engine.py
python
sha256:7011e00115e9c74d24569fed2caec6a2a6ef8fdb070d3b4715ce06e6633aaa47
feat(merge): add --explain flag with per-path decision trac…
Sonnet 4.6
minor
⚠ breaking
36 days ago
| 1 | """muse harmony engine — three-tier resolution intelligence pipeline. |
| 2 | |
| 3 | The engine evaluates incoming :class:`~muse.core.harmony.ConflictPattern` |
| 4 | instances against the harmony store and produces an :class:`EngineResult` |
| 5 | indicating what action to take. The four tiers are evaluated in order and |
| 6 | the first matching tier wins: |
| 7 | |
| 8 | Tier 1 — **Policy match** |
| 9 | Load all :class:`~muse.core.harmony.Policy` entries (scope-sorted: |
| 10 | workspace → repo → domain → file) and evaluate each condition against |
| 11 | the pattern. The first matching policy determines the action: |
| 12 | |
| 13 | - ``PREFER_OURS`` / ``PREFER_THEIRS`` at or above |
| 14 | :attr:`EngineConfig.auto_apply_threshold` → ``status="applied"``. |
| 15 | - Same actions *below* threshold → ``status="proposed"`` with |
| 16 | ``requires_confirmation=True``. |
| 17 | - ``ESCALATE`` / ``REQUIRE_HUMAN`` / ``DELEGATE`` → ``status="escalated"``. |
| 18 | |
| 19 | Tier 2 — **Exact replay** |
| 20 | Look up the best :class:`~muse.core.harmony.Resolution` for the pattern |
| 21 | (ranked by ``human_verified`` > ``confidence`` > ``applied_count``): |
| 22 | |
| 23 | - ``human_verified=True`` OR ``confidence`` ≥ threshold → ``status="applied"``, |
| 24 | :func:`~muse.core.harmony.increment_applied_count` called atomically. |
| 25 | - Below threshold → ``status="proposed"`` (``requires_confirmation=True``). |
| 26 | |
| 27 | Tier 3 — **Semantic match** |
| 28 | Call :func:`find_similar` to discover patterns with semantically similar |
| 29 | fingerprints. The :class:`HarmonyPlugin` ``similarity()`` method scores |
| 30 | each pair; matches at or above :attr:`EngineConfig.semantic_threshold` |
| 31 | are returned as proposals with ``requires_confirmation=True``. Semantic |
| 32 | matches are always proposed — never auto-applied — because they are |
| 33 | structural analogies, not exact replays. |
| 34 | |
| 35 | Tier 4 — **Escalate** |
| 36 | No match found in any tier. Returns ``status="escalated"`` with a |
| 37 | human-readable reason string. An audit entry of type |
| 38 | :attr:`~muse.core.harmony.AuditEventType.ESCALATION_RECORDED` is written. |
| 39 | |
| 40 | Plugin protocol |
| 41 | --------------- |
| 42 | |
| 43 | :class:`HarmonyPlugin` is a structural :class:`typing.Protocol` — any class |
| 44 | with a ``similarity(fp_a, fp_b) -> float`` method satisfies it without |
| 45 | explicit inheritance. :class:`DefaultPlugin` is the no-op baseline: it |
| 46 | returns ``1.0`` for identical fingerprints and ``0.0`` for all others, |
| 47 | which degrades cleanly to exact-fingerprint semantic matching. |
| 48 | |
| 49 | Plugin errors (exceptions from ``similarity()``) are caught and logged; |
| 50 | the engine falls back to :class:`DefaultPlugin` behaviour for that pair. |
| 51 | Plugin-returned similarity values outside ``[0.0, 1.0]`` are clamped before |
| 52 | comparison. |
| 53 | |
| 54 | Audit |
| 55 | ----- |
| 56 | |
| 57 | The engine writes an audit entry for every resolution applied or escalated. |
| 58 | The :data:`~muse.core.harmony.AuditEventType.RESOLUTION_APPLIED` event |
| 59 | records which resolution was replayed; the |
| 60 | :data:`~muse.core.harmony.AuditEventType.ESCALATION_RECORDED` event records |
| 61 | why no automatic resolution was possible. |
| 62 | |
| 63 | Thread safety |
| 64 | ------------- |
| 65 | |
| 66 | All writes delegate to :mod:`muse.core.harmony` functions which use atomic |
| 67 | ``os.replace`` — safe for concurrent engine calls on the same repo. |
| 68 | """ |
| 69 | |
| 70 | import logging |
| 71 | import pathlib |
| 72 | from dataclasses import dataclass, field |
| 73 | from typing import Protocol, runtime_checkable |
| 74 | |
| 75 | from muse.core.types import JsonObject, JsonValue, Manifest, short_id |
| 76 | from muse.core.validation import validate_object_id |
| 77 | |
| 78 | from .audit import append_audit |
| 79 | from .fingerprint import ( |
| 80 | _now_utc, |
| 81 | blob_fingerprint, |
| 82 | compute_pattern_id, |
| 83 | compute_resolution_id, |
| 84 | compute_semantic_fingerprint, |
| 85 | ) |
| 86 | from .patterns import list_patterns, load_pattern, record_pattern |
