feat(chat): add agent orchestration system with plan_and_execute
Introduces TaskPlanner and AgentSpec so Pyra can decompose multi-step tasks into sequential steps, each executed with a focused sub-agent context rather than the full conversation history. - plugins/base.py: AgentSpec dataclass + agent_spec() on Protocol/BasePlugin - plugins/registry.py: register_builtin, get_agent, list_agents - chat/planner.py: TaskPlanner with plan approval, per-step tool-use loop, verification call, and agent-aware routing - chat/session.py: wires plan_and_execute as a built-in tool after load_all - chat/history.py: planning hint in system prompt + dynamic agents listing Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
curo1305
@@ -0,0 +1,221 @@
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import TYPE_CHECKING, Any
|
||||
|
||||
import litellm
|
||||
from rich.panel import Panel
|
||||
|
||||
from pyra.chat.renderer import (
|
||||
console,
|
||||
render_error,
|
||||
render_info,
|
||||
render_streaming_response,
|
||||
render_text_response,
|
||||
)
|
||||
from pyra.setup.providers import get_provider
|
||||
from pyra.vault.reader import get_key
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from pyra.config.schema import PyraConfig
|
||||
from pyra.plugins.executor import ToolExecutor
|
||||
from pyra.plugins.registry import PluginRegistry
|
||||
|
||||
_STEP_SYSTEM_BASE = """\
|
||||
You are Pyra, executing one step of a multi-step plan.
|
||||
Security constraints:
|
||||
- You cannot access ~/.pyra/vault/ — it is physically blocked by the application.
|
||||
- You cannot execute shell commands — use the provided tools instead.
|
||||
- You cannot read or modify files outside ~/.pyra/memory/ directly.
|
||||
Work only on the assigned step. Use available tools if needed.
|
||||
Clearly describe what you accomplished when finished.
|
||||
"""
|
||||
|
||||
_VERIFY_SYSTEM = (
|
||||
"You evaluate task step outcomes. "
|
||||
"Reply only with the single word SUCCESS or FAILURE."
|
||||
)
|
||||
|
||||
|
||||
class TaskPlanner:
|
||||
def __init__(self, cfg: PyraConfig, registry: PluginRegistry, executor: ToolExecutor) -> None:
|
||||
self._cfg = cfg
|
||||
self._registry = registry
|
||||
self._executor = executor
|
||||
|
||||
def make_tool_handler(self):
|
||||
def handle(task: str, steps: list) -> str:
|
||||
return self._run_plan(task, steps)
|
||||
return handle
|
||||
|
||||
def _run_plan(self, task: str, steps: list) -> str:
|
||||
normalised = [
|
||||
s if isinstance(s, dict) else {"description": s}
|
||||
for s in steps
|
||||
]
|
||||
|
||||
if not self._ask_plan_approval(task, normalised):
|
||||
return "Plan declined by user."
|
||||
|
||||
previous_results: list[str] = []
|
||||
summaries: list[str] = []
|
||||
n = len(normalised)
|
||||
|
||||
for i, step in enumerate(normalised):
|
||||
desc = step.get("description", f"Step {i + 1}")
|
||||
agent_name = step.get("agent")
|
||||
label = f" [{agent_name}]" if agent_name else ""
|
||||
render_info(f"[Plan] Step {i + 1}/{n}{label}: {desc}")
|
||||
|
||||
try:
|
||||
output = self._execute_step(desc, agent_name, task, previous_results, n)
|
||||
except Exception as exc:
|
||||
render_error(f"[Plan] Step {i + 1} error: {exc}")
|
||||
return f"Plan failed at step {i + 1} ({desc}): {exc}"
|
||||
|
||||
if not self._verify_step(desc, output):
|
||||
render_error(f"[Plan] Step {i + 1} failed verification.")
|
||||
return (
|
||||
f"Plan failed at step {i + 1} ({desc}): "
|
||||
f"output did not pass verification.\n{output[:500]}"
|
||||
)
|
||||
|
||||
summary = output[:800].strip()
|
||||
previous_results.append(summary)
|
||||
summaries.append(f"Step {i + 1} ({desc}): {summary}")
|
||||
render_info(f"[Plan] Step {i + 1} ✓")
|
||||
|
||||
render_info("[Plan] All steps completed successfully.")
