raga-cli 0.8.0 → 0.9.0

package/dist/bundle/raga-runtime.js

@@ -1,3 +1,3 @@
 #!/usr/bin/env node
 import { createRequire as __piCreateRequire } from "node:module"; const require = __piCreateRequire(import.meta.url);
-import{setupCli}from"./chunks/chunk-YMZUNB25.js";import{main}from"./chunks/chunk-AI2NMHE7.js";import"./chunks/chunk-J22MF7YW.js";import"./chunks/chunk-3EGDWOKT.js";import"./chunks/chunk-IOPE2TJ4.js";import"./chunks/chunk-5AGIVGSV.js";import"./chunks/chunk-RP2WKNT7.js";import"./chunks/chunk-RFPFZTK2.js";import"./chunks/chunk-P2QCWBFM.js";import"./chunks/chunk-6QGV3IWD.js";import"./chunks/chunk-KNY3PXQI.js";import"./chunks/chunk-USJZMAIR.js";import"./chunks/chunk-TMTOQ7H7.js";import"./chunks/chunk-UKVNB6C4.js";import"./chunks/chunk-3V45WARH.js";import"./chunks/chunk-7DDKXS2Q.js";import"./chunks/chunk-IXDQLPKS.js";import"./chunks/chunk-LPXNMA2W.js";import"./chunks/chunk-H76KZMOI.js";import"./chunks/chunk-JXWJOIH4.js";setupCli();main(process.argv.slice(2));
+import{setupCli}from"./chunks/chunk-OPPYO2IE.js";import{main}from"./chunks/chunk-F7LFZTHT.js";import"./chunks/chunk-NP6TH6UT.js";import"./chunks/chunk-JQDZ4Z7L.js";import"./chunks/chunk-ENUUAC3O.js";import"./chunks/chunk-5AGIVGSV.js";import"./chunks/chunk-RP2WKNT7.js";import"./chunks/chunk-RFPFZTK2.js";import"./chunks/chunk-P2QCWBFM.js";import"./chunks/chunk-6QGV3IWD.js";import"./chunks/chunk-KNY3PXQI.js";import"./chunks/chunk-USJZMAIR.js";import"./chunks/chunk-TMTOQ7H7.js";import"./chunks/chunk-UKVNB6C4.js";import"./chunks/chunk-3V45WARH.js";import"./chunks/chunk-7DDKXS2Q.js";import"./chunks/chunk-IXDQLPKS.js";import"./chunks/chunk-LPXNMA2W.js";import"./chunks/chunk-H76KZMOI.js";import"./chunks/chunk-JXWJOIH4.js";setupCli();main(process.argv.slice(2));

package/dist/bundle/rpc-entry.js

@@ -1,3 +1,3 @@
 #!/usr/bin/env node
 import { createRequire as __piCreateRequire } from "node:module"; const require = __piCreateRequire(import.meta.url);
-import{configureHttpDispatcher,main}from"./chunks/chunk-AI2NMHE7.js";import"./chunks/chunk-J22MF7YW.js";import"./chunks/chunk-3EGDWOKT.js";import"./chunks/chunk-IOPE2TJ4.js";import"./chunks/chunk-5AGIVGSV.js";import"./chunks/chunk-RP2WKNT7.js";import"./chunks/chunk-RFPFZTK2.js";import"./chunks/chunk-P2QCWBFM.js";import"./chunks/chunk-6QGV3IWD.js";import{APP_NAME}from"./chunks/chunk-KNY3PXQI.js";import"./chunks/chunk-USJZMAIR.js";import"./chunks/chunk-TMTOQ7H7.js";import"./chunks/chunk-UKVNB6C4.js";import"./chunks/chunk-3V45WARH.js";import"./chunks/chunk-7DDKXS2Q.js";import"./chunks/chunk-IXDQLPKS.js";import"./chunks/chunk-LPXNMA2W.js";import"./chunks/chunk-H76KZMOI.js";import"./chunks/chunk-JXWJOIH4.js";process.title=`${APP_NAME}-rpc`;process.env.PI_CODING_AGENT="true";process.env.AI_AGENT="pi";process.emitWarning=(()=>{});configureHttpDispatcher();main(["--mode","rpc",...process.argv.slice(2)]);
+import{configureHttpDispatcher,main}from"./chunks/chunk-F7LFZTHT.js";import"./chunks/chunk-NP6TH6UT.js";import"./chunks/chunk-JQDZ4Z7L.js";import"./chunks/chunk-ENUUAC3O.js";import"./chunks/chunk-5AGIVGSV.js";import"./chunks/chunk-RP2WKNT7.js";import"./chunks/chunk-RFPFZTK2.js";import"./chunks/chunk-P2QCWBFM.js";import"./chunks/chunk-6QGV3IWD.js";import{APP_NAME}from"./chunks/chunk-KNY3PXQI.js";import"./chunks/chunk-USJZMAIR.js";import"./chunks/chunk-TMTOQ7H7.js";import"./chunks/chunk-UKVNB6C4.js";import"./chunks/chunk-3V45WARH.js";import"./chunks/chunk-7DDKXS2Q.js";import"./chunks/chunk-IXDQLPKS.js";import"./chunks/chunk-LPXNMA2W.js";import"./chunks/chunk-H76KZMOI.js";import"./chunks/chunk-JXWJOIH4.js";process.title=`${APP_NAME}-rpc`;process.env.PI_CODING_AGENT="true";process.env.AI_AGENT="pi";process.emitWarning=(()=>{});configureHttpDispatcher();main(["--mode","rpc",...process.argv.slice(2)]);

package/docs/cli.md

@@ -270,6 +270,12 @@ See [Pi Packages](packages.md) for source formats, filtering, installation, and
 
 Add `--local` or `-l` to `install`, `remove`, `uninstall`, or `config` to use project settings instead of global settings.
 
