agentdebugx 0.2.10__tar.gz → 0.2.11__tar.gz

{agentdebugx-0.2.10 → agentdebugx-0.2.11}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agentdebugx
-Version: 0.2.10
+Version: 0.2.11
 Summary: Portable error analysis, tracing, and recovery framework for agentic AI systems. Import as `agentdebug`.
 License: MIT
 License-File: LICENSE

{agentdebugx-0.2.10 → agentdebugx-0.2.11}/docs/23_status_v0_2.md

@@ -31,6 +31,8 @@ the forward-looking plan; this doc is the rear-view mirror.
 | DeepDebug | `agentdebug.deep.DeepDebugAnalyzer` | ✅ stable | full loop + silent LLM |
 | Cascade view | `agentdebug.traceback.format_traceback` | ✅ stable | cascade + step-order + ANSI + empty |
 | Detectors | `agentdebug.detectors.RepeatedToolCall / RepeatedState / StepCountLimit` | ✅ **new 0.2.2** | threshold + window + budget |
+| Detectors | `agentdebug.detectors.TopicDriftDetector` (embedding cosine vs goal) | ✅ **new 0.2.11** | stub-embedder ranking + no-goal short-circuit + embedder-raises safe + threshold boundary |
+| LLM client | `agentdebug.llm.OpenAICompatClient.embed()` (+ `EmbeddingClient` Protocol) | ✅ **new 0.2.11** | mocked-httpx POST /v1/embeddings + empty-input short-circuit |
 | Hub bundle | `agentdebug.hub.Bundle / pack_bundle / unpack_bundle` | ✅ stable | round-trip |
 | Hub scrubber | `agentdebug.hub.Scrubber` | ✅ stable | 12 redactions + idempotent |
 | Hub backends | `LocalHubBackend`, `GitHubBackend`, `HuggingFaceBackend` | ✅ stable | local-bare-git + local |
@@ -49,8 +51,7 @@ across 32 source files.
 
 | Doc | Component | Why deferred | Realistic ship |
 |---|---|---|---|
-| [06_detectors.md](./06_detectors.md) | `trajectory_perplexity` (TrajAD) | needs token-level LM perplexity API or embedding model + baseline calibration | v0.3 |
-| [06_detectors.md](./06_detectors.md) | `topic_drift` (embedding cosine) | needs embedding client; consider reusing `OpenAICompatClient` `/embeddings` | v0.3 |
+| [06_detectors.md](./06_detectors.md) | `trajectory_perplexity` (TrajAD) | needs token-level LM perplexity API; v0.3 |
 | [06_detectors.md](./06_detectors.md) | LTL spec monitors | requires user-supplied spec or LLM-synthesized monitors; gated on RV research | v1.2 |
 | [07_attribution.md](./07_attribution.md) | `CounterfactualAttributor` — *real* replay variant | true re-rollout requires framework-specific replay surface; the v0.2.7 LLM-simulated variant ships now, the real-replay variant is gated on adapter support (LangGraph checkpointer / OpenHands rewind) | v0.4 |
 | [07_attribution.md](./07_attribution.md) | `SBFLAttributor` — *corpus* | shipped in 0.2.8 (`tarantula`/`ochiai`/`dstar`); awaiting paired-trace adoption to gather a useful corpus in production | corpus tooling deferred to v0.4 |

{agentdebugx-0.2.10 → agentdebugx-0.2.11}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "agentdebugx"
-version = "0.2.10"
+version = "0.2.11"
 description = "Portable error analysis, tracing, and recovery framework for agentic AI systems. Import as `agentdebug`."
 authors = ["ULab @ UIUC <ulab@illinois.edu>"]
 license = "MIT"

{agentdebugx-0.2.10 → agentdebugx-0.2.11}/src/agentdebug/__init__.py

@@ -28,6 +28,7 @@ from agentdebug.detectors import (
     RepeatedStateDetector,
     RepeatedToolCallDetector,
     StepCountLimitDetector,
+    TopicDriftDetector,
     default_detectors,
     run_detectors,
 )
@@ -82,6 +83,7 @@ __all__ = [
     'SBFLAttributor',
     'SelfRefineLoop',
     'StepByStepAttributor',
+    'TopicDriftDetector',
     'StepCountLimitDetector',
     'VerifierSpec',
     'build_cascade',
@@ -108,4 +110,4 @@ __all__ = [
     'get_failure_mode',
 ]
 
-__version__ = '0.2.10'
+__version__ = '0.2.11'

{agentdebugx-0.2.10 → agentdebugx-0.2.11}/src/agentdebug/detectors.py

@@ -15,7 +15,7 @@ from __future__ import annotations
 
 import logging
 from dataclasses import dataclass
-from typing import List, Optional, Protocol
+from typing import Any, List, Optional, Protocol
 
 from agentdebug.models import (
     AgentEvent,
@@ -273,12 +273,106 @@ def _suggestion(mode: FailureMode) -> Optional[str]:
     return None
 