| 87 | from .policies import list_policies, match_policy |
| 88 | from .resolutions import best_resolution, increment_applied_count, list_resolutions, save_resolution |
| 89 | from .types import ( |
| 90 | AgentProvenance, |
| 91 | AuditEventType, |
| 92 | ConflictPattern, |
| 93 | ConflictType, |
| 94 | PolicyAction, |
| 95 | Resolution, |
| 96 | ResolutionProposal, |
| 97 | ResolutionStrategy, |
| 98 | ) |
| 99 | |
| 100 | logger = logging.getLogger(__name__) |
| 101 | |
| 102 | # --------------------------------------------------------------------------- |
| 103 | # Open string-constant namespace |
| 104 | # --------------------------------------------------------------------------- |
| 105 | |
| 106 | class EngineStatus: |
| 107 | """Possible outcomes of a :func:`resolve` call. |
| 108 | |
| 109 | ``APPLIED`` — a resolution was auto-applied (confidence ≥ threshold or |
| 110 | human_verified, or policy fired at or above threshold). |
| 111 | ``PROPOSED`` — a candidate was found but requires confirmation before |
| 112 | applying (confidence below threshold, or semantic match). |
| 113 | ``ESCALATED`` — no match found in any tier; human or specialist-agent |
| 114 | intervention is required. |
| 115 | """ |
| 116 | |
| 117 | APPLIED = "applied" |
| 118 | PROPOSED = "proposed" |
| 119 | ESCALATED = "escalated" |
| 120 | |
| 121 | # --------------------------------------------------------------------------- |
| 122 | # Plugin protocol |
| 123 | # --------------------------------------------------------------------------- |
| 124 | |
| 125 | @runtime_checkable |
| 126 | class HarmonyPlugin(Protocol): |
| 127 | """Structural protocol for domain-specific similarity scoring. |
| 128 | |
| 129 | Any class implementing ``similarity(fp_a, fp_b) -> float`` satisfies this |
| 130 | protocol without explicit inheritance. The method must return a value in |
| 131 | ``[0.0, 1.0]``; values outside this range are clamped by the engine. |
| 132 | |
| 133 | Domain plugins (MIDI, code, …) override ``similarity`` to implement |
| 134 | richer comparisons — e.g. structural graph edit distance on AST |
| 135 | fingerprints, or pitch-class vector cosine similarity for MIDI tracks. |
| 136 | """ |
| 137 | |
| 138 | def similarity(self, fp_a: str, fp_b: str) -> float: |
| 139 | """Return a similarity score in ``[0.0, 1.0]`` between two fingerprints.""" |
| 140 | ... |
| 141 | |
| 142 | class DefaultPlugin: |
| 143 | """No-op plugin — exact fingerprint match only. |
| 144 | |
| 145 | Returns ``1.0`` when the two fingerprints are identical (same semantic |
| 146 | content), ``0.0`` otherwise. This degrades semantic matching to exact |
| 147 | fingerprint replay, which is always safe. |
| 148 | """ |
| 149 | |
| 150 | def similarity(self, fp_a: str, fp_b: str) -> float: |
| 151 | """Return 1.0 for identical fingerprints, 0.0 for all others.""" |
| 152 | return 1.0 if fp_a == fp_b else 0.0 |
| 153 | |
| 154 | # --------------------------------------------------------------------------- |
| 155 | # Engine configuration |
| 156 | # --------------------------------------------------------------------------- |
| 157 | |
| 158 | @dataclass(frozen=True) |
| 159 | class EngineConfig: |
| 160 | """Thresholds and limits that govern engine behaviour. |
| 161 | |
| 162 | ``auto_apply_threshold`` |
| 163 | Minimum confidence required to auto-apply a resolution without |
| 164 | confirmation. Resolutions below this threshold are *proposed*. |
| 165 | Human-verified resolutions bypass this check and are always applied. |
| 166 | |
| 167 | ``semantic_threshold`` |
| 168 | Minimum :class:`HarmonyPlugin` similarity score for a pattern to be |
| 169 | considered a semantic match. Pairs below this threshold are ignored. |
| 170 | |
| 171 | ``max_proposals`` |
| 172 | Maximum number of semantic proposals returned by :func:`find_similar`. |
| 173 | Higher values increase recall at the cost of latency on large stores. |
| 174 | """ |
| 175 | |
| 176 | auto_apply_threshold: float = 0.85 |
| 177 | semantic_threshold: float = 0.70 |
| 178 | max_proposals: int = 5 |
| 179 | |
| 180 | # --------------------------------------------------------------------------- |
| 181 | # Engine result |
| 182 | # --------------------------------------------------------------------------- |
| 183 | |
| 184 | @dataclass(frozen=True) |
| 185 | class EngineResult: |
