codespy-ai 2.0.0__tar.gz → 2.0.1__tar.gz

{codespy_ai-2.0.0 → codespy_ai-2.0.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: codespy-ai
-Version: 2.0.0
+Version: 2.0.1
 Summary: Code review agent powered by DSPy
 License: MIT
 License-File: LICENSE

{codespy_ai-2.0.0 → codespy_ai-2.0.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "codespy-ai"
-version = "2.0.0"
+version = "2.0.1"
 description = "Code review agent powered by DSPy"
 readme = "README.md"
 license = "MIT"

{codespy_ai-2.0.0 → codespy_ai-2.0.1}/src/codespy/__init__.py

@@ -1,3 +1,3 @@
 """codespy - Code review agent powered by DSPy."""
 
-__version__ = "2.0.0"
+__version__ = "2.0.1"

{codespy_ai-2.0.0 → codespy_ai-2.0.1}/src/codespy/agents/context_safe.py

@@ -4,6 +4,7 @@ Provides transparent fallback to RLM (Recursive Language Model) when inputs
 exceed the model's context window. Used to wrap all DSPy signature modules.
 """
 
+import enum
 import logging
 import re
 
@@ -19,6 +20,47 @@ _RE_CONTEXT_LENGTH = re.compile(r"maximum context length is \d+ tokens", re.IGNO
 # Below this, context rot is not a concern regardless of ratio.
 _MIN_RLM_THRESHOLD_TOKENS = 8192
 
+# Replaces dspy.RLM's stock `reasoning` desc ("Think step-by-step: what do you
+# know? What remains? Plan your next action."), which on its own triggers Bedrock
+# content_filter on Claude Opus 5.5.
+RLM_REASONING_DESC = "Short plan for the next step."
+
+
+def _plain(value):  # type: ignore[no-untyped-def]
+    """Replace Enum members with their values, recursively.
+
+    dspy's PythonInterpreter injects inputs as `name = repr(value)`; an Enum repr
+    (`<IssueCategory.BUG: 'bug'>`) is not valid Python, so every REPL execution
+    fails with an empty SyntaxError.
+    """
+    if isinstance(value, enum.Enum):
+        return value.value
+    if isinstance(value, (list, tuple, set)):
+        return [_plain(v) for v in value]
+    if isinstance(value, dict):
+        return {_plain(k): _plain(v) for k, v in value.items()}
+    return value
+
+
+class CodespyRLM(dspy.RLM):
+    """dspy.RLM adjusted for codespy.
+
+    - Neutral action-signature `reasoning` description (Bedrock content_filter).
+    - Enum inputs passed as plain values (REPL variable injection).
+    """
+
+    def forward(self, **kwargs):  # type: ignore[no-untyped-def]
+        return super().forward(**{k: _plain(v) for k, v in kwargs.items()})
+
+    async def aforward(self, **kwargs):  # type: ignore[no-untyped-def]
+        return await super().aforward(**{k: _plain(v) for k, v in kwargs.items()})
+
+    def _build_signatures(self):  # type: ignore[no-untyped-def]
+        action_sig, extract_sig = super()._build_signatures()
+        if "reasoning" in action_sig.output_fields:
+            action_sig = action_sig.with_updated_fields("reasoning", desc=RLM_REASONING_DESC)
+        return action_sig, extract_sig
+
 
 def estimate_context_overflow(model: str, max_tokens: int, input_text: str) -> bool:
     """Estimate whether input + max_tokens would exceed the model's context window.
@@ -197,9 +239,10 @@ class ContextSafe(dspy.Module):
 
     def _create_rlm_fallback(self) -> dspy.RLM:
         """Create RLM with current signature and tools."""
-        return dspy.RLM(
+        return CodespyRLM(
             self._get_current_signature(),
             tools=self._tools,
             max_iters=self._max_iters,
             max_llm_calls=self._max_llm_calls,
+            verbose=logger.isEnabledFor(logging.DEBUG),
         )

codespy_ai-2.0.1/src/codespy/agents/diagnostic_adapter.py

@@ -0,0 +1,328 @@
+"""Diagnostic adapter for logging empty LM response details.
+
+Wraps TwoStepAdapter to capture diagnostics when the main LM returns an empty
+or null response, which helps diagnose why Claude occasionally returns empty
+text with hidden thinking content.
+"""
+
+import json
+import logging
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any
+
+import dspy  # type: ignore[import-untyped]
+from dspy.adapters.two_step_adapter import TwoStepAdapter  # type: ignore[import-untyped]
+from dspy.adapters.base import AdapterParseError  # type: ignore[import-untyped]
+
+from codespy.config import get_settings
+
+logger = logging.getLogger(__name__)
+
+
+def _getattr_or_key(obj: Any, name: str, default: Any = None) -> Any:
+    """Get attribute or dict key, tolerating both access patterns."""
+    if hasattr(obj, name):
+        return getattr(obj, name, default)
+    if isinstance(obj, dict):
+        return obj.get(name, default)
+    return default
+
+
+def _safe_repr(value: Any, max_len: int = 500) -> str:
+    """Safe repr with length limit."""
+    try:
+        s = repr(value)
+        if len(s) > max_len:
+            return s[:max_len] + f"... ({len(s)} chars total)"
+        return s
+    except Exception:
+        return "<repr failed>"
+
+
+def _get_finish_reason(entry: dict) -> str:
+    """Extract finish_reason from history entry."""
+    resp = entry.get("response")
+    if resp is None:
+        return "unknown"
+
+    resp_class_name = type(resp).__name__
+    if resp_class_name == "Response":
+        # lm15 style Response object
+        return _getattr_or_key(resp, "finish_reason", "unknown")
+    else:
+        # Standard dict-style response
+        choices = _getattr_or_key(resp, "choices", [])
+        if choices and len(choices) > 0:
+            choice = choices[0]
+            return _getattr_or_key(choice, "finish_reason", "unknown")
+    return "unknown"
+
+
+def _write_refusal_dump(entry: dict, model: str, lm_kwargs: dict, signature_name: str) -> Path | None:
+    """Write a refusal dump to disk for later analysis.
+
+    Args:
+        entry: The history entry containing messages and response.
+        model: The model name.
+        lm_kwargs: The LM kwargs (will have API keys stripped).
+        signature_name: The signature name for the dump filename.
+
+    Returns:
+        Path to the written dump file, or None if writing failed.
+    """
+    try:
+        settings = get_settings()
+        cache_dir = Path(settings.review.cache_dir).expanduser().resolve()
+        refusals_dir = cache_dir / "refusals"
+        refusals_dir.mkdir(parents=True, exist_ok=True)
+
+        timestamp = datetime.now(timezone.utc).strftime("%Y%m%dT%H%M%S")
+        safe_sig = signature_name.replace("/", "_").replace("\\", "_")[:50]
+        dump_path = refusals_dir / f"{timestamp}-{safe_sig}.json"
+
+        # Strip API keys from lm_kwargs
+        safe_kwargs = {}
+        for key, value in lm_kwargs.items():
+            lower_key = key.lower()
+            # Redact actual secret keys, but not legitimate settings like max_tokens
+            if lower_key.endswith("_key") or lower_key.endswith("_secret") or lower_key.endswith("_token") or lower_key == "api_key":
+                safe_kwargs[key] = "<redacted>"
+            else:
+                safe_kwargs[key] = value
+
+        dump = {
+            "model": model,
+            "lm_kwargs": safe_kwargs,
+            "messages": entry.get("messages", []),
+            "response": entry.get("response"),
+        }
+
+        with open(dump_path, "w", encoding="utf-8") as f:
+            json.dump(dump, f, indent=2, default=str)
+
+        logger.warning("Empty LM response: refusal dump written to %s", dump_path)
+        return dump_path
+    except Exception as e:
+        logger.debug("Empty LM response: failed to write refusal dump: %s", e)
+        return None
+
+
+def _log_empty_response(lm: Any, signature: Any) -> None:
+    """Log detailed diagnostics for an empty LM response.
+
+    Wraps everything in try/except so diagnostics can never mask the original error.
+    """
+    try:
+        model = getattr(lm, "model", "unknown")
+        max_tokens = getattr(lm, "kwargs", {}).get("max_tokens") if hasattr(lm, "kwargs") else None
+        reasoning_effort = getattr(lm, "kwargs", {}).get("reasoning_effort") if hasattr(lm, "kwargs") else None
+
+        # Get output field names from signature
+        output_fields: list[str] = []
+        signature_name = "unknown"
+        if signature is not None:
+            try:
+                signature_name = getattr(signature, "__name__", type(signature).__name__)
+                if hasattr(signature, "output_fields"):
+                    output_fields = list(getattr(signature, "output_fields", []))
+                elif hasattr(signature, "fields"):
+                    # Try to identify output fields from signature fields
+                    fields = getattr(signature, "fields", {})
+                    if isinstance(fields, dict):
+                        output_fields = [k for k, v in fields.items() if getattr(v, "is_input", False) is False]
+            except Exception:
+                pass
+
+        logger.warning(
+            "Empty LM response: model=%s max_tokens=%s reasoning_effort=%s output_fields=%s",
+            model, max_tokens, reasoning_effort, output_fields
+        )
+
+        # Read from LM history
+        history = getattr(lm, "history", None)
+        if not history:
+            logger.warning("Empty LM response: no LM history available")
+            return
+
+        try:
+            entry = history[-1]
+        except IndexError:
+            logger.warning("Empty LM response: LM history is empty")
+            return
+
+        if not isinstance(entry, dict):
+            logger.warning("Empty LM response: unexpected history entry type %s", type(entry).__name__)
+            return
+
+        # Log usage info
+        usage = entry.get("usage")
+        if usage:
+            prompt_tokens = _getattr_or_key(usage, "prompt_tokens") or _getattr_or_key(usage, "input_tokens", 0)
+            completion_tokens = _getattr_or_key(usage, "completion_tokens") or _getattr_or_key(usage, "output_tokens", 0)
+            reasoning_tokens = _getattr_or_key(usage, "reasoning_tokens", 0)
+            logger.warning(
+                "Empty LM response: usage prompt_tokens=%s completion_tokens=%s reasoning_tokens=%s",
+                prompt_tokens, completion_tokens, reasoning_tokens
+            )
+        else:
+            logger.warning("Empty LM response: no usage info in history entry")
+
+        # Log response details
+        resp = entry.get("response")
+        if resp is None:
+            logger.warning("Empty LM response: no response object in history entry")
+            return
+
+        # Handle lm15 Response object vs dict
+        resp_class_name = type(resp).__name__
+        if resp_class_name == "Response":
+            # lm15 style Response object
+            finish_reason = _getattr_or_key(resp, "finish_reason", "unknown")
+            message = _getattr_or_key(resp, "message", None)
+            parts_info: list[str] = []
+            if message and hasattr(message, "parts"):
+                parts = getattr(message, "parts", [])
+                parts_info = [type(p).__name__ for p in parts]
+            usage_obj = _getattr_or_key(resp, "usage", None)
+            logger.warning(
+                "Empty LM response: lm15 Response finish_reason=%s message_parts=%s usage=%s",
+                finish_reason, parts_info, usage_obj
+            )
+        else:
+            # Standard dict-style response
+            choices = _getattr_or_key(resp, "choices", [])
+            if choices and len(choices) > 0:
+                choice = choices[0]
+                finish_reason = _getattr_or_key(choice, "finish_reason", "unknown")
+                message = _getattr_or_key(choice, "message", {})
+                content = _getattr_or_key(message, "content", "")
+
+                logger.warning(
+                    "Empty LM response: finish_reason=%s content=%s",
+                    finish_reason, _safe_repr(content, 500)
+                )
+
+                # Check for reasoning_content (Anthropic)
+                reasoning_content = _getattr_or_key(message, "reasoning_content", None)
+                if reasoning_content:
+                    rc_len = len(str(reasoning_content))
+                    logger.warning(
+                        "Empty LM response: reasoning_content len=%s first_300=%s",
+                        rc_len, _safe_repr(str(reasoning_content)[:300], 300)
+                    )
+                else:
+                    # Check for thinking_blocks
+                    thinking_blocks = _getattr_or_key(message, "thinking_blocks", None)
+                    if thinking_blocks:
+                        block_types = [type(b).__name__ for b in thinking_blocks]
+                        logger.warning(
+                            "Empty LM response: thinking_blocks count=%s types=%s",