+### Update raga
+
+When using the `raga` command, run `raga update` in your terminal to update the CLI from the latest published `raga-cli` npm release. `raga update --self` does the same thing. Restart raga afterward. If an older CLI cannot update itself, use `npm install -g raga-cli@latest`.
+
+Use `raga update --extensions` for installed packages, `raga update --models` for model catalogs, or `raga update --all` for the CLI and packages. Model catalog refreshes do not download local model weights. Run `raga update --help` for all options. Settings, authentication, and sessions remain in the separate raga user directory.
+
 ### Update Pi or packages
 
 Running `pi update` without a target updates Pi itself.

package/docs/compaction.md

@@ -414,6 +414,30 @@ pi.on("session_before_tree", async (event, ctx) => {
 
 See `SessionBeforeTreeEvent` and `TreePreparation` in the types file.
 
+## Archive compaction
+
+Set `"compaction": { "strategy": "archive" }` to compact without a summarization request. Manual `/compact` and automatic compaction both use the strategy. An extension that returns its own compaction from `session_before_compact` still takes precedence.
+
+On compaction, raga:
+
+1. Writes the active branch, root to leaf, to a plain-text log beside the session file: `<session>.archive-<last entry id>.log`. An in-memory session writes to a private temporary directory.
+2. Appends a compaction entry that keeps no earlier entries (`firstKeptEntryId` is the entry's own ID). `details` is `{ strategy: "archive", archivePath, archivedEntries }`.
+3. Uses a fixed message as the compaction summary. It names the archive path and repeats the latest user request verbatim.
+4. Adds a `session_archive` section to the system prompt of later requests. It gives the path, explains the record format, and tells the agent to search the log with `rg` and read only the surrounding records. The section names the newest archive on the active branch. It disappears when the newest compaction on the branch is not an archive.
+
+Each record in the log starts with a header line:
+
+```text
+=== 3 tool_result bash call_1 4f2a1c9e 2026-10-30T12:01:44.000Z
+57 tests passed
+```
+
+The header holds the record number, the kind (`user`, `assistant`, `tool_result`, `user_bash`, `custom`, `compaction`, `branch_summary`, `model_change`, or `system_prompt_update`), any tool name and call ID, the session entry ID, and the timestamp. Assistant records include tool calls and their arguments. Thinking blocks, images, and system prompt contents are omitted. Text longer than 8,000 characters keeps its first 5,000 and last 2,000 characters and names the session entry that holds the full text.
+
+The JSONL session file remains the source of truth. The log is a copy written once per compaction. Delete it with the session. It contains everything the session recorded, including any secrets that appeared in tool output.
+
+`customInstructions` from `/compact <instructions>` have no effect in archive mode.
+
 ## Settings
 
 Configure compaction in `~/.pi/agent/settings.json` or `<project-dir>/.raga/settings.json`:
@@ -422,6 +446,7 @@ Configure compaction in `~/.pi/agent/settings.json` or `<project-dir>/.raga/sett
 {
   "compaction": {
     "enabled": true,
+    "strategy": "summary",
     "reserveTokens": 16384,
     "keepRecentTokens": 20000
   }
@@ -431,6 +456,7 @@ Configure compaction in `~/.pi/agent/settings.json` or `<project-dir>/.raga/sett
 | Setting | Default | Description |
 |---------|---------|-------------|
 | `enabled` | `true` | Enable auto-compaction |
+| `strategy` | `"summary"` | `"summary"` asks a model for a summary. `"archive"` writes a log and keeps no earlier context; see [Archive compaction](#archive-compaction) |
 | `reserveTokens` | `16384` | Free context tokens below which auto-compaction starts |
 | `keepRecentTokens` | `20000` | Recent tokens to keep (not summarized) |
 

package/docs/sdk.md

@@ -104,7 +104,8 @@ Each boundary can be supplied explicitly:
 - `sessionManager` supplies persistent or in-memory conversation history.
 - `resourceLoader` supplies extensions, skills, prompt templates, themes, and context files.
 - `tools`, `noTools`, `excludeTools`, and `customTools` control the active tool set.
-- `instance` names a raga instance root (`.raga/` directory) or an opened `HarnessPin`. Without it, `createAgentSession` uses `RAGA_INSTANCE_ROOT` or the nearest `.raga/instance.json` above `cwd`. Unless you pass `settingsManager` or `resourceLoader`, an instance session loads settings and resources only from its pinned revision. It stores sessions under `.raga/state/sessions/`, runs in RLM mode with a bubblewrap-sandboxed kernel, and reads credentials and models from the default user agent directory even when `agentDir` is set. A session manager that already records a pin reopens that revision. `instanceMemory: false` turns live memory retrieval off. `sandbox` supplies an explicit sandbox policy for trusted callers such as evaluators. See [raga instances](../../../docs/INSTANCE-PLAN.md).
+- `instance` names a raga instance root (`.raga/` directory) or an opened `HarnessPin`. Without it, `createAgentSession` uses `RAGA_INSTANCE_ROOT` or the nearest `.raga/instance.json` above `cwd`. Unless you pass `settingsManager` or `resourceLoader`, an instance session loads settings and resources only from its pinned revision. It stores sessions under `.raga/state/sessions/`, runs in RLM mode with a bubblewrap-sandboxed kernel, and reads credentials and models from the default user agent directory even when `agentDir` is set. A session manager that already records a pin reopens that revision. `instanceMemory` takes a `MemoryService` from `createMemoryService`, `createProxyMemoryService`, or `readOnlyMemoryService`; by default `RAGA_MEMORY_BACKEND` selects one (local unless set), and `false` turns live memory retrieval off. `createAgentSessionServices` accepts the same option and passes its selection on. Writing `rlm_query` children inherit that service; read-only children receive a read-only view, and no child can upgrade a read-only service. `sandbox` supplies an explicit sandbox policy for trusted callers such as evaluators. See [raga instances](../../../docs/INSTANCE-PLAN.md#memory-backends).
+- `RAGA_SANDBOX_READ_ONLY` is a host-only, platform-path-delimited list of extra read-only dependency roots for instance sessions and children. Use existing canonical absolute paths outside the project and instance; overlapping ancestors, relative paths, and symlinks are rejected. Normal sandbox safety checks still apply. This does not select the RLM interpreter; `RAGA_PYTHON` does.
 
