leapflow 0.0.0__py3-none-any.whl

leapflow/__init__.py

@@ -0,0 +1,6 @@
+"""LEAP Agent — Learning and Evolving from Actual Practice."""
+
+from leapflow.config import load_config
+from leapflow.version import __version__
+
+__all__ = ["__version__", "load_config"]

leapflow/__main__.py

@@ -0,0 +1,13 @@
+"""CLI entrypoint for LEAP Agent.
+
+Subcommands:
+    leap learn  — Interactive learning mode (record → distill)
+    leap run    — Execute a skill by trigger match or explicit name
+    leap skills — List / inspect learned skills
+    leap chat   — Conversational mode (original single-turn or interactive)
+"""
+
+from leapflow.cli import main
+
+if __name__ == "__main__":
+    raise SystemExit(main())

leapflow/analysis/__init__.py

@@ -0,0 +1,17 @@
+"""Offline analysis layer — synthesis, abstraction, segmentation, and distillation pipeline."""
+
+from leapflow.analysis.abstractor import ActionAbstractor
+from leapflow.analysis.synthesis import PlatformSynthesisPass
+
+__all__ = [
+    "ActionAbstractor",
+    "ImitationPipeline",
+    "PlatformSynthesisPass",
+]
+
+
+def __getattr__(name: str):
+    if name == "ImitationPipeline":
+        from leapflow.analysis.pipeline import ImitationPipeline
+        return ImitationPipeline
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")

