feat: add the hierarchical Execution engine on LangGraph

Increment 4 of .scratch/solution-architect-agent/spec.md, test-first.

- app/agents/runtime.py: framework-neutral AgentRuntime interface
  (RoleSpec, AgentTask, ToolSpec, outcomes); fake_runtime.py replays
  scripted tool calls through the real tools with step budgets.
- app/execution/graph.py: Orchestrator turn fans out to parallel
  Specialist tasks (Send), pauses with interrupt() for questions,
  Approvals and cloud-read permission, and finalizes or ends with
  warnings. Only the Orchestrator turn writes the Ledger.
- app/execution/tools.py: Orchestrator tools, Specialist submit and
  needs_input, and the cloud-read permission gate.
- app/execution/engine.py: ExecutionEngine.start/resume returning an
  Interaction; the single interface for A2A, /generate and eval.
- Recorders (in-memory, DatabaseManager), configurable checkpointer
  (memory, sqlite, postgres), engine settings, langgraph checkpoint
  dependencies.
- 19 new tests; full suite 121 passed.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
2026-09-16 09:38:13 +01:00
parent c8410946b8
commit d9214c671a
15 changed files with 2143 additions and 1 deletions

1
app/agents/__init__.py Normal file
View File

@@ -0,0 +1 @@
"""Agent runtimes: the framework-neutral interface agents run behind, and its implementations."""

View File

@@ -0,0 +1,89 @@
"""A scripted AgentRuntime for deterministic tests and evaluation dry runs.
Each role has a queue of turns; a turn is a list of tool calls, or a function of the
task that returns one. Calls go through the real ToolSpecs, so every rule the tools
enforce is exercised exactly as a model would meet it.
"""
from __future__ import annotations
from collections import deque
from collections.abc import Callable, Mapping, Sequence
from dataclasses import dataclass, field
from typing import Any
from app.agents.runtime import (
AgentTask,
BudgetExhausted,
Outcome,
RoleSpec,
Stopped,
ToolSpec,
)
@dataclass(frozen=True)
class Call:
tool: str
args: Mapping[str, Any] = field(default_factory=dict)
Turn = Sequence[Call] | Callable[[AgentTask], Sequence[Call]]
@dataclass(frozen=True)
class RunRecord:
role: str
task: AgentTask
tools: tuple[str, ...]
results: tuple[tuple[str, str], ...] # (tool name, message)
outcome: Outcome
class ScriptExhausted(AssertionError):
"""A role was run more times than the test scripted."""
class FakeAgentRuntime:
def __init__(self, turns: Mapping[str, Sequence[Turn]]) -> None:
self._turns = {role: deque(role_turns) for role, role_turns in turns.items()}
self.runs: list[RunRecord] = []
def remaining(self, role: str) -> int:
return len(self._turns.get(role, ()))
async def run(
self, role: RoleSpec, task: AgentTask, tools: Sequence[ToolSpec]
) -> Outcome:
queue = self._turns.get(role.name)
if not queue:
raise ScriptExhausted(f"No scripted turn left for role {role.name!r}")
turn = queue.popleft()
calls = turn(task) if callable(turn) else turn
by_name = {tool.name: tool for tool in tools}
results: list[tuple[str, str]] = []
outcome: Outcome = Stopped(
"The script ended without a tool that ends the turn."
)
for index, call in enumerate(calls):
if index >= role.budget:
outcome = BudgetExhausted(
f"{role.name} used all {role.budget} tool calls before finishing."
)
break
if call.tool not in by_name:
raise ValueError(
f"{role.name} called {call.tool!r}, which it was not offered: "
f"{sorted(by_name)}"
)
result = await by_name[call.tool].handler(**call.args)
results.append((call.tool, result.message))
if result.outcome is not None:
outcome = result.outcome
break
self.runs.append(
RunRecord(role.name, task, tuple(by_name), tuple(results), outcome)
)
return outcome

103
app/agents/runtime.py Normal file
View File

@@ -0,0 +1,103 @@
"""The agent-runtime interface (ADR-0003).
Everything an Orchestrator or Specialist needs crosses this seam as plain data:
a RoleSpec, an AgentTask and ToolSpecs in; an Outcome back. No framework types.
"""
from __future__ import annotations
from collections.abc import Awaitable, Callable, Mapping, Sequence
from dataclasses import dataclass, field
from typing import Any, Protocol
@dataclass(frozen=True)
class RoleSpec:
"""One agent role, parsed from agents/<role>.md."""
name: str
instructions: str
budget: int
skills: tuple[str, ...] = ()
tool_tiers: tuple[str, ...] = ()
mcp_servers: tuple[str, ...] = ()
model: str | None = None
@dataclass(frozen=True)
class AgentTask:
"""What one run of an agent is asked to do, with the context it may use."""
role: str
prompt: str
context: Mapping[str, Any] = field(default_factory=dict)
deliverable: str | None = None
# --- outcomes ----------------------------------------------------------------------
@dataclass(frozen=True)
class Submitted:
"""A Specialist's contract-complete Deliverable."""
content: str
attachments: Mapping[str, str] = field(default_factory=dict)
@dataclass(frozen=True)
class NeedsInput:
"""The agent can't continue without the user. `permission` names a tool tier to grant."""
questions: tuple[str, ...]
permission: str | None = None
@dataclass(frozen=True)
class BudgetExhausted:
summary: str
@dataclass(frozen=True)
class Stopped:
"""The agent stopped without calling a tool that ends its turn."""
message: str
@dataclass(frozen=True)
class TurnEnded:
"""The Orchestrator ended its turn by calling `tool`; the tool recorded its effect."""
tool: str
Outcome = Submitted | NeedsInput | BudgetExhausted | Stopped | TurnEnded
# --- tools -------------------------------------------------------------------------
@dataclass(frozen=True)
class ToolResult:
"""What a tool returns to the agent. A non-None `outcome` ends the agent's run."""
message: str
outcome: Outcome | None = None
@dataclass(frozen=True)
class ToolSpec:
name: str
description: str
parameters: Mapping[str, Any] # JSON schema for the keyword arguments
handler: Callable[..., Awaitable[ToolResult]]
tier: str = "general" # general | cloud_read | execution
class AgentRuntime(Protocol):
async def run(
self, role: RoleSpec, task: AgentTask, tools: Sequence[ToolSpec]
) -> Outcome:
"""Run one ReAct loop until a tool ends the turn, the budget is spent, or it stops."""
...

View File

@@ -53,6 +53,14 @@ class Settings(BaseSettings):
CLOUD_PROVIDER: str = "gcp"
MCP_ENABLE_PREVIEW: bool = True
# Hierarchical Execution Engine (LangGraph state, ReAct agents)
ORCHESTRATOR_BUDGET: int = 40
SPECIALIST_BUDGET: int = 15
MAX_REVISIONS: int = 3
MAX_ORCHESTRATOR_TURNS: int = 100
CHECKPOINTER: str = "sqlite" # memory | sqlite | postgres
CHECKPOINT_DB_PATH: str = "checkpoints.db"
# Google Developer Knowledge MCP Server Settings
DEVELOPER_KNOWLEDGE_MCP_URL: str = "https://developerknowledge.googleapis.com/mcp"
DEVELOPER_KNOWLEDGE_MCP_ENABLED: bool = True

View File

@@ -0,0 +1,38 @@
"""The configured LangGraph checkpointer that makes paused Executions resumable."""
from __future__ import annotations
from collections.abc import AsyncIterator
from contextlib import asynccontextmanager
from langgraph.checkpoint.base import BaseCheckpointSaver
from langgraph.checkpoint.memory import InMemorySaver
from app.config import Settings
@asynccontextmanager
async def open_checkpointer(settings: Settings) -> AsyncIterator[BaseCheckpointSaver]:
"""Yield a checkpointer per CHECKPOINTER: memory | sqlite | postgres."""
kind = settings.CHECKPOINTER
if kind == "memory":
yield InMemorySaver()
elif kind == "sqlite":
from langgraph.checkpoint.sqlite.aio import AsyncSqliteSaver
async with AsyncSqliteSaver.from_conn_string(
settings.CHECKPOINT_DB_PATH
) as saver:
yield saver
elif kind == "postgres":
from langgraph.checkpoint.postgres.aio import AsyncPostgresSaver
async with AsyncPostgresSaver.from_conn_string(
settings.get_database_url()
) as saver:
await saver.setup()
yield saver
else:
raise ValueError(
f"CHECKPOINTER must be memory, sqlite or postgres, not {kind!r}"
)

145
app/execution/engine.py Normal file
View File

@@ -0,0 +1,145 @@
"""ExecutionEngine: start and resume Executions.
The one interface the A2A server, /generate and the evaluation harness use. Each
call runs the Execution until it needs the user or finishes, and returns an
Interaction describing which.
"""
from __future__ import annotations
import uuid
from collections.abc import Mapping
from dataclasses import dataclass
from pathlib import Path
from typing import Any
from langchain_core.runnables import RunnableConfig
from langgraph.checkpoint.base import BaseCheckpointSaver
from langgraph.types import Command
from app.agents.runtime import AgentRuntime, RoleSpec
from app.execution.deliverables import DEPENDENCY_GRAPH
from app.execution.graph import ToolFactory, build_execution_graph
from app.execution.ledger import Ledger
from app.execution.recorder import ExecutionRecorder
from app.execution.tools import ORCHESTRATOR
_RECURSION_LIMIT = 10_000 # loops are bounded by max_orchestrator_turns instead
@dataclass(frozen=True)
class Interaction:
execution_id: str
status: str # awaiting_user | completed | completed_with_warnings
prompt: Mapping[str, Any] | None # questions | approval | permission
state: Mapping[str, Any]
class ExecutionNotAwaitingUser(RuntimeError):
pass
class ExecutionEngine:
def __init__(
self,
*,
runtime: AgentRuntime,
roles: Mapping[str, RoleSpec],
recorder: ExecutionRecorder,
checkpointer: BaseCheckpointSaver,
output_root: Path,
max_revisions: int = 3,
max_orchestrator_turns: int = 100,
tool_factory: ToolFactory | None = None,
) -> None:
required = {
ORCHESTRATOR,
*(spec.specialist for spec in DEPENDENCY_GRAPH.values()),
}
missing = sorted(required - set(roles))
if missing:
raise ValueError(f"Missing roles: {', '.join(missing)}")
self._max_revisions = max_revisions
self._graph = build_execution_graph(
runtime=runtime,
roles=roles,
recorder=recorder,
output_root=output_root,
max_orchestrator_turns=max_orchestrator_turns,
tool_factory=tool_factory,
).compile(checkpointer=checkpointer)
async def start(
self,
workflow_request: str,
*,
session_id: str | None = None,
pre_approved: bool = False,
allow_cloud_read: bool | None = None,
) -> Interaction:
execution_id = str(uuid.uuid4())
if pre_approved and allow_cloud_read is None:
allow_cloud_read = False
initial: dict[str, Any] = {
"execution_id": execution_id,
"session_id": session_id or str(uuid.uuid4()),
"workflow_request": workflow_request,
"pre_approved": pre_approved,
"allow_cloud_read": allow_cloud_read,
"ledger": Ledger(
pre_approved=pre_approved, max_revisions=self._max_revisions
).to_dict(),
"deliverables": {},
"conversation": [{"from": "user", "text": workflow_request}],
"notes": [],
"pending_questions": [],
"awaiting_questions": [],
"answers": {},
"pending_permission": None,
"prompt": None,
"outcomes": [],
"outcomes_applied": 0,
"events_recorded": 0,
"turns": 0,
"status": "running",
"warnings": [],
}
await self._graph.ainvoke(initial, self._config(execution_id))
return await self.interaction(execution_id)
async def resume(self, execution_id: str, reply: Mapping[str, Any]) -> Interaction:
"""Answer the Execution's current prompt: {"reply"} | {"approval"} | {"permission"}."""
current = await self.interaction(execution_id)
if current.status != "awaiting_user":
raise ExecutionNotAwaitingUser(
f"Execution {execution_id} is {current.status}, not awaiting the user"
)
await self._graph.ainvoke(
Command(resume=dict(reply)), self._config(execution_id)
)
return await self.interaction(execution_id)
async def interaction(self, execution_id: str) -> Interaction:
snapshot = await self._graph.aget_state(self._config(execution_id))
if not snapshot.values:
raise KeyError(f"Unknown Execution {execution_id}")
if snapshot.interrupts:
return Interaction(
execution_id,
"awaiting_user",
snapshot.interrupts[0].value,
snapshot.values,
)
return Interaction(
execution_id,
snapshot.values.get("status", "running"),
None,
snapshot.values,
)
@staticmethod
def _config(execution_id: str) -> RunnableConfig:
return {
"configurable": {"thread_id": execution_id},
"recursion_limit": _RECURSION_LIMIT,
}

478
app/execution/graph.py Normal file
View File

@@ -0,0 +1,478 @@
"""The LangGraph that holds an Execution (ADR-0003).
START -> orchestrator_turn
orchestrator_turn -> specialist_task x N (Send, parallel) -> orchestrator_turn
orchestrator_turn -> await_user (interrupt) -> orchestrator_turn
orchestrator_turn -> finalize -> END
orchestrator_turn -> end_with_warnings -> END
Only orchestrator_turn and await_user change the Ledger; parallel Specialist tasks
append outcomes that the next Orchestrator turn applies in order.
"""
from __future__ import annotations
from collections.abc import Callable, Mapping, Sequence
from pathlib import Path
from typing import Any, cast
from langgraph.graph import END, START, StateGraph
from langgraph.types import Command, Send, interrupt
from app.agents.runtime import (
AgentRuntime,
AgentTask,
BudgetExhausted,
NeedsInput,
Outcome,
RoleSpec,
Stopped,
Submitted,
ToolSpec,
TurnEnded,
)
from app.execution.deliverables import DEPENDENCY_GRAPH, PHASE_ORDER, DeliverableKind
from app.execution.ledger import Ledger, LedgerRefusal
from app.execution.recorder import ExecutionRecorder
from app.execution.state import ExecutionState
from app.execution.tools import ORCHESTRATOR, OrchestratorToolbox, specialist_tools
ToolFactory = Callable[[RoleSpec, str | None], Sequence[ToolSpec]]
NO_USER_REPLY = (
"No user is available. State your assumption, record it in the Deliverable under "
"'Assumptions', and proceed."
)
GUIDE_FILENAME = "solution-architecture-guide.md"
def _no_tools(role: RoleSpec, deliverable: str | None) -> Sequence[ToolSpec]:
return ()
def build_execution_graph(
*,
runtime: AgentRuntime,
roles: Mapping[str, RoleSpec],
recorder: ExecutionRecorder,
output_root: Path,
max_orchestrator_turns: int = 100,
tool_factory: ToolFactory | None = None,
) -> StateGraph:
tool_factory = tool_factory or _no_tools
async def record_progress(
state: ExecutionState, ledger: Ledger, status: str | None = None
) -> int:
recorded = state.get("events_recorded", 0)
await recorder.record_events(state["execution_id"], ledger.events[recorded:])
if status:
await recorder.record_status(
state["execution_id"],
session_id=state["session_id"],
workflow_request=state["workflow_request"],
status=status,
current_phase=_current_phase(ledger),
details={
"revisions": sum(
ledger.revisions(kind) for kind in DEPENDENCY_GRAPH
),
"warnings": state.get("warnings", []),
},
)
return len(ledger.events)
async def orchestrator_turn(state: ExecutionState) -> Command:
ledger = Ledger.from_dict(state["ledger"])
deliverables = dict(state.get("deliverables", {}))
pending_questions = list(state.get("pending_questions", []))
notes = list(state.get("notes", []))
pending_permission = state.get("pending_permission")
warnings = list(state.get("warnings", []))
outcomes = state.get("outcomes", [])
for record in outcomes[state.get("outcomes_applied", 0) :]:
kind = DeliverableKind(record["kind"])
try:
if record["type"] == "submitted":
ledger.submit(kind, record["attempt"], actor=record["role"])
deliverables[kind.value] = {
"content": record["content"],
"attachments": record["attachments"],
}
await recorder.record_submission(
state["execution_id"],
kind.value,
record["content"],
record["attachments"],
)
continue
ledger.release(kind, record["attempt"], reason=record["type"])
except LedgerRefusal as exc:
notes.append(f"Discarded a superseded result for {kind}: {exc}")
continue
if record["type"] == "needs_input" and record.get("permission"):
if state.get("allow_cloud_read") is None:
requested = (pending_permission or {}).get("requested_by", [])
pending_permission = {
"tier": record["permission"],
"requested_by": [*requested, kind.value],
}
continue
questions = record.get("questions") or [
(
f"The {record['role']} stopped: {record['summary']} "
"How should it proceed?"
)
]
pending_questions.append({"kind": kind.value, "questions": questions})
turns = state.get("turns", 0) + 1
update: dict[str, Any] = {
"deliverables": deliverables,
"pending_questions": pending_questions,
"pending_permission": pending_permission,
"outcomes_applied": len(outcomes),
"turns": turns,
}
def finish_with(goto: str, **extra: Any) -> Command:
return Command(
update={**update, "ledger": ledger.to_dict(), "notes": notes, **extra},
goto=goto,
)
if pending_permission and not state.get("pre_approved"):
update["events_recorded"] = await record_progress(
state, ledger, "awaiting_user"
)
return finish_with("await_user", prompt={"permission": pending_permission})
if turns > max_orchestrator_turns:
warnings.append(
f"Stopped after {max_orchestrator_turns} Orchestrator turns without "
"finishing."
)
return finish_with("end_with_warnings", warnings=warnings)
toolbox = OrchestratorToolbox(ledger, deliverables, pending_questions)
task = AgentTask(
role=ORCHESTRATOR,
prompt="Decide the next step for this Execution using your tools.",
context=_orchestrator_context(state, ledger, pending_questions, notes),
)
outcome = await runtime.run(roles[ORCHESTRATOR], task, toolbox.tools())
notes = []
pre_approved = bool(state.get("pre_approved"))
conversation = list(state.get("conversation", []))
def ask(prompt: dict[str, Any], **extra: Any) -> Command:
conversation.append({"from": ORCHESTRATOR, "text": _render_prompt(prompt)})
return finish_with(
"await_user", prompt=prompt, conversation=conversation, **extra
)
if isinstance(outcome, TurnEnded) and outcome.tool == "delegate":
sends = [
Send(
"specialist_task",
_specialist_input(state, ledger, deliverables, kind, attempt),
)
for kind, attempt in toolbox.started.items()
]
update["events_recorded"] = await record_progress(state, ledger)
return Command(
update={**update, "ledger": ledger.to_dict(), "notes": notes},
goto=sends,
)
if isinstance(outcome, TurnEnded) and outcome.tool == "finish":
update["events_recorded"] = await record_progress(state, ledger)
return finish_with("finalize")
if isinstance(outcome, TurnEnded) and outcome.tool == "escalate":
warnings.append(toolbox.escalation or "Escalated without detail.")
update["events_recorded"] = await record_progress(state, ledger)
return finish_with("end_with_warnings", warnings=warnings)
if isinstance(outcome, TurnEnded) and outcome.tool == "request_approval":
update["events_recorded"] = await record_progress(
state, ledger, None if pre_approved else "awaiting_user"
)
return ask({"approval": toolbox.approval})
if isinstance(outcome, TurnEnded) and outcome.tool == "ask_user":
update["events_recorded"] = await record_progress(
state, ledger, None if pre_approved else "awaiting_user"
)
return ask(
{"questions": toolbox.questions},
pending_questions=[],
awaiting_questions=pending_questions,
)
message = _describe_unfinished(outcome)
if pre_approved:
warnings.append(f"The Orchestrator could not continue: {message}")
update["events_recorded"] = await record_progress(state, ledger)
return finish_with("end_with_warnings", warnings=warnings)
update["events_recorded"] = await record_progress(
state, ledger, "awaiting_user"
)
return ask({"questions": [{"from": ORCHESTRATOR, "question": message}]})
async def specialist_task(task_input: dict[str, Any]) -> dict[str, Any]:
role = roles[task_input["role"]]
kind = DeliverableKind(task_input["kind"])
tools = specialist_tools(
kind,
extra=tool_factory(role, kind.value),
allow_cloud_read=task_input["allow_cloud_read"],
)
task = AgentTask(
role=role.name,
prompt=f"Produce the {kind} Deliverable.",
context=task_input["context"],
deliverable=kind.value,
)
outcome = await runtime.run(role, task, tools)
return {
"outcomes": [
_outcome_record(role.name, kind, task_input["attempt"], outcome)
]
}
async def await_user(state: ExecutionState) -> Command:
prompt = state["prompt"] or {}
if state.get("pre_approved"):
reply = _automatic_reply(prompt)
else:
reply = interrupt(prompt)
ledger = Ledger.from_dict(state["ledger"])
conversation = list(state.get("conversation", []))
update: dict[str, Any] = {"prompt": None}
if "approval" in prompt:
phase = prompt["approval"]["phase"]
decision = reply.get("approval") if isinstance(reply, Mapping) else None
if (
isinstance(decision, Mapping)
and decision.get("phase") == phase
and decision.get("decision") in {"approve", "changes"}
):
phase_id = next(p for p in PHASE_ORDER if p.value == phase)
if decision["decision"] == "approve":
ledger.approve(phase_id, actor="user")
text = f"Approved {phase}."
else:
feedback = str(decision.get("feedback", ""))
ledger.request_changes(phase_id, feedback, actor="user")
text = f"Requested changes to {phase}: {feedback}"
conversation.append({"from": "user", "text": text})
else:
conversation.append({"from": "user", "text": _reply_text(reply)})
elif "permission" in prompt:
tier = prompt["permission"]["tier"]
permission = reply.get("permission") if isinstance(reply, Mapping) else None
grant = permission.get(tier) if isinstance(permission, Mapping) else None
if isinstance(grant, bool):
update["allow_cloud_read"] = grant
update["pending_permission"] = None
verdict = "granted" if grant else "denied"
conversation.append(
{
"from": "user",
"text": f"Permission to use {tier} tools {verdict}.",
}
)
else:
conversation.append({"from": "user", "text": _reply_text(reply)})
else:
text = _reply_text(reply)
conversation.append({"from": "user", "text": text})
answers = {
kind: list(items) for kind, items in state.get("answers", {}).items()
}
for asked in state.get("awaiting_questions", []):
answers.setdefault(asked["kind"], []).append(
{"questions": list(asked["questions"]), "reply": text}
)
update.update(answers=answers, awaiting_questions=[])
update["events_recorded"] = await record_progress(state, ledger)
update.update(ledger=ledger.to_dict(), conversation=conversation)
return Command(update=update, goto="orchestrator_turn")
async def finalize(state: ExecutionState) -> dict[str, Any]:
ledger = Ledger.from_dict(state["ledger"])
_write_outputs(
output_root / "deliverables" / "executions" / state["execution_id"],
state.get("deliverables", {}),
)
await record_progress(state, ledger, "completed")
return {"status": "completed"}
async def end_with_warnings(state: ExecutionState) -> dict[str, Any]:
ledger = Ledger.from_dict(state["ledger"])
await record_progress(state, ledger, "completed_with_warnings")
return {"status": "completed_with_warnings"}
graph = StateGraph(ExecutionState)
graph.add_node("orchestrator_turn", orchestrator_turn)
# Specialist tasks receive their Send payload, not the Execution state.
graph.add_node("specialist_task", cast(Any, specialist_task))
graph.add_node("await_user", await_user)
graph.add_node("finalize", finalize)
graph.add_node("end_with_warnings", end_with_warnings)
graph.add_edge(START, "orchestrator_turn")
graph.add_edge("specialist_task", "orchestrator_turn")
graph.add_edge("finalize", END)
graph.add_edge("end_with_warnings", END)
return graph
# --- helpers ---------------------------------------------------------------------------
def _current_phase(ledger: Ledger) -> str:
for phase in PHASE_ORDER:
if ledger.phase_status(phase).value != "approved":
return phase.value
return PHASE_ORDER[-1].value
def _orchestrator_context(
state: ExecutionState,
ledger: Ledger,
pending_questions: Sequence[Mapping[str, Any]],
notes: Sequence[str],
) -> dict[str, Any]:
return {
"workflow_request": state["workflow_request"],
"pre_approved": bool(state.get("pre_approved")),
"allow_cloud_read": state.get("allow_cloud_read"),
"deliverables": {
kind.value: {
"phase": spec.phase.value,
"status": ledger.status(kind).value,
"revisions": ledger.revisions(kind),
"feedback": list(ledger.feedback(kind)),
}
for kind, spec in DEPENDENCY_GRAPH.items()
},
"phases": {
phase.value: ledger.phase_status(phase).value for phase in PHASE_ORDER
},
"ready_work": [kind.value for kind in ledger.ready_work()],
"phases_ready_for_approval": [
p.value for p in ledger.phases_ready_for_approval()
],
"pending_questions": list(pending_questions),
"conversation": list(state.get("conversation", []))[-20:],
"notes": list(notes),
}
def _specialist_input(
state: ExecutionState,
ledger: Ledger,
deliverables: Mapping[str, Mapping[str, Any]],
kind: DeliverableKind,
attempt: int,
) -> dict[str, Any]:
spec = DEPENDENCY_GRAPH[kind]
return {
"role": spec.specialist,
"kind": kind.value,
"attempt": attempt,
"allow_cloud_read": state.get("allow_cloud_read"),
"context": {
"workflow_request": state["workflow_request"],
"pre_approved": bool(state.get("pre_approved")),
"dependencies": {
dependency.value: dict(deliverables[dependency.value])
for dependency in spec.depends_on
},
"feedback": list(ledger.feedback(kind)),
"answers": list(state.get("answers", {}).get(kind.value, [])),
"attempt": attempt,
},
}
def _outcome_record(
role: str, kind: DeliverableKind, attempt: int, outcome: Outcome
) -> dict[str, Any]:
base = {"role": role, "kind": kind.value, "attempt": attempt}
if isinstance(outcome, Submitted):
return {
**base,
"type": "submitted",
"content": outcome.content,
"attachments": dict(outcome.attachments),
}
if isinstance(outcome, NeedsInput):
return {
**base,
"type": "needs_input",
"questions": list(outcome.questions),
"permission": outcome.permission,
}
if isinstance(outcome, BudgetExhausted):
return {**base, "type": "budget_exhausted", "summary": outcome.summary}
return {**base, "type": "stopped", "summary": _describe_unfinished(outcome)}
def _describe_unfinished(outcome: Outcome) -> str:
if isinstance(outcome, BudgetExhausted):
return f"I paused: {outcome.summary} How should I proceed?"
if isinstance(outcome, Stopped):
return outcome.message
if isinstance(outcome, NeedsInput):
return " ".join(outcome.questions)
return f"Ended with an unexpected outcome: {type(outcome).__name__}."
def _automatic_reply(prompt: Mapping[str, Any]) -> dict[str, Any]:
if "approval" in prompt:
return {
"approval": {"phase": prompt["approval"]["phase"], "decision": "approve"}
}
if "permission" in prompt:
return {"permission": {prompt["permission"]["tier"]: False}}
return {"reply": NO_USER_REPLY}
def _reply_text(reply: object) -> str:
if isinstance(reply, str):
return reply
if isinstance(reply, Mapping) and isinstance(reply.get("reply"), str):
return reply["reply"]
return str(reply)
def _render_prompt(prompt: Mapping[str, Any]) -> str:
if "approval" in prompt:
approval = prompt["approval"]
return f"Please approve {approval['phase']}: {approval['summary']}"
if "permission" in prompt:
return f"May I use {prompt['permission']['tier']} tools?"
return "\n".join(item["question"] for item in prompt.get("questions", []))
def _write_outputs(
out_dir: Path, deliverables: Mapping[str, Mapping[str, Any]]
) -> None:
out_dir.mkdir(parents=True, exist_ok=True)
guide = deliverables.get(DeliverableKind.SOLUTION_GUIDE.value)
if guide:
(out_dir / GUIDE_FILENAME).write_text(guide["content"], encoding="utf-8")
for kind, deliverable in deliverables.items():
path = out_dir / "deliverables" / f"{kind}.md"
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(deliverable["content"], encoding="utf-8")
for name, content in deliverable.get("attachments", {}).items():
attachment = out_dir / name
attachment.parent.mkdir(parents=True, exist_ok=True)
attachment.write_text(content, encoding="utf-8")

156
app/execution/recorder.py Normal file
View File

@@ -0,0 +1,156 @@
"""Recording an Execution's progress for audit and reporting.
Two adapters sit behind ExecutionRecorder: InMemoryRecorder for tests, and
DatabaseRecorder, which reuses the existing DatabaseManager tables until the
Execution store (review candidate 03) replaces them.
"""
from __future__ import annotations
import asyncio
import uuid
from collections import defaultdict
from collections.abc import Mapping, Sequence
from typing import Any, Protocol
from app.execution.ledger import LedgerEvent
SESSION_AGENT_NAME = "solution-architect"
class ExecutionRecorder(Protocol):
async def record_events(
self, execution_id: str, events: Sequence[LedgerEvent]
) -> None: ...
async def record_submission(
self,
execution_id: str,
kind: str,
content: str,
attachments: Mapping[str, str],
) -> None: ...
async def record_status(
self,
execution_id: str,
*,
session_id: str,
workflow_request: str,
status: str,
current_phase: str,
details: Mapping[str, Any],
) -> None: ...
class InMemoryRecorder:
def __init__(self) -> None:
self.events: dict[str, list[LedgerEvent]] = defaultdict(list)
self.submissions: dict[str, list[dict[str, Any]]] = defaultdict(list)
self.statuses: dict[str, list[dict[str, Any]]] = defaultdict(list)
async def record_events(
self, execution_id: str, events: Sequence[LedgerEvent]
) -> None:
self.events[execution_id].extend(events)
async def record_submission(
self,
execution_id: str,
kind: str,
content: str,
attachments: Mapping[str, str],
) -> None:
self.submissions[execution_id].append(
{"kind": kind, "content": content, "attachments": dict(attachments)}
)
async def record_status(
self,
execution_id: str,
*,
session_id: str,
workflow_request: str,
status: str,
current_phase: str,
details: Mapping[str, Any],
) -> None:
self.statuses[execution_id].append(
{
"session_id": session_id,
"workflow_request": workflow_request,
"status": status,
"current_phase": current_phase,
"details": dict(details),
}
)
class DatabaseRecorder:
"""Writes to DatabaseManager (blocking) off the event loop."""
def __init__(self, db_manager: Any = None) -> None:
if db_manager is None:
from app.database import get_db_manager
db_manager = get_db_manager()
self._db = db_manager
async def record_events(
self, execution_id: str, events: Sequence[LedgerEvent]
) -> None:
for event in events:
await asyncio.to_thread(
self._db.record_orchestrator_log,
log_id=str(uuid.uuid4()),
execution_id=execution_id,
iteration=event.sequence,
review_status=event.type,
reviewer_agent=event.actor,
feedback=f"{event.subject}: {event.detail}".rstrip(": "),
)
async def record_submission(
self,
execution_id: str,
kind: str,
content: str,
attachments: Mapping[str, str],
) -> None:
files = {f"deliverables/{kind}.md": content, **attachments}
for file_path, file_content in files.items():
await asyncio.to_thread(
self._db.save_artifact,
artifact_id=str(uuid.uuid4()),
execution_id=execution_id,
artifact_type=kind,
file_path=file_path,
content=file_content,
)
async def record_status(
self,
execution_id: str,
*,
session_id: str,
workflow_request: str,
status: str,
current_phase: str,
details: Mapping[str, Any],
) -> None:
await asyncio.to_thread(
self._db.save_workflow_execution,
execution_id=execution_id,
session_id=session_id,
status=status,
current_phase=current_phase,
loop_count=int(details.get("revisions", 0)),
request_summary=workflow_request,
results=dict(details),
)
await asyncio.to_thread(
self._db.save_session,
session_id=session_id,
agent_name=SESSION_AGENT_NAME,
state_data={"latest_execution_id": execution_id, "status": status},
)

