Rediseño tool results + compactación por step + integración Docker
- Tool results completos en conversación (como Claude Code/Cursor) en vez de resúmenes en system prompt - Parser multi-tool: trackea tool calls por tool_call_id para OpenAI streaming interleaved - Deduplicación por fingerprint + detección de loop cuando todos los calls de un step son duplicados - Compactación inteligente por step: el orquestador decide cuándo comprimir steps anteriores (cambio de agente o >3 steps) - stdio.js lee URLs del .acai como fallback (local_web_url, local_forge_host) - Buffer MCP aumentado a 1MB para respuestas grandes - Dockerfile adaptado para build context desde raíz del proyecto Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Jordan Diaz
@@ -2,6 +2,7 @@
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import hashlib
|
||||
import json
|
||||
import logging
|
||||
import time
|
||||
@@ -47,6 +48,10 @@ class BaseAgent:
|
||||
) -> dict[str, Any]:
|
||||
"""Run the agent's execution loop.
|
||||
|
||||
Uses real conversation messages with complete tool results,
|
||||
like professional agentic tools (Claude Code, Cursor).
|
||||
Compaction happens at the step level, not per tool result.
|
||||
|
||||
Returns a result dict with keys: content, artifacts, tool_executions.
|
||||
"""
|
||||
artifacts: list[ArtifactSummary] = await self.memory.list_artifacts(
|
||||
@@ -54,15 +59,18 @@ class BaseAgent:
|
||||
)
|
||||
tool_executions: list[ToolExecution] = []
|
||||
accumulated_content = ""
|
||||
working_items: list[dict[str, Any]] = []
|
||||
# Real conversation history: assistant messages + tool results
|
||||
conversation: list[dict[str, Any]] = []
|
||||
tool_fingerprints: dict[str, ToolExecution] = {}
|
||||
all_duplicates_streak = 0 # consecutive steps where ALL calls are duplicates
|
||||
|
||||
for step in range(max_steps):
|
||||
# Build context — NEVER includes raw tool output
|
||||
# Build context with real conversation
|
||||
ctx = await self.context.build_context(
|
||||
session=session,
|
||||
agent=self.profile,
|
||||
artifacts=artifacts,
|
||||
working_items=working_items,
|
||||
conversation=conversation,
|
||||
)
|
||||
|
||||
# Prepare tool definitions
|
||||
@@ -77,7 +85,7 @@ class BaseAgent:
|
||||
|
||||
full_text = ""
|
||||
tool_calls: list[dict[str, Any]] = []
|
||||
current_tool: dict[str, Any] | None = None
|
||||
active_tools: dict[str, dict[str, Any]] = {}
|
||||
|
||||
async for chunk in self.model.stream(
|
||||
messages=ctx.to_messages(),
|
||||
@@ -96,35 +104,40 @@ class BaseAgent:
|
||||
session_id=session.session_id,
|
||||
)
|
||||
|
||||
if chunk.tool_name and (current_tool is None or not current_tool.get("name")):
|
||||
current_tool = {
|
||||
"id": chunk.tool_call_id,
|
||||
"name": chunk.tool_name,
|
||||
"arguments": "",
|
||||
}
|
||||
await self.sse.emit(
|
||||
EventType.TOOL_STARTED,
|
||||
{"tool": chunk.tool_name, "step": step},
|
||||
session_id=session.session_id,
|
||||
)
|
||||
if chunk.tool_name and chunk.tool_call_id:
|
||||
if chunk.tool_call_id not in active_tools:
|
||||
active_tools[chunk.tool_call_id] = {
|
||||
"id": chunk.tool_call_id,
|
||||
"name": chunk.tool_name,
|
||||
"arguments": "",
|
||||
}
|
||||
await self.sse.emit(
|
||||
EventType.TOOL_STARTED,
|
||||
{"tool": chunk.tool_name, "step": step},
|
||||
session_id=session.session_id,
|
||||
)
|
||||
|
||||
if chunk.tool_arguments and current_tool is not None and not chunk.finish_reason:
|
||||
# Accumulate partial argument chunks (NOT the final one)
|
||||
current_tool["arguments"] += chunk.tool_arguments
|
||||
if chunk.tool_arguments and chunk.tool_call_id and not chunk.finish_reason:
|
||||
tool = active_tools.get(chunk.tool_call_id)
|
||||
if tool:
|
||||