leapflow/analysis/abstractor.py

@@ -0,0 +1,304 @@
+"""Multi-level action abstraction for trajectory analysis.
+
+Abstraction levels:
+    L0 (Raw)     → individual events from EventBus
+    L1 (Grouped) → consecutive same-type actions merged (e.g. keystrokes → type_text)
+    L2 (Pattern) → recognized action patterns (e.g. copy + switch + paste → transfer_data)
+    L3 (Intent)  → LLM-inferred user intent (optional, requires LLMProvider)
+
+Each level is a composable AbstractionPass (Pipeline pattern).
+"""
+
+from __future__ import annotations
+
+import logging
+from abc import ABC, abstractmethod
+from typing import Any, Dict, List, Optional, Sequence
+
+from leapflow.analysis.patterns import PatternLibrary
+from leapflow.domain.trajectory import (
+    ActionType,
+    RawAction,
+    SemanticAction,
+    TrajectoryStep,
+)
+
+logger = logging.getLogger(__name__)
+
+
+# ── Abstraction pass protocol ──
+
+
+class AbstractionPass(ABC):
+    """A single transformation in the abstraction pipeline."""
+
+    @abstractmethod
+    def apply(
+        self,
+        actions: List[SemanticAction],
+        steps: Optional[List[TrajectoryStep]] = None,
+    ) -> List[SemanticAction]:
+        """Transform a list of semantic actions into a (usually shorter) list.
+
+        Args:
+            actions: Current semantic actions to transform.
+            steps: Optional raw trajectory steps for passes that need
+                   access to state context (e.g., visual frames).
+        """
+
+
+# ── L0→L1: Rule-based grouping ──
+
+
+class GroupingPass(AbstractionPass):
+    """Merge consecutive low-level actions of the same type.
+
+    - Sequential type events → single "type_text" with merged text_content
+    - Sequential file modifications on the same target → single "batch_modify"
+    - Sequential clipboard changes → keep only the last one
+    """
+
+    def apply(
+        self,
+        actions: List[SemanticAction],
+        steps: Optional[List[TrajectoryStep]] = None,
+    ) -> List[SemanticAction]:
+        if len(actions) <= 1:
+            return actions
+
+        result: List[SemanticAction] = []
+        i = 0
+        while i < len(actions):
+            # Special case: merge consecutive type events with char data
+            if actions[i].action_name == "ui.type":
+                type_end = self._find_type_group_end(actions, i)
+                if type_end > i + 1:
+                    result.append(self._merge_type_group(actions[i:type_end]))
+                    i = type_end
+                    continue
+
+            group_end = self._find_group_end(actions, i)
+            if group_end > i + 1:
+                result.append(self._merge_group(actions[i:group_end]))
+                i = group_end
+            else:
+                result.append(actions[i])
+                i += 1
+        return result
+
+    def _find_type_group_end(self, actions: List[SemanticAction], start: int) -> int:
+        """Find end of a consecutive ui.type run within the same app."""
+        app = actions[start].parameters.get("app_bundle_id", "")
+        end = start + 1
+        while end < len(actions) and actions[end].action_name == "ui.type":
+            if actions[end].parameters.get("app_bundle_id", "") != app:
+                break
+            end += 1
+        return end
+
+    @staticmethod
+    def _merge_type_group(group: List[SemanticAction]) -> SemanticAction:
+        """Merge consecutive type actions, concatenating char fields into text_content."""
+        first, last = group[0], group[-1]
+        chars = [a.parameters.get("char", "") for a in group]
+        text_content = "".join(chars)
+        params: Dict[str, Any] = {
+            "app_bundle_id": first.parameters.get("app_bundle_id", ""),
+            "keystroke_count": len(group),
+        }
+        if text_content:
+            params["text_content"] = text_content
+        return SemanticAction(
+            action_name="type_text",
+            description=f"Type '{text_content[:50]}'" if text_content else f"Type {len(group)} keys",
+            parameters=params,
+            raw_action_range=(first.raw_action_range[0], last.raw_action_range[1]),
+            confidence=1.0,
+        )
+
+    def _find_group_end(self, actions: List[SemanticAction], start: int) -> int:
+        """Find the end index of a mergeable group starting at `start`."""
+        base = actions[start]
+        end = start + 1
+        while end < len(actions):
+            curr = actions[end]
+            if not self._can_merge(base, curr):
+                break
+            end += 1
+        return end
+
+    @staticmethod
+    def _can_merge(a: SemanticAction, b: SemanticAction) -> bool:
+        if a.action_name != b.action_name:
+            return False
+        mergeable = {"file.modify", "file.create", "file.rename", "file.move", "clipboard.copy"}
+        return a.action_name in mergeable
+
+    @staticmethod
+    def _merge_group(group: List[SemanticAction]) -> SemanticAction:
+        first, last = group[0], group[-1]
+        count = len(group)
+        merged_params = dict(first.parameters)
+        merged_params["count"] = count
+        return SemanticAction(
+            action_name=f"batch_{first.action_name.split('.')[-1]}",
+            description=f"{first.action_name} x{count}",
+            parameters=merged_params,
+            raw_action_range=(first.raw_action_range[0], last.raw_action_range[1]),
+            confidence=min(a.confidence for a in group),
+        )
+
+
+# ── L1→L2: Pattern matching ──
+
+
+class PatternPass(AbstractionPass):
+    """Pattern-based action abstraction using an extensible PatternLibrary.
+
+    Replaces hardcoded patterns with a YAML-driven, wildcard-capable
+    pattern matching engine that supports runtime addition of new patterns.
+    """
+
+    def __init__(self, library: Optional[PatternLibrary] = None) -> None:
+        self._library = library or PatternLibrary.default()
+
+    def apply(
+        self,
+        actions: List[SemanticAction],
+        steps: Optional[List[TrajectoryStep]] = None,
+    ) -> List[SemanticAction]:
+        if len(actions) < 2:
+            return actions
+
+        matches = self._library.match(actions)
+        if not matches:
+            return list(actions)
+
+        # Build result: replace matched spans with semantic actions, keep rest
+        result: List[SemanticAction] = []
+        consumed = 0
+
+        for m in matches:
+            # Append unmatched actions before this match
+            result.extend(actions[consumed : m.start_idx])
+            # Create the abstracted semantic action
+            first = actions[m.start_idx]
+            last = actions[m.end_idx - 1]
+            params: Dict[str, Any] = {}
+            # Merge original params from matched window
+            for a in actions[m.start_idx : m.end_idx]:
+                params.update(a.parameters)
+            # Overlay extracted params from pattern rules
+            params.update(m.extracted_params)
+
+            result.append(
+                SemanticAction(
+                    action_name=m.pattern.semantic_name,
+                    description=m.pattern.description or m.pattern.semantic_name,
+                    parameters=params,
+                    raw_action_range=(
+                        first.raw_action_range[0],
+                        last.raw_action_range[1],
+                    ),
+                    confidence=m.match_confidence,
+                )
+            )
+            consumed = m.end_idx
+
+        # Append remaining unmatched actions
+        result.extend(actions[consumed:])
+        return result
+
+
+# ── Orchestrator ──
+
+
+class ActionAbstractor:
+    """Multi-level action abstraction pipeline.
+
+    By default runs L0→L1→L2 (rule-based only, no LLM cost).
+    """
+
+    def __init__(
+        self,
+        passes: Sequence[AbstractionPass] | None = None,
+        *,
+        platform_hint: str = "darwin",
+    ) -> None:
+        if passes is not None:
+            self._passes = list(passes)
+        else:
+            from leapflow.analysis.denoise import DenoisePass
+            from leapflow.analysis.synthesis import PlatformSynthesisPass
+            self._passes = [
+                DenoisePass(),
+                PlatformSynthesisPass.for_platform(platform_hint),
+                GroupingPass(),
+                PatternPass(),
+            ]
+
+    def abstract(self, steps: List[TrajectoryStep]) -> List[SemanticAction]:
+        """Run the abstraction pipeline on raw trajectory steps."""
+        actions = [self._step_to_semantic(i, step) for i, step in enumerate(steps)]
+        for p in self._passes:
+            actions = p.apply(actions, steps=steps)
+        return actions
+
+    @staticmethod
+    def _step_to_semantic(idx: int, step: TrajectoryStep) -> SemanticAction:
+        """Convert a single TrajectoryStep to an initial SemanticAction (L0)."""
+        a = step.action
+        params: Dict[str, Any] = {}
+
+        if a.target:
+            params["target"] = a.target
+        if a.target_label:
+            params["target_label"] = a.target_label
+        if a.app_bundle_id:
+            params["app_bundle_id"] = a.app_bundle_id
+        if a.action_type in (ActionType.FILE_CREATE, ActionType.FILE_MODIFY,
+                             ActionType.FILE_DELETE, ActionType.FILE_RENAME):
+            params["path"] = a.target
+        if a.action_type == ActionType.CLIPBOARD_COPY:
+            params["text"] = a.params.get("text", "")
+        if a.action_type == ActionType.UI_TYPE:
+            char = a.params.get("char")
+            if char:
+                params["char"] = char
+        if a.action_type == ActionType.UI_DRAG:
+            for k in ("start_x", "start_y", "end_x", "end_y",
+                      "end_role", "end_label", "cross_app", "start_app", "end_app"):
+                if k in a.params:
+                    params[k] = a.params[k]
+
+        return SemanticAction(
+            action_name=a.action_type.value,
+            description=_describe_action(a),
+            parameters=params,
+            raw_action_range=(idx, idx + 1),
+            confidence=1.0,
+        )
+
+
+def _describe_action(a: RawAction) -> str:
+    """Generate a human-readable description of a raw action."""
+    at = a.action_type
+    if at == ActionType.APP_SWITCH:
+        return f"Switch to {a.app_name or a.app_bundle_id}"
+    if at in (ActionType.FILE_CREATE, ActionType.FILE_MODIFY,
+              ActionType.FILE_DELETE, ActionType.FILE_RENAME):
+        verb = at.value.split(".")[-1].capitalize()
+        return f"{verb} {a.target}"
+    if at == ActionType.CLIPBOARD_COPY:
+        length = a.params.get("text_length", 0)
+        return f"Copy to clipboard ({length} chars)"
+    if at == ActionType.UI_CLICK:
+        return f"Click {a.target_label or a.target or 'element'}"
+    if at == ActionType.UI_TYPE:
+        char = a.params.get("char", "")
+        return f"Type '{char}'" if char else "Type key"
+    if at == ActionType.UI_DRAG:
+        label = a.target_label or "element"
+        end_label = a.params.get("end_label", "") or "target"
+        return f"Drag {label} to {end_label}"
+    return f"{at.value} {a.target or ''}"

leapflow/analysis/causal.py

@@ -0,0 +1,91 @@
+"""Causal chain extraction from semantic action sequences.
+
+Identifies only the actions that contribute to the final observable state,
+filtering out detours, distractions, and non-contributing operations.
+Used by the distillation layer to produce minimal, goal-directed skill steps.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import FrozenSet, List, Set
+
+from leapflow.domain.trajectory import SemanticAction
+
+logger = logging.getLogger(__name__)
+
+_EFFECT_ACTIONS: FrozenSet[str] = frozenset({
+    "file.create", "file.modify", "file.delete", "file.rename", "file.move",
+    "clipboard.copy", "batch_modify", "batch_rename", "batch_delete",
+    "batch_move", "batch_move_to_folder", "move_to_new_folder",
+})
+
+_SUPPORT_ACTIONS: FrozenSet[str] = frozenset({
+    "app.switch", "open_file_dialog", "transfer_data",
+    "create_and_edit", "download_organize",
+})
+
+
+class CausalChainAnalyzer:
+    """Extract causally-necessary steps from a semantic action sequence.
+
+    Algorithm:
+        1. Forward scan: identify effect actions (file ops, clipboard writes)
+        2. For each effect action, back-trace to the nearest un-claimed
+           support action (app.switch, open_file_dialog, etc.)
+        3. Return only the causal subset, preserving original order.
+    """
+
+    def __init__(
+        self,
+        *,
+        effect_actions: FrozenSet[str] = _EFFECT_ACTIONS,
+        support_actions: FrozenSet[str] = _SUPPORT_ACTIONS,
+    ) -> None:
+        self._effects = effect_actions
+        self._supports = support_actions
+
+    def extract_causal_chain(
+        self, actions: List[SemanticAction]
+    ) -> List[SemanticAction]:
+        if len(actions) <= 2:
+            return list(actions)
+
+        causal_indices = self._find_causal_indices(actions)
+        if not causal_indices:
+            return list(actions)
+
+        result = [a for i, a in enumerate(actions) if i in causal_indices]
+        if len(result) < len(actions):
+            logger.debug(
+                "causal_chain: %d → %d actions (removed %d non-causal)",
+                len(actions), len(result), len(actions) - len(result),
+            )
+        return result
+
+    def _find_causal_indices(self, actions: List[SemanticAction]) -> Set[int]:
+        causal: List[int] = []
+        claimed: Set[int] = set()
+
+        for i, action in enumerate(actions):
+            if action.action_name in self._effects:
+                causal.append(i)
+                claimed.add(i)
+                self._backtrace_support(actions, i, causal, claimed)
+
+        return set(causal)
+
+    def _backtrace_support(
+        self,
+        actions: List[SemanticAction],
+        effect_idx: int,
+        causal: List[int],
+        claimed: Set[int],
+    ) -> None:
+        for j in range(effect_idx - 1, -1, -1):
+            if j in claimed:
+                break
+            if actions[j].action_name in self._supports:
+                causal.append(j)
+                claimed.add(j)
+                break

