Coverage for little_loops / fsm / schema.py: 37%
549 statements
« prev ^ index » next coverage.py v7.12.0, created at 2026-06-15 17:27 -0500
« prev ^ index » next coverage.py v7.12.0, created at 2026-06-15 17:27 -0500
1"""FSM loop schema dataclasses.
3This module defines the type-safe dataclasses that represent FSM loop
4definitions. These match the universal FSM schema described in the
5design documentation.
7The schema supports:
8- Multiple evaluator types (exit_code, output_numeric, etc.)
9- Two-layer transition system (evaluate + route)
10- Both shorthand (on_success/on_failure) and full routing syntax
11- Context variables and captured values
12- LLM evaluation configuration
13"""
15from __future__ import annotations
17from dataclasses import dataclass, field
18from typing import Any, Literal
20# Default LLM model for structured evaluation
21DEFAULT_LLM_MODEL: str = "sonnet"
24@dataclass
25class EvaluateConfig:
26 """Evaluator configuration for action result interpretation.
28 The evaluator determines how to interpret an action's output and
29 produce a verdict string for routing.
31 Attributes:
32 type: Evaluator type. One of:
33 - exit_code: Map exit codes to verdicts (default for shell)
34 - output_numeric: Compare numeric output to target
35 - output_json: Extract and compare JSON path value
36 - output_contains: Pattern matching on stdout
37 - convergence: Compare current vs previous value toward target
38 - llm_structured: Use LLM with structured output (default for slash)
39 operator: Comparison operator (eq, ne, lt, le, gt, ge)
40 target: Target value for comparison
41 tolerance: Acceptable distance from target (for convergence)
42 pattern: Pattern string for output_contains
43 negate: If True, invert the match result (output_contains)
44 path: JSON path for output_json (jq-style)
45 prompt: Custom prompt for llm_structured
46 schema: Custom JSON schema for llm_structured response
47 min_confidence: Minimum confidence threshold for llm_structured
48 uncertain_suffix: If True, append _uncertain to low-confidence verdicts
49 source: Override default source (current action output)
50 previous: Previous value reference for convergence
51 direction: Optimization direction for convergence (minimize/maximize)
52 scope: Paths to limit git diff to for diff_stall evaluator
53 max_stall: Consecutive no-change iterations before failure (diff_stall)
54 pairs: List of producer/consumer pair dicts for contract evaluator
55 """
57 type: Literal[
58 "exit_code",
59 "output_numeric",
60 "output_json",
61 "output_contains",
62 "convergence",
63 "diff_stall",
64 "action_stall",
65 "llm_structured",
66 "mcp_result",
67 "harbor_scorer",
68 "comparator",
69 "contract",
70 ]
71 operator: str | None = None
72 target: int | float | str | None = None
73 tolerance: float | str | None = None # str for interpolation (e.g., "${context.tolerance}")
74 pattern: str | None = None
75 negate: bool = False
76 path: str | None = None
77 prompt: str | None = None
78 schema: dict[str, Any] | None = None
79 min_confidence: float = 0.5
80 uncertain_suffix: bool = False
81 source: str | None = None
82 previous: str | None = None
83 direction: Literal["minimize", "maximize"] = "minimize"
84 scope: list[str] | None = None # for diff_stall: limit git diff to these paths
85 max_stall: int = 1 # for diff_stall: consecutive no-change iterations before failure
86 track: list[str] | None = None # for action_stall: context keys to track (default: ["action"])
87 max_repeat: int = 2 # for action_stall: consecutive identical iterations before failure
88 baseline_path: str | None = None # for comparator: path to .loops/baselines/<loop>/ dir
89 auto_promote: bool = False # for comparator: write output to baseline on yes verdict
90 min_pairs: int = 1 # for comparator: number of blind A/B comparisons to run
91 pairs: list[dict] | None = None # for contract: list of producer/consumer pair dicts
93 def to_dict(self) -> dict[str, Any]:
94 """Convert to dictionary for JSON/YAML serialization."""
95 result: dict[str, Any] = {"type": self.type}
97 # Only include non-None optional fields
98 if self.operator is not None:
99 result["operator"] = self.operator
100 if self.target is not None:
101 result["target"] = self.target
102 if self.tolerance is not None:
103 result["tolerance"] = self.tolerance
104 if self.pattern is not None:
105 result["pattern"] = self.pattern
106 if self.negate:
107 result["negate"] = self.negate
108 if self.path is not None:
109 result["path"] = self.path
110 if self.prompt is not None:
111 result["prompt"] = self.prompt
112 if self.schema is not None:
113 result["schema"] = self.schema
114 if self.min_confidence != 0.5:
115 result["min_confidence"] = self.min_confidence
116 if self.uncertain_suffix:
117 result["uncertain_suffix"] = self.uncertain_suffix
118 if self.source is not None:
119 result["source"] = self.source
120 if self.previous is not None:
121 result["previous"] = self.previous
122 if self.direction != "minimize":
123 result["direction"] = self.direction
124 if self.scope is not None:
125 result["scope"] = self.scope
126 if self.max_stall != 1:
127 result["max_stall"] = self.max_stall
128 if self.track is not None:
129 result["track"] = self.track
130 if self.max_repeat != 2:
131 result["max_repeat"] = self.max_repeat
132 if self.baseline_path is not None:
133 result["baseline_path"] = self.baseline_path
134 if self.auto_promote:
135 result["auto_promote"] = self.auto_promote
136 if self.min_pairs != 1:
137 result["min_pairs"] = self.min_pairs
138 if self.pairs is not None:
139 result["pairs"] = self.pairs
141 return result
143 @classmethod
144 def from_dict(cls, data: dict[str, Any]) -> EvaluateConfig:
145 """Create from dictionary (JSON/YAML deserialization)."""
