localcode 0.3.4__tar.gz → 0.3.5__tar.gz

{localcode-0.3.4/src/localcode.egg-info → localcode-0.3.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: localcode
-Version: 0.3.4
+Version: 0.3.5
 Summary: High-performance AI coding on consumer hardware.
 Author: LocalCode contributors
 License-Expression: Apache-2.0

{localcode-0.3.4 → localcode-0.3.5}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "localcode"
-version = "0.3.4"
+version = "0.3.5"
 description = "High-performance AI coding on consumer hardware."
 readme = "README.md"
 requires-python = ">=3.10"

{localcode-0.3.4 → localcode-0.3.5}/src/localcode/__init__.py

@@ -1,3 +1,3 @@
 __all__ = ["__version__"]
 
-__version__ = "0.3.4"
+__version__ = "0.3.5"

{localcode-0.3.4 → localcode-0.3.5}/src/localcode/agent/loop.py

@@ -715,22 +715,22 @@ def run_agent_loop(
                     list(getattr(_tool_exec_state, "files_read", {}) or {}),
                     progress_ledger_budget_chars(_win_tokens),
                 )
-                if _ledger:
-                    model_messages = list(model_messages) + [
-                        {"role": "user", "content": _ledger}
-                    ]
             except Exception:
-                pass
-            # Inject the working-memory checklist (todo_write tool): the model
-            # always sees its own plan (done / in-progress / left) so it
-            # advances instead of repeating. Appended LAST, ephemeral, empty-safe.
+                _ledger = ""
+            # Working-memory checklist (todo_write): model sees its own plan.
+            _todo_note = ""
             try:
                 from ..tools.todo_write import render_todo_reminder
-                _todos = list(getattr(app.session, "todos", []) or [])
-                _todo_note = render_todo_reminder(_todos)
-                if _todo_note:
+                _todo_note = render_todo_reminder(list(getattr(app.session, "todos", []) or []))
+            except Exception:
+                _todo_note = ""
+            # Inject ledger + todo as ONE trailing SYSTEM message (not role:user
+            # — as a user turn the model re-greeted every round). Ephemeral.
+            try:
+                _ctx_block = "\n\n".join(b for b in (_ledger, _todo_note) if b)
+                if _ctx_block:
                     model_messages = list(model_messages) + [
-                        {"role": "user", "content": _todo_note}
+                        {"role": "system", "content": _ctx_block}
                     ]
             except Exception:
                 pass

{localcode-0.3.4 → localcode-0.3.5}/src/localcode/agent/prompts.py

@@ -164,9 +164,12 @@ def model_identity_line(model: str) -> str:
     # to its training priors otherwise answers "I'm Gemma by Google". So: lead
     # with LocalCode, disclose the model only on request.
     return (
-        f"You are LocalCode — that is your name and how you refer to yourself in "
-        f"conversation (say \"I'm LocalCode\", never lead with the model name). "
-        f"Do NOT volunteer the underlying model in greetings or ordinary replies. "
+        f"Your name is LocalCode (never lead with the underlying model name). "
+        f"Do NOT introduce or re-introduce yourself. Never begin a message with "
+        f"\"I'm LocalCode\", a greeting, or a restatement of what you're doing — "
+        f"the user already knows who you are and what the task is. Just continue "
+        f"the work: take the next action directly. Only state your name if the "
+        f"user explicitly asks who you are. "
         f"The model currently powering you is \"{friendly}\"; report it as EXACTLY "
         f"\"{friendly}\" — and only — when the user specifically asks which "
         f"model / version / quant you are, or asks to name something (a file, "