leapflow/analysis/consensus.py

@@ -0,0 +1,111 @@
+"""Cross-trajectory consensus distillation.
+
+When a user naturally performs the same task multiple times, each recording
+will contain different noise (typos, detours, hesitations). The Longest
+Common Subsequence across all trajectories converges on the essential steps,
+filtering noise that varies between demonstrations.
+
+    Demo 1: [A, B, X, C, D]    (X = mistake)
+    Demo 2: [A, B, C, Y, D]    (Y = distraction)
+    Demo 3: [A, Z, B, C, D]    (Z = hesitation)
+    LCS(1,2,3) = [A, B, C, D]  ← clean skill
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING, Dict, List, Optional, Sequence
+
+from leapflow.domain.skill_types import DistillationCandidate
+
+if TYPE_CHECKING:
+    from leapflow.analysis.pipeline import ImitationPipeline
+
+logger = logging.getLogger(__name__)
+
+
+class MultiTrajectoryDistiller:
+    """Extract consensus skills from multiple demonstrations of the same task."""
+
+    def __init__(
+        self,
+        pipeline: "ImitationPipeline",
+        *,
+        min_trajectories: int = 2,
+    ) -> None:
+        self._pipeline = pipeline
+        self._min_count = min_trajectories
+
+    async def distill_consensus(
+        self, trajectory_ids: List[str]
+    ) -> Optional[DistillationCandidate]:
+        if len(trajectory_ids) < self._min_count:
+            return None
+
+        all_episode_actions: List[List[str]] = []
+        for tid in trajectory_ids:
+            episodes = await self._pipeline.analyze(tid)
+            if not episodes:
+                continue
+            best = max(episodes, key=lambda e: len(e.semantic_actions))
+            action_names = [a.action_name for a in best.semantic_actions]
+            all_episode_actions.append(action_names)
+
+        if len(all_episode_actions) < self._min_count:
+            return None
+
+        consensus = all_episode_actions[0]
+        for other in all_episode_actions[1:]:
+            consensus = _lcs(consensus, other)
+
+        if len(consensus) < 2:
+            return None
+
+        total = len(all_episode_actions)
+        step_frequency: Dict[str, int] = {}
+        for actions in all_episode_actions:
+            for a in set(actions):
+                step_frequency[a] = step_frequency.get(a, 0) + 1
+
+        confidence = sum(
+            step_frequency.get(s, 0) / total for s in consensus
+        ) / len(consensus)
+
+        logger.info(
+            "consensus: %d trajectories → %d steps (confidence=%.2f)",
+            len(trajectory_ids), len(consensus), confidence,
+        )
+
+        return DistillationCandidate(
+            title=f"Consensus skill ({len(trajectory_ids)} demos)",
+            trigger_phrases=[],
+            steps=consensus,
+            confidence=min(confidence, 0.95),
+        )
+
+
+def _lcs(a: Sequence[str], b: Sequence[str]) -> List[str]:
+    """Longest common subsequence of two string sequences."""
+    m, n = len(a), len(b)
+    if m == 0 or n == 0:
+        return []
+    dp = [[0] * (n + 1) for _ in range(m + 1)]
+    for i in range(1, m + 1):
+        for j in range(1, n + 1):
+            if a[i - 1] == b[j - 1]:
+                dp[i][j] = dp[i - 1][j - 1] + 1
+            else:
+                dp[i][j] = max(dp[i - 1][j], dp[i][j - 1])
+    result: List[str] = []
+    i, j = m, n
+    while i > 0 and j > 0:
+        if a[i - 1] == b[j - 1]:
+            result.append(a[i - 1])
+            i -= 1
+            j -= 1
+        elif dp[i - 1][j] >= dp[i][j - 1]:
+            i -= 1
+        else:
+            j -= 1
+    result.reverse()
+    return result