 Use `DefaultResourceLoader` when you want standard discovery with selected overrides. Supply a custom `ResourceLoader` when the host owns resource storage and discovery completely.
 

package/docs/settings.md

@@ -83,6 +83,7 @@ CLI tool options override this setting for one invocation. `--tools` with only `
 | Setting | Type | Default | Description |
 |---|---|---|---|
 | `compaction.enabled` | boolean | `true` | Enable automatic compaction. |
+| `compaction.strategy` | `"summary" \| "archive"` | `"summary"` | `summary` replaces earlier context with an LLM summary. `archive` writes the earlier conversation to a plain-text log and keeps none of it in context, without a model call. See [Archive compaction](compaction.md#archive-compaction). |
 | `compaction.reserveTokens` | number | `16384` | Free context tokens below which automatic compaction starts. |
 | `compaction.keepRecentTokens` | number | `20000` | Recent tokens retained without summarization. |
 | `compaction.modelOverrides` | object | None | Per-model token settings keyed by exact `provider/modelId`. |

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "raga-cli",
-	"version": "0.8.0",
+	"version": "0.9.0",
 	"description": "Raga, a coding agent for long tasks with a persistent Python kernel (a fork of pi)",
 	"type": "module",
 	"license": "MIT",

package/python/raga_rlm/__init__.py

@@ -206,19 +206,35 @@ class Context:
 
 
 class Memory:
-    """Scoped retrieval and unreviewed observations. Entries never grant permissions."""
+    """Scoped retrieval and unreviewed observations. Entries never grant permissions.
 
-    async def search(self, query: str, *, task_type: str | None = None, max_chars: int = 4000, limit: int = 5) -> dict[str, Any]:
+    Search returns reviewed guidance only. `include_unverified=True` also returns unreviewed and hypothesis
+    entries, such as observations, with their `trust` and `epistemicStatus` labels. An observation is bound to
+    this harness revision unless `portable=True`; its origin records the revision either way.
+    """
+
+    async def search(
+        self,
+        query: str,
+        *,
+        task_type: str | None = None,
+        max_chars: int = 4000,
+        limit: int = 5,
+        include_unverified: bool = False,
+    ) -> dict[str, Any]:
         return await host_request({
             "type": "memory.search",
             **_drop_none(query=query, task_type=task_type, max_chars=max_chars, limit=limit),
+            "include_unverified": include_unverified,
         })
 
     async def get(self, entry_id: str) -> dict[str, Any]:
         return await host_request({"type": "memory.get", "id": entry_id})
 
-    async def observe(self, text: str, evidence: list[str]) -> dict[str, Any]:
-        return await host_request({"type": "memory.observe", "text": text, "evidence": evidence})
+    async def observe(self, text: str, evidence: list[str], *, portable: bool = False) -> dict[str, Any]:
+        """Record an unreviewed hypothesis. Returns {"entry", "ingestion"}: the archived record and whether the
+        selected search backend accepted it."""
+        return await host_request({"type": "memory.observe", "text": text, "evidence": evidence, "portable": portable})
 
 
 class Instance:

package/python/raga_rlm/mlx_models.py

@@ -0,0 +1,265 @@
+"""Private local model manager protocol. Model specs arrive on stdin, never argv.
+
+No mode installs uv or fetches executable model code. Only download uses network.
+"""
+from __future__ import annotations
+
+import argparse
+import importlib
+import importlib.metadata
+import json
+import os
+from pathlib import Path, PurePosixPath
+import platform
+import re
+import shutil
+import subprocess
+import sys
+import time
+import threading
+
+DEPENDENCIES = {"mlx-lm": "0.32.2", "huggingface-hub": "2.4.0", "hf-xet": "1.7.1"}
+OUTPUT_LOCK = threading.Lock()
+
+
+class ModelError(Exception):
+    """A safe, user-facing diagnostic (no hub exceptions or credentials)."""
+
+
+def emit(**fields: object) -> None:
+    with OUTPUT_LOCK:
+        print(json.dumps(fields), flush=True)
+
+
+def validate(spec: dict) -> None:
+    if not isinstance(spec.get("id"), str) or not re.fullmatch(r"[\w.-]+/[\w.-]+", spec["id"]):
+        raise ModelError("Invalid Hugging Face model id")
+    if not isinstance(spec.get("revision"), str) or not re.fullmatch(r"[0-9a-fA-F]{40}", spec["revision"]):
+        raise ModelError("Model revision must be a pinned 40-character commit SHA")
+    files = spec.get("files")
+    if not isinstance(files, list) or not files or not all(isinstance(f, str) for f in files):
+        raise ModelError("An exact model file allowlist is required")
+    for name in files:
+        path = PurePosixPath(name)
+        if path.is_absolute() or ".." in path.parts or str(path) != name or any(c in name for c in "*?[]\\"):
+            raise ModelError("Model files must be exact relative paths, not patterns")
+    if spec.get("format") == "gguf":
+        if not all(f.endswith(".gguf") for f in files):
+            raise ModelError("GGUF selection must contain only exact GGUF files")
+        model_file = spec.get("modelFile", files[0])
+        if model_file not in files:
+            raise ModelError("GGUF modelFile must be in the selected files")
+        context = spec.get("contextWindow")
+        if type(context) is not int or not 1 <= context <= 16384:
+            raise ModelError("GGUF contextWindow must be between 1 and 16384")
+        for name in files:
+            shard = re.fullmatch(r"(.+)-(\d{5})-of-(\d{5})\.gguf", name)
+            if shard:
+                prefix, number, count = shard.groups()
+                total = int(count)
+                if not 1 <= int(number) <= total <= len(files) or any(
+                    f"{prefix}-{i:05d}-of-{count}.gguf" not in files for i in range(1, total + 1)
+                ):
+                    raise ModelError("GGUF selection must include all shards")
+        first = re.fullmatch(r"(.+)-(\d{5})-of-(\d{5})\.gguf", model_file)
+        if first and first[2] != "00001":
+            raise ModelError("GGUF modelFile must be the first shard")
+        return
+    if spec.get("format") is not None:
+        raise ModelError("Unsupported model format")
+    if "config.json" not in files or not any(f.endswith(".safetensors") for f in files):
+        raise ModelError("Model files must include config.json and safetensors weights")
+
+
+def backend() -> str:
+    value = os.environ.get("RAGA_LOCAL_MODEL_BACKEND", "mlx")
+    if value not in ("mlx", "llama.cpp"):
+        raise ModelError("Unknown local model backend")
+    return value
+
+
+def llama_server() -> str:
+    instructions = ("llama-server >= v0.8.0 / b11960 is required. Install or upgrade llama.cpp "
+                    "and put llama-server on PATH, or set LLAMA_SERVER_PATH.")
+    binary = shutil.which(os.environ.get("LLAMA_SERVER_PATH", "llama-server"))
+    if binary is None:
+        raise ModelError(instructions)
+    try:
+        result = subprocess.run([binary, "--version"], capture_output=True, text=True, timeout=15, check=True)
+    except (OSError, subprocess.SubprocessError) as exc:
+        raise ModelError(instructions) from exc
+    text = result.stdout + result.stderr
+    build = re.search(r"\bversion:\s*b?(\d+)(?![\d.])\b", text, re.I) or re.search(r"\bb(\d+)\b", text, re.I)
+    release = re.search(r"(?:\bversion:\s*v?|\bv)(\d+)\.(\d+)\.(\d+)\b", text)
+    if not (int(build[1]) >= 11960 if build else
+            release and tuple(map(int, release.groups())) >= (0, 8, 0)):
+        raise ModelError(instructions)
+    return binary
+
+
+def probe() -> None:
+    gguf = backend() == "llama.cpp"
+    if gguf:
+        if platform.system() not in ("Darwin", "Linux"):
+            raise ModelError("llama.cpp requires macOS or Linux (POSIX process groups)")
+        llama_server()
+    elif platform.system() != "Darwin" or platform.machine() != "arm64":
+        raise ModelError("MLX requires macOS on Apple Silicon (arm64)")
+    for name, version in DEPENDENCIES.items():
+        if gguf and name == "mlx-lm":
+            continue
+        if importlib.metadata.version(name) != version:
+            raise ModelError("Local model dependencies need installation or repair")
+    modules = ("huggingface_hub", "hf_xet") if gguf else ("mlx.core", "mlx_lm.server", "huggingface_hub", "hf_xet")
+    for module in modules:
+        importlib.import_module(module)
+
+
+def cached_snapshot(spec: dict) -> Path | None:
+    """Check every selected file without network, including empty/broken blobs."""
+    from huggingface_hub import hf_hub_download, snapshot_download
+    try:
+        root = Path(snapshot_download(repo_id=spec["id"], revision=spec["revision"],
+                                      allow_patterns=spec["files"], local_files_only=True))
+        for name in spec["files"]:
+            path = Path(hf_hub_download(repo_id=spec["id"], revision=spec["revision"],
+                                        filename=name, local_files_only=True))
+            if not path.is_file() or path.stat().st_size == 0 or not (root / name).is_file() or (root / name).stat().st_size == 0:
+                return None
+        # A partial selection must not omit weight shards referenced by the index.
+        index = root / "model.safetensors.index.json"
+        if spec.get("format") != "gguf" and index.is_file():
+            shards = set(json.loads(index.read_text())["weight_map"].values())
+            if not shards.issubset(spec["files"]):
+                return None
+        return root
+    except (OSError, ValueError, KeyError):
+        return None
+    except Exception:
+        # Hub cache misses use library-specific exception classes. Never echo auth errors.
+        return None
+
+
+def preflight(root: Path) -> None:
+    from mlx_lm.utils import _get_classes
+    config = json.loads((root / "config.json").read_text())
+    model_type = config.get("model_type", "")
+    if not isinstance(model_type, str) or not re.fullmatch(r"[a-zA-Z0-9_]+", model_type):
+        raise ModelError("Unsupported MLX model architecture")
+    architectures = config.get("architectures") or []
+    if isinstance(architectures, str):
+        architectures = [architectures]
+    if "lfm3" in model_type.lower().replace("_", "") or any(
+        "lfm3" in str(arch).lower().replace("_", "") for arch in architectures
+    ):
+        raise ModelError("Unsupported MLX architecture: LFM3. Official mlx-lm support is pending.")
+    try:
+        _get_classes(config)
+    except (ImportError, ValueError, KeyError) as exc:
+        raise ModelError(f"Unsupported MLX architecture: {model_type}. mlx-lm support may be pending (including LFM3).") from exc
+
+
+def progress_class():
+    from huggingface_hub.utils import tqdm
+
+    class Progress(tqdm):
+        def __init__(self, *args, **kwargs):
+            self.last_report = 0.0
+            super().__init__(*args, **kwargs)
+
+        def display(self, msg=None, pos=None):
+            now = time.monotonic()
+            if now - self.last_report < 0.2:
+                return
+            self.last_report = now
+            if self.unit != "B" or not self.desc.startswith("Downloading"):
+                return
+            total = self.total or 0
+            rate = self.format_dict.get("rate") or 0
+            size = f"{self.n / 1048576:.1f} / {total / 1048576:.1f} MiB"
+            emit(message=f"Downloading {size} · {rate / 1048576:.1f} MiB/s")
+
+    return Progress
+
+
+def download(spec: dict) -> None:
+    os.environ["HF_XET_HIGH_PERFORMANCE"] = "1"
+    from huggingface_hub import snapshot_download
+    import hf_xet  # noqa: F401 -- ensure the pinned transfer backend is available
+    # Downloading is allowed even when mlx-lm cannot yet run the architecture (e.g. LFM3).
+    emit(message="Downloading selected model files via Hugging Face cache")
+    snapshot_download(repo_id=spec["id"], revision=spec["revision"], allow_patterns=spec["files"],
+                      max_workers=8, tqdm_class=progress_class())
+    if cached_snapshot(spec) is None:
+        raise ModelError("Download incomplete: selected model files are missing or empty")
+    emit(message="Model download complete", percent=100)
+
+
+def serve(spec: dict, port: int) -> None:
+    # Set before importing HF/transformers/mlx_lm: their constants are import-time.
+    os.environ["HF_HUB_OFFLINE"] = "1"
+    os.environ["TRANSFORMERS_OFFLINE"] = "1"
+    root = cached_snapshot(spec)
+    if root is None:
+        raise ModelError("Selected model files are not fully cached; download first")
+    root = root.resolve()
+    if backend() == "llama.cpp":
+        if spec.get("format") != "gguf":
+            raise ModelError("llama.cpp requires a GGUF model spec")
+        binary = llama_server()
+        model = root / spec.get("modelFile", spec["files"][0])
+        emit(modelPath=str(model))
+        os.execv(binary, [binary, "--model", str(model), "--host", "127.0.0.1", "--port", str(port),
+                          "--alias", spec["id"], "--ctx-size", str(spec["contextWindow"]), "--jinja",
+                          "--temp", "0.1", "--top-k", "50", "--repeat-penalty", "1.1"])
+        return
+    if spec.get("format") == "gguf":
+        raise ModelError("GGUF models require the llama.cpp backend")
+    preflight(root)
+    emit(modelPath=str(root))
+    from mlx_lm import server
+
+    class LocalModelProvider(server.ModelProvider):
+        def __init__(self, cli_args: argparse.Namespace) -> None:
+            # mlx-lm 0.32.2 otherwise loads in a background thread, and /models
+            # can advertise the path before weights/tokenizer have loaded.
+            cli_args.trust_remote_code = False
+            super().__init__(cli_args)
+            self._model_map[spec["id"]] = str(root)
+            self.load_default()
+
+    server.ModelProvider = LocalModelProvider
+    # No --trust-remote-code: pinned mlx-lm defaults to false for tokenizers.
+    sys.argv = ["mlx_lm.server", "--model", str(root), "--host", "127.0.0.1", "--port", str(port)]
+    server.main()
+
+
+def main() -> int:
+    try:
+        mode = sys.argv[1]
+        if mode == "probe":
+            probe()
+            emit(installed=True)
+            return 0
+        spec = json.load(sys.stdin)
+        validate(spec)
+        if mode == "cache":
+            os.environ["HF_HUB_OFFLINE"] = "1"
+            emit(cached=cached_snapshot(spec) is not None)
+        elif mode == "download":
+            download(spec)
+        elif mode == "serve":
+            serve(spec, int(sys.argv[2]))
+        else:
+            raise ModelError("Unknown local model helper mode")
+        return 0
+    except ModelError as exc:
+        emit(error=str(exc))
+    except Exception as exc:
+        # Deliberately omit exception text: requests/hub errors can include credentials.
+        emit(error=f"Local model operation failed ({type(exc).__name__}); check installation, cache, and Hugging Face access")
+    return 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

