localcode 0.3.6__tar.gz → 0.3.7__tar.gz

{localcode-0.3.6/src/localcode.egg-info → localcode-0.3.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: localcode
-Version: 0.3.6
+Version: 0.3.7
 Summary: High-performance AI coding on consumer hardware.
 Author: LocalCode contributors
 License-Expression: Apache-2.0

{localcode-0.3.6 → localcode-0.3.7}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "localcode"
-version = "0.3.6"
+version = "0.3.7"
 description = "High-performance AI coding on consumer hardware."
 readme = "README.md"
 requires-python = ">=3.10"

{localcode-0.3.6 → localcode-0.3.7}/src/localcode/__init__.py

@@ -1,3 +1,3 @@
 __all__ = ["__version__"]
 
-__version__ = "0.3.6"
+__version__ = "0.3.7"

{localcode-0.3.6 → localcode-0.3.7}/src/localcode/agent/loop.py

@@ -40,6 +40,7 @@ __all__ = ["run_agent_loop"]
 
 from .constants import (
     MAX_ROUNDS,
+    CHURN_FILE_WRITE_LIMIT,
     MAX_OUTPUT_TOKENS,
     MAX_THINKING_SECONDS,
     MAX_THINKING_CHARS,
@@ -70,6 +71,7 @@ from .prompts import (
 )
 from .prompt_context import build_agent_system_prompt
 from .recovery import (
+    rewrite_hard_stop,
     StallMode,
     detect_stall,
     nudge_for,
@@ -1141,33 +1143,13 @@ def run_agent_loop(
             _ephemeral_nudge_indices.append(len(messages) - 1)
 
         # ── Stalled-round recovery ──
-        # Three failure modes end a round with the model NOT actually
-        # done, but the user watching a frozen screen:
-        #
-        # (A) Empty round — long reasoning, then server closes the
-        #     stream with NO content and NO tool_calls. Images 107/108.
-        #
-        # (B) Intent-without-action — model emits ONE sentence of
-        #     forward-looking narration ("I'll build the web app now.
-        #     Let me create this.") and stops, with no tool call. Image
-        #     112. Happens because IQ2/IQ3 sometimes exits the reasoning
-        #     channel with just a sign-off instead of an action.
-        #
-        # (C) Gave-up-after-rejection — most recent tool call returned
-        #     a "REJECTED" / "Error" string AND the round ended with no
-        #     follow-up tool call. The model hit a snag and bailed
-        #     instead of reading the rejection's actionable feedback
-        #     and retrying. This was image 123: write_file got
-        #     REJECTED for syntax error, model stopped completely.
-        #
-        # Stall detection + nudge. Logic lives in agent/recovery.py
-        # (T0.1-d split). `stall` is None for a productive round,
-        # otherwise a `StallMode` enum that drives both the telemetry
-        # label and the per-mode nudge text.
-        # Gated on Feature.AUTO_NUDGE_RECOVERY. Disabled → stalls end
-        # the turn silently (no synthetic SYSTEM: nudge), which is the
-        # pre-recovery behaviour and what eval wants when A/B-ing the
-        # nudge feature's actual contribution to task completion.
+        # Three failure modes end a round with the model NOT done: (A) empty
+        # round (long reasoning, stream closes with no content/tools); (B)
+        # intent-without-action (one narration sentence, no tool call); (C)
+        # gave-up-after-rejection (last tool returned REJECTED/Error, round
+        # ended with no retry). Detection + nudge live in agent/recovery.py;
+        # `stall` is None for a productive round else a StallMode. Gated on
+        # Feature.AUTO_NUDGE_RECOVERY (off → stalls end the turn silently).
         from ..features import Feature, is_enabled as _is_enabled
         stall = detect_stall(
             tool_calls=tool_calls,
@@ -1508,11 +1490,22 @@ def run_agent_loop(
                     "smallest targeted edit and verify it."
                 )
 
+            # HARD rewrite-stop: the churn NUDGE (limit 3) only advises — logs
+            # showed a model rewrite one file 16x while 25-34 nudges fired.
+            # Past 2x the nudge limit, REJECT further full rewrites (see
+            # recovery.rewrite_hard_stop). Key on the RAW path like the counter.
+            _rewrite_limit_stub = None
+            if tool_name in ("write_file", "append_file", "multi_edit") and isinstance(args, dict):
+                _rw_path = args.get("path") or args.get("file_path") or ""
+                _rewrite_limit_stub = rewrite_hard_stop(_rw_path, _file_write_counts)
             # Execute (timed — wall-clock added to _round_tool_exec_ms
             # so round_end can show the model what fraction of its
             # round time was spent waiting on tools vs LLM).
             _tool_started_at = time.monotonic()
-            if _edit_sequence_stub is not None:
+            if _rewrite_limit_stub is not None:
+                from ..tools import ToolResult as _ToolResult
+                _tool_result_obj = _ToolResult(text=_rewrite_limit_stub, ok=False, facts={"tool": tool_name, "ok": False, "repeated_failed_call": True, "rewrite_hard_stop": True})
+            elif _edit_sequence_stub is not None:
                 from ..tools import ToolResult as _ToolResult
                 _tool_result_obj = _ToolResult(text=_edit_sequence_stub, ok=False, facts={"tool": tool_name, "ok": False, "edit_sequence": "missing_context"})
             elif _oversize_stub is not None:

{localcode-0.3.6 → localcode-0.3.7}/src/localcode/agent/recovery.py

@@ -127,25 +127,12 @@ def detect_stall(
     if not content:
         return StallMode.EMPTY
 
