This repository has been archived on 2026-03-24. You can view files and clone it. You cannot open issues or pull requests or push a commit.
Files
Timmy-time-dashboard/src/timmy/agentic_loop.py

360 lines
12 KiB
Python

"""Agentic loop — multi-step task execution with progress tracking.
Provides `run_agentic_loop()`, the engine behind the `plan_and_execute` tool.
When the model recognises a task needs 3+ sequential steps, it calls
`plan_and_execute(task)` which spawns this loop in the background.
Flow:
1. Planning — ask the model to break the task into numbered steps
2. Execution — run each step sequentially, feeding results forward
3. Adaptation — on failure, ask the model to adapt the plan
4. Summary — ask the model to summarise what was accomplished
Progress is broadcast via WebSocket so the dashboard can show live updates.
"""
from __future__ import annotations
import asyncio
import logging
import re
import threading
import time
import uuid
from collections.abc import Callable
from dataclasses import dataclass, field
logger = logging.getLogger(__name__)
# ---------------------------------------------------------------------------
# Data structures
# ---------------------------------------------------------------------------
@dataclass
class AgenticStep:
"""Result of a single step in the agentic loop."""
step_num: int
description: str
result: str
status: str # "completed" | "failed" | "adapted"
duration_ms: int
@dataclass
class AgenticResult:
"""Final result of the entire agentic loop."""
task_id: str
task: str
summary: str
steps: list[AgenticStep] = field(default_factory=list)
status: str = "completed" # "completed" | "partial" | "failed"
total_duration_ms: int = 0
# ---------------------------------------------------------------------------
# Agent factory
# ---------------------------------------------------------------------------
_loop_agent = None
_loop_agent_lock = threading.Lock()
def _get_loop_agent():
"""Create a fresh agent for the agentic loop.
Returns the same type of agent as `create_timmy()` but with a
dedicated session so it doesn't pollute the main chat history.
"""
global _loop_agent
if _loop_agent is None:
with _loop_agent_lock:
if _loop_agent is None:
from timmy.agent import create_timmy
_loop_agent = create_timmy()
return _loop_agent
# ---------------------------------------------------------------------------
# Plan parser
# ---------------------------------------------------------------------------
_STEP_RE = re.compile(r"^\s*(\d+)[.)]\s*(.+)$", re.MULTILINE)
def _parse_steps(plan_text: str) -> list[str]:
"""Extract numbered steps from the model's planning output."""
matches = _STEP_RE.findall(plan_text)
if matches:
return [desc.strip() for _, desc in matches]
# Fallback: split on newlines, ignore blanks
return [line.strip() for line in plan_text.strip().splitlines() if line.strip()]
# ---------------------------------------------------------------------------
# Extracted helpers
# ---------------------------------------------------------------------------
def _extract_content(run_result) -> str:
"""Extract text content from an agent run result."""
return run_result.content if hasattr(run_result, "content") else str(run_result)
def _clean(text: str) -> str:
"""Clean a model response using session's response cleaner."""
from timmy.session import _clean_response
return _clean_response(text)
async def _plan_task(
agent, task: str, session_id: str, max_steps: int
) -> tuple[list[str], bool] | str:
"""Run the planning phase — returns (steps, was_truncated) or error string."""
plan_prompt = (
f"Break this task into numbered steps (max {max_steps}). "
f"Return ONLY a numbered list, nothing else.\n\n"
f"Task: {task}"
)
try:
plan_run = await asyncio.to_thread(
agent.run, plan_prompt, stream=False, session_id=f"{session_id}_plan"
)
plan_text = _extract_content(plan_run)
except Exception as exc: # broad catch intentional: agent.run can raise any error
logger.error("Agentic loop: planning failed: %s", exc)
return f"Planning failed: {exc}"
steps = _parse_steps(plan_text)
if not steps:
return "Planning produced no steps."
planned_count = len(steps)
steps = steps[:max_steps]
return steps, planned_count > len(steps)
async def _execute_step(
agent,
task: str,
step_desc: str,
step_num: int,
total_steps: int,
recent_results: list[str],
session_id: str,
) -> AgenticStep:
"""Execute a single step, returning an AgenticStep."""
step_start = time.monotonic()
context = (
f"Task: {task}\n"
f"Step {step_num}/{total_steps}: {step_desc}\n"
f"Recent progress: {recent_results[-2:] if recent_results else []}\n\n"
f"Execute this step and report what you did."
)
step_run = await asyncio.to_thread(
agent.run, context, stream=False, session_id=f"{session_id}_step{step_num}"
)
step_result = _clean(_extract_content(step_run))
return AgenticStep(
step_num=step_num,
description=step_desc,
result=step_result,
status="completed",
duration_ms=int((time.monotonic() - step_start) * 1000),
)
async def _adapt_step(
agent,
step_desc: str,
step_num: int,
error: Exception,
step_start: float,
session_id: str,
) -> AgenticStep:
"""Attempt adaptation after a step failure."""
adapt_prompt = (
f"Step {step_num} failed with error: {error}\n"
f"Original step was: {step_desc}\n"
f"Adapt the plan and try an alternative approach for this step."