146 return cls(
147 type=data["type"],
148 operator=data.get("operator"),
149 target=data.get("target"),
150 tolerance=data.get("tolerance"),
151 pattern=data.get("pattern"),
152 negate=data.get("negate", False),
153 path=data.get("path"),
154 prompt=data.get("prompt"),
155 schema=data.get("schema"),
156 min_confidence=data.get("min_confidence", 0.5),
157 uncertain_suffix=data.get("uncertain_suffix", False),
158 source=data.get("source"),
159 previous=data.get("previous"),
160 direction=data.get("direction", "minimize"),
161 scope=data.get("scope"),
162 max_stall=data.get("max_stall", 1),
163 track=data.get("track"),
164 max_repeat=data.get("max_repeat", 2),
165 baseline_path=data.get("baseline_path"),
166 auto_promote=data.get("auto_promote", False),
167 min_pairs=data.get("min_pairs", 1),
168 pairs=data.get("pairs"),
169 )
172@dataclass
173class RouteConfig:
174 """Routing table configuration for verdict-to-state mapping.
176 Maps verdict strings from evaluators to next state names.
178 Attributes:
179 routes: Mapping from verdict string to next state name
180 default: Default state for unmatched verdicts (the "_" key)
181 error: State for evaluation/execution errors (the "_error" key)
182 """
184 routes: dict[str, str] = field(default_factory=dict)
185 default: str | None = None
186 error: str | None = None
188 def to_dict(self) -> dict[str, Any]:
189 """Convert to dictionary for JSON/YAML serialization."""
190 result = dict(self.routes)
191 if self.default is not None:
192 result["_"] = self.default
193 if self.error is not None:
194 result["_error"] = self.error
195 return result
197 @classmethod
198 def from_dict(cls, data: dict[str, Any]) -> RouteConfig:
199 """Create from dictionary (JSON/YAML deserialization)."""
200 routes = {k: v for k, v in data.items() if not k.startswith("_")}
201 return cls(
202 routes=routes,
203 default=data.get("_"),
204 error=data.get("_error"),
205 )
208@dataclass
209class ParameterSpec:
210 """Specification for a single loop input parameter.
212 Declares a typed input that callers bind via the 'with:' block on sub-loop states.
214 Attributes:
215 type: Parameter type. One of: string, integer, number, boolean, enum, path.
216 required: If True, callers must supply this parameter in 'with:' (mutually
217 exclusive with 'default').
218 default: Default value used when the caller does not bind the parameter.
219 Only valid when required is False.
220 description: Human-readable description of the parameter.
221 values: Allowed values for 'enum' type parameters.
222 """
224 type: str
225 required: bool = False
226 default: Any = None
227 description: str | None = None
228 values: list[Any] | None = None
230 def to_dict(self) -> dict[str, Any]:
231 """Convert to dictionary for JSON/YAML serialization."""
232 result: dict[str, Any] = {"type": self.type}
233 if self.required:
234 result["required"] = self.required
235 if self.default is not None:
236 result["default"] = self.default
237 if self.description is not None:
238 result["description"] = self.description
239 if self.values is not None:
240 result["values"] = self.values
241 return result
243 @classmethod
244 def from_dict(cls, data: dict[str, Any]) -> ParameterSpec:
245 """Create from dictionary (JSON/YAML deserialization)."""
246 return cls(
247 type=data["type"],
248 required=data.get("required", False),
249 default=data.get("default"),
250 description=data.get("description"),
251 values=data.get("values"),
252 )
255@dataclass
256class ThrottleConfig:
257 """Per-state tool-call progressive throttling configuration.
259 Counts successful tool calls within a single state visit and escalates
260 restrictions to self-throttle runaway states before provider limits are reached.
262 Thresholds fall back to executor module defaults when not set:
263 - normal_max: _DEFAULT_THROTTLE_NORMAL_MAX (3) — calls 1..normal_max pass through
264 - warn_max: _DEFAULT_THROTTLE_WARN_MAX (8) — calls at warn_max inject a warning event
265 - hard_max: _DEFAULT_THROTTLE_HARD_MAX (12) — calls at hard_max route to on_throttle_hard
266 - calls > hard_max: hard stop, loop marked stuck
268 States with type="learning" (FEAT-1283) are exempt from the hard_max hard-stop because
269 they legitimately make N tool calls per visit (one per unproven target). The warn_max
270 warning still applies so users can see the state is doing significant work.
271 """
273 normal_max: int | None = None
274 warn_max: int | None = None
275 hard_max: int | None = None
277 def to_dict(self) -> dict[str, Any]:
278 """Convert to dictionary for JSON/YAML serialization."""
279 result: dict[str, Any] = {}
280 if self.normal_max is not None:
281 result["normal_max"] = self.normal_max
282 if self.warn_max is not None:
283 result["warn_max"] = self.warn_max
284 if self.hard_max is not None:
285 result["hard_max"] = self.hard_max
286 return result
288 @classmethod
289 def from_dict(cls, data: dict[str, Any]) -> ThrottleConfig:
290 """Create from dictionary (JSON/YAML deserialization)."""
291 return cls(
292 normal_max=data.get("normal_max"),
293 warn_max=data.get("warn_max"),
294 hard_max=data.get("hard_max"),
295 )
298@dataclass
299class LearningConfig:
300 """Per-state configuration for FEAT-1283 `type: learning` dispatch.
302 Declares the list of external-API/SDK targets a learning state must prove
303 against the learning-tests registry (ENH-1282) before advancing. On a
304 missing or stale record, the state invokes `/ll:explore-api <target>` up to
305 `max_retries` times before transitioning to a blocked target.
307 Attributes:
308 targets: Ordered list of target identifiers (e.g. "Anthropic SDK
309 streaming"). Each is slugified internally when looking up the
310 registry record. All targets must reach status="proven" for the
311 state to advance via on_yes.
312 targets_csv: Runtime-interpolated comma-separated target string (e.g.
313 "${context.targets}"). Resolved and CSV-split by the executor at
314 execution time. Alternative to the static ``targets`` list for loops
315 that receive targets as a context CSV string. Exactly one of
316 ``targets`` (non-empty) or ``targets_csv`` must be set.
317 max_retries: Maximum number of `/ll:explore-api` invocations per target
318 before the state routes to on_blocked / on_no with reason
319 ``retries_exhausted``. Counts only re-exploration attempts; the
320 initial registry lookup is free. Distinct from ENH-1115's throttle
321 counter, which measures tool-call volume; learning states are
322 already exempt from throttle hard_max via FSMExecutor._check_throttle.
323 max_retries_expr: Runtime-interpolated retry limit (e.g.
324 "${context.max_retries}"). Resolved via interpolate() and int()-cast
325 at execution time. Takes precedence over ``max_retries`` when set.
326 """
328 targets: list[str] = field(default_factory=list)
329 targets_csv: str | None = None
330 max_retries: int = 2
331 max_retries_expr: str | None = None
333 def to_dict(self) -> dict[str, Any]:
334 """Convert to dictionary for JSON/YAML serialization."""
335 result: dict[str, Any] = {"targets": list(self.targets), "max_retries": self.max_retries}
336 if self.targets_csv is not None:
337 result["targets_csv"] = self.targets_csv
338 if self.max_retries_expr is not None:
339 result["max_retries_expr"] = self.max_retries_expr
340 return result
342 @classmethod
343 def from_dict(cls, data: dict[str, Any]) -> LearningConfig:
344 """Create from dictionary (JSON/YAML deserialization)."""
345 return cls(
346 targets=list(data.get("targets") or []),
347 targets_csv=data.get("targets_csv") or None,
348 max_retries=int(data.get("max_retries", 2)),
349 max_retries_expr=data.get("max_retries_expr") or None,
350 )
353@dataclass
354class StateConfig:
355 """Configuration for a single FSM state.
357 States can have actions, evaluators, and routing. Supports both
358 shorthand (on_success/on_failure) and full routing table syntax.
360 Attributes:
361 action: Command to execute (shell, slash command, or "server/tool-name" for mcp_tool)
362 action_type: How to execute the action (prompt, slash_command, shell, mcp_tool).
363 If None, uses heuristic: / prefix = slash_command, else = shell.
364 params: MCP tool arguments (only used with action_type: mcp_tool). Supports
365 ${variable} interpolation in string values.
366 evaluate: Evaluator configuration for result interpretation
367 route: Full routing table (verdict -> state mapping)
368 on_yes: Shorthand for yes verdict routing
369 on_no: Shorthand for no verdict routing
370 on_error: Shorthand for error verdict routing
371 on_partial: Shorthand for partial verdict routing
372 next: Unconditional transition (no evaluation)
373 terminal: If True, this is an end state
374 capture: Variable name to store action output
375 timeout: Action-level timeout in seconds
376 on_maintain: State to transition to when maintain=True and loop completes
377 max_retries: Max consecutive re-entries before transitioning to on_retry_exhausted.
378 A value of N allows N retries after the initial execution (N+1 total entries).
379 Requires on_retry_exhausted to also be set.
380 on_retry_exhausted: State to transition to when max_retries consecutive re-entries
381 are exceeded. Required when max_retries is set.
382 max_rate_limit_retries: Max consecutive 429/rate-limit in-place retries for this
383 state before transitioning to on_rate_limit_exhausted. Requires
384 on_rate_limit_exhausted to also be set.
385 on_rate_limit_exhausted: State to transition to when max_rate_limit_retries
386 consecutive rate-limit retries are exceeded. Required when
387 max_rate_limit_retries is set.
388 rate_limit_backoff_base_seconds: Base seconds for exponential backoff between
389 rate-limit retries; actual sleep is base * 2^(attempt-1) + uniform(0, base).
390 Defaults to 30.
391 rate_limit_max_wait_seconds: Total wall-clock budget (seconds) for rate-limit
392 handling in this state before routing to on_rate_limit_exhausted. When unset,
393 defaults from commands.rate_limits.max_wait_seconds (21600 = 6h).
394 rate_limit_long_wait_ladder: Backoff ladder (seconds) for the long-wait tier used
395 once the short-tier retry budget is spent. Each entry is the sleep before the
396 next retry attempt. When unset, defaults from
397 commands.rate_limits.long_wait_ladder ([300, 900, 1800, 3600]).
398 loop: Name of a loop YAML to execute as a sub-FSM. Mutually exclusive with action.
399 context_passthrough: When True, pass parent context variables to child loop and
400 merge child captures back into parent context. Legacy escape hatch; prefer
401 'with_' for explicit named bindings.
402 with_: Explicit parameter bindings for sub-loop calls (YAML key: 'with'). Maps
403 declared child parameter names to parent expressions. Only valid when 'loop'
404 is set. Mutually exclusive with context_passthrough.
405 fragment_name: Original fragment name from the 'fragment:' YAML key. Populated by
406 resolve_fragments; None for non-fragment states. Used for validation and debugging.