|
||||
body = "\n\n".join(summaries)
|
||||
result = f"Plan completed successfully.\n\n{body}"
|
||||
return result[:3900]
|
||||
|
||||
def _execute_step(
|
||||
self,
|
||||
desc: str,
|
||||
agent_name: str | None,
|
||||
task: str,
|
||||
previous_results: list[str],
|
||||
total: int,
|
||||
) -> str:
|
||||
step_num = len(previous_results) + 1
|
||||
agent_info = self._registry.get_agent(agent_name) if agent_name else None
|
||||
|
||||
if agent_info:
|
||||
agent_spec, agent_tools = agent_info
|
||||
system_prompt = agent_spec.system_prompt
|
||||
tools = agent_tools
|
||||
else:
|
||||
system_prompt = _STEP_SYSTEM_BASE
|
||||
tools = [t for t in self._registry.get_all_tools() if t.name != "plan_and_execute"]
|
||||
|
||||
messages: list[dict[str, Any]] = [
|
||||
{"role": "system", "content": system_prompt},
|
||||
{"role": "user", "content": self._step_user_msg(task, step_num, total, desc, previous_results)},
|
||||
]
|
||||
tools_spec = [
|
||||
{
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": t.name,
|
||||
"description": t.description,
|
||||
"parameters": t.parameters,
|
||||
},
|
||||
}
|
||||
for t in tools
|
||||
]
|
||||
base_kw = self._base_kwargs()
|
||||
litellm.suppress_debug_info = True
|
||||
|
||||
if not tools_spec:
|
||||
stream = litellm.completion(**base_kw, messages=messages, stream=True)
|
||||
return render_streaming_response(stream)
|
||||
|
||||
for _ in range(5):
|
||||
resp = litellm.completion(
|
||||
**base_kw,
|
||||
messages=messages,
|
||||
tools=tools_spec,
|
||||
tool_choice="auto",
|
||||
stream=False,
|
||||
)
|
||||
msg = resp.choices[0].message
|
||||
if not msg.tool_calls:
|
||||
text = msg.content or ""
|
||||
render_text_response(text)
|
||||
return text
|
||||
messages.append({
|
||||
"role": "assistant",
|
||||
"content": msg.content,
|
||||
"tool_calls": [
|
||||
{
|
||||
"id": tc.id,
|
||||
"type": "function",
|
||||
"function": {"name": tc.function.name, "arguments": tc.function.arguments},
|
||||
}
|
||||
for tc in msg.tool_calls
|
||||
],
|
||||
})
|
||||
results = self._executor.execute_tool_call_batch(msg.tool_calls)
|
||||
for r in results:
|
||||
messages.append({"role": "tool", "tool_call_id": r["tool_call_id"], "content": r["result"]})
|
||||
|
||||
return "Step exceeded maximum tool iterations."
|
||||
|
||||
def _verify_step(self, desc: str, output: str) -> bool:
|
||||
try:
|
||||
resp = litellm.completion(
|
||||
**self._base_kwargs(),
|
||||
messages=[
|
||||
{"role": "system", "content": _VERIFY_SYSTEM},
|
||||
{"role": "user", "content": f"Step: {desc}\n\nOutput:\n{output[:1000]}"},
|
||||
],
|
||||
stream=False,
|
||||
)
|
||||
text = (resp.choices[0].message.content or "").upper()
|
||||
return "SUCCESS" in text
|
||||
except Exception:
|
||||
return True
|
||||
|
||||
def _base_kwargs(self) -> dict:
|
||||
provider = get_provider(self._cfg.ai.provider_id)
|
||||
api_key = get_key(self._cfg.ai.provider_id) if provider.requires_key else "local"
|
||||
kw: dict = {
|
||||
"model": f"{provider.litellm_prefix}{self._cfg.ai.model}",
|
||||
"api_key": api_key,
|
||||
}
|
||||
if self._cfg.ai.base_url:
|
||||
kw["api_base"] = self._cfg.ai.base_url
|
||||
return kw
|
||||
|
||||
def _step_user_msg(
|
||||
self,
|
||||
task: str,
|
||||
step_num: int,
|
||||
total: int,
|
||||
desc: str,
|
||||
previous_results: list[str],
|
||||
) -> str:
|
||||
lines = [f"Overall task: {task}", "", f"Step {step_num}/{total}: {desc}"]
|
||||
if previous_results:
|
||||
lines += ["", "Results from previous steps:"]
|
||||
for i, r in enumerate(previous_results, 1):
|
||||
lines.append(f" Step {i}: {r}")
|
||||
return "\n".join(lines)
|
||||
|
||||
def _ask_plan_approval(self, task: str, steps: list[dict]) -> bool:
|
||||
lines = [f"[bold]Task:[/bold] {task}", "", "[bold]Steps:[/bold]"]
|
||||
for i, step in enumerate(steps, 1):
|
||||
desc = step.get("description", "")
|
||||
agent = step.get("agent", "")
|
||||
suffix = f" [dim][{agent}][/dim]" if agent else ""
|
||||
lines.append(f" {i}. {desc}{suffix}")
|
||||
console.print(Panel(
|
||||
"\n".join(lines),
|
||||
title="[bold cyan]Pyra — Multi-Step Plan[/bold cyan]",
|
||||
border_style="cyan",
|
||||
))
|
||||
try:
|
||||
answer = console.input("[bold]Execute this plan?[/bold] [dim][y/N][/dim] ").strip().lower()
|
||||
except (KeyboardInterrupt, EOFError):
|
||||
return False
|
||||
return answer == "y"
|
||||
Reference in New Issue
Block a user