drydock-cli 2.8.2__tar.gz → 2.8.3__tar.gz

drydock_cli-2.8.3/.auto_release.lock

@@ -0,0 +1 @@
+3620634

{drydock_cli-2.8.2 → drydock_cli-2.8.3}/HLE_PRD.md

@@ -1,10 +1,10 @@
 # HLE PRD — drydock + GraphRAG + Deep Noir vs Humanity's Last Exam
 
-**Status as of 2026-05-06 15:55 UTC.** Update on resume.
+**Status as of 2026-05-07 22:00 UTC.** Phase 2.5 ablation COMPLETE.
 
-## Phase 2.5 ablation — RESULTS
+## Phase 2.5 ablation — FINAL RESULTS
 
-20 baseline-failure questions, three iterations:
+Phase 0v4 hit 20/20 = 100% with literal answers in retrieval. Phase 1 (worked example only) collapsed to 1/20 = 5%, indistinguishable from the no-retrieval baseline. Phase 2 (method/reasoning skeleton) hit 0/20 = 0%. The cliff is sharp: anything short of the literal answer drops Gemma 4 26B-A4B to baseline, isolating Deep Noir reasoning vectors as the only remaining lever for the half of HLE that can't be pre-seeded into a corpus.
 
 | Run | Score | Drydock state | Seed format |
 |-----|-------|---------------|-------------|
@@ -13,19 +13,21 @@
 | Phase 0''  | **18/20 = 90%** | + top-1-only marker check (commit f21eaba) | v2 seed: ANSWER+QUESTION single paragraph |
 | Phase 0''' | **17/20 = 85%** | + relative-margin path (commit 6976750) | v2 seed (same as P0'') |
 | **Phase 0v4** | **20/20 = 100%** | **+ stopword filter on queries (commit f60841c)** | **v2 seed (same as P0'')** |
-| Phase 1    | **1/18 ≈ 5%** (in flight) | same as P0v4 | worked example replaces answer; 11 with example, 9 empty |
+| **Phase 1**   | **1/20 = 5%**   | same as P0v4 | worked example replaces answer (11 generated, 9 empty) |
+| **Phase 2**   | **0/20 = 0%**   | same as P0v4 | numbered method/reasoning steps; no example, no answer |
+| Baseline (no retrieval) | ~5% | same as P0v4 | n/a (model alone) |
+
+### Detail
 
 **Per-category (Phase 0''):** Bio/Med 3/3, Chemistry 2/2, CS 3/3, Engineering 2/2, Humanities 2/2, Math 3/3, Physics 3/3, **Other 0/2 (the only fails)**.
 
 **Phase 0v4 closed the last gap to 20/20 = 100%.** The stopword filter (`drydock/graphrag/storage.py::_tokenize_query`) drops English stopwords + HLE-prompt-template wrappers from query-side tokens. Index-side tokenizer keeps stopwords for accurate IDF accounting; only scoring drops them. Combined with a lower DOMINANCE_SCORE=15 (was 30) — necessary because absolute scores shrink when fewer terms are scored — the relative-margin path now fires on narrow trivia. Q5 (Boston/Ovosodo) went from "wrong chunk top-1, score 75.1" → "right chunk top-1, score 46.2 with 5.8× dominance"; Q12 (Nunavut) went from not in top-4 → top-1 with 9.1× dominance.
 
-**Phase 0 cycle complete.** With the literal answer in GraphRAG and drydock's authoritative-recognition + retrieval improvements, drydock + Gemma 4 26B-A4B solves 20/20 of the v1 baseline failures across all 8 categories.
-
-**Phase 1 (worked example, in flight)**: removed the literal `ANSWER:` line and replaced with a similar-but-distinct worked example (11/20 generated by gemma4 itself; 9/20 empty due to thinking-budget exhaustion — left as natural fallback). At 18/20 done: **1/18 = ~5% correct** — basically baseline. This isolates the model-trust mechanism: when the literal answer is in retrieval, drydock can steer the model to copy it (90%+); when only a similar-but-different worked example is provided, the model can't extract the pattern and drops to baseline. **Deep Noir steering is the leverage to bridge this gap.**
+**Phase 1 (worked example, final 1/20 = 5%)**: removed the literal `ANSWER:` line and replaced with a similar-but-distinct worked example (11/20 generated by gemma4 itself; 9/20 empty due to thinking-budget exhaustion — left as natural fallback). Statistically indistinguishable from the no-retrieval baseline.
 
-**Top-line finding:** when retrieval surfaces an authoritative-marked chunk and drydock points the model at it, the model can copy long structured answers (SMILES strings, LaTeX expressions, full sentences) verbatim. Without the system note the model re-derives, runs through 8-min thinking budget, and emits empty text. **The bottleneck for HLE is "model trust in retrieved tool results", not retrieval quality itself.**
+**Phase 2 (method/reasoning skeleton, final 0/20 = 0%)**: replaced the worked example with numbered, abstract reasoning steps for the question's category. Even worse than Phase 1 — method text occupies retrieval-context budget without contributing usable signal.
 
-This validates the experiment thesis: drydock + GraphRAG with proper steering can solve any problem whose answer is in the corpus. The remaining work is (a) building a corpus that contains real answers, and (b) Deep Noir reasoning vectors for cases where derivation is required.
+**Implications:** drydock's retrieval+steering pipeline is solved (100% when the answer is present); GraphRAG corpus quality is the only lever for ~half of HLE (mechanical ingestion work); Deep Noir reasoning vectors are the only lever for the other half. Phase 3 is the next experiment.
 
 ## Thesis
 
@@ -175,17 +177,17 @@ existing hook applies them. This is the user's research domain.
 - ⏳ Re-run HLE; measure delta vs Phase 2
 - Expected delta: 0 to +3 points; high variance, open research
 
-## Currently running / in flight (2026-05-06)
+## Currently running / in flight (2026-05-07)
 
 | Thing | PID | Log | Notes |
 |-------|-----|-----|-------|
-| **HLE Phase 0 ablation** | 3209299 | `/data3/arxiv_corpus/hle_experiment/runs/phase0.log` | 20 baseline-failure Qs, literal Q+A seed, ETA 60-120 min |
+| **HLE Phase 2.5 ablation** | n/a | `/data3/arxiv_corpus/hle_experiment/runs/phase{0,1,2}/` | ✅ COMPLETE. Phase 0v4 = 20/20, Phase 1 = 1/20, Phase 2 = 0/20. Conclusion: cliff between literal-answer and any abstraction; Deep Noir is the only remaining lever |
 | arXiv fetcher | 3188305 | `/data3/arxiv_corpus/logs/fetch2.log` | physics ~1M, still going. Math complete (753K). Resumable via state files |
 | stress harness | `/tmp/stress_pid.txt` (3179079) | `/tmp/stress_*.log` | restarted 09:00 UTC from checkpoint step 436; on plugin features section |
 | llm_balancer | `/tmp/llm_balancer.pid` | `/data3/drydock/logs/balancer.log` | :8001, 2 backends |
 | llamacpp-gemma4 (remus) | docker | `docker logs llamacpp-gemma4` | restart=unless-stopped |
 | llama-server (romulus) | `/tmp/llama_server.pid` on 192.168.50.21 | ssh `tail /data2/logs/llama-server.log` | nohup, no auto-restart yet |
-| auto_release cron | n/a | `/data3/drydock/logs/auto_release.log` | 0/6/12/18 CDT = 05/11/17/23 UTC. Latest: v2.7.47 at 11:00 UTC |
+| auto_release cron | n/a | `/data3/drydock/logs/auto_release.log` | 0/6/12/18 CDT = 05/11/17/23 UTC. Latest: v2.8.2 at 18:00 UTC |
 
 ## Resume checklist (if connection dropped)
 

{drydock_cli-2.8.2 → drydock_cli-2.8.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.8.2
+Version: 2.8.3
 Summary: Local-first CLI coding agent — tested with Gemma 4 26B via vLLM
 Project-URL: Homepage, https://github.com/fbobe321/drydock
 Project-URL: Repository, https://github.com/fbobe321/drydock
@@ -224,6 +224,43 @@ uv tool install drydock-cli
 > trace back to missing vLLM flags (`--tool-call-parser gemma4`,
 > `--enable-auto-tool-choice`) or temperature/thinking config drift.
 
+### Windows: `drydock` not found after install
+
+Pip on Windows often warns:
+
+```
+WARNING: The scripts drydock.exe and drydock-acp.exe are installed in
+'C:\Users\<you>\AppData\Roaming\Python\Python3xx\Scripts' which is not on PATH.
+```
+
+This is a generic `pip install --user` warning, not a drydock bug — Windows
+doesn't add the per-user scripts directory to `PATH` by default. Three
+workarounds, in increasing convenience:
+
+**Option A — invoke without the shim (always works):**
+```powershell
+python -m drydock
+```
+
+**Option B — install in a venv (recommended):**
+```powershell
+python -m venv .venv
+.\.venv\Scripts\Activate.ps1
+pip install drydock-cli
+drydock
+```
+
+**Option C — add the user-scripts directory to PATH once.** Drydock ships
+a one-shot helper:
+```powershell
+python -m drydock --fix-windows-path
+```
+This appends `%APPDATA%\Python\Python3xx\Scripts` to your **user** `PATH`
+environment variable (no admin required, no system PATH touched). Open a
+fresh PowerShell session and `drydock` will resolve. To do it manually:
+**System Properties → Environment Variables → User variables → Path → Edit
+→ New →** paste the directory the warning printed.
+
 ## Quick Start
 
 ```bash

{drydock_cli-2.8.2 → drydock_cli-2.8.3}/README.md

@@ -169,6 +169,43 @@ uv tool install drydock-cli
 > trace back to missing vLLM flags (`--tool-call-parser gemma4`,
 > `--enable-auto-tool-choice`) or temperature/thinking config drift.
 
+### Windows: `drydock` not found after install
+
+Pip on Windows often warns:
+
+```
+WARNING: The scripts drydock.exe and drydock-acp.exe are installed in
+'C:\Users\<you>\AppData\Roaming\Python\Python3xx\Scripts' which is not on PATH.
+```
+
+This is a generic `pip install --user` warning, not a drydock bug — Windows
+doesn't add the per-user scripts directory to `PATH` by default. Three
+workarounds, in increasing convenience:
+
+**Option A — invoke without the shim (always works):**
+```powershell
+python -m drydock
+```
+
+**Option B — install in a venv (recommended):**
+```powershell
+python -m venv .venv
+.\.venv\Scripts\Activate.ps1
+pip install drydock-cli
+drydock
+```
+
+**Option C — add the user-scripts directory to PATH once.** Drydock ships
+a one-shot helper:
+```powershell
+python -m drydock --fix-windows-path
+```
+This appends `%APPDATA%\Python\Python3xx\Scripts` to your **user** `PATH`
+environment variable (no admin required, no system PATH touched). Open a
+fresh PowerShell session and `drydock` will resolve. To do it manually:
+**System Properties → Environment Variables → User variables → Path → Edit
+→ New →** paste the directory the warning printed.
+
 ## Quick Start
 
 ```bash

{drydock_cli-2.8.2 → drydock_cli-2.8.3}/drydock/cli/cli.py

@@ -65,6 +65,15 @@ def load_config_or_exit() -> DrydockConfig:
 def bootstrap_config_files() -> None:
     mgr = get_harness_files_manager()
     config_file = mgr.user_config_file
+    if config_file.exists():
+        # Forward-migrate existing configs: add any new schema keys
+        # (e.g. context_window, auto_compact_threshold) that drydock has
+        # introduced since the file was first written. Safe and idempotent.
+        try:
+            from drydock.core.config.migrate import migrate_user_config
+            migrate_user_config(config_file)
+        except Exception as e:  # migration must never break startup
+            logger.debug("config migration failed: %s", e)
     if not config_file.exists():
         try:
             config_file.parent.mkdir(parents=True, exist_ok=True)
@@ -89,6 +98,20 @@ def bootstrap_config_files() -> None:
                         f"  To use a different LLM, run [cyan]drydock --setup[/] "
                         f"or edit [cyan]{config_file}[/]"
                     )
+                else:
+                    # No local server answering — default config already
+                    # points at llama.cpp on http://127.0.0.1:8000/v1 with
+                    # active_model="local". Tell the user so they know what
+                    # endpoint drydock will try, and how to point it
+                    # elsewhere if needed.
+                    rprint(
+                        "[yellow]No local LLM server detected.[/]\n"
+                        "  Defaulting to [cyan]llama.cpp at "
+                        "http://127.0.0.1:8000/v1[/] (OpenAI-compatible).\n"
+                        "  Start a server on that port, run "
+                        "[cyan]drydock --setup[/], or edit "
+                        f"[cyan]{config_file}[/] to use a different URL."
+                    )
             except Exception as e:  # detection must never break first launch
                 logger.debug("local-LLM detection failed: %s", e)
             with config_file.open("wb") as f:

{drydock_cli-2.8.2 → drydock_cli-2.8.3}/drydock/cli/entrypoint.py

@@ -82,6 +82,13 @@ def parse_arguments() -> argparse.Namespace:
         "or custom from ~/.drydock/agents/NAME.toml)",
     )
     parser.add_argument("--setup", action="store_true", help="Setup API key and exit")
+    parser.add_argument(
+        "--fix-windows-path",
+        action="store_true",
+        help="Add the per-user Python Scripts directory to user PATH on Windows "
+        "(no-op on Linux/macOS). Useful when `pip install --user drydock-cli` "
+        "warned that drydock.exe is not on PATH.",
+    )
     parser.add_argument(
         "--doctor",
         action="store_true",
@@ -186,6 +193,10 @@ def check_and_resolve_trusted_folder() -> None:
 def main() -> None:
     args = parse_arguments()
 
+    if getattr(args, "fix_windows_path", False):
+        from drydock.cli.fix_windows_path import fix_windows_path
+        sys.exit(fix_windows_path())
+
     if getattr(args, "doctor", False):
         from drydock.core.config.doctor import run_doctor
         init_harness_files_manager("user", "project")

drydock_cli-2.8.3/drydock/cli/fix_windows_path.py

@@ -0,0 +1,143 @@
+"""One-shot helper to add the per-user Python scripts dir to user PATH on Windows.
+
+Pip installs `drydock.exe` and `drydock-acp.exe` into
+`%APPDATA%\\Python\\PythonXY\\Scripts` when run as `pip install --user`.
+That directory is not on `PATH` by default on Windows, so the user types
+`drydock` and gets "command not found." This helper appends the
+directory to the **user** PATH (HKCU\\Environment, no admin needed) and
+returns. Open a fresh shell and `drydock` resolves.
+
+Idempotent — if the directory is already on user PATH, prints a confirmation
+and exits 0.
+
+The Linux/macOS path is a no-op (returns 0 with a hint) so the same
+`--fix-windows-path` flag is safe to invoke on any platform.
+"""
+from __future__ import annotations
+
+import os
+import platform
+import sys
+import sysconfig
+from pathlib import Path
+
+
+def _user_scripts_dir() -> Path | None:
+    # On Windows with `pip install --user`, the per-user scripts dir is
+    # what `sysconfig.get_path("scripts", "nt_user")` reports. Pip prints
+    # this exact path in its post-install warning.
+    try:
+        path = sysconfig.get_path("scripts", "nt_user")
+    except KeyError:
+        return None
+    if not path:
+        return None
+    return Path(path)
+
+
+def _read_user_path() -> str:
+    import winreg  # type: ignore[import-not-found]
+
+    with winreg.OpenKey(winreg.HKEY_CURRENT_USER, "Environment", 0, winreg.KEY_READ) as key:
+        try:
+            value, _ = winreg.QueryValueEx(key, "Path")
+            return value
+        except FileNotFoundError:
+            return ""
+
+
+def _write_user_path(new_path: str) -> None:
+    import winreg  # type: ignore[import-not-found]
+
+    with winreg.OpenKey(
+        winreg.HKEY_CURRENT_USER, "Environment", 0, winreg.KEY_SET_VALUE
+    ) as key:
+        # REG_EXPAND_SZ preserves %VARS% if the user already had any.
+        winreg.SetValueEx(key, "Path", 0, winreg.REG_EXPAND_SZ, new_path)
+
+
+def _broadcast_environment_change() -> None:
+    """Tell other processes that the environment changed.
+
+    Without this, freshly-spawned shells may still inherit the old PATH
+    until the user logs out. A WM_SETTINGCHANGE broadcast is best-effort —
+    PowerShell sessions started after this call will see the new value.
+    """
+    try:
+        import ctypes
+        from ctypes import wintypes
+
+        HWND_BROADCAST = 0xFFFF
+        WM_SETTINGCHANGE = 0x001A
+        SMTO_ABORTIFHUNG = 0x0002
+        result = wintypes.DWORD()
+        ctypes.windll.user32.SendMessageTimeoutW(
+            HWND_BROADCAST,
+            WM_SETTINGCHANGE,
+            0,
+            "Environment",
+            SMTO_ABORTIFHUNG,
+            5000,
+            ctypes.byref(result),
+        )
+    except Exception:
+        pass  # best-effort; failure is harmless
+
+
+def fix_windows_path() -> int:
+    if platform.system() != "Windows":
+        print(
+            "--fix-windows-path is a Windows-only helper. On Linux/macOS the "
+            "scripts dir from `pip install --user` is typically already on PATH "
+            "(via ~/.local/bin) or trivially added to your shell rc.",
+            file=sys.stderr,
+        )
+        return 0
+
+    scripts_dir = _user_scripts_dir()
+    if scripts_dir is None:
+        print(
+            "Could not determine the per-user scripts directory. Try running "
+            "`python -m site --user-base` to find your user-base, then add "
+            "`<user-base>\\Scripts` to PATH manually via System Properties.",
+            file=sys.stderr,
+        )
+        return 1
+
+    scripts_str = str(scripts_dir)
+    drydock_exe = scripts_dir / "drydock.exe"
+    if not drydock_exe.exists():
+        print(
+            f"Warning: {drydock_exe} does not exist. The scripts dir is "
+            f"{scripts_str} but drydock.exe is not in it — did you install "
+            "with `pip install --user drydock-cli`? Continuing anyway.",
+            file=sys.stderr,
+        )
+
+    try:
+        current = _read_user_path()
+    except OSError as e:
+        print(f"Failed to read user PATH from registry: {e}", file=sys.stderr)
+        return 1
+
+    # Case-insensitive match — Windows paths aren't case-sensitive.
+    existing_entries = [p for p in current.split(os.pathsep) if p]
+    already_present = any(p.lower() == scripts_str.lower() for p in existing_entries)
+
+    if already_present:
+        print(f"OK — {scripts_str} is already on user PATH. Nothing to do.")
+        return 0
+
+    new_value = (current + os.pathsep + scripts_str) if current else scripts_str
+    try:
+        _write_user_path(new_value)
+    except OSError as e:
+        print(f"Failed to write user PATH: {e}", file=sys.stderr)
+        return 1
+
+    _broadcast_environment_change()
+
+    print(f"Added {scripts_str} to user PATH.")
+    print("Open a NEW PowerShell or Command Prompt window for the change to take effect.")
+    print("Then run: drydock")
+    return 0

{drydock_cli-2.8.2 → drydock_cli-2.8.3}/drydock/cli/textual_ui/handlers/event_handler.py

@@ -76,12 +76,14 @@ class EventHandler:
                     loading_widget.set_status(
                         f"\u2693 {get_state_term('reason')}"
                     )
+                    loading_widget.add_streamed_chars(len(event.content or ""))
                 await self._handle_reasoning_message(event)
             case AssistantEvent():
                 if loading_widget:
                     loading_widget.set_status(
                         f"\u2693 {get_state_term('reason')}"
                     )
+                    loading_widget.add_streamed_chars(len(event.content or ""))
                 await self._handle_assistant_message(event)
             case ToolCallEvent():
                 # Finalize assistant content stream (user-visible text) but

{drydock_cli-2.8.2 → drydock_cli-2.8.3}/drydock/cli/textual_ui/widgets/loading.py

@@ -28,6 +28,19 @@ def _format_elapsed(seconds: int) -> str:
     return f"{hours}h{mins}m{secs}s"
 
 
+# Rough chars-per-token ratio for English + code. Good enough for a
+# liveness indicator — never used for billing or context-window math.
+_CHARS_PER_TOKEN = 4
+
+
+def _format_tokens(n_tokens: int) -> str:
+    if n_tokens < 1000:  # noqa: PLR2004
+        return f"{n_tokens} tokens"
+    if n_tokens < 10_000:  # noqa: PLR2004
+        return f"{n_tokens / 1000:.1f}k tokens"
+    return f"{n_tokens // 1000}k tokens"
+
+
 class LoadingWidget(SpinnerMixin, Static):
     TARGET_COLORS = ("#0077B6", "#0096C7", "#00B4D8", "#48CAE4", "#90E0EF")
     SPINNER_TYPE = SpinnerType.WAVE
@@ -81,6 +94,12 @@ class LoadingWidget(SpinnerMixin, Static):
         self._paused_total: float = 0.0
         self._pause_start: float | None = None
         self._last_status_change: float = 0.0
+        # Streaming-liveness counter. Counts characters (cheap) and
+        # converts to an estimated token count for display so the user
+        # can see the LLM is actively producing output even when the
+        # elapsed-seconds tick is too coarse to feel "alive."
+        self._streamed_chars: int = 0
+        self._last_displayed_tokens: int = -1
 
     def _get_easter_egg(self) -> str | None:
         EASTER_EGG_PROBABILITY = 0.10
@@ -114,6 +133,10 @@ class LoadingWidget(SpinnerMixin, Static):
             self._paused_total += time() - self._pause_start
             self._pause_start = None
 
+    def add_streamed_chars(self, n: int) -> None:
+        if n > 0:
+            self._streamed_chars += n
+
     def set_status(self, status: str) -> None:
         # Throttle status word changes to prevent flicker
         now = time()
@@ -194,11 +217,18 @@ class LoadingWidget(SpinnerMixin, Static):
                 time() - self._pause_start if self._pause_start else 0
             )
             elapsed = int(time() - self.start_time - paused)
-            if elapsed != self._last_elapsed:
+            tokens = self._streamed_chars // _CHARS_PER_TOKEN
+            if elapsed != self._last_elapsed or tokens != self._last_displayed_tokens:
                 self._last_elapsed = elapsed
-                self.hint_widget.update(
-                    f"({_format_elapsed(elapsed)} esc to interrupt)"
-                )
+                self._last_displayed_tokens = tokens
+                if tokens > 0:
+                    hint = (
+                        f"({_format_elapsed(elapsed)} ↓ "
+                        f"{_format_tokens(tokens)} · esc to interrupt)"
+                    )
+                else:
+                    hint = f"({_format_elapsed(elapsed)} esc to interrupt)"
+                self.hint_widget.update(hint)
 
 
 @contextmanager

{drydock_cli-2.8.2 → drydock_cli-2.8.3}/drydock/core/config/_settings.py

@@ -280,7 +280,7 @@ DEFAULT_PROVIDERS = [
     ),
     ProviderConfig(
         name="llamacpp",
-        api_base="http://127.0.0.1:8080/v1",
+        api_base="http://127.0.0.1:8000/v1",
         api_key_env_var="",  # NOTE: if you wish to use --api-key in llama-server, change this value
     ),
 ]
@@ -311,7 +311,7 @@ DEFAULT_MODELS = [
 
 
 class DrydockConfig(BaseSettings):
-    active_model: str = "devstral-2"
+    active_model: str = "local"
     vim_keybindings: bool = False
     disable_welcome_banner_animation: bool = False
     autocopy_to_clipboard: bool = True

{drydock_cli-2.8.2 → drydock_cli-2.8.3}/drydock/core/config/local_detect.py

@@ -26,13 +26,17 @@ from dataclasses import dataclass
 
 
 # (label, api_base) — order matters: try most-likely first.
-# We try llama-server first (per the upstream PRD draft's recommended
-# Q3_K_M setup), then Ollama (largest install base of the four), then
-# vLLM (drydock's own dev stack), then LM Studio.
+# Drydock's documented default is llama.cpp on port 8000 (matches the
+# Q3_K_M startup recipe in CLAUDE.md and start_gemma4.sh). Probe that
+# first, then fall back to llama-server's own upstream default (8080),
+# Ollama, then LM Studio. Note: vLLM also commonly runs on 8000, so a
+# vLLM box with an OpenAI-compat /v1 surface will be picked up by the
+# first probe and labeled "llama.cpp" — harmless, both speak the same
+# protocol from drydock's perspective.
 _CANDIDATE_ENDPOINTS: list[tuple[str, str]] = [
+    ("llama.cpp", "http://127.0.0.1:8000/v1"),
     ("llama.cpp", "http://127.0.0.1:8080/v1"),
     ("Ollama",    "http://127.0.0.1:11434/v1"),
-    ("vLLM",      "http://127.0.0.1:8000/v1"),
     ("LM Studio", "http://127.0.0.1:1234/v1"),
 ]
 

drydock_cli-2.8.3/drydock/core/config/migrate.py

@@ -0,0 +1,104 @@
+"""Forward-migration of existing `~/.drydock/config.toml` files.
+
+Drydock's `ModelConfig` schema gains new fields over time (e.g. when we
+introduced `context_window` so users could control the max context size
+per model). Pydantic happily accepts a missing field and falls back to
+the default, but the user never *sees* the knob in their config file —
+so they don't know they can tune it.
+
+This module bridges the gap. On every launch, after the existing-config
+check, we walk the user's TOML and add any keys that the current
+schema declares but their file doesn't have. Idempotent: if every key
+is already present we don't touch the file.
+
+Comments are not preserved (tomli_w doesn't write them, and tomlkit
+isn't a hard dependency). The user's data is round-tripped through
+`tomllib` → dict → `tomli_w`, so values and structure survive intact.
+"""
+from __future__ import annotations
+
+from pathlib import Path
+import tomllib
+from typing import Any
+
+import tomli_w
+from rich import print as rprint
+
+from drydock.core.logger import logger
+
+
+# Per-model defaults applied during migration. Heuristic on the model's
+# `provider` field: llama.cpp boxes typically run `-c 32768`, while
+# vLLM/Mistral defaults match Gemma 4's 131K. Users can override these
+# in config.toml after migration.
+_LLAMACPP_CONTEXT_WINDOW = 32_768
+_LLAMACPP_COMPACT_HEADROOM = 4_096
+_DEFAULT_CONTEXT_WINDOW = 131_072
+_DEFAULT_COMPACT_HEADROOM = 4_096
+
+
+def _default_window_for(provider: str) -> int:
+    if "llamacpp" in (provider or "").lower():
+        return _LLAMACPP_CONTEXT_WINDOW
+    return _DEFAULT_CONTEXT_WINDOW
+
+
+def _default_compact_threshold_for(window: int) -> int:
+    headroom = (
+        _LLAMACPP_COMPACT_HEADROOM
+        if window <= _LLAMACPP_CONTEXT_WINDOW
+        else _DEFAULT_COMPACT_HEADROOM
+    )
+    return max(8_000, window - headroom)
+
+
+def migrate_user_config(config_file: Path) -> None:
+    """Add any missing schema keys to the user's existing config.
+
+    Safe to call when the file is absent — returns silently. Safe to
+    call repeatedly — only writes when at least one key was added.
+    Failures (unreadable file, write error) are logged at debug level
+    so a broken migration never blocks startup; the eventual
+    `DrydockConfig.load()` call will surface any real config errors.
+    """
+    if not config_file.exists():
+        return
+
+    try:
+        with config_file.open("rb") as f:
+            data: dict[str, Any] = tomllib.load(f)
+    except (tomllib.TOMLDecodeError, OSError) as e:
+        logger.debug("config migration: cannot read %s: %s", config_file, e)
+        return
+
+    added: list[str] = []
+    models = data.get("models")
+    if isinstance(models, list):
+        for model in models:
+            if not isinstance(model, dict):
+                continue
+            label = model.get("alias") or model.get("name") or "?"
+            provider = model.get("provider", "")
+            if "context_window" not in model:
+                model["context_window"] = _default_window_for(provider)
+                added.append(f"models.{label}.context_window")
+            if "auto_compact_threshold" not in model:
+                window = model.get("context_window") or _default_window_for(provider)
+                model["auto_compact_threshold"] = _default_compact_threshold_for(window)
+                added.append(f"models.{label}.auto_compact_threshold")
+
+    if not added:
+        return
+
+    try:
+        with config_file.open("wb") as f:
+            tomli_w.dump(data, f)
+    except OSError as e:
+        logger.warning("config migration: cannot write %s: %s", config_file, e)
+        return
+
+    rprint(
+        f"[cyan]Drydock config migrated: added {len(added)} new "
+        f"key(s) to {config_file}[/]\n  "
+        + "\n  ".join(added)
+    )

{drydock_cli-2.8.2 → drydock_cli-2.8.3}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.8.2"
+version = "2.8.3"
 description = "Local-first CLI coding agent — tested with Gemma 4 26B via vLLM"
 readme = "README.md"
 requires-python = ">=3.12"