)
adapt_run = await asyncio.to_thread(
agent.run, adapt_prompt, stream=False, session_id=f"{session_id}_adapt{step_num}"
)
adapt_result = _clean(_extract_content(adapt_run))
return AgenticStep(
step_num=step_num,
description=f"[Adapted] {step_desc}",
result=adapt_result,
status="adapted",
duration_ms=int((time.monotonic() - step_start) * 1000),
)
def _summarize(result: AgenticResult, total_steps: int, was_truncated: bool) -> None:
"""Fill in summary and final status on the result object (mutates in place)."""
completed = sum(1 for s in result.steps if s.status == "completed")
adapted = sum(1 for s in result.steps if s.status == "adapted")
failed = sum(1 for s in result.steps if s.status == "failed")
parts = [f"Completed {completed}/{total_steps} steps"]
if adapted:
parts.append(f"{adapted} adapted")
if failed:
parts.append(f"{failed} failed")
result.summary = f"{result.task}: {', '.join(parts)}."
if was_truncated or len(result.steps) < total_steps or failed:
result.status = "partial"
else:
result.status = "completed"
# ---------------------------------------------------------------------------
# Core loop
# ---------------------------------------------------------------------------
async def run_agentic_loop(
task: str,
*,
session_id: str = "agentic",
max_steps: int = 0,
on_progress: Callable | None = None,
) -> AgenticResult:
"""Execute a multi-step task with planning, execution, and adaptation.
Args:
task: Full description of the task to execute.
session_id: Agno session_id for conversation continuity.
max_steps: Max steps to execute (0 = use config default).
on_progress: Optional async callback(description, step_num, total_steps).
Returns:
AgenticResult with steps, summary, and status.
"""
from config import settings
if max_steps <= 0:
max_steps = getattr(settings, "max_agent_steps", 10)
task_id = str(uuid.uuid4())[:8]
start_time = time.monotonic()
agent = _get_loop_agent()
result = AgenticResult(task_id=task_id, task=task, summary="")
# Phase 1: Planning
plan = await _plan_task(agent, task, session_id, max_steps)
if isinstance(plan, str):
result.status = "failed"
result.summary = plan
result.total_duration_ms = int((time.monotonic() - start_time) * 1000)
return result
steps, was_truncated = plan
total_steps = len(steps)
await _broadcast_progress(
"agentic.plan_ready",
{"task_id": task_id, "task": task, "steps": steps, "total": total_steps},
)
# Phase 2: Execution
completed_results: list[str] = []
for i, step_desc in enumerate(steps, 1):
step_start = time.monotonic()
try:
step = await _execute_step(
agent,
task,
step_desc,
i,
total_steps,
completed_results,
session_id,
)
result.steps.append(step)
completed_results.append(f"Step {i}: {step.result[:200]}")
await _broadcast_progress(
"agentic.step_complete",
{
"task_id": task_id,
"step": i,
"total": total_steps,
"description": step_desc,
"result": step.result[:200],
},
)
if on_progress:
await on_progress(step_desc, i, total_steps)
except Exception as exc: # broad catch intentional: agent.run can raise any error
logger.warning("Agentic loop step %d failed: %s", i, exc)
try:
step = await _adapt_step(agent, step_desc, i, exc, step_start, session_id)
result.steps.append(step)
completed_results.append(f"Step {i} (adapted): {step.result[:200]}")
await _broadcast_progress(
"agentic.step_adapted",
{
"task_id": task_id,
"step": i,
"total": total_steps,
"description": step_desc,
"error": str(exc),
"adaptation": step.result[:200],
},
)
if on_progress:
await on_progress(f"[Adapted] {step_desc}", i, total_steps)
except Exception as adapt_exc: # broad catch intentional
logger.error("Agentic loop adaptation also failed: %s", adapt_exc)
result.steps.append(
AgenticStep(
step_num=i,
description=step_desc,
result=f"Failed: {exc}; Adaptation also failed: {adapt_exc}",
status="failed",
duration_ms=int((time.monotonic() - step_start) * 1000),
)
)
completed_results.append(f"Step {i}: FAILED")
# Phase 3: Summary
_summarize(result, total_steps, was_truncated)
result.total_duration_ms = int((time.monotonic() - start_time) * 1000)
await _broadcast_progress(
"agentic.task_complete",
{
"task_id": task_id,
"status": result.status,
"steps_completed": len(result.steps),
"summary": result.summary[:300],
"duration_ms": result.total_duration_ms,
},
)
return result
# ---------------------------------------------------------------------------
# WebSocket broadcast helper
# ---------------------------------------------------------------------------
async def _broadcast_progress(event: str, data: dict) -> None:
"""Broadcast agentic loop progress via WebSocket (best-effort)."""
try:
from infrastructure.ws_manager.handler import ws_manager
await ws_manager.broadcast(event, data)
except (ImportError, AttributeError, ConnectionError, RuntimeError) as exc:
logger.warning("Agentic loop broadcast failed: %s", exc)
logger.debug("Agentic loop: WS broadcast failed for %s", event)