package/schemas/keybindings.schema.json

@@ -195,6 +195,54 @@
       "$ref": "#/$defs/KeybindingValue",
       "description": "Scroll viewport to bottom"
     },
+    "app.diff.commits": {
+      "$ref": "#/$defs/KeybindingValue",
+      "description": "Diff dashboard: browse commits"
+    },
+    "app.diff.working": {
+      "$ref": "#/$defs/KeybindingValue",
+      "description": "Diff dashboard: working-tree changes"
+    },
+    "app.diff.staged": {
+      "$ref": "#/$defs/KeybindingValue",
+      "description": "Diff dashboard: staged changes"
+    },
+    "app.diff.file": {
+      "$ref": "#/$defs/KeybindingValue",
+      "description": "Diff dashboard: toggle diff and full file"
+    },
+    "app.diff.zoom": {
+      "$ref": "#/$defs/KeybindingValue",
+      "description": "Diff dashboard: toggle fullscreen viewer"
+    },
+    "app.diff.refresh": {
+      "$ref": "#/$defs/KeybindingValue",
+      "description": "Diff dashboard: refresh"
+    },
+    "app.diff.nextCategory": {
+      "$ref": "#/$defs/KeybindingValue",
+      "description": "Diff dashboard: next category"
+    },
+    "app.diff.previousCategory": {
+      "$ref": "#/$defs/KeybindingValue",
+      "description": "Diff dashboard: previous category"
+    },
+    "app.diff.left": {
+      "$ref": "#/$defs/KeybindingValue",
+      "description": "Diff dashboard: scroll left"
+    },
+    "app.diff.right": {
+      "$ref": "#/$defs/KeybindingValue",
+      "description": "Diff dashboard: scroll right"
+    },
+    "app.diff.start": {
+      "$ref": "#/$defs/KeybindingValue",
+      "description": "Diff dashboard: first line"
+    },
+    "app.diff.end": {
+      "$ref": "#/$defs/KeybindingValue",
+      "description": "Diff dashboard: last line"
+    },
     "app.interrupt": {
       "$ref": "#/$defs/KeybindingValue",
       "description": "Cancel or abort"