407 fragment_bindings: Parameter bindings for fragment references (YAML key 'with' on a
408 fragment state, renamed to 'fragment_bindings' by resolve_fragments). Distinct from
409 with_ (sub-loop bindings) and params (MCP tool args). Only valid when fragment_name
410 is set.
411 fragment_parameters: Parsed ParameterSpec declarations from the fragment definition's
412 'parameters:' block. Populated by resolve_fragments for validation and runtime checks.
413 """
415 action: str | None = None
416 action_type: str | None = None
417 params: dict[str, Any] = field(default_factory=dict)
418 evaluate: EvaluateConfig | None = None
419 route: RouteConfig | None = None
420 on_yes: str | None = None
421 on_no: str | None = None
422 on_error: str | None = None
423 on_partial: str | None = None
424 on_blocked: str | None = None
425 next: str | None = None
426 terminal: bool = False
427 capture: str | None = None
428 append_to_messages: str | None = None
429 timeout: int | None = None
430 on_maintain: str | None = None
431 max_retries: int | None = None
432 on_retry_exhausted: str | None = None
433 retryable_exit_codes: list[int] | None = None
434 max_rate_limit_retries: int | None = None
435 on_rate_limit_exhausted: str | None = None
436 rate_limit_backoff_base_seconds: int | None = None
437 rate_limit_max_wait_seconds: int | None = None
438 rate_limit_long_wait_ladder: list[int] | None = None
439 loop: str | None = None
440 context_passthrough: bool = False
441 with_: dict[str, Any] = field(default_factory=dict)
442 fragment_name: str | None = None
443 fragment_bindings: dict[str, Any] = field(default_factory=dict)
444 fragment_parameters: dict[str, ParameterSpec] = field(default_factory=dict)
445 agent: str | None = None
446 tools: list[str] | None = None
447 model: str | None = None
448 extra_routes: dict[str, str] = field(default_factory=dict)
449 type: str | None = None
450 throttle: ThrottleConfig | None = None
451 on_throttle_hard: str | None = None
452 learning: LearningConfig | None = None
454 def to_dict(self) -> dict[str, Any]:
455 """Convert to dictionary for JSON/YAML serialization."""
456 result: dict[str, Any] = {}
458 if self.action is not None:
459 result["action"] = self.action
460 if self.action_type is not None:
461 result["action_type"] = self.action_type
462 if self.params:
463 result["params"] = self.params
464 if self.evaluate is not None:
465 result["evaluate"] = self.evaluate.to_dict()
466 if self.route is not None:
467 result["route"] = self.route.to_dict()
468 if self.on_yes is not None:
469 result["on_yes"] = self.on_yes
470 if self.on_no is not None:
471 result["on_no"] = self.on_no
472 if self.on_error is not None:
473 result["on_error"] = self.on_error
474 if self.on_partial is not None:
475 result["on_partial"] = self.on_partial
476 if self.on_blocked is not None:
477 result["on_blocked"] = self.on_blocked
478 if self.next is not None:
479 result["next"] = self.next
480 if self.terminal:
481 result["terminal"] = self.terminal
482 if self.capture is not None:
483 result["capture"] = self.capture
484 if self.append_to_messages is not None:
485 result["append_to_messages"] = self.append_to_messages
486 if self.timeout is not None:
487 result["timeout"] = self.timeout
488 if self.on_maintain is not None:
489 result["on_maintain"] = self.on_maintain
490 if self.max_retries is not None:
491 result["max_retries"] = self.max_retries
492 if self.on_retry_exhausted is not None:
493 result["on_retry_exhausted"] = self.on_retry_exhausted
494 if self.retryable_exit_codes is not None:
495 result["retryable_exit_codes"] = self.retryable_exit_codes
496 if self.max_rate_limit_retries is not None:
497 result["max_rate_limit_retries"] = self.max_rate_limit_retries
498 if self.on_rate_limit_exhausted is not None:
499 result["on_rate_limit_exhausted"] = self.on_rate_limit_exhausted
500 if self.rate_limit_backoff_base_seconds is not None:
501 result["rate_limit_backoff_base_seconds"] = self.rate_limit_backoff_base_seconds
502 if self.rate_limit_max_wait_seconds is not None:
503 result["rate_limit_max_wait_seconds"] = self.rate_limit_max_wait_seconds
504 if self.rate_limit_long_wait_ladder is not None:
505 result["rate_limit_long_wait_ladder"] = self.rate_limit_long_wait_ladder
506 if self.loop is not None:
507 result["loop"] = self.loop
508 if self.context_passthrough:
509 result["context_passthrough"] = self.context_passthrough
510 if self.with_:
511 result["with"] = self.with_
512 if self.fragment_bindings:
513 result["fragment_bindings"] = self.fragment_bindings
514 if self.agent is not None:
515 result["agent"] = self.agent
516 if self.tools is not None:
517 result["tools"] = self.tools
518 if self.model is not None:
519 result["model"] = self.model
520 for verdict, target in self.extra_routes.items():
521 result[f"on_{verdict}"] = target
522 if self.type is not None:
523 result["type"] = self.type
524 if self.throttle is not None:
525 result["throttle"] = self.throttle.to_dict()
526 if self.on_throttle_hard is not None:
527 result["on_throttle_hard"] = self.on_throttle_hard
528 if self.learning is not None:
529 result["learning"] = self.learning.to_dict()
531 return result
533 @classmethod
534 def from_dict(cls, data: dict[str, Any]) -> StateConfig:
535 """Create from dictionary (JSON/YAML deserialization)."""