+                            len(thinking_blocks), block_types
+                        )
+                    else:
+                        logger.warning("Empty LM response: no reasoning_content or thinking_blocks")
+
+                # Check for tool_calls
+                tool_calls = _getattr_or_key(message, "tool_calls", None)
+                if tool_calls:
+                    logger.warning("Empty LM response: tool_calls count=%s", len(tool_calls))
+                else:
+                    logger.warning("Empty LM response: no tool_calls")
+
+                # Log provider_specific_fields if present
+                provider_fields = _getattr_or_key(message, "provider_specific_fields", None)
+                if provider_fields:
+                    logger.warning("Empty LM response: provider_specific_fields keys=%s", list(provider_fields.keys()) if isinstance(provider_fields, dict) else type(provider_fields).__name__)
+            else:
+                logger.warning("Empty LM response: no choices in response")
+
+        # Log request size
+        messages = entry.get("messages", [])
+        request_size = len(str(messages))
+        logger.warning("Empty LM response: request_size_chars=%s", request_size)
+
+        # If finish_reason is content_filter, write a refusal dump
+        finish_reason = _get_finish_reason(entry)
+        if finish_reason == "content_filter":
+            lm_kwargs = getattr(lm, "kwargs", {}) if hasattr(lm, "kwargs") else {}
+            _write_refusal_dump(entry, model, lm_kwargs, signature_name)
+
+    except Exception as diag_exc:
+        # Never let diagnostics mask the original error
+        logger.debug("Empty LM response: diagnostic logging failed: %s", diag_exc, exc_info=True)
+
+
+def _get_finish_reason_from_lm(lm: Any) -> str | None:
+    """Extract finish_reason from LM history entry."""
+    history = getattr(lm, "history", None)
+    if not history:
+        return None
+    try:
+        entry = history[-1]
+        if not isinstance(entry, dict):
+            return None
+        return _get_finish_reason(entry)
+    except Exception:
+        return None
+
+
+class DiagnosticTwoStepAdapter(TwoStepAdapter):
+    """TwoStepAdapter that logs diagnostics on empty LM responses.
+
+    Wraps the parent's __call__ and __acall__ to catch AdapterParseError for empty/null
+    responses and log detailed diagnostics before re-raising the original
+    exception unchanged.
+    """
+
+    def _check_and_log_empty_response(
+        self, exc: AdapterParseError, lm: dspy.LM, signature: Any
+    ) -> None:
+        """Check if exception is for empty response and log diagnostics.
+
+        If finish_reason is content_filter, raises a more descriptive error.
+        """
+        # Only log for empty/null response errors
+        if "empty or null response" in str(exc).lower():
+            _log_empty_response(lm, signature)
+            # If finish_reason is content_filter, raise a more descriptive error
+            finish_reason = _get_finish_reason_from_lm(lm)
+            if finish_reason == "content_filter":
+                raise AdapterParseError(
+                    adapter_name=getattr(self, "__class__.__name__", "DiagnosticTwoStepAdapter"),
+                    signature=signature,
+                    lm_response="",
+                    message="model refused (content_filter)",
+                ) from exc
+
+    def __call__(
+        self,
+        lm: dspy.LM,
+        lm_kwargs: dict[str, Any],
+        signature: Any,
+        demos: list[dict[str, Any]],
+        inputs: dict[str, Any],
+    ) -> dict[str, Any]:
+        """Call parent adapter, logging diagnostics on empty response."""
+        try:
+            return super().__call__(lm, lm_kwargs, signature, demos, inputs)
+        except AdapterParseError as exc:
+            self._check_and_log_empty_response(exc, lm, signature)
+            # Always re-raise the original exception unchanged
+            raise
+
+    async def acall(
+        self,
+        lm: dspy.LM,
+        lm_kwargs: dict[str, Any],
+        signature: Any,
+        demos: list[dict[str, Any]],
+        inputs: dict[str, Any],
+    ) -> dict[str, Any]:
+        """Async call parent adapter, logging diagnostics on empty response."""
+        try:
+            return await super().acall(lm, lm_kwargs, signature, demos, inputs)
+        except AdapterParseError as exc:
+            self._check_and_log_empty_response(exc, lm, signature)
+            # Always re-raise the original exception unchanged
+            raise

{codespy_ai-2.0.0 → codespy_ai-2.0.1}/src/codespy/agents/dspy_config.py

@@ -7,8 +7,8 @@ from typing import Any
 
 import dspy  # type: ignore[import-untyped]
 import litellm  # type: ignore[import-untyped]
-from dspy.adapters.two_step_adapter import TwoStepAdapter  # type: ignore[import-untyped]
 
+from codespy.agents.diagnostic_adapter import DiagnosticTwoStepAdapter
 from codespy.config import Settings, get_settings
 from codespy.config_memory import MEMORY_CONSOLIDATION, MEMORY_MENTAL_MODELS, MEMORY_RECALL, MEMORY_RETAIN, REFLECTION_MODULES, LLMSettings
 from codespy.config_utils import secret_value
@@ -216,7 +216,7 @@ def lm_context(name: str) -> LMScope:
     llm_config = settings.get_llm_config(name)
     lm = new_lm(settings, llm_config)
     extraction_lm = new_extraction_lm(settings)
-    ctx = dspy.context(lm=lm, adapter=TwoStepAdapter(extraction_lm))
+    ctx = dspy.context(lm=lm, adapter=DiagnosticTwoStepAdapter(extraction_lm))
     return LMScope(ctx=ctx, lm=lm, extraction_lm=extraction_lm)
 
 
@@ -270,7 +270,7 @@ def configure_dspy(settings: Settings) -> None:
 
     dspy.settings.configure(
         lm=lm,
-        adapter=TwoStepAdapter(extraction_lm),  # TwoStepAdapter solves ChatAdapter parsing failures
+        adapter=DiagnosticTwoStepAdapter(extraction_lm),  # TwoStepAdapter solves ChatAdapter parsing failures
     )
 
     # Enable memory-only caching for LLM calls (no disk caching)

{codespy_ai-2.0.0 → codespy_ai-2.0.1}/src/codespy/agents/review/code_review/agent.py

@@ -14,14 +14,16 @@ from codespy.agents.context_safe import ContextSafe
 from codespy.agents.memory.hippocampus import ContextMemory, Hippocampus
 from codespy.agents.memory.hippocampus.episode import submit_episode_save
 from codespy.agents.memory.prefrontal import with_prefrontal_memory
-from codespy.agents.review.models import Issue, IssueCategory, ReviewContext
-from codespy.agents.review.scope.models import ScopeResult
 from codespy.agents.review.helpers import (
+    filter_by_confidence,
     issues_to_markdown,
     make_scope_relative,
     resolve_scope_root,
     restore_repo_paths,
 )
+from codespy.agents.review.models import Issue, IssueCategory, ReviewContext
+from codespy.agents.review.scope.models import ScopeResult
+from codespy.agents.rlm_tools import build_rlm_agent, log_rlm_outcome, run_rlm
 from codespy.config import get_settings
 from codespy.config_memory import get_cerebral, get_episode_store, get_prefrontal
 from codespy.tools.mcp_utils import cleanup_mcp_contexts, connect_mcp_server
@@ -129,6 +131,7 @@ class CodeReviewSignature(dspy.Signature):
     - description: ≤25 words, imperative tone, no filler ("Fix X", "Rename Y to Z")
     - No polite or conversational language
     - Do not populate code_snippet — use line numbers instead
+    - Always set confidence: ≥0.9 verified with tools; 0.7–0.9 strong evidence; <0.7 weak
     """
 
     scope: ScopeResult = dspy.InputField(
@@ -282,15 +285,9 @@ class CodeReviewer(dspy.Module):
             sig = with_prefrontal_memory(CodeReviewSignature) if review_context.prefrontal_memory else CodeReviewSignature
             recall_tool = pf.recall_tool() if pf else None
             agent_tools = [*tools, recall_tool] if recall_tool else tools
-            agent = ContextSafe(
-                dspy.RLM(
-                    sig,
-                    tools=agent_tools,
-                    max_iters=max_iters,
-                    max_llm_calls=self._settings.get_max_llm_calls("code_review"),
-                ),
+            agent = build_rlm_agent(
                 sig,
-                tools=agent_tools,
+                agent_tools,
                 name="code_review",
                 max_iters=max_iters,
                 max_llm_calls=self._settings.get_max_llm_calls("code_review"),
@@ -336,16 +333,11 @@ class CodeReviewer(dspy.Module):
                         initial_memory=scope_initial_memory,
                         topics=topics,
                     )
-                result = await agent.acall(
-                    scope=scoped,
-                    categories=categories,
-                    **pf_kwargs,
+                result = await run_rlm(agent, scope=scoped, categories=categories, **pf_kwargs)
+                log_rlm_outcome(result, "code_review", scope.subroot)
+                issues = filter_by_confidence(
+                    result.issues, self._settings.review.min_confidence, "code_review", scope.subroot
                 )
-                issues = [
-                    issue
-                    for issue in (result.issues or [])
-                    if issue.confidence >= self._settings.review.min_confidence
-                ]
                 if hippo is not None:
                     await hippo.aobserve(result)
                     # Fire-and-forget background episode save

{codespy_ai-2.0.0 → codespy_ai-2.0.1}/src/codespy/agents/review/doc/agent.py

@@ -19,6 +19,7 @@ from codespy.agents.review.scope.models import ScopeResult
 from codespy.agents.review.doc.doc_extractor import extract_documentation
 from codespy.agents.review.helpers import (
     build_patches,
+    filter_by_confidence,
     issues_to_markdown,
     make_scope_relative,
     resolve_scope_root,
@@ -78,6 +79,7 @@ class DocReviewSignature(dspy.Signature):
     - description: ≤25 words, imperative tone ("Update X section", "Add Y to README")
     - Empty list if documentation is up to date. No approval text ("LGTM", "looks good")
     - No polite or conversational language
+    - Always set confidence: ≥0.9 verified with tools; 0.7–0.9 strong evidence; <0.7 weak
     """
 
     patches: str = dspy.InputField(
@@ -261,11 +263,9 @@ class DocReviewer(dspy.Module):
                     categories=[IssueCategory.DOCUMENTATION],
                     **pf_kwargs,
                 )
-                issues = [
-                    issue
-                    for issue in (result.issues or [])
-                    if issue.confidence >= self._settings.review.min_confidence
-                ]
+                issues = filter_by_confidence(
+                    result.issues, self._settings.review.min_confidence, "doc", scope.subroot
+                )
                 if hippo is not None:
                     await hippo.aobserve(result)
                     # Fire-and-forget background episode save

{codespy_ai-2.0.0 → codespy_ai-2.0.1}/src/codespy/agents/review/helpers.py

@@ -267,3 +267,36 @@ def scope_package_names(scope: ScopeResult | Any) -> list[str]:
     if manifest is not None and getattr(manifest, "package_name", None):
         return [manifest.package_name]
     return []
+
+
+def filter_by_confidence(
+    issues: list[Issue] | None,
+    min_confidence: float,
+    module: str,
+    scope_subroot: str,
+) -> list[Issue]:
+    """Filter issues by confidence threshold with logging.
+
+    Args:
+        issues: Raw issues list (may be None)
+        min_confidence: Minimum confidence threshold (0.0-1.0)
+        module: Module name for logging (e.g., 'code_review')
+        scope_subroot: Scope identifier for logging
+
+    Returns:
+        Filtered issues with confidence >= min_confidence
+    """
+    if issues is None:
+        logger.info("%s scope %s: 0 raw issues, 0 kept (min_confidence=%.2f)",
+                    module, scope_subroot, min_confidence)
+        return []
+
+    raw_count = len(issues)
+    kept = [issue for issue in issues if issue.confidence >= min_confidence]
+    kept_count = len(kept)
+
+    logger.info(
+        "%s scope %s: %d raw issues, %d kept (min_confidence=%.2f)",
+        module, scope_subroot, raw_count, kept_count, min_confidence
+    )
+    return kept

{codespy_ai-2.0.0 → codespy_ai-2.0.1}/src/codespy/agents/review/models.py

@@ -128,7 +128,12 @@ class Issue(BaseModel):
     cwe_id: str | None = Field(
         default=None, description="CWE ID for security issues (e.g., CWE-79)"
     )
-    confidence: float = Field(default=0.8, ge=0.0, le=1.0, description="Confidence score (0-1)")
+    confidence: float = Field(
+        default=0.8,
+        ge=0.0,
+        le=1.0,
+        description="Confidence score (0-1). ≥0.9 verified with tools; 0.7–0.9 strong evidence; <0.7 weak"
+    )
 
     @property
     def location(self) -> str:

{codespy_ai-2.0.0 → codespy_ai-2.0.1}/src/codespy/agents/review/scope/agent.py

@@ -18,13 +18,13 @@ import dspy  # type: ignore[import-untyped]
 from pydantic import BaseModel, Field
 
 from codespy.agents import SignatureContext, get_cost_tracker
-from codespy.agents.context_safe import ContextSafe
 from codespy.agents.memory.hippocampus import ContextMemory, Hippocampus
 from codespy.agents.memory.hippocampus.episode import submit_episode_save
 from codespy.agents.memory.prefrontal import build_facets, with_prefrontal_memory
 from codespy.agents.review.models import ReviewContext
 from codespy.agents.review.scope.manifest_parser import extract_package_name
 from codespy.agents.review.scope.models import PackageManifest, ScopeResult, ScopeType
+from codespy.agents.rlm_tools import build_rlm_agent, log_rlm_outcome, run_rlm
 from codespy.config import get_settings
 from codespy.config_memory import get_cerebral, get_episode_store, get_prefrontal
 from codespy.tools.git.client import get_client
@@ -1060,15 +1060,9 @@ class ScopeResolver(dspy.Module):
             )
             recall_tool = pf.recall_tool() if pf else None
             agent_tools = [*tools, recall_tool] if recall_tool else tools
-            agent = ContextSafe(
-                dspy.RLM(
-                    sig,
-                    tools=agent_tools,
-                    max_iters=max_iters,
-                    max_llm_calls=self._settings.get_max_llm_calls("scope"),
-                ),
+            agent = build_rlm_agent(
                 sig,
-                tools=agent_tools,
+                agent_tools,
                 name="scope",
                 max_iters=max_iters,
                 max_llm_calls=self._settings.get_max_llm_calls("scope"),
@@ -1123,7 +1117,8 @@ class ScopeResolver(dspy.Module):
                         run_id=run_id,
                         initial_memory=scope_initial_memory,
                     )
-                result = await agent.acall(
+                result = await run_rlm(
+                    agent,
                     candidates=candidates_str,
                     orphan_files=[f.filename for f in orphans],
                     pr_title=pr.title or "No title",
@@ -1131,6 +1126,7 @@ class ScopeResolver(dspy.Module):
                     project_instructions=project_instructions,
                     **pf_kwargs,
                 )
+                log_rlm_outcome(result, "scope", pr.repo_full_name)
                 if hippo is not None:
                     await hippo.aobserve(result)