| 186 | """Outcome of a single :func:`resolve` call. |
| 187 | |
| 188 | ``status`` |
| 189 | One of :class:`EngineStatus` constants — ``"applied"``, ``"proposed"``, |
| 190 | or ``"escalated"``. |
| 191 | |
| 192 | ``pattern_id`` |
| 193 | The ``pattern_id`` of the :class:`~muse.core.harmony.ConflictPattern` |
| 194 | that was evaluated. |
| 195 | |
| 196 | ``proposal`` |
| 197 | Populated for ``"applied"`` and ``"proposed"`` statuses. Contains the |
| 198 | :class:`~muse.core.harmony.ResolutionProposal` that the engine produced. |
| 199 | ``None`` for ``"escalated"``. |
| 200 | |
| 201 | ``applied_resolution_id`` |
| 202 | Set when an existing :class:`~muse.core.harmony.Resolution` was replayed |
| 203 | exactly (exact-replay tier). ``None`` for policy-driven applications |
| 204 | (no pre-existing resolution) and all other statuses. |
| 205 | |
| 206 | ``escalation_reason`` |
| 207 | Human-readable explanation populated for ``"escalated"`` results. |
| 208 | ``None`` for other statuses. |
| 209 | """ |
| 210 | |
| 211 | status: str |
| 212 | pattern_id: str |
| 213 | proposal: ResolutionProposal | None = None |
| 214 | applied_resolution_id: str | None = None |
| 215 | escalation_reason: str | None = None |
| 216 | |
| 217 | # --------------------------------------------------------------------------- |
| 218 | # find_similar |
| 219 | # --------------------------------------------------------------------------- |
| 220 | |
| 221 | def find_similar( |
| 222 | root: pathlib.Path, |
| 223 | pattern: ConflictPattern, |
| 224 | *, |
| 225 | plugin: HarmonyPlugin | None = None, |
| 226 | config: EngineConfig | None = None, |
| 227 | ) -> list[ResolutionProposal]: |
| 228 | """Search the harmony store for patterns semantically similar to *pattern*. |
| 229 | |
| 230 | Iterates all stored patterns, calls *plugin*``similarity()`` on each pair |
| 231 | of semantic fingerprints, and returns proposals for patterns whose |
| 232 | similarity score meets :attr:`EngineConfig.semantic_threshold` and that |
| 233 | have at least one saved resolution. |
| 234 | |
| 235 | Results are sorted by ``confidence`` descending (best proposal first) and |
| 236 | capped at :attr:`EngineConfig.max_proposals`. |
| 237 | |
| 238 | Args: |
| 239 | root: Repository root. |
| 240 | pattern: The incoming conflict pattern to match against. |
| 241 | plugin: Domain plugin for similarity scoring. Defaults to |
| 242 | :class:`DefaultPlugin`. |
| 243 | config: Engine configuration. Defaults to :class:`EngineConfig`. |
| 244 | |
| 245 | Returns: |
| 246 | List of :class:`~muse.core.harmony.ResolutionProposal` instances, |
| 247 | sorted by confidence descending. Empty if no matches exceed the |
| 248 | threshold. |
| 249 | """ |
| 250 | cfg = config or EngineConfig() |
| 251 | plug = plugin or DefaultPlugin() |
| 252 | |
| 253 | try: |
| 254 | all_patterns = list_patterns(root) |
| 255 | except Exception: |
| 256 | return [] |
| 257 | |
| 258 | proposals: list[ResolutionProposal] = [] |
| 259 | |
| 260 | for p in all_patterns: |
| 261 | if p.pattern_id == pattern.pattern_id: |
| 262 | continue # skip self |
| 263 | |
| 264 | try: |
| 265 | raw_sim = plug.similarity(pattern.semantic_fingerprint, p.semantic_fingerprint) |
| 266 | sim = max(0.0, min(1.0, float(raw_sim))) |
| 267 | except Exception as exc: |
| 268 | logger.warning("⚠️ harmony engine: plugin.similarity() raised %s — skipping", exc) |
| 269 | sim = 0.0 |
| 270 | |
| 271 | if sim < cfg.semantic_threshold: |
| 272 | continue |
| 273 | |
| 274 | best = best_resolution(root, p.pattern_id) |
| 275 | if best is None: |
| 276 | continue |
| 277 | |
| 278 | proposals.append( |
| 279 | ResolutionProposal( |
| 280 | pattern_id=pattern.pattern_id, |
| 281 | strategy=ResolutionStrategy.SEMANTIC_PROPOSAL, |
| 282 | proposed_action=PolicyAction.PREFER_OURS, |
| 283 | confidence=round(best.confidence * sim, 4), |
| 284 | rationale=( |
| 285 | f"Semantically similar to pattern {p.pattern_id} " |
| 286 | f"(similarity={sim:.2f}): {best.rationale}" |
| 287 | ), |
| 288 | similar_pattern_id=p.pattern_id, |
| 289 | similarity=sim, |
| 290 | requires_confirmation=True, |
| 291 | ) |
| 292 | ) |
| 293 | |
| 294 | proposals.sort(key=lambda pr: pr.confidence, reverse=True) |
| 295 | return proposals[: cfg.max_proposals] |
| 296 | |
| 297 | # --------------------------------------------------------------------------- |