package/schemas/settings.schema.json

@@ -104,6 +104,20 @@
           "type": "boolean",
           "default": true
         },
+        "strategy": {
+          "anyOf": [
+            {
+              "type": "string",
+              "const": "summary"
+            },
+            {
+              "type": "string",
+              "const": "archive"
+            }
+          ],
+          "description": "\"summary\" replaces earlier context with an LLM summary. \"archive\" writes the earlier conversation to a plain-text log, drops it from context without a model call, and tells the agent to search the log.",
+          "default": "summary"
+        },
         "reserveTokens": {
           "type": "integer",
           "minimum": 0,

package/dist/bundle/chunks/chunk-C4IJNE7U.js

@@ -1,2 +0,0 @@
-import { createRequire as __piCreateRequire } from "node:module"; const require = __piCreateRequire(import.meta.url);
-import{AgentSession,AgentSessionRuntime,AssistantMessageComponent,BashExecutionComponent,BorderedLoader,BranchSummaryMessageComponent,CURRENT_SESSION_VERSION,CompactionSummaryMessageComponent,CredentialSynchronizationError,CustomEditor,CustomMessageComponent,DEFAULT_COMPACTION_SETTINGS,DefaultPackageManager,DefaultResourceLoader,DynamicBorder,ExtensionEditorComponent,ExtensionInputComponent,ExtensionRunner,ExtensionSelectorComponent,FooterComponent,InstanceMemory,InteractiveMode,LoginDialogComponent,ModelRegistry,ModelRuntime,ModelSelectorComponent,NoMistakesBackend,OAuthSelectorComponent,RpcClient,SessionManager,SessionSelectorComponent,SettingsManager,SettingsSelectorComponent,SkillInvocationMessageComponent,ThinkingSelectorComponent,ToolExecutionComponent,TreeSelectorComponent,UserMessageComponent,UserMessageSelectorComponent,VIRTUAL_MODEL_STATE_ENTRY,activateRevision,buildContextEntries,buildSessionContext,buildSessionProjection,calculateContextTokens,checkSandbox,collectEntriesForBranchSummary,commitCandidate,compact,convertToLlm,convertToPng,copyToClipboard,createAgentSession,createAgentSessionFromServices,createAgentSessionRuntime,createAgentSessionServices,createBashTool,createBashToolDefinition,createCandidate,createCodemodeExtension,createCodingTools,createEditTool,createEditToolDefinition,createEventBus,createExtensionRuntime,createFindTool,createFindToolDefinition,createGrepTool,createGrepToolDefinition,createLocalBashOperations,createLocalPowerShellOperations,createLsTool,createLsToolDefinition,createMcpExtension,createPowerShellTool,createPowerShellToolDefinition,createReadOnlyTools,createReadTool,createReadToolDefinition,createSyntheticSourceInfo,createToolSearchExtension,createWriteTool,createWriteToolDefinition,defineTool,detectSupportedImageMimeTypeFromFile,discoverAndLoadExtensions,estimateTokens,findCutPoint,findTurnStartIndex,formatDimensionNote,formatSkillsForPrompt,generateBranchSummary,generateDiffString,generateSummary,generateSummaryWithUsage,generateUnifiedPatch,getBundledHarnessPath,getCandidate,getLastAssistantUsage,getLatestCompactionEntry,getNoMistakesBackend,initializeInstance,inspectBundledSkills,inspectCandidate,instancePaths,isBashToolResult,isEditToolResult,isFindToolResult,isGrepToolResult,isLsToolResult,isPowerShellToolResult,isReadToolResult,isToolCallEventType,isWriteToolResult,listRevisions,loadProjectContextFiles,loadSkills,loadSkillsFromDir,main,migrateProjectConfig,migrateSessionEntries,openInstance,parseArgs,parseFrontmatter,parseSessionEntries,parseSkillBlock,prepareBranchEntries,proposeSkill,readInstance,refineInstance,renderDiff,resizeImage,resolveCliModel,resolveModelScopeWithDiagnostics,rollbackInstance,runModelCheckpoint,runModelRefinement,runPrintMode,runRpcMode,sandboxCommand,serializeConversation,sessionEntryToContextMessages,shouldCompact,stripFrontmatter,validateSkill,verifySkillRevision,withFileMutationQueue,wrapRegisteredTool,wrapRegisteredTools}from"./chunk-AI2NMHE7.js";import{DEFAULT_MAX_BYTES,DEFAULT_MAX_LINES,formatReadLineRange,formatSize,truncateHead,truncateLine,truncateTail,truncateToVisualLines}from"./chunk-J22MF7YW.js";import{ArminComponent}from"./chunk-3EGDWOKT.js";import{Container,SelectList,Theme,getAvailableThemes,getLanguageFromPath,getMarkdownTheme,getSelectListTheme,getSettingsListTheme,highlightCode,initTheme,keyHint,keyText,rawKeyHint}from"./chunk-IOPE2TJ4.js";import{ProjectTrustStore,hasTrustRequiringProjectResources}from"./chunk-RP2WKNT7.js";import{readStoredCredential}from"./chunk-RFPFZTK2.js";import{getPowerShellConfig,getShellConfig}from"./chunk-P2QCWBFM.js";import{CONFIG_DIR_NAME,PROJECT_CONFIG_DIR_NAME,VERSION,getAgentDir,getDocsPath,getExamplesPath,getPackageDir,getReadmePath}from"./chunk-KNY3PXQI.js";import{__export}from"./chunk-JXWJOIH4.js";var dist_exports={};__export(dist_exports,{AgentSession:()=>AgentSession,AgentSessionRuntime:()=>AgentSessionRuntime,ArminComponent:()=>ArminComponent,AssistantMessageComponent:()=>AssistantMessageComponent,BashExecutionComponent:()=>BashExecutionComponent,BorderedLoader:()=>BorderedLoader,BranchSummaryMessageComponent:()=>BranchSummaryMessageComponent,CONFIG_DIR_NAME:()=>CONFIG_DIR_NAME,CURRENT_SESSION_VERSION:()=>CURRENT_SESSION_VERSION,CompactionSummaryMessageComponent:()=>CompactionSummaryMessageComponent,CredentialSynchronizationError:()=>CredentialSynchronizationError,CustomEditor:()=>CustomEditor,CustomMessageComponent:()=>CustomMessageComponent,DEFAULT_COMPACTION_SETTINGS:()=>DEFAULT_COMPACTION_SETTINGS,DEFAULT_MAX_BYTES:()=>DEFAULT_MAX_BYTES,DEFAULT_MAX_LINES:()=>DEFAULT_MAX_LINES,DefaultPackageManager:()=>DefaultPackageManager,DefaultResourceLoader:()=>DefaultResourceLoader,DynamicBorder:()=>DynamicBorder,ExtensionEditorComponent:()=>ExtensionEditorComponent,ExtensionInputComponent:()=>ExtensionInputComponent,ExtensionRunner:()=>ExtensionRunner,ExtensionSelectorComponent:()=>ExtensionSelectorComponent,FooterComponent:()=>FooterComponent,InstanceMemory:()=>InstanceMemory,InteractiveMode:()=>InteractiveMode,LoginDialogComponent:()=>LoginDialogComponent,ModelRegistry:()=>ModelRegistry,ModelRuntime:()=>ModelRuntime,ModelSelectorComponent:()=>ModelSelectorComponent,NoMistakesBackend:()=>NoMistakesBackend,OAuthSelectorComponent:()=>OAuthSelectorComponent,PROJECT_CONFIG_DIR_NAME:()=>PROJECT_CONFIG_DIR_NAME,ProjectTrustStore:()=>ProjectTrustStore,RpcClient:()=>RpcClient,SessionManager:()=>SessionManager,SessionSelectorComponent:()=>SessionSelectorComponent,SettingsManager:()=>SettingsManager,SettingsSelectorComponent:()=>SettingsSelectorComponent,ShowImagesSelectorComponent:()=>ShowImagesSelectorComponent,SkillInvocationMessageComponent:()=>SkillInvocationMessageComponent,Theme:()=>Theme,ThemeSelectorComponent:()=>ThemeSelectorComponent,ThinkingSelectorComponent:()=>ThinkingSelectorComponent,ToolExecutionComponent:()=>ToolExecutionComponent,TreeSelectorComponent:()=>TreeSelectorComponent,UserMessageComponent:()=>UserMessageComponent,UserMessageSelectorComponent:()=>UserMessageSelectorComponent,VERSION:()=>VERSION,VIRTUAL_MODEL_STATE_ENTRY:()=>VIRTUAL_MODEL_STATE_ENTRY,activateRevision:()=>activateRevision,buildContextEntries:()=>buildContextEntries,buildSessionContext:()=>buildSessionContext,buildSessionProjection:()=>buildSessionProjection,calculateContextTokens:()=>calculateContextTokens,checkSandbox:()=>checkSandbox,collectEntriesForBranchSummary:()=>collectEntriesForBranchSummary,commitCandidate:()=>commitCandidate,compact:()=>compact,convertToLlm:()=>convertToLlm,convertToPng:()=>convertToPng,copyToClipboard:()=>copyToClipboard,createAgentSession:()=>createAgentSession,createAgentSessionFromServices:()=>createAgentSessionFromServices,createAgentSessionRuntime:()=>createAgentSessionRuntime,createAgentSessionServices:()=>createAgentSessionServices,createBashTool:()=>createBashTool,createBashToolDefinition:()=>createBashToolDefinition,createCandidate:()=>createCandidate,createCodemodeExtension:()=>createCodemodeExtension,createCodingTools:()=>createCodingTools,createEditTool:()=>createEditTool,createEditToolDefinition:()=>createEditToolDefinition,createEventBus:()=>createEventBus,createExtensionRuntime:()=>createExtensionRuntime,createFindTool:()=>createFindTool,createFindToolDefinition:()=>createFindToolDefinition,createGrepTool:()=>createGrepTool,createGrepToolDefinition:()=>createGrepToolDefinition,createLocalBashOperations:()=>createLocalBashOperations,createLocalPowerShellOperations:()=>createLocalPowerShellOperations,createLsTool:()=>createLsTool,createLsToolDefinition:()=>createLsToolDefinition,createMcpExtension:()=>createMcpExtension,createPowerShellTool:()=>createPowerShellTool,createPowerShellToolDefinition:()=>createPowerShellToolDefinition,createReadOnlyTools:()=>createReadOnlyTools,createReadTool:()=>createReadTool,createReadToolDefinition:()=>createReadToolDefinition,createSyntheticSourceInfo:()=>createSyntheticSourceInfo,createToolSearchExtension:()=>createToolSearchExtension,createWriteTool:()=>createWriteTool,createWriteToolDefinition:()=>createWriteToolDefinition,defineTool:()=>defineTool,detectSupportedImageMimeTypeFromFile:()=>detectSupportedImageMimeTypeFromFile,discoverAndLoadExtensions:()=>discoverAndLoadExtensions,estimateTokens:()=>estimateTokens,findCutPoint:()=>findCutPoint,findTurnStartIndex:()=>findTurnStartIndex,formatDimensionNote:()=>formatDimensionNote,formatReadLineRange:()=>formatReadLineRange,formatSize:()=>formatSize,formatSkillsForPrompt:()=>formatSkillsForPrompt,generateBranchSummary:()=>generateBranchSummary,generateDiffString:()=>generateDiffString,generateSummary:()=>generateSummary,generateSummaryWithUsage:()=>generateSummaryWithUsage,generateUnifiedPatch:()=>generateUnifiedPatch,getAgentDir:()=>getAgentDir,getBundledHarnessPath:()=>getBundledHarnessPath,getCandidate:()=>getCandidate,getDocsPath:()=>getDocsPath,getExamplesPath:()=>getExamplesPath,getLanguageFromPath:()=>getLanguageFromPath,getLastAssistantUsage:()=>getLastAssistantUsage,getLatestCompactionEntry:()=>getLatestCompactionEntry,getMarkdownTheme:()=>getMarkdownTheme,getNoMistakesBackend:()=>getNoMistakesBackend,getPackageDir:()=>getPackageDir,getPowerShellConfig:()=>getPowerShellConfig,getReadmePath:()=>getReadmePath,getSelectListTheme:()=>getSelectListTheme,getSettingsListTheme:()=>getSettingsListTheme,getShellConfig:()=>getShellConfig,hasTrustRequiringProjectResources:()=>hasTrustRequiringProjectResources,highlightCode:()=>highlightCode,initTheme:()=>initTheme,initializeInstance:()=>initializeInstance,inspectBundledSkills:()=>inspectBundledSkills,inspectCandidate:()=>inspectCandidate,instancePaths:()=>instancePaths,isBashToolResult:()=>isBashToolResult,isEditToolResult:()=>isEditToolResult,isFindToolResult:()=>isFindToolResult,isGrepToolResult:()=>isGrepToolResult,isLsToolResult:()=>isLsToolResult,isPowerShellToolResult:()=>isPowerShellToolResult,isReadToolResult:()=>isReadToolResult,isToolCallEventType:()=>isToolCallEventType,isWriteToolResult:()=>isWriteToolResult,keyHint:()=>keyHint,keyText:()=>keyText,listRevisions:()=>listRevisions,loadProjectContextFiles:()=>loadProjectContextFiles,loadSkills:()=>loadSkills,loadSkillsFromDir:()=>loadSkillsFromDir,main:()=>main,migrateProjectConfig:()=>migrateProjectConfig,migrateSessionEntries:()=>migrateSessionEntries,openInstance:()=>openInstance,parseArgs:()=>parseArgs,parseFrontmatter:()=>parseFrontmatter,parseSessionEntries:()=>parseSessionEntries,parseSkillBlock:()=>parseSkillBlock,prepareBranchEntries:()=>prepareBranchEntries,proposeSkill:()=>proposeSkill,rawKeyHint:()=>rawKeyHint,readInstance:()=>readInstance,readStoredCredential:()=>readStoredCredential,refineInstance:()=>refineInstance,renderDiff:()=>renderDiff,resizeImage:()=>resizeImage,resolveCliModel:()=>resolveCliModel,resolveModelScopeWithDiagnostics:()=>resolveModelScopeWithDiagnostics,rollbackInstance:()=>rollbackInstance,runModelCheckpoint:()=>runModelCheckpoint,runModelRefinement:()=>runModelRefinement,runPrintMode:()=>runPrintMode,runRpcMode:()=>runRpcMode,sandboxCommand:()=>sandboxCommand,serializeConversation:()=>serializeConversation,sessionEntryToContextMessages:()=>sessionEntryToContextMessages,shouldCompact:()=>shouldCompact,stripFrontmatter:()=>stripFrontmatter,truncateHead:()=>truncateHead,truncateLine:()=>truncateLine,truncateTail:()=>truncateTail,truncateToVisualLines:()=>truncateToVisualLines,validateSkill:()=>validateSkill,verifySkillRevision:()=>verifySkillRevision,withFileMutationQueue:()=>withFileMutationQueue,wrapRegisteredTool:()=>wrapRegisteredTool,wrapRegisteredTools:()=>wrapRegisteredTools});var SHOW_IMAGES_SELECT_LIST_LAYOUT={minPrimaryColumnWidth:12,maxPrimaryColumnWidth:32},ShowImagesSelectorComponent=class extends Container{selectList;constructor(currentValue,onSelect,onCancel){super();let items=[{value:"yes",label:"Yes",description:"Show images inline in terminal"},{value:"no",label:"No",description:"Show text placeholder instead"}];this.addChild(new DynamicBorder),this.selectList=new SelectList(items,5,getSelectListTheme(),SHOW_IMAGES_SELECT_LIST_LAYOUT),this.selectList.setSelectedIndex(currentValue?0:1),this.selectList.onSelect=item=>{onSelect(item.value==="yes")},this.selectList.onCancel=()=>{onCancel()},this.addChild(this.selectList),this.addChild(new DynamicBorder)}getSelectList(){return this.selectList}};var THEME_SELECT_LIST_LAYOUT={minPrimaryColumnWidth:12,maxPrimaryColumnWidth:32},ThemeSelectorComponent=class extends Container{selectList;onPreview;constructor(currentTheme,onSelect,onCancel,onPreview){super(),this.onPreview=onPreview;let themes=getAvailableThemes(),themeItems=themes.map(name=>({value:name,label:name,description:name===currentTheme?"(current)":void 0}));this.addChild(new DynamicBorder),this.selectList=new SelectList(themeItems,10,getSelectListTheme(),THEME_SELECT_LIST_LAYOUT);let currentIndex=themes.indexOf(currentTheme);currentIndex!==-1&&this.selectList.setSelectedIndex(currentIndex),this.selectList.onSelect=item=>{onSelect(item.value)},this.selectList.onCancel=()=>{onCancel()},this.selectList.onSelectionChange=item=>{this.onPreview(item.value)},this.addChild(this.selectList),this.addChild(new DynamicBorder)}getSelectList(){return this.selectList}};export{ShowImagesSelectorComponent,ThemeSelectorComponent,dist_exports};