536 evaluate = None
537 if "evaluate" in data:
538 evaluate = EvaluateConfig.from_dict(data["evaluate"])
540 route = None
541 if "route" in data:
542 route = RouteConfig.from_dict(data["route"])
544 throttle = None
545 if "throttle" in data:
546 throttle = ThrottleConfig.from_dict(data["throttle"])
548 learning = None
549 if "learning" in data:
550 learning = LearningConfig.from_dict(data["learning"])
552 _known_on_keys = {
553 "on_yes",
554 "on_success",
555 "on_no",
556 "on_failure",
557 "on_error",
558 "on_partial",
559 "on_blocked",
560 "on_maintain",
561 "on_retry_exhausted",
562 "on_rate_limit_exhausted",
563 "on_throttle_hard",
564 }
565 extra_routes = {
566 key[3:]: val
567 for key, val in data.items()
568 if key.startswith("on_") and key not in _known_on_keys and isinstance(val, str)
569 }
571 return cls(
572 action=data.get("action"),
573 action_type=data.get("action_type"),
574 params=data.get("params", {}),
575 evaluate=evaluate,
576 route=route,
577 on_yes=data.get("on_yes") or data.get("on_success"),
578 on_no=data.get("on_no") or data.get("on_failure"),
579 on_error=data.get("on_error"),
580 on_partial=data.get("on_partial"),
581 on_blocked=data.get("on_blocked"),
582 next=data.get("next"),
583 terminal=data.get("terminal", False),
584 capture=data.get("capture"),
585 append_to_messages=data.get("append_to_messages"),
586 timeout=data.get("timeout"),
587 on_maintain=data.get("on_maintain"),
588 max_retries=data.get("max_retries"),
589 on_retry_exhausted=data.get("on_retry_exhausted"),
590 retryable_exit_codes=data.get("retryable_exit_codes"),
591 max_rate_limit_retries=data.get("max_rate_limit_retries"),
592 on_rate_limit_exhausted=data.get("on_rate_limit_exhausted"),
593 rate_limit_backoff_base_seconds=data.get("rate_limit_backoff_base_seconds"),
594 rate_limit_max_wait_seconds=data.get("rate_limit_max_wait_seconds"),
595 rate_limit_long_wait_ladder=data.get("rate_limit_long_wait_ladder"),
596 loop=data.get("loop"),
597 context_passthrough=data.get("context_passthrough", False),
598 with_=data.get("with", {}),
599 agent=data.get("agent"),
600 tools=data.get("tools"),
601 model=data.get("model"),
602 extra_routes=extra_routes,
603 type=data.get("type"),
604 throttle=throttle,
605 on_throttle_hard=data.get("on_throttle_hard"),
606 learning=learning,
607 fragment_name=data.get("fragment_name"),
608 fragment_bindings=data.get("fragment_bindings", {}),
609 fragment_parameters={
610 name: ParameterSpec.from_dict(spec)
611 for name, spec in data.get("fragment_parameters", {}).items()
612 },
613 )
615 def get_referenced_states(self) -> set[str]:
616 """Get all state names referenced by this state configuration.
618 Returns:
619 Set of state names that this state can transition to.
620 """
621 refs: set[str] = set()
623 if self.on_yes is not None:
624 refs.add(self.on_yes)
625 if self.on_no is not None:
626 refs.add(self.on_no)
627 if self.on_error is not None:
628 refs.add(self.on_error)
629 if self.on_partial is not None:
630 refs.add(self.on_partial)
631 if self.on_blocked is not None:
632 refs.add(self.on_blocked)
633 if self.next is not None:
634 refs.add(self.next)
635 if self.on_maintain is not None:
636 refs.add(self.on_maintain)
637 if self.on_retry_exhausted is not None:
638 refs.add(self.on_retry_exhausted)
639 if self.on_rate_limit_exhausted is not None:
640 refs.add(self.on_rate_limit_exhausted)
641 if self.on_throttle_hard is not None:
642 refs.add(self.on_throttle_hard)
643 if self.route is not None:
644 refs.update(self.route.routes.values())
645 if self.route.default is not None:
646 refs.add(self.route.default)
647 if self.route.error is not None:
648 refs.add(self.route.error)
649 refs.update(self.extra_routes.values())
651 return refs
654@dataclass
655class LLMConfig:
656 """LLM evaluation configuration.
658 Settings for the llm_structured evaluator.
660 Attributes:
661 enabled: If False, disable LLM evaluation entirely
662 model: Model identifier for LLM calls
663 max_tokens: Maximum tokens for evaluation response
664 timeout: Timeout for LLM calls in seconds
665 """
667 enabled: bool = True
668 model: str = DEFAULT_LLM_MODEL
669 max_tokens: int = 256
670 timeout: int = 1800
672 def to_dict(self) -> dict[str, Any]:
673 """Convert to dictionary for JSON/YAML serialization."""
674 result: dict[str, Any] = {}
676 if not self.enabled:
677 result["enabled"] = self.enabled
678 if self.model != DEFAULT_LLM_MODEL:
679 result["model"] = self.model
680 if self.max_tokens != 256:
681 result["max_tokens"] = self.max_tokens
682 if self.timeout != 1800:
683 result["timeout"] = self.timeout
685 return result if result else {}
687 @classmethod
688 def from_dict(cls, data: dict[str, Any]) -> LLMConfig:
689 """Create from dictionary (JSON/YAML deserialization)."""
690 return cls(
691 enabled=data.get("enabled", True),
692 model=data.get("model", DEFAULT_LLM_MODEL),
693 max_tokens=data.get("max_tokens", 256),
694 timeout=data.get("timeout", 1800),
695 )
698@dataclass
699class LoopConfigOverrides:
700 """Per-loop ll-config overrides embedded in the loop YAML definition.
702 All fields are optional (None = use global ll-config default).
703 Precedence: CLI flags > YAML config block > global ll-config > schema defaults.
705 Attributes:
706 handoff_threshold: Override for LL_HANDOFF_THRESHOLD env var (1-100)
707 readiness_threshold: Override for commands.confidence_gate.readiness_threshold (1-100)
708 outcome_threshold: Override for commands.confidence_gate.outcome_threshold (1-100)
709 max_continuations: Override for automation.max_continuations (>=1)
710 """
712 handoff_threshold: int | None = None
713 readiness_threshold: int | None = None
714 outcome_threshold: int | None = None
715 max_continuations: int | None = None
717 def to_dict(self) -> dict[str, Any]:
718 """Convert to dictionary for JSON/YAML serialization (skip-if-None)."""
719 result: dict[str, Any] = {}
721 if self.handoff_threshold is not None:
722 result["handoff_threshold"] = self.handoff_threshold
724 confidence_gate: dict[str, Any] = {}
725 if self.readiness_threshold is not None:
726 confidence_gate["readiness_threshold"] = self.readiness_threshold
727 if self.outcome_threshold is not None:
728 confidence_gate["outcome_threshold"] = self.outcome_threshold
729 if confidence_gate:
730 result["commands"] = {"confidence_gate": confidence_gate}
732 if self.max_continuations is not None:
733 result["automation"] = {"max_continuations": self.max_continuations}
735 return result
737 @classmethod
738 def from_dict(cls, data: dict[str, Any]) -> LoopConfigOverrides:
739 """Create from dictionary (JSON/YAML deserialization)."""
740 commands = data.get("commands", {})
741 confidence_gate = commands.get("confidence_gate", {}) if isinstance(commands, dict) else {}
742 automation = data.get("automation", {})
743 continuation = data.get("continuation", {})
745 max_continuations = None
746 if isinstance(automation, dict) and "max_continuations" in automation:
747 max_continuations = automation["max_continuations"]
748 elif isinstance(continuation, dict) and "max_continuations" in continuation:
749 max_continuations = continuation["max_continuations"]
751 return cls(
752 handoff_threshold=data.get("handoff_threshold"),
753 readiness_threshold=confidence_gate.get("readiness_threshold")
754 if isinstance(confidence_gate, dict)
755 else None,
756 outcome_threshold=confidence_gate.get("outcome_threshold")
757 if isinstance(confidence_gate, dict)
758 else None,
759 max_continuations=max_continuations,
760 )
763@dataclass
764class CommandEntry:
765 """A single command entry for the loop's Commands display section.
767 Attributes:
768 cmd: Full command string to display (e.g., "ll-loop run my-loop --param x=1")
769 comment: Short description shown as a comment (e.g., "run with parameter x")
770 """
772 cmd: str
773 comment: str
776@dataclass
777class TargetStateSpec:
778 """Per-state targeting specification for harness-optimize APO (ENH-1552).
780 Names a single FSM state inside a target loop file and associates it with
781 the examples file and eval fragment used during that state's optimization.
783 Attributes:
784 name: State name within the target loop
785 examples_file: Path to the examples YAML file for this state
786 eval_fragment: Eval fragment identifier used during optimization
787 """
789 name: str
790 examples_file: str
791 eval_fragment: str
793 def to_dict(self) -> dict[str, Any]:
794 """Convert to dictionary for JSON/YAML serialization."""
795 return {
796 "name": self.name,
797 "examples_file": self.examples_file,
798 "eval": self.eval_fragment,
799 }
801 @classmethod
802 def from_dict(cls, data: dict[str, Any]) -> TargetStateSpec:
803 """Create from dictionary (JSON/YAML deserialization)."""
804 return cls(
805 name=data["name"],
806 examples_file=data["examples_file"],
807 eval_fragment=data["eval"],
808 )
811@dataclass
812class TargetFileSpec:
813 """Per-file targeting specification for harness-optimize APO (ENH-1552).
815 Associates a loop YAML file (or glob pattern) with the list of states
816 to optimize within that file.
818 Attributes:
819 file: Explicit path to a loop YAML file (mutually exclusive with glob)
820 glob: Glob pattern matching one or more loop YAML files
821 states: States within the matched file(s) to target
822 """
824 file: str | None = None
825 glob: str | None = None
826 states: list[TargetStateSpec] = field(default_factory=list)
828 def to_dict(self) -> dict[str, Any]:
829 """Convert to dictionary for JSON/YAML serialization."""
830 result: dict[str, Any] = {}
831 if self.file is not None:
832 result["file"] = self.file
833 if self.glob is not None:
834 result["glob"] = self.glob
835 if self.states:
836 result["states"] = [s.to_dict() for s in self.states]
837 return result
839 @classmethod
840 def from_dict(cls, data: dict[str, Any]) -> TargetFileSpec:
841 """Create from dictionary (JSON/YAML deserialization)."""
842 return cls(
843 file=data.get("file"),
844 glob=data.get("glob"),
845 states=[TargetStateSpec.from_dict(s) for s in (data.get("states") or [])],
846 )
849@dataclass
850class RepeatedFailureConfig:
851 """Configuration for the FSM stall detector (FEAT-1637).
853 When N consecutive iterations produce an identical
854 `(state_name, exit_code, eval_verdict)` triple, the FSM either
855 aborts the run or routes to a configured recovery state.
857 Attributes:
858 window: Consecutive iterations with identical triple required to
859 fire (default 3).
860 on_repeated_failure: Either the literal ``"abort"`` (terminate
861 with ``terminated_by="stall_detected"``) or the name of a
862 declared state to route to.
863 """
865 window: int = 3
866 on_repeated_failure: str = "abort"
867 progress_paths: list[str] = field(default_factory=list)
868 exclude_paths: list[str] = field(default_factory=list)
870 def to_dict(self) -> dict[str, Any]:
871 """Convert to dictionary for JSON/YAML serialization (skip-if-default)."""
872 result: dict[str, Any] = {}
873 if self.window != 3:
874 result["window"] = self.window
875 if self.on_repeated_failure != "abort":
876 result["on_repeated_failure"] = self.on_repeated_failure
877 if self.progress_paths:
878 result["progress_paths"] = self.progress_paths
879 if self.exclude_paths:
880 result["exclude_paths"] = self.exclude_paths
881 return result
883 @classmethod
884 def from_dict(cls, data: dict[str, Any]) -> RepeatedFailureConfig:
885 """Create from dictionary (JSON/YAML deserialization)."""
886 return cls(
887 window=data.get("window", 3),
888 on_repeated_failure=data.get("on_repeated_failure", "abort"),
889 progress_paths=data.get("progress_paths", []),
890 exclude_paths=data.get("exclude_paths", []),
891 )
894@dataclass
895class CircuitConfig:
896 """Top-level ``circuit:`` block grouping loop-level safety knobs.
898 Currently exposes ``repeated_failure`` (the stall detector). Future
899 safety knobs (e.g. global timeouts, panic-stop guards) should be
900 added here rather than as separate top-level keys.
902 Attributes:
903 repeated_failure: Stall-detector configuration, or None to disable.
904 """
906 repeated_failure: RepeatedFailureConfig | None = None
908 def to_dict(self) -> dict[str, Any]:
909 """Convert to dictionary for JSON/YAML serialization (skip-if-None)."""
910 result: dict[str, Any] = {}
911 if self.repeated_failure is not None:
912 rf_dict = self.repeated_failure.to_dict()
913 # Always emit the key when configured (even if all fields are defaults)
914 # so the round-trip preserves "repeated_failure was set" intent.
915 result["repeated_failure"] = rf_dict
916 return result
918 @classmethod
919 def from_dict(cls, data: dict[str, Any]) -> CircuitConfig:
920 """Create from dictionary (JSON/YAML deserialization)."""
921 rf = None
922 if "repeated_failure" in data and data["repeated_failure"] is not None:
923 rf = RepeatedFailureConfig.from_dict(data["repeated_failure"])
924 return cls(repeated_failure=rf)
927@dataclass
928class FSMLoop:
929 """Complete FSM loop definition.
931 The main dataclass representing a loop configuration.
933 Attributes:
934 name: Unique loop identifier
935 initial: Starting state name
936 states: Mapping from state name to StateConfig
937 context: User-defined shared variables
938 scope: Paths this loop operates on (for concurrency control). Supports ${context.<var>} template variables that are resolved at runtime.
939 max_iterations: Safety limit for loop iterations
940 backoff: Seconds between iterations
941 timeout: Max total runtime in seconds (loop-level)
942 maintain: If True, restart after completion
943 llm: LLM evaluation configuration
944 on_handoff: Behavior when handoff signal detected (pause/spawn/terminate)
945 commands: Optional override for the Commands section in ll-loop show
946 """
948 name: str
949 initial: str
950 states: dict[str, StateConfig]
951 description: str | None = None
952 context: dict[str, Any] = field(default_factory=dict)
953 parameters: dict[str, ParameterSpec] = field(default_factory=dict)
954 scope: list[str] = field(default_factory=list)
955 max_iterations: int = 50
956 on_max_iterations: str | None = None
957 max_edge_revisits: int = 100
958 backoff: float | None = None
959 timeout: int | None = None
960 default_timeout: int | None = None
961 maintain: bool = False
962 llm: LLMConfig = field(default_factory=LLMConfig)
963 on_handoff: Literal["pause", "spawn", "terminate"] = "pause"
964 input_key: str = "input"
965 config: LoopConfigOverrides | None = None
966 category: str = ""
967 labels: list[str] = field(default_factory=list)
968 # Audience axis for `ll-loop list` filtering (orthogonal to `category`, which
969 # is topical). "public" = user-facing (default); "internal" = delegated-only
970 # sub-loop, never run directly; "example" = demo/template. See ENH (loop-list
971 # visibility tiers).
972 visibility: str = "public"
973 required_inputs: list[str] = field(default_factory=list)
974 commands: list[CommandEntry] = field(default_factory=list)
975 targets: list[TargetFileSpec] = field(default_factory=list)
976 circuit: CircuitConfig | None = None
977 meta_self_eval_ok: bool = False
978 shared_state_ok: bool = False
979 partial_route_ok: bool = False
980 artifact_versioning: bool = False
981 artifact_versioning_ok: bool = False
982 generator_fix_ok: bool = False
983 # Populated from the raw `import:` list by from_dict(); not serialized by to_dict()
984 imports: list[str] = field(default_factory=list)
986 def to_dict(self) -> dict[str, Any]:
987 """Convert to dictionary for JSON/YAML serialization."""
988 result: dict[str, Any] = {
989 "name": self.name,
990 "initial": self.initial,
991 "states": {name: state.to_dict() for name, state in self.states.items()},
992 }
994 if self.description is not None:
995 result["description"] = self.description
996 if self.context:
997 result["context"] = self.context
998 if self.parameters:
999 result["parameters"] = {name: spec.to_dict() for name, spec in self.parameters.items()}
1000 if self.scope:
1001 result["scope"] = self.scope
1002 if self.max_iterations != 50:
1003 result["max_iterations"] = self.max_iterations
1004 if self.max_edge_revisits != 100:
1005 result["max_edge_revisits"] = self.max_edge_revisits
1006 if self.backoff is not None:
1007 result["backoff"] = self.backoff
1008 if self.timeout is not None:
1009 result["timeout"] = self.timeout
1010 if self.default_timeout is not None:
1011 result["default_timeout"] = self.default_timeout
1012 if self.maintain:
1013 result["maintain"] = self.maintain
1014 if self.on_handoff != "pause":
1015 result["on_handoff"] = self.on_handoff
1016 if self.on_max_iterations is not None:
1017 result["on_max_iterations"] = self.on_max_iterations
1018 if self.input_key != "input":
1019 result["input_key"] = self.input_key
1021 llm_dict = self.llm.to_dict()
1022 if llm_dict:
1023 result["llm"] = llm_dict
1025 if self.config is not None:
1026 config_dict = self.config.to_dict()
1027 if config_dict:
1028 result["config"] = config_dict
1030 if self.category:
1031 result["category"] = self.category
1032 if self.labels:
1033 result["labels"] = self.labels
1034 if self.visibility and self.visibility != "public":
1035 result["visibility"] = self.visibility
1036 if self.required_inputs:
1037 result["required_inputs"] = self.required_inputs
1038 if self.commands:
1039 result["commands"] = [{"cmd": e.cmd, "comment": e.comment} for e in self.commands]
1040 if self.targets:
1041 result["targets"] = [t.to_dict() for t in self.targets]
1043 if self.circuit is not None:
1044 circuit_dict = self.circuit.to_dict()
1045 if circuit_dict:
1046 result["circuit"] = circuit_dict
1048 if self.meta_self_eval_ok:
1049 result["meta_self_eval_ok"] = self.meta_self_eval_ok
1050 if self.shared_state_ok:
1051 result["shared_state_ok"] = self.shared_state_ok
1052 if self.partial_route_ok:
1053 result["partial_route_ok"] = self.partial_route_ok
1054 if self.artifact_versioning:
1055 result["artifact_versioning"] = self.artifact_versioning
1056 if self.artifact_versioning_ok:
1057 result["artifact_versioning_ok"] = self.artifact_versioning_ok
1058 if self.generator_fix_ok:
1059 result["generator_fix_ok"] = self.generator_fix_ok
1061 return result
1063 @classmethod
1064 def from_dict(cls, data: dict[str, Any]) -> FSMLoop:
1065 """Create from dictionary (JSON/YAML deserialization)."""
1066 states = {
1067 name: StateConfig.from_dict(state_data)
1068 for name, state_data in data.get("states", {}).items()
1069 }
1071 llm = LLMConfig()
1072 if "llm" in data:
1073 llm = LLMConfig.from_dict(data["llm"])
1075 loop_config = None
1076 if "config" in data:
1077 loop_config = LoopConfigOverrides.from_dict(data["config"])
1079 circuit = None
1080 if "circuit" in data and data["circuit"] is not None:
1081 circuit = CircuitConfig.from_dict(data["circuit"])
1083 parameters = {
1084 name: ParameterSpec.from_dict(spec) for name, spec in data.get("parameters", {}).items()
1085 }
1087 return cls(
1088 name=data["name"],
1089 initial=data["initial"],
1090 states=states,
1091 description=data.get("description"),
1092 context=data.get("context", {}),
1093 parameters=parameters,
1094 scope=data.get("scope", []),
1095 max_iterations=data.get("max_iterations", 50),
1096 on_max_iterations=data.get("on_max_iterations"),
1097 max_edge_revisits=data.get("max_edge_revisits", 100),
1098 backoff=data.get("backoff"),
1099 timeout=data.get("timeout"),
1100 default_timeout=data.get("default_timeout"),
1101 maintain=data.get("maintain", False),
1102 llm=llm,
1103 on_handoff=data.get("on_handoff", "pause"),
1104 input_key=data.get("input_key", "input"),
1105 config=loop_config,
1106 category=data.get("category", ""),
1107 labels=data.get("labels", []),
1108 visibility=data.get("visibility", "public"),
1109 required_inputs=data.get("required_inputs", []),
1110 commands=[CommandEntry(**e) for e in data.get("commands", [])],
1111 targets=[TargetFileSpec.from_dict(t) for t in (data.get("targets") or [])],
1112 circuit=circuit,
1113 meta_self_eval_ok=data.get("meta_self_eval_ok", False),
1114 shared_state_ok=data.get("shared_state_ok", False),
1115 partial_route_ok=data.get("partial_route_ok", False),
1116 artifact_versioning=data.get("artifact_versioning", False),
1117 artifact_versioning_ok=data.get("artifact_versioning_ok", False),
1118 generator_fix_ok=data.get("generator_fix_ok", False),
1119 imports=data.get("import", []),
1120 )
1122 def get_all_state_names(self) -> set[str]:
1123 """Get all defined state names.
1125 Returns:
1126 Set of all state names in this FSM.
1127 """
1128 return set(self.states.keys())
1130 def get_terminal_states(self) -> set[str]:
1131 """Get all terminal state names.
1133 Returns:
1134 Set of state names where terminal=True.
1135 """
1136 return {name for name, state in self.states.items() if state.terminal}
1138 def get_all_referenced_states(self) -> set[str]:
1139 """Get all state names referenced by transitions.
1141 This includes the initial state and all states referenced
1142 in routing configurations.
1144 Returns:
1145 Set of all referenced state names.
1146 """
1147 refs: set[str] = {self.initial}
1148 for state in self.states.values():
1149 refs.update(state.get_referenced_states())
1150 if self.on_max_iterations is not None:
1151 refs.add(self.on_max_iterations)
1152 return refs