| 298 | # Core engine |
| 299 | # --------------------------------------------------------------------------- |
| 300 | |
| 301 | def resolve( |
| 302 | root: pathlib.Path, |
| 303 | pattern: ConflictPattern, |
| 304 | config: EngineConfig | None = None, |
| 305 | *, |
| 306 | plugin: HarmonyPlugin | None = None, |
| 307 | actor: AgentProvenance | None = None, |
| 308 | ) -> EngineResult: |
| 309 | """Run the three-tier resolution pipeline for *pattern*. |
| 310 | |
| 311 | Evaluates tiers in order — policy → exact replay → semantic → escalate — |
| 312 | and returns the first successful outcome. |
| 313 | |
| 314 | Args: |
| 315 | root: Repository root (need not contain the pattern — the engine |
| 316 | escalates gracefully if the store is absent or empty). |
| 317 | pattern: Incoming conflict pattern to resolve. |
| 318 | config: Engine configuration overrides. Defaults to |
| 319 | :class:`EngineConfig`. |
| 320 | plugin: Domain plugin for semantic similarity. Defaults to |
| 321 | :class:`DefaultPlugin`. |
| 322 | actor: Attribution for audit entries. Defaults to |
| 323 | :func:`~muse.core.harmony.AgentProvenance.human`. |
| 324 | |
| 325 | Returns: |
| 326 | :class:`EngineResult` with ``status``, ``proposal``, and optional |
| 327 | ``applied_resolution_id`` / ``escalation_reason``. |
| 328 | """ |
| 329 | cfg = config or EngineConfig() |
| 330 | plug = plugin or DefaultPlugin() |
| 331 | act = actor or AgentProvenance.human() |
| 332 | |
| 333 | # ------------------------------------------------------------------ |
| 334 | # Tier 1 — Policy |
| 335 | # ------------------------------------------------------------------ |
| 336 | try: |
| 337 | policies = list_policies(root) |
| 338 | matched = match_policy(policies, pattern) |
| 339 | except Exception as exc: |
| 340 | logger.warning("⚠️ harmony engine: policy load failed (%s) — skipping tier 1", exc) |
| 341 | matched = None |
| 342 | |
| 343 | if matched is not None: |
| 344 | action = matched.action |
| 345 | confidence = matched.confidence |
| 346 | |
| 347 | if action in (PolicyAction.PREFER_OURS, PolicyAction.PREFER_THEIRS): |
| 348 | proposal = ResolutionProposal( |
| 349 | pattern_id=pattern.pattern_id, |
| 350 | strategy=ResolutionStrategy.POLICY, |
| 351 | proposed_action=action, |
| 352 | confidence=confidence, |
| 353 | rationale=( |
| 354 | f"Policy '{matched.policy_id}' fired: {matched.description}" |
| 355 | ), |
| 356 | policy_id=matched.policy_id, |
| 357 | requires_confirmation=confidence < cfg.auto_apply_threshold, |
| 358 | ) |
| 359 | if confidence >= cfg.auto_apply_threshold: |
| 360 | _write_applied_audit(root, pattern.pattern_id, None, act) |
| 361 | return EngineResult( |
| 362 | status=EngineStatus.APPLIED, |
| 363 | pattern_id=pattern.pattern_id, |
| 364 | proposal=proposal, |
| 365 | ) |
| 366 | else: |
| 367 | return EngineResult( |
| 368 | status=EngineStatus.PROPOSED, |
| 369 | pattern_id=pattern.pattern_id, |
| 370 | proposal=proposal, |
| 371 | ) |
| 372 | |
| 373 | # ESCALATE / REQUIRE_HUMAN / DELEGATE → escalate |
| 374 | if action == PolicyAction.DELEGATE: |
| 375 | reason = ( |
| 376 | f"Policy '{matched.policy_id}' delegates to agent " |
| 377 | f"'{matched.delegate_to or 'unknown'}'" |
| 378 | ) |
| 379 | elif action == PolicyAction.REQUIRE_HUMAN: |
| 380 | reason = f"Policy '{matched.policy_id}' requires human resolution" |
| 381 | else: |
| 382 | target = matched.escalate_to or "unspecified" |
| 383 | reason = ( |
| 384 | f"Policy '{matched.policy_id}' requires escalation to {target}" |
| 385 | ) |
| 386 | _write_escalation_audit(root, pattern.pattern_id, act, reason) |
| 387 | return EngineResult( |
| 388 | status=EngineStatus.ESCALATED, |
| 389 | pattern_id=pattern.pattern_id, |
| 390 | escalation_reason=reason, |
| 391 | ) |
| 392 | |
| 393 | # ------------------------------------------------------------------ |
| 394 | # Tier 2 — Exact replay |
| 395 | # ------------------------------------------------------------------ |
| 396 | try: |
| 397 | best = best_resolution(root, pattern.pattern_id) |
| 398 | except Exception as exc: |
| 399 | logger.warning("⚠️ harmony engine: exact-replay lookup failed (%s)", exc) |
| 400 | best = None |
| 401 | |
| 402 | if best is not None: |
| 403 | proposal = ResolutionProposal( |
| 404 | pattern_id=pattern.pattern_id, |
| 405 | strategy=ResolutionStrategy.EXACT_REPLAY, |
| 406 | proposed_action=PolicyAction.PREFER_OURS, |
| 407 | confidence=best.confidence, |
| 408 | rationale=f"Exact replay of resolution {best.resolution_id}", |
| 409 | requires_confirmation=not ( |
| 410 | best.human_verified or best.confidence >= cfg.auto_apply_threshold |
| 411 | ), |
| 412 | ) |
| 413 | if best.human_verified or best.confidence >= cfg.auto_apply_threshold: |
| 414 | try: |
| 415 | increment_applied_count(root, pattern.pattern_id, best.resolution_id) |
| 416 | except Exception as exc: |
| 417 | logger.warning("⚠️ harmony engine: increment_applied_count failed: %s", exc) |
| 418 | _write_applied_audit(root, pattern.pattern_id, best.resolution_id, act) |
| 419 | return EngineResult( |
| 420 | status=EngineStatus.APPLIED, |
| 421 | pattern_id=pattern.pattern_id, |
| 422 | proposal=proposal, |
| 423 | applied_resolution_id=best.resolution_id, |
| 424 | ) |
| 425 | else: |
| 426 | return EngineResult( |
| 427 | status=EngineStatus.PROPOSED, |
| 428 | pattern_id=pattern.pattern_id, |
| 429 | proposal=proposal, |
| 430 | ) |
| 431 | |
| 432 | # ------------------------------------------------------------------ |
| 433 | # Tier 3 — Semantic match |
| 434 | # ------------------------------------------------------------------ |
| 435 | try: |
| 436 | similar = find_similar(root, pattern, plugin=plug, config=cfg) |
| 437 | except Exception as exc: |
| 438 | logger.warning("⚠️ harmony engine: semantic search failed (%s)", exc) |
| 439 | similar = [] |
| 440 | |
| 441 | if similar: |
| 442 | best_proposal = similar[0] |
| 443 | return EngineResult( |
| 444 | status=EngineStatus.PROPOSED, |
| 445 | pattern_id=pattern.pattern_id, |
| 446 | proposal=best_proposal, |
| 447 | ) |
| 448 | |
| 449 | # ------------------------------------------------------------------ |
| 450 | # Tier 4 — Escalate |
| 451 | # ------------------------------------------------------------------ |
| 452 | reason = ( |
| 453 | f"No policy, exact resolution, or semantic match found for " |
| 454 | f"pattern {pattern.pattern_id} " |
| 455 | f"(path='{pattern.path}', domain='{pattern.domain}')" |
| 456 | ) |
| 457 | _write_escalation_audit(root, pattern.pattern_id, act, reason) |
| 458 | return EngineResult( |
| 459 | status=EngineStatus.ESCALATED, |
| 460 | pattern_id=pattern.pattern_id, |
| 461 | escalation_reason=reason, |
| 462 | ) |
| 463 | |
| 464 | # --------------------------------------------------------------------------- |
| 465 | # Audit helpers |
| 466 | # --------------------------------------------------------------------------- |
| 467 | |
| 468 | def _write_applied_audit( |
| 469 | root: pathlib.Path, |
| 470 | pattern_id: str, |
| 471 | resolution_id: str | None, |
| 472 | actor: AgentProvenance, |
| 473 | ) -> None: |
| 474 | """Write a RESOLUTION_APPLIED audit entry, swallowing any IO errors.""" |
| 475 | try: |
| 476 | append_audit( |
| 477 | root, |
| 478 | AuditEventType.RESOLUTION_APPLIED, |
| 479 | actor, |
| 480 | pattern_id=pattern_id, |
| 481 | resolution_id=resolution_id, |
| 482 | ) |
| 483 | except Exception as exc: |
| 484 | logger.warning("⚠️ harmony engine: audit write (applied) failed: %s", exc) |
| 485 | |
| 486 | def _write_escalation_audit( |
| 487 | root: pathlib.Path, |
| 488 | pattern_id: str, |
| 489 | actor: AgentProvenance, |
| 490 | reason: str, |
| 491 | ) -> None: |
| 492 | """Write an ESCALATION_RECORDED audit entry, swallowing any IO errors.""" |
| 493 | try: |
| 494 | append_audit( |
| 495 | root, |
| 496 | AuditEventType.ESCALATION_RECORDED, |
| 497 | actor, |
| 498 | pattern_id=pattern_id, |
| 499 | metadata={"reason": reason}, |
| 500 | ) |
| 501 | except Exception as exc: |
| 502 | logger.warning("⚠️ harmony engine: audit write (escalation) failed: %s", exc) |
| 503 | |
| 504 | # --------------------------------------------------------------------------- |
| 505 | # Config helpers |
| 506 | # --------------------------------------------------------------------------- |
| 507 | |
| 508 | def _read_config_float(root: pathlib.Path, key: str, default: float) -> float: |
| 509 | """Read a float from .muse/config.toml by dotted key; return *default* if absent. |
| 510 | |
| 511 | Traverses nested TOML tables using dot-separated *key* components. |
| 512 | Silently returns *default* on any parse or type error. |
| 513 | """ |