tool["arguments"] += chunk.tool_arguments
|
||||
|
||||
if chunk.finish_reason == "tool_use" and current_tool is not None and current_tool.get("name"):
|
||||
# Final chunk carries complete arguments — use those if
|
||||
# partial accumulation is empty, otherwise use accumulated
|
||||
final_args = current_tool["arguments"] or chunk.tool_arguments or ""
|
||||
if chunk.finish_reason == "tool_use" and chunk.tool_call_id:
|
||||
tool = active_tools.pop(chunk.tool_call_id, None)
|
||||
if not tool:
|
||||
tool = {
|
||||
"id": chunk.tool_call_id,
|
||||
"name": chunk.tool_name or "",
|
||||
"arguments": "",
|
||||
}
|
||||
final_args = tool["arguments"] or chunk.tool_arguments or ""
|
||||
try:
|
||||
args = json.loads(final_args) if final_args else {}
|
||||
except json.JSONDecodeError:
|
||||
logger.warning("Failed to parse tool args: %s", final_args[:200])
|
||||
args = {}
|
||||
current_tool["parsed_arguments"] = args
|
||||
logger.debug("Tool call finalized: %s args=%s", current_tool["name"], json.dumps(args)[:200])
|
||||
tool_calls.append(current_tool)
|
||||
current_tool = None
|
||||
tool["parsed_arguments"] = args
|
||||
tool_calls.append(tool)
|
||||
|
||||
if chunk.finish_reason == "end_turn":
|
||||
break
|
||||
@@ -133,24 +146,90 @@ class BaseAgent:
|
||||
|
||||
# If no tool calls, we're done
|
||||
if not tool_calls:
|
||||
# Add final assistant message to conversation
|
||||
if full_text:
|
||||
conversation.append({"role": "assistant", "content": full_text})
|
||||
break
|
||||
|
||||
# Execute tool calls
|
||||
# Add assistant message with tool calls to conversation
|
||||
# (OpenAI format: assistant message carries tool_calls)
|
||||
assistant_msg: dict[str, Any] = {"role": "assistant"}
|
||||
if full_text:
|
||||
assistant_msg["content"] = full_text
|
||||
assistant_msg["tool_calls"] = [
|
||||
{
|
||||
"id": tc["id"],
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": tc["name"],
|
||||
"arguments": json.dumps(tc.get("parsed_arguments", {})),
|
||||
},
|
||||
}
|
||||
for tc in tool_calls
|
||||
]
|
||||
conversation.append(assistant_msg)
|
||||
|
||||
# Execute tool calls and add COMPLETE results to conversation
|
||||
duplicates_this_step = 0
|
||||
for tc in tool_calls:
|
||||
fp_raw = f"{tc['name']}:{json.dumps(tc.get('parsed_arguments', {}), sort_keys=True)}"
|
||||
fp = hashlib.md5(fp_raw.encode()).hexdigest()
|
||||
|
||||
if fp in tool_fingerprints:
|
||||
prev_exec = tool_fingerprints[fp]
|
||||
tool_executions.append(prev_exec)
|
||||
duplicates_this_step += 1
|
||||
# Return cached result as tool message
|
||||
conversation.append({
|
||||
"role": "tool",
|
||||
"tool_call_id": tc["id"],
|
||||
"content": f"[DUPLICADO] Ya ejecutada con mismos argumentos. Resultado: {prev_exec.raw_output[:2000]}",
|
||||
})
|
||||
logger.warning("Duplicate tool call skipped: %s (fingerprint: %s)", tc["name"], fp[:8])
|
||||
continue
|
||||
|
||||
tool_exec = await self._execute_tool(
|
||||
session=session,
|
||||
tool_name=tc["name"],
|
||||
arguments=tc.get("parsed_arguments", {}),
|
||||
artifacts=artifacts,
|
||||
)
|
||||
tool_fingerprints[fp] = tool_exec
|
||||
tool_executions.append(tool_exec)
|
||||
|
||||
# Add summarised result to working context (NEVER raw)
|
||||
working_items.append({
|
||||
"role": "tool_result",
|
||||
"content": f"[{tc['name']}] {tool_exec.result_summary}",
|
||||
# COMPLETE result in conversation (truncated to safe limit)
|
||||
conversation.append({
|
||||
"role": "tool",
|
||||
"tool_call_id": tc["id"],
|
||||
"content": tool_exec.raw_output[:8000] if tool_exec.raw_output else tool_exec.result_summary,
|
||||
})
|
||||
|
||||
# Loop detection: if ALL tool calls in this step were duplicates
|
||||
if duplicates_this_step == len(tool_calls):
|
||||
all_duplicates_streak += 1
|
||||
if all_duplicates_streak >= 2:
|
||||
logger.warning("Loop detected: %d consecutive steps with all duplicate calls. Breaking.", all_duplicates_streak)
|
||||
conversation.append({
|
||||
"role": "user",
|
||||
"content": "[SISTEMA] Se detectaron llamadas repetidas. Ya tienes toda la información necesaria. Genera tu respuesta final ahora.",
|
||||
})
|
||||
# One more chance to generate a final response
|
||||
ctx = await self.context.build_context(
|
||||
session=session, agent=self.profile,
|
||||
artifacts=artifacts, conversation=conversation,
|
||||
)
|
||||
async for chunk in self.model.stream(
|
||||
messages=ctx.to_messages(),
|
||||
config=config,
|
||||
):
|
||||
if chunk.delta:
|
||||
accumulated_content += chunk.delta
|
||||
if chunk.finish_reason:
|
||||
break
|
||||
break
|
||||
else:
|
||||
all_duplicates_streak = 0
|
||||
|
||||
return {
|
||||
"content": accumulated_content,
|
||||
"artifacts": artifacts,
|
||||
@@ -200,6 +279,7 @@ class BaseAgent:
|
||||
|
||||
tool_exec.status = ToolExecutionStatus.COMPLETED
|
||||
tool_exec.result_summary = artifact.summary
|
||||
tool_exec.raw_output = raw_output[:8000]
|
||||
tool_exec.duration_ms = duration
|
||||
|
||||
await self.sse.emit(
|
||||
|
||||
@@ -16,7 +16,7 @@ from ..context.engine import ContextEngine
|
||||
from ..mcp.manager import MCPManager
|
||||
from ..memory.store import MemoryStore
|
||||
from ..models.agent import AgentRole
|
||||
from ..models.session import SessionState, SessionStatus, TaskStatus
|
||||
from ..models.session import SessionState, SessionStatus, TaskState, TaskStatus
|
||||
from ..streaming.sse import SSEEmitter, EventType
|
||||
from .agents.coder import CoderAgent, create_coder_profile
|
||||
from .agents.collector import CollectorAgent, create_collector_profile
|
||||
@@ -181,6 +181,10 @@ class OrchestratorEngine:
|
||||
for artifact in step_result.get("artifacts", []):
|
||||
task.facts_extracted.extend(artifact.facts[:5])
|
||||
|
||||
# Decide if previous steps should be compacted
|
||||
if i > 0:
|
||||
self._maybe_compact_previous_steps(task, current_index=i)
|
||||
|
||||
except Exception as e:
|
||||
logger.error("Step %d failed: %s", i + 1, e)
|
||||
step.status = TaskStatus.FAILED
|
||||
@@ -323,6 +327,35 @@ class OrchestratorEngine:
|
||||
task.task_id, len(task.facts_extracted), len(tools_used), len(task_artifacts),
|
||||
)
|
||||
|
||||
def _maybe_compact_previous_steps(
|
||||
self, task: TaskState, current_index: int
|
||||
) -> None:
|
||||
"""Decide if previous steps should be compacted. Deterministic rules."""
|
||||
current_step = task.plan[current_index]
|
||||
|
||||
for i in range(current_index):
|
||||
prev = task.plan[i]
|
||||
if prev.compacted or prev.status != TaskStatus.COMPLETED:
|
||||
continue
|
||||
|
||||
# Rule 1: Change of agent role → previous steps are a different focus
|
||||
if prev.agent_role != current_step.agent_role:
|
||||
prev.compacted = True
|
||||
logger.info(
|
||||
"Compacted step %d (%s) — agent changed to %s",
|
||||
i + 1, prev.agent_role, current_step.agent_role,
|
||||
)
|
||||
continue
|
||||
|
||||
# Rule 2: More than 3 completed non-compacted steps → compact oldest
|
||||
non_compacted = [
|
||||
s for s in task.plan[:current_index]
|
||||
if s.status == TaskStatus.COMPLETED and not s.compacted
|
||||
]
|
||||
if len(non_compacted) > 3:
|
||||
non_compacted[0].compacted = True
|
||||
logger.info("Compacted oldest step to stay within budget")
|
||||
|
||||
def _create_agent(self, role: AgentRole) -> PlannerAgent | CoderAgent | CollectorAgent | ReviewerAgent:
|
||||
"""Instantiate a subagent for the given role."""
|
||||
profile = self._profiles[role]
|
||||
|
||||
Reference in New Issue
Block a user