{localcode-0.3.4 → localcode-0.3.5}/src/localcode/memory_guard.py

@@ -209,3 +209,61 @@ def recommended_jetsam_limit_mb() -> int:
         total_mb = 16 * 1024
     reserve = 3500 if total_mb < 24 * 1024 else 6000
     return max(2048, total_mb - reserve)
+
+
+def should_disable_metal_residency(model_path: str, ctx_size: int) -> bool:
+    """Whether to set GGML_METAL_NO_RESIDENCY=1 for this launch.
+
+    Root cause of the recurring mid-stream SIGKILL (-9, no crash report, tens
+    of GB of system RAM still free): on Apple Silicon, llama.cpp's Metal
+    residency sets keep the model weights + KV cache + compute buffers WIRED
+    for ~3 minutes (ggml-metal-device.m). During a long generation the wired
+    total grows, and when a decode-step spike crosses the GPU's
+    `recommendedMaxWorkingSetSize` (~75% of RAM — a hard ceiling independent of
+    free system RAM), the kernel kills the process outright. This is the exact
+    bug class in ggml-org/llama.cpp#16646 on the same M4 Max / 128 GB hardware.
+
+    Setting GGML_METAL_NO_RESIDENCY=1 makes those buffers EVICTABLE instead of
+    wired, so the OS pages under pressure rather than killing — turning a hard
+    crash into (at worst) mild slowdown, and only when actually near the cap.
+
+    Gated to at-risk launches so the common small-context fast path is
+    untouched: large context (>=128K, where KV + compute-buffer growth is the
+    trigger) OR a model whose weights alone are a large fraction of the wired
+    cap. No user action, no sudo — we control the spawn env.
+    """
+    try:
+        r = subprocess.run(["sysctl", "-n", "hw.memsize"],
+                           capture_output=True, text=True, timeout=2)
+        total_mb = int(r.stdout.strip()) // (1024 * 1024)
+    except Exception:
+        return False
+    # Metal's wired ceiling is ~75% of physical RAM.
+    wired_cap_mb = int(total_mb * 0.75)
+    if ctx_size >= 131072:
+        return True
+    try:
+        import os as _os
+        model_mb = _os.path.getsize(model_path) // (1024 * 1024)
+    except Exception:
+        model_mb = 0
+    # Weights alone over ~55% of the wired cap → KV + compute will breach it.
+    return model_mb > int(wired_cap_mb * 0.55)
+
+
+def metal_residency_env(model_path: str, cmd: list) -> dict:
+    """Return {GGML_METAL_NO_RESIDENCY: '1'} when this launch is at risk of the
+    wired-cap SIGKILL, else {}. macOS-only; parses --ctx-size from the command.
+    """
+    import sys as _sys
+    if _sys.platform != "darwin":
+        return {}
+    ctx = 0
+    for i, tok in enumerate(cmd):
+        if tok == "--ctx-size" and i + 1 < len(cmd):
+            try:
+                ctx = int(cmd[i + 1])
+            except Exception:
+                ctx = 0
+            break
+    return {"GGML_METAL_NO_RESIDENCY": "1"} if should_disable_metal_residency(model_path, ctx) else {}

{localcode-0.3.4 → localcode-0.3.5}/src/localcode/server_manager.py

@@ -299,12 +299,10 @@ class ServerManager:
         self._lock = threading.Lock()
         self._last_exit_code: Optional[int] = None  # disconnect diagnostics
         self._last_death_was_pressure: bool = False
-        # Idle auto-suspend: track wall-clock of the most recent chat
-        # activity. A background watchdog thread shuts the server down
-        # after `_idle_timeout_s` seconds of inactivity to stop the GPU
-        # from cooking the laptop while the user is reading replies.
-        # The next chat call transparently respawns via _restart_server.
-        # 0 disables; default 10 min. Override with env LOCALCODE_IDLE_SUSPEND_S.
+        # Idle auto-suspend: a watchdog shuts the server down after
+        # `_idle_timeout_s` of inactivity (stops the GPU cooking the laptop
+        # while the user reads); the next chat call respawns via
+        # _restart_server. 0 disables; default 10 min. Env: LOCALCODE_IDLE_SUSPEND_S.
         import os as _os
         try:
             self._idle_timeout_s: float = float(
@@ -394,10 +392,8 @@ class ServerManager:
             _prev_model = self._model_path  # capture before shutdown clears it
             self._shutdown_locked()
             # Wait for the OLD server's memory to release before spawning the
-            # new one: the kernel takes 1-3 s to free wired Metal memory after
-            # the child reaps, and spawning into that window double-commits and
-            # trips the pressure monitor on a 16 GB Mac. Best-effort — targets
-            # ~11 GB free, times out at 8 s, proceeds anyway.
+            # new one (kernel takes 1-3 s to free wired Metal memory; spawning
+            # into that window double-commits). Best-effort: ~11 GB target, 8 s.
             if had_prior:
                 # Only a genuine model change is "model_switch"; a same-model
                 # relaunch (crash/disconnect recovery) is "server_restart".
@@ -422,6 +418,15 @@ class ServerManager:
             # AGX driver may read the env at first Metal touch — setting it in
             # the parent env before spawn is the belt-and-suspenders approach.
             env.setdefault("AGX_RELAX_CDM_CTXSTORE_TIMEOUT", "1")
+            # Disable Metal residency sets on at-risk launches → prevents the wired-cap SIGKILL.
+            try:
+                from .memory_guard import metal_residency_env
+                _renv = metal_residency_env(str(model_path), cmd)
+                if _renv:
+                    env.update(_renv)
+                    _lifecycle_log("metal_no_residency", model=str(model_path))
+            except Exception:
+                pass
             # Capture the server's OWN stdout/stderr to server.log (this runtime
             # path used to send both to /dev/null, discarding llama-server's
             # final ggml/Metal error on the recurring -9). Append per-spawn.
@@ -461,11 +466,10 @@ class ServerManager:
             except Exception:
                 pass
 