| 514 | config_path = root / ".muse" / "config.toml" |
| 515 | if not config_path.is_file(): |
| 516 | return default |
| 517 | try: |
| 518 | import tomllib |
| 519 | with config_path.open("rb") as fh: |
| 520 | raw = tomllib.load(fh) |
| 521 | val: JsonValue = raw |
| 522 | for part in key.split("."): |
| 523 | if not isinstance(val, dict): |
| 524 | return default |
| 525 | val = val.get(part, ...) # type: ignore[union-attr] |
| 526 | if val is ...: |
| 527 | return default |
| 528 | return float(val) if isinstance(val, (int, float)) else default # type: ignore[arg-type] |
| 529 | except Exception: |
| 530 | return default |
| 531 | |
| 532 | |
| 533 | # --------------------------------------------------------------------------- |
| 534 | # High-level integration helpers (merge and commit integration) |
| 535 | # --------------------------------------------------------------------------- |
| 536 | |
| 537 | def auto_apply( |
| 538 | root: pathlib.Path, |
| 539 | conflict_paths: list[str], |
| 540 | ours_manifest: Manifest, |
| 541 | theirs_manifest: Manifest, |
| 542 | domain: str, |
| 543 | plugin: "MuseDomainPlugin", |
| 544 | *, |
| 545 | explain_decisions: JsonObject | None = None, |
| 546 | ) -> tuple[Manifest, list[str]]: |
| 547 | """Attempt to auto-resolve conflicts using saved harmony resolutions. |
| 548 | |
| 549 | For each path in *conflict_paths*: |
| 550 | |
| 551 | 1. Validate the path stays within the repository root (path traversal guard). |
| 552 | 2. Compute blob and semantic fingerprints from the ours/theirs object IDs. |
| 553 | 3. Look up the best :class:`~muse.core.harmony.Resolution` for the computed |
| 554 | pattern ID. |
| 555 | 4. If found: restore the resolution blob to the working tree and collect |
| 556 | the ``{path: outcome_blob}`` mapping. |
| 557 | 5. If not found: record the pattern for future learning and leave the path |
| 558 | for manual resolution. |
| 559 | |
| 560 | Args: |
| 561 | root: Repository root. |
| 562 | conflict_paths: Workspace-relative POSIX paths that conflicted. |
| 563 | ours_manifest: ``{path: object_id}`` for the "ours" snapshot. |
| 564 | theirs_manifest: ``{path: object_id}`` for the "theirs" snapshot. |
| 565 | domain: Active domain name string. |
| 566 | plugin: Active domain plugin instance. |
| 567 | |
| 568 | Returns: |
| 569 | ``(resolved, remaining)`` where: |
| 570 | |
| 571 | - ``resolved`` maps path → outcome_blob for auto-resolved paths. |
| 572 | - ``remaining`` is the list of paths still requiring manual resolution. |
| 573 | """ |
| 574 | from muse.core.object_store import restore_object |
| 575 | |
| 576 | resolved: dict[str, str] = {} |
| 577 | remaining: list[str] = [] |
| 578 | |
| 579 | for path in conflict_paths: |
| 580 | # Conflict paths from the code domain are symbol addresses of the form |
| 581 | # "file.py::SymbolName". Manifests are keyed by file path, so extract |
| 582 | # the file portion for all manifest lookups. The full address is kept |
| 583 | # as the canonical path for pattern storage and fingerprinting so that |
| 584 | # two different symbols in the same file produce distinct patterns. |
| 585 | file_path = path.split("::")[0] |
| 586 | |
| 587 | # Guard against path traversal from a tampered MERGE_STATE.json. |
| 588 | # Use the file portion for the filesystem check. |
| 589 | # resolve() normalises any ".." components and symlinks before the |
| 590 | # relative_to() check, so "../traversal.py" is caught correctly. |
| 591 | try: |
| 592 | (root / file_path).resolve().relative_to(root.resolve()) |
| 593 | except ValueError: |
| 594 | logger.warning( |
| 595 | "⚠️ harmony: path traversal attempt in auto_apply — skipping %r", |
| 596 | path, |
| 597 | ) |
| 598 | remaining.append(path) |
| 599 | continue |
| 600 | |
| 601 | ours_id = ours_manifest.get(file_path, "") |
| 602 | theirs_id = theirs_manifest.get(file_path, "") |
| 603 | |
| 604 | if not ours_id or not theirs_id: |
| 605 | # One side deleted the file — cannot fingerprint, leave for manual. |
| 606 | remaining.append(path) |
| 607 | continue |
| 608 | |
| 609 | blob_fp = blob_fingerprint(ours_id, theirs_id) |
| 610 | semantic_fp = compute_semantic_fingerprint(path, ours_id, theirs_id, plugin, root) |
| 611 | pattern_id = compute_pattern_id(path, blob_fp, semantic_fp) |
| 612 | |
| 613 | best = best_resolution(root, pattern_id) |
| 614 | if best is not None: |
