From 8ea8807ced96fc8d8ac57fcc4fa166922393fd06 Mon Sep 17 00:00:00 2001 From: Liang Jiaqing Date: Fri, 10 Jul 2026 11:40:22 +0800 Subject: [PATCH] Improve session controls and compact LLM logs --- frontends/conductor.py | 34 ++++++++++++++++++++++++++-------- frontends/worldline.py | 16 ++++++++++------ llmcore.py | 11 ++++++++--- 3 files changed, 44 insertions(+), 17 deletions(-) diff --git a/frontends/conductor.py b/frontends/conductor.py index 092744e5..7b739e84 100644 --- a/frontends/conductor.py +++ b/frontends/conductor.py @@ -42,6 +42,20 @@ def _apply_desktop_model(agent: "GenericAgent") -> None: print(f"[conductor] failed to apply desktop model #{no}: {e}", file=sys.stderr) +def _select_llm(agent: "GenericAgent", llm: Any) -> bool: + if llm is None or str(llm).strip() == "": return False + q = str(llm).strip() + if isinstance(llm, int) or q.isdigit(): agent.next_llm(int(q)); return True + q = q.lower(); agent.load_llm_sessions() + for i, c in enumerate(agent.llmclients): + vals = [] + for fn in (lambda: agent.get_llm_name(c), lambda: agent.get_llm_name(c, model=True), lambda: c.backend.name, lambda: c.backend.model): + try: vals.append(str(fn()).lower()) + except Exception: pass + if any(q in v for v in vals): agent.next_llm(i); return True + raise ValueError(f"llm not found: {llm}") + + @asynccontextmanager async def lifespan(app: FastAPI): # 服务启动（事件循环已就绪）：捕获 loop 供工作线程跨线程推 WS 广播，并起主agent @@ -62,6 +76,7 @@ class ChatIn(BaseModel): class StartSubagentIn(BaseModel): prompt: str + llm: Any = None class ApprovalIn(BaseModel): prompt: str @@ -70,6 +85,7 @@ class ApprovalIn(BaseModel): class SubagentActionIn(BaseModel): action: str = "intervene" # intervene | abort | kill msg: str = "" + llm: Any = None @dataclass class SubAgentState: @@ -215,13 +231,13 @@ class SubagentPool: s.agent.task_queue.put("EXIT") with self.lock: self.subagents.pop(sid, None) if to_abort: push_cards() - def start_subagent(self, prompt: str) -> dict: + def start_subagent(self, prompt: str, llm: Any = None) -> dict: sid = short_id() agent = GenericAgent() agent.inc_out = True agent.verbose = False agent.no_print = True - _apply_desktop_model(agent) + if not _select_llm(agent, llm): _apply_desktop_model(agent) th = start_agent_runner(agent, f"subagent-{sid}") state = SubAgentState(id=sid, agent=agent, prompt=prompt, status="running", thread=th) with self.lock: self.subagents[sid] = state @@ -233,10 +249,11 @@ class SubagentPool: threading.Thread(target=monitor_display_queue, args=(sid, dq, True), name=f"monitor-{sid}", daemon=True).start() push_cards() return {"id": sid, "status": "running"} - def input_subagent(self, sid: str, msg: str) -> dict: + def input_subagent(self, sid: str, msg: str, llm: Any = None) -> dict: with self.lock: s = self.subagents.get(sid) if not s: return {"error": "subagent not found", "id": sid} if s.status == "running": return {"error": "subagent is still running, cannot input/reply. Start a new subagent instead.", "id": sid} + _select_llm(s.agent, llm) s.prompt = msg s.reply = "" s.status = "running" @@ -257,10 +274,10 @@ READMES = { Conductor API\tBase: http://{HOST}:{PORT} POST /chat\tbody: {{"msg": "..."}}\t给用户发消息 -POST /subagent\tbody: {{"prompt": "..."}}\t启动新subagent，返回 {{"id": "xxx"}} +POST /subagent\tbody: {{"prompt": "..."}}\t启动新subagent，返回 {{"id": "xxx"}}；指定模型加参数llm(数字/名称) POST /approval\tbody: {{"prompt": "...", "source": "..."}}\t推一条待批任务到前端(后端不存)，用户同意则直接派发为subagent POST /subagent/{{id}}\tbody: {{"action": "keyinfo", "msg": "..."}}\t注入key_info（agent下轮可见） -POST /subagent/{{id}}\tbody: {{"action": "input", "msg": "..."}}\t开新一轮任务（agent停下后追加） +POST /subagent/{{id}}\tbody: {{"action": "input", "msg": "..."}}\t开新一轮任务；指定模型加参数llm(数字/名称) POST /subagent/{{id}}\tbody: {{"action": "stop"}}\t中断执行但保留（可继续input/reply） GET /chat?last=N\t返回最近N条对话（默认20） GET /subagent\t返回 {{"items": [...]}}\t查看所有subagent状态 @@ -452,7 +469,7 @@ INSTR_DISPATCHED = "Task received. I'll handle THIS TASK from here. You MUST to @app.post("/subagent") def api_start_subagent(body: StartSubagentIn): - result = pool.start_subagent(body.prompt) + result = pool.start_subagent(body.prompt, body.llm) result["instruction"] = INSTR_DISPATCHED return result @@ -466,7 +483,7 @@ def api_subagent_action(sid: str, body: SubagentActionIn): result["instruction"] = "Received. I'll incorporate this. You MUST to do other task or end your reply." return result if action in ("input", "reply", "append", "message", "msg"): - result = pool.input_subagent(sid, body.msg) + result = pool.input_subagent(sid, body.msg, body.llm) result["instruction"] = INSTR_DISPATCHED return result if action in ("abort", "stop"): @@ -479,10 +496,11 @@ def api_subagent_action(sid: str, body: SubagentActionIn): @app.get("/chat") def api_get_chat(last: int = 20): + items = [m.copy() for m in chat_messages[-last:]] for m in chat_messages: if m.get("role") == "user" and not m.get("read"): m["read"] = True schedule_broadcast({"type": "chat_read"}) - return {"items": chat_messages[-last:]} + return {"items": items} @app.post("/chat") def api_chat(body: ChatIn): diff --git a/frontends/worldline.py b/frontends/worldline.py index 551248bf..de458f4c 100644 --- a/frontends/worldline.py +++ b/frontends/worldline.py @@ -1081,10 +1081,9 @@ def render_native_history(history, ts=None): `_parse_native_history(_pairs(render_native_history(h))) == h`. Byte-for-byte matches llmcore's `_write_llm_log` for the tool backend: - Prompt = `json.dumps(merged, ensure_ascii=False, indent=2)`; Response = - `str(content_blocks)` which for a list equals `repr(...)` (survives - True/False/None that JSON would emit as true/false/null and break - `ast.literal_eval`). Only complete (user, assistant) pairs are emitted.""" + Prompt/Response each put one content block per line, with their wrappers + attached to the first/last lines. Response keeps Python repr so + True/False/None survive `ast.literal_eval`. Only complete pairs are emitted.""" ts = ts or datetime.now().strftime('%Y-%m-%d %H:%M:%S') out, i, n = [], 0, len(history) while i < n: @@ -1092,8 +1091,13 @@ def render_native_history(history, ts=None): if isinstance(u, dict) and u.get('role') == 'user' and i + 1 < n: a = history[i + 1] if isinstance(a, dict) and a.get('role') == 'assistant': - out.append(f"=== Prompt === {ts}\n{json.dumps(u, ensure_ascii=False, indent=2)}\n\n") - out.append(f"=== Response === {ts}\n{a.get('content')!r}\n\n") + prompt_blocks = u.get('content', []) + prompt_raw = '{"role": "user", "content": [\n' + ",\n".join( + json.dumps(b, ensure_ascii=False) for b in prompt_blocks) + "]}" + out.append(f"=== Prompt === {ts}\n{prompt_raw}\n\n") + blocks = a.get('content', []) + raw = "[" + ",\n".join(repr(b) for b in blocks) + "]" + out.append(f"=== Response === {ts}\n{raw}\n\n") i += 2 continue i += 1 diff --git a/llmcore.py b/llmcore.py index c9a976db..0e83c065 100644 --- a/llmcore.py +++ b/llmcore.py @@ -557,6 +557,7 @@ class BaseSession: self.service_tier = _enum('service_tier', {'auto', 'default', 'priority', 'flex'}) self.thinking_type = _enum('thinking_type', {'adaptive', 'enabled', 'disabled'}) self.thinking_budget_tokens = cfg.get('thinking_budget_tokens') + self.omit_thinking = cfg.get('omit_thinking', False) # Exclude thinking from session history mode = str(cfg.get('api_mode', 'chat_completions')).strip().lower().replace('-', '_') self.api_mode = 'responses' if mode in ('responses', 'response') else 'chat_completions' self.temperature = cfg.get('temperature', 1) @@ -735,7 +736,9 @@ class NativeClaudeSession(BaseSession): except StopIteration as e: content_blocks = e.value or [] if content_blocks and (_injected := _ensure_text_block(content_blocks)): yield _injected if content_blocks and not (len(content_blocks) == 1 and content_blocks[0].get("text", "").startswith("!!!Error:")): - self.history.append({"role": "assistant", "content": content_blocks}) + history_blocks = content_blocks + if self.omit_thinking: history_blocks = [b for b in content_blocks if b.get("type") != "thinking"] + self.history.append({"role": "assistant", "content": history_blocks}) text_parts = [b["text"] for b in content_blocks if b.get("type") == "text"] content = "\n".join(text_parts).strip() tool_calls = [MockToolCall(b["name"], b.get("input", {}), id=b.get("id", "")) for b in content_blocks if b.get("type") == "tool_use"] @@ -748,7 +751,8 @@ class NativeClaudeSession(BaseSession): if think_match: thinking = think_match.group(1).strip() content = re.sub(think_pattern, "", content, flags=re.DOTALL) - return MockResponse(thinking, content, tool_calls, str(content_blocks)) + raw = "[" + ",\n".join(repr(b) for b in content_blocks) + "]" + return MockResponse(thinking, content, tool_calls, raw) class NativeOAISession(NativeClaudeSession): native_ua = "codex_exec/0.139.0 (Windows 10.0.26200; x86_64) unknown (codex_exec; 0.139.0)" @@ -1060,7 +1064,8 @@ class NativeToolClient: final_content = tool_result_blocks + filtered_content if not final_content: final_content = [{"type": "text", "text": "."}] merged = {"role": "user", "content": final_content} - _write_llm_log('Prompt', json.dumps(merged, ensure_ascii=False, indent=2), self.log_path) + prompt_raw = '{"role": "user", "content": [\n' + ",\n".join(json.dumps(b, ensure_ascii=False) for b in final_content) + "]}" + _write_llm_log('Prompt', prompt_raw, self.log_path) gen = self.backend.ask(merged) try: while True: