File size: 20,225 Bytes
921d377 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 | """
WorkflowRunner β the in-house orchestrator that chains small,
single-decision prompts into an end-to-end workflow (stage-1
autoplan, stage-2 autogen, etc.).
Why not LangChain / LangGraph / CrewAI?
* LangChain adds a ~40 MB dependency tree for what is, at its
core, a `for step in steps: call_llm(step); validate; retry`
loop. Its retry + output-parsing abstractions would still need
a shim to talk to our Enterprise Settings resolver and our
YAML prompt library.
* LangGraph's strength is cyclical graphs with conditional
branching; our workflows are straight DAGs (mostly linear,
occasionally a fan-out over scenes). The framework tax isn't
justified for 7 nodes.
* CrewAI is multi-agent-role oriented; we have one role ("the
planner") running seven specialised prompts, not seven agents
collaborating. Using CrewAI would force us to pretend each
prompt is an agent, which just obscures the data flow.
What this module does
---------------------
Given a list of ``Step``s and a starting context dict:
1. For each step, build the prompt variables by calling
``step.build_vars(context)``.
2. Render the prompt via ``PromptLibrary``.
3. Dispatch via ``llm_router.call_prompt`` (Enterprise Settings
chat model, policy-driven timeout).
4. Parse the assistant content with ``step.parse``.
5. Validate the parsed value with ``step.validate``.
6. On parse/validate failure, retry up to ``policy.retries``.
7. On final failure, consult ``policy.fallback``:
'abort' β raise StepFailure (workflow halts)
other str β let ``step.on_fallback`` interpret (returns
a safe default, or raises if step.build_vars
can't cope).
8. Store the value under ``step.output_key`` in the context so
later steps can read it.
9. Emit events via ``on_event`` hook so the frontend spinner
can display real progress (REV-5 wires this to SSE).
Events
------
The runner emits four event kinds:
workflow_started {workflow, step_ids, started_at_ms}
step_started {step_id, prompt_id, attempt=1}
step_completed {step_id, prompt_id, attempt, duration_ms,
preview} # first 80 chars of output
step_failed {step_id, prompt_id, attempt, reason,
fatal} # fatal=True only if abort
workflow_completed {workflow, duration_ms, ok}
``preview`` is a short string derived from ``str(value)[:80]`` so
operators can eyeball "did the LLM actually say something sensible?"
without exposing full content.
Design choices
--------------
* **Pure Python, no framework tax.** ~180 LoC, no extra deps.
* **Context is a plain dict.** Steps read/write it freely. No
immutable-state ceremony; this module is intentionally small.
* **No global state.** The runner holds no caches. Same library,
same context, same result β easy to test.
* **Parallel fan-out is deferred.** Today each Step runs
sequentially (simplest, deterministic). REV-4 introduces a
``ParallelFanout`` wrapper when per-scene script generation
needs it. Premature abstraction risk is not worth it now.
"""
from __future__ import annotations
import asyncio
import inspect
import logging
import time
from dataclasses import dataclass, field
from typing import (
Any, Awaitable, Callable, Dict, List, Mapping,
Optional, Sequence,
)
from ..llm_router import call_prompt
from ..prompts import (
PromptLibrary,
PromptLibraryError,
PromptPolicy,
RenderedPrompt,
default_library,
)
log = logging.getLogger(__name__)
# ββ Errors βββββββββββββββββββββββββββββββββββββββββββββββββββββ
class StepFailure(Exception):
"""Raised when a step exhausts retries and its fallback is
``abort``. Carries enough context for the route layer to
produce a useful error payload.
"""
def __init__(
self,
*,
step_id: str,
prompt_id: str,
reason: str,
attempts: int,
) -> None:
self.step_id = step_id
self.prompt_id = prompt_id
self.reason = reason
self.attempts = attempts
super().__init__(
f"Step {step_id!r} ({prompt_id}) failed after {attempts} attempt(s): {reason}"
)
# ββ Data shapes ββββββββββββββββββββββββββββββββββββββββββββββββ
@dataclass(frozen=True)
class WorkflowEvent:
"""Structured telemetry emitted per step transition.
The frontend SSE stream serialises these to JSON; tests assert
on the ``kind`` + payload keys. Payloads are plain dicts so
we don't leak dataclass types across the wire.
"""
kind: str # 'workflow_started' | 'step_started' | ...
ts_ms: int # epoch-relative-ish, for ordering
payload: Dict[str, Any] = field(default_factory=dict)
@dataclass(frozen=True)
class StepResult:
"""Outcome of one completed step. Stored on the WorkflowResult
so tests can introspect the full trace.
"""
step_id: str
prompt_id: str
value: Any
attempts: int
used_fallback: bool
duration_ms: int
@dataclass
class WorkflowResult:
"""Final context + per-step trace for a completed run.
``aborted`` is True if a StepFailure halted the workflow β in
that case ``steps`` contains every step that *did* run up to
and including the failing one.
"""
context: Dict[str, Any]
steps: List[StepResult] = field(default_factory=list)
events: List[WorkflowEvent] = field(default_factory=list)
aborted: bool = False
error: Optional[str] = None
started_at_ms: int = 0
duration_ms: int = 0
# ββ Step definition ββββββββββββββββββββββββββββββββββββββββββββ
# Type aliases keep the Step signature readable.
VarBuilder = Callable[[Mapping[str, Any]], Mapping[str, Any]]
Parser = Callable[[str], Any]
Validator = Callable[[Any], Optional[str]] # returns err msg or None
FallbackFn = Callable[[Mapping[str, Any], Optional[str]], Any]
# (context, policy.fallback) -> value
@dataclass(frozen=True)
class Step:
"""One prompt in a workflow.
Fields
------
step_id
Stable ID for logs/events (often equals the prompt id's
trailing segment, e.g. ``classify_mode``).
prompt_id
Full library id (``autoplan.classify_mode``).
output_key
Where in ``context`` the validated value lands.
build_vars
``context -> variables dict`` for the prompt renderer.
Keep it pure β no I/O.
parse
``raw content -> parsed value``. Raise ValueError on
parse failure; the runner treats it like a validation
miss and retries.
validate
``value -> None | error message``. Returning a string
triggers retry (and eventually fallback).
fallback
``(context, fallback_token) -> default value``. Called
only when retries are exhausted AND the policy fallback
is NOT ``abort``. Raise StepFailure if no safe default
is producible.
temperature, max_tokens
Per-step generation knobs. Short enums want temp=0.0;
creative prose wants 0.5ish.
"""
step_id: str
prompt_id: str
output_key: str
build_vars: VarBuilder
parse: Parser = field(default=lambda s: s.strip())
validate: Validator = field(default=lambda v: None)
fallback: Optional[FallbackFn] = None
temperature: float = 0.3
max_tokens: int = 350
# Per-step model override (Ollama model id). Empty string means
# "use the server default model". Threaded into chat_ollama by
# the runner so individual workflow steps can route around the
# default model β used by Mature (gated) experiences to pick an
# uncensored / abliterated model the operator selected in the
# wizard's Step 0 picker.
model: str = ""
# ββ Event hook type ββββββββββββββββββββββββββββββββββββββββββββ
EventHook = Callable[[WorkflowEvent], None]
# ββ Helper: content extractor ββββββββββββββββββββββββββββββββββ
def extract_content(response: Any) -> str:
"""Pull the assistant content out of an OpenAI-style envelope.
Returns empty string on any shape we don't recognise β the
caller's validator then produces the "empty output" retry.
"""
if not isinstance(response, dict):
return ""
choices = response.get("choices")
if not isinstance(choices, list) or not choices:
return ""
first = choices[0]
if not isinstance(first, dict):
return ""
msg = first.get("message") or {}
if isinstance(msg, dict):
content = msg.get("content")
if isinstance(content, str):
return content.strip()
return ""
# ββ Runner βββββββββββββββββββββββββββββββββββββββββββββββββββββ
class WorkflowRunner:
"""Sequentially run a list of ``Step``s against a shared
context. Stateless across runs β instantiate once per
process is fine.
"""
def __init__(
self,
*,
library: Optional[PromptLibrary] = None,
) -> None:
self._library = library or default_library()
async def run(
self,
*,
workflow: str,
steps: Sequence[Step],
context: Optional[Dict[str, Any]] = None,
on_event: Optional[EventHook] = None,
) -> WorkflowResult:
"""Execute ``steps`` in order. Returns a ``WorkflowResult``
with the final context, per-step trace, and event log.
A StepFailure is caught: the workflow is marked aborted
and returned (instead of raised). Callers that want the
exception style should check ``result.aborted`` and re-raise.
"""
ctx: Dict[str, Any] = dict(context or {})
events: List[WorkflowEvent] = []
def _emit(ev: WorkflowEvent) -> None:
events.append(ev)
if on_event is not None:
try:
# Backward-compat hook bridge:
# * New runner-native hooks expect one arg:
# on_event(WorkflowEvent)
# * Legacy planner/generator hooks expect two args:
# on_event(kind, payload)
#
# We support both so route-level SSE hooks can be
# passed through unchanged without dropping workflow
# events due to signature mismatch.
payload = dict(ev.payload or {})
argc: Optional[int] = None
has_varargs = False
try:
sig = inspect.signature(on_event)
params = list(sig.parameters.values())
has_varargs = any(
p.kind == inspect.Parameter.VAR_POSITIONAL
for p in params
)
argc = len([
p for p in params
if p.kind in (
inspect.Parameter.POSITIONAL_ONLY,
inspect.Parameter.POSITIONAL_OR_KEYWORD,
)
])
except Exception:
argc = None
if has_varargs or (argc is not None and argc >= 2):
on_event(ev.kind, payload)
elif argc == 0:
on_event()
else:
# Default path: WorkflowEvent callback.
on_event(ev)
except Exception: # noqa: BLE001
# Hooks must never break the workflow.
log.exception("workflow event hook raised β ignored")
started = _now_ms()
result = WorkflowResult(
context=ctx, steps=[], events=events,
started_at_ms=started,
)
_emit(WorkflowEvent(
kind="workflow_started", ts_ms=_now_ms(),
payload={
"workflow": workflow,
"step_ids": [s.step_id for s in steps],
"started_at_ms": started,
},
))
try:
for step in steps:
step_result = await self._run_step(step, ctx, _emit)
result.steps.append(step_result)
ctx[step.output_key] = step_result.value
except StepFailure as exc:
result.aborted = True
result.error = str(exc)
log.warning("workflow_aborted workflow=%s reason=%s", workflow, exc)
finally:
result.duration_ms = max(0, _now_ms() - started)
_emit(WorkflowEvent(
kind="workflow_completed", ts_ms=_now_ms(),
payload={
"workflow": workflow,
"duration_ms": result.duration_ms,
"ok": not result.aborted,
},
))
return result
# ββ Per-step execution ββββββββββββββββββββββββββββββββ
async def _run_step(
self,
step: Step,
ctx: Mapping[str, Any],
emit: Callable[[WorkflowEvent], None],
) -> StepResult:
try:
policy = self._library.policy(step.prompt_id)
except PromptLibraryError as exc:
emit(WorkflowEvent(
kind="step_failed", ts_ms=_now_ms(),
payload={
"step_id": step.step_id, "prompt_id": step.prompt_id,
"attempt": 1, "reason": f"prompt-load: {exc}", "fatal": True,
},
))
raise StepFailure(
step_id=step.step_id, prompt_id=step.prompt_id,
reason=f"prompt load failed: {exc}", attempts=1,
) from exc
max_attempts = max(1, int(policy.retries) + 1)
last_reason: str = ""
step_started = _now_ms()
for attempt in range(1, max_attempts + 1):
emit(WorkflowEvent(
kind="step_started", ts_ms=_now_ms(),
payload={
"step_id": step.step_id, "prompt_id": step.prompt_id,
"attempt": attempt,
},
))
t0 = _now_ms()
try:
rendered = self._library.render(
step.prompt_id, **dict(step.build_vars(ctx)),
)
except PromptLibraryError as exc:
last_reason = f"render: {exc}"
emit(_fail_event(step, attempt, last_reason, fatal=False))
break # can't retry a deterministic render failure
try:
response = await call_prompt(
rendered, policy,
temperature=float(step.temperature),
max_tokens=int(step.max_tokens),
model_override=(step.model or None),
)
except asyncio.TimeoutError:
last_reason = f"timeout after {policy.timeout_s:.1f}s"
emit(_fail_event(step, attempt, last_reason, fatal=False))
continue
except Exception as exc: # noqa: BLE001
last_reason = f"llm: {exc.__class__.__name__}: {str(exc)[:180]}"
emit(_fail_event(step, attempt, last_reason, fatal=False))
continue
content = extract_content(response)
if not content:
last_reason = "empty response"
emit(_fail_event(step, attempt, last_reason, fatal=False))
continue
try:
parsed = step.parse(content)
except Exception as exc: # noqa: BLE001
last_reason = f"parse: {str(exc)[:180]}"
emit(_fail_event(step, attempt, last_reason, fatal=False))
continue
verdict = step.validate(parsed)
if verdict is not None:
last_reason = f"validate: {verdict}"
emit(_fail_event(step, attempt, last_reason, fatal=False))
continue
# Success.
duration = max(0, _now_ms() - t0)
emit(WorkflowEvent(
kind="step_completed", ts_ms=_now_ms(),
payload={
"step_id": step.step_id, "prompt_id": step.prompt_id,
"attempt": attempt, "duration_ms": duration,
"preview": _preview(parsed),
},
))
return StepResult(
step_id=step.step_id, prompt_id=step.prompt_id,
value=parsed, attempts=attempt,
used_fallback=False,
duration_ms=max(0, _now_ms() - step_started),
)
# All attempts exhausted. Consult fallback policy.
fallback = (policy.fallback or "").strip().lower()
if fallback == "abort" or not fallback:
emit(_fail_event(step, max_attempts, last_reason, fatal=True))
raise StepFailure(
step_id=step.step_id, prompt_id=step.prompt_id,
reason=last_reason or "unknown failure",
attempts=max_attempts,
)
if step.fallback is None:
# Policy said non-abort but the Step didn't wire a
# fallback callable β treat as fatal so we don't
# silently ship garbage.
reason = f"{last_reason} (no Step.fallback configured for '{fallback}')"
emit(_fail_event(step, max_attempts, reason, fatal=True))
raise StepFailure(
step_id=step.step_id, prompt_id=step.prompt_id,
reason=reason, attempts=max_attempts,
)
try:
value = step.fallback(ctx, policy.fallback)
except StepFailure:
raise
except Exception as exc: # noqa: BLE001
reason = f"fallback raised: {exc}"
emit(_fail_event(step, max_attempts, reason, fatal=True))
raise StepFailure(
step_id=step.step_id, prompt_id=step.prompt_id,
reason=reason, attempts=max_attempts,
) from exc
emit(WorkflowEvent(
kind="step_completed", ts_ms=_now_ms(),
payload={
"step_id": step.step_id, "prompt_id": step.prompt_id,
"attempt": max_attempts, "duration_ms": 0,
"preview": _preview(value), "used_fallback": True,
},
))
return StepResult(
step_id=step.step_id, prompt_id=step.prompt_id,
value=value, attempts=max_attempts,
used_fallback=True,
duration_ms=max(0, _now_ms() - step_started),
)
# ββ Private helpers ββββββββββββββββββββββββββββββββββββββββββββ
def _now_ms() -> int:
return int(time.time() * 1000)
def _preview(value: Any) -> str:
try:
s = str(value)
except Exception: # noqa: BLE001
return ""
s = s.replace("\n", " ").strip()
return s if len(s) <= 80 else s[:77] + "β¦"
def _fail_event(
step: Step, attempt: int, reason: str, *, fatal: bool,
) -> WorkflowEvent:
return WorkflowEvent(
kind="step_failed", ts_ms=_now_ms(),
payload={
"step_id": step.step_id, "prompt_id": step.prompt_id,
"attempt": attempt, "reason": reason, "fatal": fatal,
},
)
|