| 615 | min_confidence = _read_config_float( |
| 616 | root, "harmony.min_auto_apply_confidence", default=0.85 |
| 617 | ) |
| 618 | if best.confidence < min_confidence: |
| 619 | logger.info( |
| 620 | "⚠️ harmony: skipping auto-apply for '%s' — " |
| 621 | "confidence %.2f < threshold %.2f", |
| 622 | path, best.confidence, min_confidence, |
| 623 | ) |
| 624 | if explain_decisions is not None: |
| 625 | explain_decisions[file_path] = { |
| 626 | "result": "below_confidence_threshold", |
| 627 | "pattern_id": pattern_id, |
| 628 | } |
| 629 | remaining.append(path) |
| 630 | continue |
| 631 | |
| 632 | dest = root / file_path |
| 633 | dest.parent.mkdir(parents=True, exist_ok=True) |
| 634 | ok = restore_object(root, best.outcome_blob, dest) |
| 635 | if ok: |
| 636 | resolved[path] = best.outcome_blob |
| 637 | logger.info("✅ harmony: auto-resolved '%s'", path) |
| 638 | if explain_decisions is not None: |
| 639 | explain_decisions[file_path] = { |
| 640 | "result": "auto_resolved", |
| 641 | "pattern_id": pattern_id, |
| 642 | } |
| 643 | try: |
| 644 | increment_applied_count(root, pattern_id, best.resolution_id) |
| 645 | except Exception as exc: |
| 646 | logger.warning( |
| 647 | "⚠️ harmony: increment_applied_count failed for '%s': %s", |
| 648 | path, exc, |
| 649 | ) |
| 650 | else: |
| 651 | logger.warning( |
| 652 | "⚠️ harmony: resolution blob %s for '%s' not in local store", |
| 653 | best.outcome_blob, path, |
| 654 | ) |
| 655 | if explain_decisions is not None: |
| 656 | explain_decisions[file_path] = { |
| 657 | "result": "no_pattern", |
| 658 | "pattern_id": pattern_id, |
| 659 | } |
| 660 | remaining.append(path) |
| 661 | else: |
| 662 | # No saved resolution — record the pattern for future learning. |
| 663 | now = _now_utc() |
| 664 | pattern = ConflictPattern( |
| 665 | pattern_id=pattern_id, |
| 666 | path=path, |
| 667 | domain=domain, |
| 668 | conflict_type=ConflictType.CONTENT, |
| 669 | blob_fingerprint=blob_fp, |
| 670 | semantic_fingerprint=semantic_fp, |
| 671 | ours_id=ours_id, |
| 672 | theirs_id=theirs_id, |
| 673 | description={}, |
| 674 | recorded_at=now, |
| 675 | recorded_by="auto_apply", |
| 676 | ) |
| 677 | record_pattern(root, pattern) |
| 678 | if explain_decisions is not None: |
| 679 | explain_decisions[file_path] = { |
| 680 | "result": "no_pattern", |
| 681 | "pattern_id": None, |
| 682 | } |
| 683 | remaining.append(path) |
| 684 | |
| 685 | return resolved, remaining |
| 686 | |
| 687 | |
| 688 | def record_resolutions( |
| 689 | root: pathlib.Path, |
| 690 | conflict_paths: list[str], |
| 691 | ours_manifest: Manifest, |
| 692 | theirs_manifest: Manifest, |
| 693 | new_manifest: Manifest, |
| 694 | domain: str, |
| 695 | plugin: "MuseDomainPlugin", |
| 696 | manually_resolved: set[str] | None = None, |
| 697 | ) -> list[str]: |
| 698 | """Record how the user resolved each conflict after a merge commit. |
| 699 | |
| 700 | Called by ``muse commit`` immediately after writing a merge commit that |
| 701 | resolves a conflicted merge. For each previously conflicting path the |
| 702 | function reads the resolution object ID from the new snapshot manifest and |
| 703 | saves it to the harmony store as a :class:`~muse.core.harmony.Resolution`. |
| 704 | |
| 705 | Paths present in *manually_resolved* (those explicitly confirmed via |
| 706 | ``muse resolve`` after manual editing) are recorded with |
| 707 | ``confidence=1.0, human_verified=True``. All other paths (resolved by |
| 708 | accepting one side with ``muse checkout --ours/--theirs``) are recorded |
| 709 | with ``confidence=0.8, human_verified=False`` — a deliberate choice, but |
| 710 | not a careful manual merge. |
| 711 | |
| 712 | When *manually_resolved* is ``None`` (legacy callers / no MERGE_STATE |
| 713 | tracking) every path is recorded with ``confidence=1.0, |
| 714 | human_verified=True`` to preserve the pre-Phase-4 behaviour. |
| 715 | |
| 716 | Args: |
| 717 | root: Repository root. |
| 718 | conflict_paths: Paths that were listed as conflicts in MERGE_STATE. |
| 719 | ours_manifest: ``{path: object_id}`` for the "ours" snapshot. |
| 720 | theirs_manifest: ``{path: object_id}`` for the "theirs" snapshot. |
| 721 | new_manifest: ``{path: object_id}`` from the committed merge snapshot. |
| 722 | domain: Active domain name string. |
| 723 | plugin: Active domain plugin instance. |
| 724 | manually_resolved: Set of conflict addresses explicitly confirmed via |
| 725 | ``muse resolve``. ``None`` → treat all as manual. |
| 726 | |
| 727 | Returns: |
| 728 | List of paths for which a resolution was successfully saved. |
| 729 | """ |
| 730 | saved: list[str] = [] |
| 731 | |
| 732 | for path in conflict_paths: |
| 733 | # Conflict paths from the code domain are symbol addresses of the form |
| 734 | # "file.py::SymbolName". Extract the file portion for manifest lookups; |
| 735 | # keep the full address as the canonical path in the harmony store. |
| 736 | file_path = path.split("::")[0] |
| 737 | |
| 738 | ours_id = ours_manifest.get(file_path, "") |
| 739 | theirs_id = theirs_manifest.get(file_path, "") |
| 740 | outcome_blob = new_manifest.get(file_path, "") |
| 741 | |
| 742 | if not ours_id or not theirs_id or not outcome_blob: |
| 743 | continue |
| 744 | |
| 745 | try: |
| 746 | validate_object_id(outcome_blob) |
| 747 | except ValueError: |
| 748 | continue |
| 749 | |
| 750 | blob_fp = blob_fingerprint(ours_id, theirs_id) |
| 751 | semantic_fp = compute_semantic_fingerprint(path, ours_id, theirs_id, plugin, root) |
| 752 | pattern_id = compute_pattern_id(path, blob_fp, semantic_fp) |
| 753 | now = _now_utc() |
| 754 | |
| 755 | # Ensure the pattern exists in the harmony store. |
| 756 | if load_pattern(root, pattern_id) is None: |
| 757 | pattern = ConflictPattern( |
| 758 | pattern_id=pattern_id, |
| 759 | path=path, |
| 760 | domain=domain, |
| 761 | conflict_type=ConflictType.CONTENT, |
| 762 | blob_fingerprint=blob_fp, |
| 763 | semantic_fingerprint=semantic_fp, |
| 764 | ours_id=ours_id, |
| 765 | theirs_id=theirs_id, |
| 766 | description={}, |
| 767 | recorded_at=now, |
| 768 | recorded_by="record_resolutions", |
| 769 | ) |
| 770 | record_pattern(root, pattern) |
| 771 | |
| 772 | # Idempotency: skip if a resolution with this outcome_blob already exists. |
| 773 | existing = list_resolutions(root, pattern_id) |
| 774 | if any(r.outcome_blob == outcome_blob for r in existing): |
| 775 | logger.debug( |
| 776 | "harmony: resolution for '%s' (pattern %s) already recorded — skipping", |
| 777 | path, short_id(pattern_id), |
| 778 | ) |
| 779 | saved.append(path) |
| 780 | continue |
| 781 | |
| 782 | # Paths in manually_resolved were carefully edited by the user via |
| 783 | # `muse resolve` — high confidence, human-verified. Paths resolved via |
| 784 | # `muse checkout --ours/--theirs` (side-pick) get lower confidence and |
| 785 | # human_verified=False. When manually_resolved is None (legacy callers) |
| 786 | # treat everything as human-verified to preserve pre-Phase-4 behaviour. |
| 787 | is_manual = manually_resolved is None or path in manually_resolved |
| 788 | confidence = 1.0 if is_manual else 0.8 |
| 789 | resolution_id = compute_resolution_id( |
| 790 | pattern_id, |
| 791 | outcome_blob, |
| 792 | ResolutionStrategy.MANUAL, |
| 793 | AgentProvenance.human(), |
| 794 | now, |
| 795 | ) |
| 796 | resolution = Resolution( |
| 797 | resolution_id=resolution_id, |
| 798 | pattern_id=pattern_id, |
| 799 | strategy=ResolutionStrategy.MANUAL, |
| 800 | policy_id=None, |
| 801 | outcome_blob=outcome_blob, |
| 802 | rationale=( |
| 803 | f"User manually resolved conflict in '{path}' via muse resolve" |
| 804 | if is_manual |
| 805 | else f"Conflict in '{path}' resolved by accepting one side (checkout --ours/--theirs)" |
| 806 | ), |
| 807 | confidence=confidence, |
| 808 | applied_count=0, |
| 809 | human_verified=is_manual, |
| 810 | resolved_at=now, |
| 811 | resolved_by=AgentProvenance.human(), |
| 812 | ) |
| 813 | try: |
| 814 | save_resolution(root, resolution) |
| 815 | saved.append(path) |
| 816 | logger.info( |
| 817 | "✅ harmony: recorded resolution for '%s' (pattern %s)", |
| 818 | path, |
| 819 | pattern_id, |
| 820 | ) |
| 821 | except Exception as exc: |
| 822 | logger.warning( |
| 823 | "⚠️ harmony: failed to save resolution for '%s': %s", |
| 824 | path, exc, |
| 825 | ) |
| 826 | |
| 827 | return saved |
File History
2 commits
sha256:7011e00115e9c74d24569fed2caec6a2a6ef8fdb070d3b4715ce06e6633aaa47
feat(merge): add --explain flag with per-path decision trac…
Sonnet 4.6
minor
⚠
36 days ago
sha256:f0cc3f3fe40f1df9a488585b5b2aef6aa4dcb0b22b5fdbf7bab7071a5b0c7f7c
feat(merge): Phase 7 — independence merge correctness + Har…
Sonnet 4.6
minor
⚠
36 days ago