 'none'}
|
|
|
|
A peer asks:
|
|
{query}
|
|
|
|
Your workspace files:
|
|
{json.dumps(workspace_files, indent=2)}
|
|
Relevant workspace file contents:
|
|
{json.dumps(workspace_context, indent=2)}
|
|
|
|
Answer concisely and concretely, grounded in YOUR actual work/artifacts. Treat implementation
|
|
artifacts as the source of truth for behavior and exception semantics; if the peer's assumption
|
|
conflicts with your work, say so.
|
|
|
|
Return ONLY JSON:
|
|
{{\"stance\": \"agree|disagree|info\", \"content\": \"<concise answer to the peer>\", \"evidence\": \"<what in your work supports this>\", \"refs\": [\"relative/file/path\"]}}"""
|
|
content = await self._complete(prompt, max_tokens=max_tokens)
|
|
result = self._parse_json_response(content)
|
|
if not result.get("content"):
|
|
result["content"] = (content or "").strip()[:1000]
|
|
result.setdefault("stance", "info")
|
|
result.setdefault("evidence", "")
|
|
result.setdefault("refs", [])
|
|
return result
|
|
|
|
def _empty_usage(self, context: Optional[dict] = None) -> dict:
|
|
plan = ((context or {}).get("orchestration_plan") or {})
|
|
billing = plan.get("billing_context") or {}
|
|
return {
|
|
"model_id": self.model,
|
|
"model_tokens": 0,
|
|
"prompt_tokens": 0,
|
|
"completion_tokens": 0,
|
|
"model_cost_usd": 0.0,
|
|
"runtime_seconds": 0.0,
|
|
"billing_source": billing.get("provider") or os.getenv("BILLING_SOURCE", "unknown"),
|
|
}
|
|
|
|
def _record_openai_usage(self, response):
|
|
usage = getattr(response, "usage", None)
|
|
if not usage:
|
|
return
|
|
prompt_tokens = int(getattr(usage, "prompt_tokens", 0) or 0)
|
|
completion_tokens = int(getattr(usage, "completion_tokens", 0) or 0)
|
|
total_tokens = int(getattr(usage, "total_tokens", 0) or prompt_tokens + completion_tokens)
|
|
self._add_usage(prompt_tokens, completion_tokens, total_tokens)
|
|
|
|
def _add_usage(self, prompt_tokens: int, completion_tokens: int, total_tokens: int):
|
|
self.usage["prompt_tokens"] += prompt_tokens
|
|
self.usage["completion_tokens"] += completion_tokens
|
|
self.usage["model_tokens"] += total_tokens
|
|
input_cost = float(os.getenv("MODEL_INPUT_COST_PER_1M", "0") or 0)
|
|
output_cost = float(os.getenv("MODEL_OUTPUT_COST_PER_1M", "0") or 0)
|
|
self.usage["model_cost_usd"] += (
|
|
prompt_tokens * input_cost / 1_000_000
|
|
+ completion_tokens * output_cost / 1_000_000
|
|
)
|
|
|
|
def _usage_payload(self, runtime_seconds: float) -> dict:
|
|
usage = dict(self.usage)
|
|
usage["runtime_seconds"] = runtime_seconds
|
|
return usage
|
|
|
|
def _model_attribution_headers(self) -> dict:
|
|
headers = {}
|
|
mapping = {
|
|
"manager_deployment_id": ["X-Agent-Manager-Deployment-ID", "X-Agnet-Manager-Deployment-ID"],
|
|
"swarm_id": ["X-Agent-Swarm-ID", "X-Agnet-Swarm-ID"],
|
|
"task_id": ["X-Agent-Task-ID", "X-Agnet-Task-ID"],
|
|
"agent_role": ["X-Agent-Agent-Role", "X-Agnet-Agent-Role"],
|
|
"correlation_id": ["X-Correlation-ID"],
|
|
"model_id": ["X-Agent-Model-ID", "X-Agnet-Model-ID"],
|
|
}
|
|
for key, header_names in mapping.items():
|
|
value = self.current_context.get(key)
|
|
if value:
|
|
for header in header_names:
|
|
headers[header] = str(value)
|
|
headers.setdefault("X-Agent-Model-ID", self.model)
|
|
headers.setdefault("X-Agnet-Model-ID", self.model)
|
|
return headers
|
|
|
|
def _subtask_handoff_enabled(self) -> bool:
|
|
return os.getenv("ENABLE_SUBTASK_HANDOFF", "false").lower() in {"1", "true", "yes"}
|
|
|
|
def _multi_agent_leaf_mode(self, context: Optional[dict]) -> bool:
|
|
return self._subtask_handoff_enabled() and (context or {}).get("workflow_mode") == "multi_agent"
|
|
|
|
def _allow_dynamic_handoff(self, context: Optional[dict]) -> bool:
|
|
if (context or {}).get("workflow_mode") != "multi_agent":
|
|
return True
|
|
return bool((context or {}).get("allow_handoff", False))
|
|
|
|
def _strip_json_fence(self, content: str) -> str:
|
|
content = content.strip()
|
|
if not content.startswith("```"):
|
|
return content
|
|
lines = content.split("\n")
|
|
if lines and lines[0].startswith("```"):
|
|
lines = lines[1:]
|
|
if lines and lines[-1].strip() == "```":
|
|
lines = lines[:-1]
|
|
return "\n".join(lines).strip()
|
|
|
|
def _parse_json_response(self, content: str) -> dict:
|
|
content = self._strip_json_fence(content)
|
|
try:
|
|
return json.loads(content)
|
|
except json.JSONDecodeError:
|
|
start = content.find("{")
|
|
end = content.rfind("}")
|
|
if start == -1 or end == -1 or end <= start:
|
|
raise
|
|
return json.loads(content[start:end + 1])
|
|
|
|
def _summarize_workspace(self, max_files: int = 80) -> list[str]:
|
|
root = Path(self.workspace_dir)
|
|
if not root.exists():
|
|
return []
|
|
ignored_dirs = {".git", "__pycache__", "node_modules", ".venv", "venv"}
|
|
files = []
|
|
for path in root.rglob("*"):
|
|
if len(files) >= max_files:
|
|
break
|
|
if not path.is_file():
|
|
continue
|
|
if any(part in ignored_dirs for part in path.relative_to(root).parts):
|
|
continue
|
|
files.append(str(path.relative_to(root)))
|
|
return sorted(files)
|
|
|
|
def _collect_workspace_context(self, max_files: int = 20, max_bytes_per_file: int = 6000) -> dict[str, str]:
|
|
# Real-repo edits need the FULL file (the LLM rewrites complete content; truncated input →
|
|
# truncated/hallucinated rewrite → wrong diff). Raise budgets, env-tunable for big repos.
|
|
max_files = int(os.getenv("AGENT_CTX_MAX_FILES", str(max_files)))
|
|
max_bytes_per_file = int(os.getenv("AGENT_CTX_MAX_BYTES", str(max_bytes_per_file)))
|
|
root = Path(self.workspace_dir)
|
|
if not root.exists():
|
|
return {}
|
|
ignored_dirs = {".git", "__pycache__", "node_modules", ".venv", "venv"}
|
|
allowed_suffixes = {".py", ".js", ".ts", ".tsx", ".jsx", ".json", ".md", ".txt", ".yaml", ".yml", ".toml", ".ini", ".cfg", ".sh"}
|
|
context = {}
|
|
for path in sorted(root.rglob("*")):
|
|
if len(context) >= max_files:
|
|
break
|
|
if not path.is_file():
|
|
continue
|
|
relative = path.relative_to(root)
|
|
if any(part in ignored_dirs for part in relative.parts):
|
|
continue
|
|
if path.suffix and path.suffix.lower() not in allowed_suffixes:
|
|
continue
|
|
try:
|
|
data = path.read_bytes()[:max_bytes_per_file]
|
|
context[str(relative)] = data.decode("utf-8")
|
|
except UnicodeDecodeError:
|
|
continue
|
|
except Exception as e:
|
|
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}")
|
|
root = Path(self.workspace_dir).resolve()
|
|
resolved = (root / file_path).resolve()
|
|
if root != resolved and root not in resolved.parents:
|
|
raise ValueError(f"Path escapes workspace: {file_path}")
|
|
return resolved
|
|
|
|
async def _apply_file_changes(self, files: list[dict]) -> dict:
|
|
result = {"files_modified": [], "files_deleted": [], "errors": []}
|
|
for file_change in files:
|
|
path = file_change.get("path")
|
|
action = file_change.get("action", "write")
|
|
try:
|
|
target = self._resolve_workspace_path(path)
|
|
if action == "write":
|
|
content = file_change.get("content")
|
|
if content is None:
|
|
raise ValueError(f"Missing content for {path}")
|
|
target.parent.mkdir(parents=True, exist_ok=True)
|
|
target.write_text(content, encoding="utf-8")
|
|
result["files_modified"].append(path)
|
|
elif action == "delete":
|
|
if target.exists():
|
|
target.unlink()
|
|
result["files_deleted"].append(path)
|
|
else:
|
|
raise ValueError(f"Unsupported file action for {path}: {action}")
|
|
except Exception as e:
|
|
logger.error(f"Failed to apply file change {path}: {e}")
|
|
result["errors"].append(f"{path}: {e}")
|
|
return result
|
|
|
|
async def read_file(self, file_path: str) -> Optional[str]:
|
|
try:
|
|
full_path = os.path.join(self.workspace_dir, file_path)
|
|
with open(full_path, "r", encoding="utf-8") as f:
|
|
return f.read()
|
|
except Exception as e:
|
|
logger.error(f"Error reading file {file_path}: {e}")
|
|
return None
|
|
|
|
async def write_file(self, file_path: str, content: str) -> bool:
|
|
try:
|
|
full_path = self._resolve_workspace_path(file_path)
|
|
os.makedirs(os.path.dirname(full_path), exist_ok=True)
|
|
with open(full_path, "w", encoding="utf-8") as f:
|
|
f.write(content)
|
|
logger.info(f"Wrote file: {file_path}")
|
|
return True
|
|
except Exception as e:
|
|
logger.error(f"Error writing file {file_path}: {e}")
|
|
return False
|