File size: 8,402 Bytes
7302343 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 | """Orchestrator loop — runs tools per tier plan, enforces budgets, early-stops.
Spec: docs/04-InvestigationEngine.md §3, docs/Specs.md §7.2.
For each tier the Strategy Selector picks a plan (FAST=2 tools, STANDARD=4,
DEEP=5+ with potential re-loop). The Orchestrator iterates the plan; each
iteration checks budgets *before* running, invokes the tool with isolated
exception handling, appends the result to the Evidence Accumulator, then
asks the convergence policy whether enough signal has accumulated to stop.
The Orchestrator owns:
- tier → tool plan mapping (overridable per call for testing)
- budget enforcement (tool count + wall-clock time)
- per-tool exception isolation: a single tool throwing must not blow
up the investigation; we record a `failure` ToolResult and keep going
- convergence policy: 1 strong-signal success on FAST, 2 on STANDARD/DEEP
- structured logging at start, per-tool, and stop
It does NOT decide the verdict — that's the Reasoner. It does NOT persist —
that's the API handler wrapping `run()` with a session.
"""
from __future__ import annotations
import time
from dataclasses import dataclass, field
from datetime import UTC, datetime
from typing import TYPE_CHECKING
import structlog
from orchestrator.tools import (
EvidenceAccumulator,
ToolContext,
ToolName,
ToolRegistry,
ToolResult,
)
if TYPE_CHECKING:
from collections.abc import Callable
from orchestrator.strategy import StrategyDecision
logger = structlog.get_logger(__name__)
# === Result =============================================================
StopReason = str # "plan_complete" | "converged" | "budget_tool" | "budget_time"
@dataclass(frozen=True)
class OrchestratorResult:
correlation_id: str
subreddit_id: str
tier: str
accumulator: EvidenceAccumulator
started_at: datetime
completed_at: datetime
total_latency_ms: int
tools_run: int
early_stopped: bool
stop_reason: StopReason
plan: list[ToolName] = field(default_factory=list)
# === Default tier plans =================================================
# Specs §7.1 budget table; STANDARD picks the 4-tool variant pending the
# open question about 4-vs-5.
_DEFAULT_PLANS: dict[str, list[ToolName]] = {
"FAST": ["policy_match", "report_velocity"],
"STANDARD": ["policy_match", "report_velocity", "user_history", "prior_actions"],
"DEEP": [
"policy_match",
"report_velocity",
"user_history",
"prior_actions",
"thread_context",
],
}
# === Orchestrator =======================================================
class Orchestrator:
"""Stateless coordinator. Safe to share across requests."""
def __init__(
self,
registry: ToolRegistry,
*,
# `perf_counter`-style clock injected so tests can drive elapsed
# time deterministically. Returns seconds.
clock: Callable[[], float] = time.perf_counter,
) -> None:
self._registry = registry
self._clock = clock
def default_plan(self, tier: str) -> list[ToolName]:
try:
return list(_DEFAULT_PLANS[tier])
except KeyError as e:
raise ValueError(f"no default plan for tier {tier!r}") from e
async def run(
self,
*,
decision: StrategyDecision,
context: ToolContext,
plan: list[ToolName] | None = None,
) -> OrchestratorResult:
chosen_plan = plan if plan is not None else self.default_plan(decision.tier)
accumulator = EvidenceAccumulator()
log = logger.bind(
correlation_id=context.correlation_id,
subreddit_id=context.subreddit_id,
tier=decision.tier,
plan_length=len(chosen_plan),
)
log.info("orchestrator.started", plan=chosen_plan)
started_at = datetime.now(UTC)
started_perf = self._clock()
tools_run = 0
stop_reason: StopReason = "plan_complete"
early_stopped = False
for tool_name in chosen_plan:
# 1. Budget pre-check.
elapsed_ms = int((self._clock() - started_perf) * 1000)
if elapsed_ms >= decision.time_budget_ms:
stop_reason = "budget_time"
early_stopped = True
log.info("orchestrator.stop", reason=stop_reason, elapsed_ms=elapsed_ms)
break
if tools_run >= decision.tool_budget:
stop_reason = "budget_tool"
early_stopped = True
log.info("orchestrator.stop", reason=stop_reason, tools_run=tools_run)
break
# 2. Resolve tool. Unregistered → record skip, keep going.
if not self._registry.has(tool_name):
accumulator.append(
ToolResult(
tool=tool_name,
status="skipped",
summary=f"tool {tool_name!r} not registered",
latency_ms=0,
)
)
tools_run += 1
log.warning("orchestrator.tool.unregistered", tool=tool_name)
continue
# 3. Run the tool with isolated exception handling.
tool = self._registry.get(tool_name)
tool_started = self._clock()
try:
result = await tool.run(context)
except Exception as exc:
latency = int((self._clock() - tool_started) * 1000)
result = ToolResult(
tool=tool_name,
status="failure",
summary=f"tool raised: {type(exc).__name__}",
latency_ms=latency,
error=str(exc),
)
log.warning(
"orchestrator.tool.raised",
tool=tool_name,
exc_type=type(exc).__name__,
)
accumulator.append(result)
tools_run += 1
log.info(
"orchestrator.tool.completed",
tool=tool_name,
status=result.status,
latency_ms=result.latency_ms,
)
# 4. Convergence check — stop early if we have enough signal.
if _converged(accumulator, decision.tier):
stop_reason = "converged"
early_stopped = True
log.info(
"orchestrator.stop",
reason=stop_reason,
tools_run=tools_run,
successful_evidence=len(accumulator.successful_entries()),
)
break
completed_at = datetime.now(UTC)
total_latency_ms = int((self._clock() - started_perf) * 1000)
log.info(
"orchestrator.completed",
tools_run=tools_run,
stop_reason=stop_reason,
total_latency_ms=total_latency_ms,
)
return OrchestratorResult(
correlation_id=context.correlation_id,
subreddit_id=context.subreddit_id,
tier=decision.tier,
accumulator=accumulator,
started_at=started_at,
completed_at=completed_at,
total_latency_ms=total_latency_ms,
tools_run=tools_run,
early_stopped=early_stopped,
stop_reason=stop_reason,
plan=chosen_plan,
)
# === Convergence policy =================================================
def _converged(acc: EvidenceAccumulator, tier: str) -> bool:
"""A simple, explainable convergence rule.
For FAST tier: 1 strong-signal success is enough.
For STANDARD / DEEP: require 2 strong signals — we want the Reasoner
to have at least two corroborating evidence rows before short-circuiting.
A "strong signal" is a successful tool result whose `detail.signal` is
`"high"`. Tools self-report this. If no tool has set it, convergence
never triggers and the plan runs to its budget.
"""
threshold = 1 if tier == "FAST" else 2
strong = sum(
1
for entry in acc.successful_entries()
if entry.detail.get("signal") == "high"
)
return strong >= threshold
__all__ = [
"Orchestrator",
"OrchestratorResult",
"StopReason",
]
|