35
app/execution/state.py Normal file
View File

@@ -0,0 +1,35 @@
"""The checkpointed state of one Execution. Plain, JSON-compatible data only."""
from __future__ import annotations
import operator
from typing import Annotated, Any, TypedDict
class ExecutionState(TypedDict, total=False):
execution_id: str
session_id: str
workflow_request: str
pre_approved: bool
allow_cloud_read: bool | None
ledger: dict[str, Any] # Ledger.to_dict()
deliverables: dict[str, dict[str, Any]] # kind -> {content, attachments}
conversation: list[dict[str, str]] # {"from": "user" | "orchestrator", "text"}
notes: list[str] # facts for the Orchestrator's next turn
pending_questions: list[dict[str, Any]] # {kind, questions} not yet asked
awaiting_questions: list[dict[str, Any]] # {kind, questions} asked, not answered
answers: dict[str, list[dict[str, Any]]] # kind -> [{questions, reply}]
pending_permission: dict[str, Any] | None # {tier, requested_by}
prompt: dict[str, Any] | None # what the user is being asked
# Specialist outcomes, appended concurrently by parallel Specialist tasks and
# applied to the Ledger by the next Orchestrator turn.
outcomes: Annotated[list[dict[str, Any]], operator.add]
outcomes_applied: int
events_recorded: int
turns: int
status: str # running | completed | completed_with_warnings
warnings: list[str]