-    # NARRATION. Real failures showed the model
-    # emitting hundreds of chars describing the edit it was about to
-    # make, then ending the round with zero tool calls.
-    #
-    # Important: do NOT classify every short answer after a tool call as
-    # narration. A question like "weather in ny?" legitimately uses one
-    # tool, then answers in ~200 chars. The old broad `<400 chars after
-    # tools` rule deleted that answer, injected an auto-nudge, and made
-    # the model call the same API repeatedly. Only intent phrases count
-    # as narration.
-    #
-    # Two complementary detectors:
-    #   • _NARRATION_INTENT_RE — "let me X" / "I'll Y" / "going to Z"
-    #     anywhere in the content (full scan, not just tail) catches
-    #     case (1) and similar variants where the commit phrase may
-    #     be mid-content rather than at the end.
-    #   • _NARRATION_PRESENT_PARTICIPLE_RE — last sentence starts with
-    #     "Updating X / Changing Y / Now writing Z" + target. Past
-    #     tense "Updated X" / "Changed Y" deliberately excluded.
+    # NARRATION detection. Only INTENT phrases count — not every short answer
+    # after a tool call (a "weather in ny?" answer legitimately runs one tool
+    # then replies in ~200 chars; the old broad rule deleted such answers and
+    # looped). Two detectors: _NARRATION_INTENT_RE ("let me X"/"I'll Y" anywhere)
+    # and _NARRATION_PRESENT_PARTICIPLE_RE (last sentence "Updating X/Now writing
+    # Z" + target; past tense "Updated X" excluded).
     stripped = content.strip()
     if bool(tools_called_prior) and stripped:
         if _NARRATION_INTENT_RE.search(stripped):
@@ -398,3 +385,19 @@ def churn_nudge_for(signal: ChurnSignal) -> str:
         "disk that you did not create are not part of your task. If you truly "
         "lack information only the user has, ask ONE focused question."
     )
+
+
+def rewrite_hard_stop(path, file_write_counts) -> str | None:
+    """Circuit-breaker for rewrite drift: reject once a path is rewritten past
+    2x the nudge limit (the nudge only advises; logs showed 16x on one file)."""
+    if not path or not isinstance(path, str):
+        return None
+    n = file_write_counts.get(path, 0)
+    if n < CHURN_FILE_WRITE_LIMIT * 2:
+        return None
+    return (
+        f"REJECTED — HARD STOP: you have rewritten {path} {n} times this turn "
+        "and it's still not right — full rewrites are drifting. Make ONE "
+        "targeted edit_file change to the broken line, or leave this file and "
+        "move on. Do NOT call write_file on this path again."
+    )

{localcode-0.3.6 → localcode-0.3.7}/src/localcode/tools/read_file.py

@@ -32,12 +32,52 @@ SCHEMA = {
 }
 
 
+def _suggest_path(missing) -> str:
+    """Suggest the real file when the model gave a typo'd/wrong-case path.
+
+    Walk up to the nearest existing ancestor dir and fuzzy-match the missing
+    filename (and its path segments) against what's actually there.
+    """
+    import difflib
+    from pathlib import Path
+    try:
+        missing = Path(missing)
+        name = missing.name
+        # Nearest existing ancestor to search in.
+        base = missing.parent
+        for _ in range(6):
+            if base.exists():
+                break
+            base = base.parent
+        if not base.exists():
+            return ""
+        candidates = []
+        for p in base.rglob("*"):
+            if p.is_file():
+                candidates.append(p)
+            if len(candidates) > 4000:
+                break
+        names = [p.name for p in candidates]
+        close = difflib.get_close_matches(name, names, n=1, cutoff=0.6)
+        if close:
+            match = next(p for p in candidates if p.name == close[0])
+            return f"Did you mean: {match} — read THAT exact path (don't guess variants)."
+    except Exception:
+        pass
+    return ""
+
+
 def execute(ctx: ToolContext, args: dict) -> str:
     if "path" not in args:
         return "Error: 'path' argument is required for read_file."
     path = ctx.repo / args["path"]
     if not path.exists():
-        return f"File not found: {args['path']}"
+        # A small model invents misspelled/wrong-case paths ("Aki" for "Anki",
+        # gitHub/github/Github) and, getting a bare "not found", retries with
+        # ANOTHER wrong variant forever (dedup can't collapse differing typos).
+        # Point it at the real nearby file so it corrects instead of guessing.
+        hint = _suggest_path(path)
+        return f"File not found: {args['path']}" + (f"\n{hint}" if hint else "")
     content = path.read_text(errors="replace")
     lines = content.splitlines()
     offset = args.get("offset", 0)

{localcode-0.3.6 → localcode-0.3.7/src/localcode.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: localcode
-Version: 0.3.6
+Version: 0.3.7
 Summary: High-performance AI coding on consumer hardware.
 Author: LocalCode contributors
 License-Expression: Apache-2.0

{localcode-0.3.6 → localcode-0.3.7}/tests/test_architecture.py

@@ -67,6 +67,9 @@ LEGACY_BIG_FILES: dict[str, int] = {
     # whole point of the extraction was to take that bulk out of runtime.py.
     "runtime_diffusion.py": 837,
     "agent/loop.py": 1675,
+    # Anti-repetition/self-conditioning recovery grew: churn nudges (rewritten
+    # forward-only), the rewrite hard-stop circuit-breaker, and stall detection.
+    "agent/recovery.py": 405,
     "server_manager.py": 667,
     "skills.py": 589,
     "tui/screens/setup.py": 946,