-            # Layer 2: userspace memory-pressure monitor. Polls
-            # kern.memorystatus_vm_pressure_level every 500 ms; on WARN
-            # transition, SIGTERMs our server before the kernel stalls.
-            # Complementary to the jetsam ceiling — acts on system-wide
-            # pressure (everything else swelling) not just our own RSS.
+            # Layer 2: userspace pressure monitor — polls
+            # kern.memorystatus_vm_pressure_level every 500 ms and SIGTERMs our
+            # server on WARN. Complements the jetsam ceiling (acts on system-
+            # wide pressure, not just our own RSS).
             try:
                 self._pressure_thread = start_pressure_monitor(
                     self._process,
@@ -474,10 +478,8 @@ class ServerManager:
             except Exception:
                 self._pressure_thread = None
             self._write_pid_file(self._process.pid)
-            # Log the FULL launch command so debugging can answer "what flags
-            # was the server using?" — subtle flags (e.g. --spec-type ngram-mod)
-            # cause repetition pathologies that look like model bugs. Truncated
-            # to 4000 chars (~80 args) to keep the event small.
+            # Log the FULL launch command (which flags was the server using?);
+            # truncated to 4000 chars (~80 args) to keep the event small.
             _flags_str = " ".join(str(c) for c in cmd)[:4000]
             _lifecycle_log("server_started", pid=self._process.pid, port=port,
                            model=Path(model_path).name,
@@ -698,10 +700,8 @@ class ServerManager:
         _lifecycle_log("pressure_kill", level=level, pid=killed_pid,
                        guard="memory_pressure", exit_code=self._last_exit_code,
                        free_mb=_system_free_memory_mb())
-        # Clear internal state so subsequent code paths see "no server
-        # running" rather than a phantom Popen handle. Safe to do from
-        # the pressure-monitor thread because `_process` reads/writes
-        # are atomic in CPython for the assignment.
+        # Clear internal state so callers see "no server" not a phantom Popen.
+        # Safe from the pressure-monitor thread (assignment is atomic in CPython).
         self._process = None
         self._model_path = None
 

{localcode-0.3.4 → localcode-0.3.5}/src/localcode/tools/write_file.py

@@ -173,6 +173,19 @@ def execute(ctx: ToolContext, args: dict) -> str:
     # catches the data.py-style "model copies in-memory stub back into
     # write_file" disaster, which was the one case where rejecting
     # was load-bearing.
+    # Log the EXACT content the model sent (post-JSON-parse) so a "the file is
+    # corrupted / write_file isn't writing what I sent" report is provable from
+    # logs in one grep, instead of a mystery. Best-effort, capped, append-only.
+    try:
+        from ..paths import global_state_dir
+        _wl = global_state_dir() / "write_content.log"
+        with open(_wl, "a", encoding="utf-8", errors="replace") as _fh:
+            _fh.write(f"\n===== write_file {args['path']} ({len(content)} chars) =====\n")
+            _fh.write(content[:20000])
+            _fh.write("\n")
+    except Exception:
+        pass
+
     existed = path.is_file()
     # Capture the prior content so an in-place rewrite renders as a diff card
     # (like edit_file/multi_edit), not a bare "Rewrote (N lines)" one-liner.

{localcode-0.3.4 → localcode-0.3.5/src/localcode.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: localcode
-Version: 0.3.4
+Version: 0.3.5
 Summary: High-performance AI coding on consumer hardware.
 Author: LocalCode contributors
 License-Expression: Apache-2.0