318
app/execution/tools.py Normal file
View File

@@ -0,0 +1,318 @@
"""Execution tools: how agents act on an Execution.
Tools enforce the Ledger's rules at the moment an agent acts, and return refusals as
messages so a ReAct agent can recover. A tool ends the agent's turn by returning an
outcome.
"""
from __future__ import annotations
from collections.abc import Mapping, Sequence
from pathlib import PurePosixPath
from typing import Any
from app.agents.runtime import (
NeedsInput,
Submitted,
ToolResult,
ToolSpec,
TurnEnded,
)
from app.execution.deliverables import (
DEPENDENCY_GRAPH,
DeliverableKind,
PhaseId,
check_submission,
deliverables_in,
)
from app.execution.ledger import Ledger, LedgerRefusal, PhaseStatus
ORCHESTRATOR = "orchestrator"
CLOUD_READ = "cloud_read"
_NO_ARGS: Mapping[str, Any] = {"type": "object", "properties": {}}
def _schema(properties: Mapping[str, Any], required: Sequence[str]) -> dict[str, Any]:
return {
"type": "object",
"properties": dict(properties),
"required": list(required),
}
def _refused(reason: object) -> ToolResult:
return ToolResult(f"Refused: {reason}")
class OrchestratorToolbox:
"""The Orchestrator's tools, bound to a working copy of the Ledger for one turn."""
def __init__(
self,
ledger: Ledger,
deliverables: Mapping[str, Mapping[str, Any]],
pending_questions: Sequence[Mapping[str, Any]],
) -> None:
self.ledger = ledger
self._deliverables = deliverables
self._pending_questions = pending_questions
self.started: dict[DeliverableKind, int] = {}
self.questions: list[dict[str, str]] = []
self.approval: dict[str, Any] | None = None
self.escalation: str | None = None
def tools(self) -> list[ToolSpec]:
return [
ToolSpec(
"ready_work",
"List Deliverables ready to delegate and Phases ready for Approval.",
_NO_ARGS,
self.ready_work,
tier="execution",
),
ToolSpec(
"delegate",
"Start Specialists on ready Deliverables, in parallel. Ends your turn.",
_schema(
{"kinds": {"type": "array", "items": {"type": "string"}}},
["kinds"],
),
self.delegate,
tier="execution",
),
ToolSpec(
"ask_user",
"Ask the user your questions together with any waiting Specialist "
"questions. Ends your turn.",
_schema(
{"questions": {"type": "array", "items": {"type": "string"}}}, []
),
self.ask_user,
tier="execution",
),
ToolSpec(
"request_approval",
"Ask the user to approve a Phase whose Deliverables are all "
"submitted. Ends your turn unless the Execution is pre-approved.",
_schema(
{"phase": {"type": "string"}, "summary": {"type": "string"}},
["phase", "summary"],
),
self.request_approval,
tier="execution",
),
ToolSpec(
"reopen",
"Revise a submitted Deliverable; everything depending on it returns "
"to draft.",
_schema(
{"kind": {"type": "string"}, "reason": {"type": "string"}},
["kind", "reason"],
),
self.reopen,
tier="execution",
),
ToolSpec(
"read_deliverable",
"Read a submitted Deliverable's content.",
_schema({"kind": {"type": "string"}}, ["kind"]),
self.read_deliverable,
tier="execution",
),
ToolSpec(
"finish",
"Complete the Execution once packaging is approved. Ends your turn.",
_NO_ARGS,
self.finish,
tier="execution",
),
]
async def ready_work(self) -> ToolResult:
ready = ", ".join(self.ledger.ready_work()) or "none"
phases = ", ".join(self.ledger.phases_ready_for_approval()) or "none"
return ToolResult(
f"Ready to delegate: {ready}. Phases ready for approval: {phases}."
)
async def delegate(self, kinds: Sequence[str]) -> ToolResult:
try:
parsed = [DeliverableKind(kind) for kind in kinds]
if not parsed:
raise LedgerRefusal("name at least one Deliverable")
started = self.ledger.start(parsed, actor=ORCHESTRATOR)
except (ValueError, LedgerRefusal) as exc:
return _refused(exc)
self.started.update(started)
return ToolResult(f"Started: {', '.join(started)}.", TurnEnded("delegate"))
async def ask_user(self, questions: Sequence[str] = ()) -> ToolResult:
waiting = sorted(
self._pending_questions,
key=lambda item: list(DEPENDENCY_GRAPH).index(
DeliverableKind(item["kind"])
),
)
combined = [
{"from": item["kind"], "question": question}
for item in waiting
for question in item["questions"]
]
combined += [{"from": ORCHESTRATOR, "question": q} for q in questions if q]
if not combined:
return _refused("there is nothing to ask the user")
self.questions = combined
return ToolResult("Asked the user.", TurnEnded("ask_user"))
async def request_approval(self, phase: str, summary: str) -> ToolResult:
try:
phase_id = PhaseId(phase)
except ValueError as exc:
return _refused(exc)
if self.ledger.phase_status(phase_id) is not PhaseStatus.AWAITING_APPROVAL:
try:
self.ledger.request_approval(phase_id, actor=ORCHESTRATOR)
except LedgerRefusal as exc:
return _refused(exc)
if self.ledger.phase_status(phase_id) is PhaseStatus.APPROVED:
return ToolResult(
f"Phase {phase_id} approved (pre-approved Execution)."
)
self.approval = {
"phase": phase_id.value,
"summary": summary,
"deliverables": [kind.value for kind in deliverables_in(phase_id)],
}
return ToolResult(
f"Asked the user to approve {phase_id}.", TurnEnded("request_approval")
)
async def reopen(self, kind: str, reason: str) -> ToolResult:
try:
result = self.ledger.reopen(
DeliverableKind(kind), reason, actor=ORCHESTRATOR
)
except (ValueError, LedgerRefusal) as exc:
return _refused(exc)
if result.escalation:
if self.ledger.pre_approved:
self.escalation = result.escalation
return ToolResult(result.escalation, TurnEnded("escalate"))
return ToolResult(
f"{result.escalation} Use ask_user to put this to the user."
)
withdrawn = ", ".join(result.withdrawn_approvals) or "none"
return ToolResult(
f"Reopened {kind}. Returned to draft: {', '.join(result.reset)}. "
f"Approvals withdrawn: {withdrawn}."
)
async def read_deliverable(self, kind: str) -> ToolResult:
deliverable = self._deliverables.get(kind)
if deliverable is None:
return ToolResult(f"{kind} has not been submitted.")
return ToolResult(deliverable["content"])
async def finish(self) -> ToolResult:
if not self.ledger.can_finish():
return _refused("packaging has not been approved")
return ToolResult("Finishing the Execution.", TurnEnded("finish"))
def _safe_relative(name: str) -> bool:
path = PurePosixPath(name)
return bool(name) and not path.is_absolute() and ".." not in path.parts
def specialist_tools(
kind: DeliverableKind,
extra: Sequence[ToolSpec],
allow_cloud_read: bool | None,
) -> list[ToolSpec]:
"""A Specialist's tools: submit, ask, and its role's tools with the cloud-read gate."""
async def submit_deliverable(
content: str, attachments: Mapping[str, str] | None = None
) -> ToolResult:
attachments = dict(attachments or {})
unsafe = [name for name in attachments if not _safe_relative(name)]
if unsafe:
return _refused(
"attachment names must be relative paths without '..': "
+ ", ".join(unsafe)
)
missing = check_submission(kind, content, attachments)
if missing:
return ToolResult(
"Not submitted. The Deliverable is missing: " + "; ".join(missing)
)
return ToolResult(f"Submitted {kind}.", Submitted(content, attachments))
async def needs_input(questions: Sequence[str]) -> ToolResult:
questions = tuple(q for q in questions if q)
if not questions:
return _refused("ask at least one question")
return ToolResult("Questions sent to the user.", NeedsInput(questions))
return [
ToolSpec(
"submit_deliverable",
f"Submit the {kind} Deliverable once it meets its contract. Ends your turn.",
_schema(
{
"content": {"type": "string"},
"attachments": {
"type": "object",
"additionalProperties": {"type": "string"},
},
},
["content"],
),
submit_deliverable,
tier="execution",
),
ToolSpec(
"needs_input",
"Ask the user questions you can't proceed without. Ends your turn.",
_schema(
{"questions": {"type": "array", "items": {"type": "string"}}},
["questions"],
),
needs_input,
tier="execution",
),
*(gate_cloud_read(tool, allow_cloud_read) for tool in extra),
]
def gate_cloud_read(tool: ToolSpec, allow_cloud_read: bool | None) -> ToolSpec:
"""Apply the Execution's cloud-read permission to a tool in that tier."""
if tool.tier != CLOUD_READ or allow_cloud_read is True:
return tool
if allow_cloud_read is False:
async def refuse(**_: Any) -> ToolResult:
return _refused(
"reading live cloud resources is not permitted for this Execution. "
"Work from what the user has told you and record the gap."
)
handler = refuse
else:
async def ask_permission(**_: Any) -> ToolResult:
return ToolResult(
"Permission to read live cloud resources is needed first.",
NeedsInput(
(
f"May I read live resources in your cloud environment? ({tool.name})",
),
permission=CLOUD_READ,
),
)
handler = ask_permission
return ToolSpec(tool.name, tool.description, tool.parameters, handler, tool.tier)