 
+class TopicDriftDetector:
+    """Embedding-based anomaly detector for goal drift.
+
+    Embed the trajectory's goal once; embed each user-facing event payload
+    (LLM_RESPONSE / PLAN / OBSERVATION outputs); flag any step whose cosine
+    similarity with the goal drops below ``threshold``.
+
+    Maps to ``FM-2.3 task_derailment`` (MAST) / ``planning.inefficient_plan``.
+    Closes the anomaly family from doc 06 alongside the existing
+    RepeatedToolCallDetector / RepeatedStateDetector.
+
+    Skipped silently if the embedding client raises or returns no vectors —
+    the rest of the detector pipeline is unaffected.
+    """
+
+    id = 'topic_drift'
+
+    def __init__(
+        self,
+        embedding_client: Any,
+        *,
+        threshold: float = 0.35,
+        max_events: int = 60,
+    ) -> None:
+        # embedding_client is duck-typed to EmbeddingClient to avoid an
+        # import cycle (detectors.py is imported from agentdebug/__init__.py).
+        self.embedding_client = embedding_client
+        self.threshold = threshold
+        self.max_events = max_events
+
+    def detect(self, trajectory: AgentTrajectory) -> List[FailureFinding]:
+        if not trajectory.goal:
+            return []
+        contextful = [
+            e for e in trajectory.events
+            if e.event_type in {
+                EventType.LLM_RESPONSE, EventType.PLAN, EventType.OBSERVATION,
+                EventType.LLM_RESPONSE.value, EventType.PLAN.value,
+                EventType.OBSERVATION.value,
+            }
+            and e.output is not None and str(e.output).strip()
+        ]
+        if not contextful:
+            return []
+        contextful = contextful[-self.max_events:]
+        texts = [trajectory.goal] + [str(e.output)[:1000] for e in contextful]
+        try:
+            vectors = self.embedding_client.embed(texts)
+        except Exception as exc:  # pragma: no cover - defensive
+            LOG.warning('topic_drift detector embed() failed: %s', exc)
+            return []
+        if not vectors or len(vectors) != len(texts):
+            return []
+        goal_vec = vectors[0]
+        findings: List[FailureFinding] = []
+        mode = SEED_FAILURE_MODES['planning.inefficient_plan']
+        for evt, evt_vec in zip(contextful, vectors[1:]):
+            sim = _cosine(goal_vec, evt_vec)
+            if sim >= self.threshold:
+                continue
+            findings.append(FailureFinding(
+                finding_id=new_id('finding'),
+                failure_mode=mode,
+                event_id=evt.event_id,
+                agent_name=evt.agent_name,
+                step_index=evt.step_index,
+                # Confidence proportional to how far below threshold we drifted.
+                confidence=min(0.95, 0.4 + (self.threshold - sim)),
+                evidence=[
+                    f'goal/output cosine={sim:.3f} < threshold={self.threshold:.2f}',
+                ],
+                suggestion=_suggestion(mode),
+                metadata={
+                    'source': self.id,
+                    'cosine_to_goal': round(sim, 4),
+                    'threshold': self.threshold,
+                },
+            ))
+        return findings
+
+
+def _cosine(a: List[float], b: List[float]) -> float:
+    import math
+    if not a or not b or len(a) != len(b):
+        return 0.0
+    dot = sum(x * y for x, y in zip(a, b))
+    na = math.sqrt(sum(x * x for x in a))
+    nb = math.sqrt(sum(y * y for y in b))
+    if na == 0 or nb == 0:
+        return 0.0
+    return dot / (na * nb)
+
+
 __all__ = [
     'Detector',
     'DetectorConfig',
     'RepeatedStateDetector',
     'RepeatedToolCallDetector',
     'StepCountLimitDetector',
+    'TopicDriftDetector',
     'default_detectors',
     'run_detectors',
 ]

{agentdebugx-0.2.10 → agentdebugx-0.2.11}/src/agentdebug/llm.py

@@ -44,6 +44,25 @@ class LLMClient(Protocol):
         ...
 
 
+class EmbeddingClient(Protocol):
+    """Subprotocol for clients that also expose ``/v1/embeddings``.
+
+    Kept separate from :class:`LLMClient` so detectors can declare a
+    narrower dependency and tests can stub embeddings without faking
+    a chat client.
+    """
+
+    embedding_model: str
+
+    def embed(
+        self,
+        texts: List[str],
+        *,
+        timeout: float = 60.0,
+    ) -> List[List[float]]:
+        ...
+
+
 class OpenAICompatClient:
     """OpenAI-compatible chat completions client.
 
@@ -63,12 +82,16 @@ class OpenAICompatClient:
         base_url: str,
         api_key: str,
         model: str,
+        embedding_model: str = 'text-embedding-3-small',
         default_max_tokens: int = 2048,
         timeout: float = 60.0,
     ) -> None:
         self.base_url = base_url.rstrip('/')
         self.api_key = api_key
         self.model = model
+        # Embeddings hit a separate endpoint with a separate model id; default
+        # to OpenAI's small embedding model since the gateway is OpenAI-compat.
+        self.embedding_model = embedding_model
         self.default_max_tokens = default_max_tokens
         self.timeout = timeout
 
@@ -132,6 +155,40 @@ class OpenAICompatClient:
         return CompletionResult(text=text, raw=data)
 
 
+    def embed(
+        self,
+        texts: List[str],
+        *,
+        timeout: Optional[float] = None,
+    ) -> List[List[float]]:
+        """OpenAI-compatible ``/v1/embeddings`` POST.
+
+        Returns a list of vectors (one per input text) in the same order.
+        Empty ``texts`` short-circuits to ``[]`` to save a network round-trip.
+        """
+        if not texts:
+            return []
+        url = f'{self.base_url}/embeddings'
+        headers = {
+            'Authorization': f'Bearer {self.api_key}',
+            'Content-Type': 'application/json',
+        }
+        body = {'model': self.embedding_model, 'input': list(texts)}
+        resp = httpx.post(
+            url, headers=headers, json=body, timeout=timeout or self.timeout
+        )
+        resp.raise_for_status()
+        data = resp.json()
+        rows = data.get('data') or []
+        out: List[List[float]] = []
+        for row in rows:
+            vec = row.get('embedding')
+            if not isinstance(vec, list):
+                continue
+            out.append([float(v) for v in vec])
+        return out
+
+
 def extract_json_block(text: str) -> Optional[Dict[str, Any]]:
     """Extract the first top-level JSON object from a possibly-fenced response."""
     if not text: