drydock-cli 3.1.6__tar.gz → 3.1.8__tar.gz

{drydock_cli-3.1.6/drydock_cli.egg-info → drydock_cli-3.1.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 3.1.6
+Version: 3.1.8
 Summary: Drydock — a local, provider-agnostic terminal coding agent for local LLMs
 Author: Frank Bobe III
 License-Expression: Apache-2.0
@@ -112,6 +112,12 @@ A full agentic CLI harness — every tool below is clean-room and dependency-fre
 - **Cross-platform** — runs natively on Linux, macOS, and **Windows via PowerShell/cmd**
   (no WSL or Git-Bash required); `/shell` shows which shell your commands run in.
 - **Loops** — `/loop <count> <prompt>` runs a prompt iteratively (Esc stops).
+- **Ratchet** — `/ratchet <goal>` keeps solving across rounds, snapshotting the
+  workspace whenever more tests pass and rolling back regressions, until it goes
+  green. The verifier auto-detects (pytest/cargo/go/npm/make); `--verify "<cmd>"`
+  overrides. Progress can't slip backward. **`--effort low|medium|high|xhigh|max`**
+  is one dial from a cheap plain pawl (low) to full evolutionary search — fan-out
+  and crossover of partial solutions — for the hardest tasks (high+).
 
 ## Slash commands
 
@@ -129,6 +135,7 @@ Typed into the prompt. The agent also knows these, so you can just **ask it**
 | `/skills new <name> <prompt>` | Create a reusable `/<name>` skill (use `$ARGS` for input) |
 | `/<name>` | Run a skill |
 | `/loop <count> <prompt>` | Repeat a prompt N times (Esc stops) |
+| `/ratchet <goal>` | Solve across rounds, snapshotting on verifier gains, rolling back regressions. `--effort low..max` scales plain-pawl→evolutionary; verifier auto-detects (`--verify "<cmd>"` to override; `--rounds N`, `--fitness auto\|exitcode\|<regex>`) |
 | `/mcp` | List connected MCP servers + their tools |
 | `/rmf bootstrap [families]` | Ingest the NIST SP 800-53 catalog (RMF automation) |
 | `/rmf-control` · `/rmf-categorize` · `/rmf-review` · `/rmf-poam` | Bundled RMF skills |
@@ -367,6 +374,54 @@ docker run -it --add-host=host.docker.internal:host-gateway \
 `-v "$PWD:/work"` mounts your project so the agent can read/edit it. Tags:
 `fbobe3/drydock:latest` and `fbobe3/drydock:<version>` (e.g. `:3.0.135`).
 
+### Serving Gemma-4-31B
+
+Drydock is provider-agnostic, but its primary target is dense **Gemma-4-31B**.
+Two proven ways to serve it on a 2×GPU box, both exposing an OpenAI-compatible
+API on `:8000` as model `gemma4`:
+
+**llama.cpp** — QAT GGUF, flexible concurrency (more slots if you need them):
+
+```
+llama-server -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf \
+  -c 65536 -np 2 --host 0.0.0.0 --port 8000 --alias gemma4
+```
+
+**vLLM** — w4a16 QAT, ~2× faster per request, 128K context:
+
+```
+docker run -d --name vllm-prod --gpus all --ipc=host --restart unless-stopped \
+  -v /data3/Models:/models -p 8000:8000 \
+  -e NCCL_P2P_DISABLE=1 \
+  vllm/vllm-openai:v0.26.0 \
+  --model /models/gemma-4-31B-it-qat-w4a16-ct \
+  --served-model-name gemma4 \
+  --tensor-parallel-size 2 \
+  --max-model-len 131072 \
+  --max-num-seqs 2 \
+  --gpu-memory-utilization 0.97 \
+  --kv-cache-dtype fp8 \
+  --tool-call-parser gemma4 \
+  --enable-auto-tool-choice
+```
+
+Point drydock at either with
+`--provider vllm --base-url http://<host>:8000/v1 --model gemma4`.
+
+**vLLM-on-Gemma-4 notes** (drydock handles these for you as of **3.1.7**):
+
+- **`skip_special_tokens: false`** is sent on every request — without it, a turn
+  truncated at `max_tokens` (tool-call / reasoning tails) comes back with *empty*
+  content, which looks like a refusal. If `finish_reason == "length"`, retry with
+  a larger `max_tokens` (≥ 2000 for reasoning-heavy calls).
+- **Reasoning is off by default.** Enable per-request via
+  `chat_template_kwargs.enable_thinking`; the vLLM reasoning parser is disabled,
+  so `<|channel>thought …` markers arrive inside `content` (stripped client-side).
+- **`--max-num-seqs 2`** caps concurrency at 2 — the cost of fitting 128K context
+  into the VRAM. Use llama.cpp if you need more concurrent sessions than context.
+- Tool-calling is the standard OpenAI format; `/props` is llama.cpp-only (vLLM
+  404s it — drydock falls back to `/v1/models` for the context length).
+
 ## Using it
 
 Type a task and press **Enter**. Drydock reads/writes/edits files and runs

{drydock_cli-3.1.6 → drydock_cli-3.1.8}/README.md

@@ -88,6 +88,12 @@ A full agentic CLI harness — every tool below is clean-room and dependency-fre
 - **Cross-platform** — runs natively on Linux, macOS, and **Windows via PowerShell/cmd**
   (no WSL or Git-Bash required); `/shell` shows which shell your commands run in.
 - **Loops** — `/loop <count> <prompt>` runs a prompt iteratively (Esc stops).
+- **Ratchet** — `/ratchet <goal>` keeps solving across rounds, snapshotting the
+  workspace whenever more tests pass and rolling back regressions, until it goes
+  green. The verifier auto-detects (pytest/cargo/go/npm/make); `--verify "<cmd>"`
+  overrides. Progress can't slip backward. **`--effort low|medium|high|xhigh|max`**
+  is one dial from a cheap plain pawl (low) to full evolutionary search — fan-out
+  and crossover of partial solutions — for the hardest tasks (high+).
 
 ## Slash commands
 
@@ -105,6 +111,7 @@ Typed into the prompt. The agent also knows these, so you can just **ask it**
 | `/skills new <name> <prompt>` | Create a reusable `/<name>` skill (use `$ARGS` for input) |
 | `/<name>` | Run a skill |
 | `/loop <count> <prompt>` | Repeat a prompt N times (Esc stops) |
+| `/ratchet <goal>` | Solve across rounds, snapshotting on verifier gains, rolling back regressions. `--effort low..max` scales plain-pawl→evolutionary; verifier auto-detects (`--verify "<cmd>"` to override; `--rounds N`, `--fitness auto\|exitcode\|<regex>`) |
 | `/mcp` | List connected MCP servers + their tools |
 | `/rmf bootstrap [families]` | Ingest the NIST SP 800-53 catalog (RMF automation) |
 | `/rmf-control` · `/rmf-categorize` · `/rmf-review` · `/rmf-poam` | Bundled RMF skills |
@@ -343,6 +350,54 @@ docker run -it --add-host=host.docker.internal:host-gateway \
 `-v "$PWD:/work"` mounts your project so the agent can read/edit it. Tags:
 `fbobe3/drydock:latest` and `fbobe3/drydock:<version>` (e.g. `:3.0.135`).
 
+### Serving Gemma-4-31B
+
+Drydock is provider-agnostic, but its primary target is dense **Gemma-4-31B**.
+Two proven ways to serve it on a 2×GPU box, both exposing an OpenAI-compatible
+API on `:8000` as model `gemma4`:
+
+**llama.cpp** — QAT GGUF, flexible concurrency (more slots if you need them):
+
+```
+llama-server -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf \
+  -c 65536 -np 2 --host 0.0.0.0 --port 8000 --alias gemma4
+```
+
+**vLLM** — w4a16 QAT, ~2× faster per request, 128K context:
+
+```
+docker run -d --name vllm-prod --gpus all --ipc=host --restart unless-stopped \
+  -v /data3/Models:/models -p 8000:8000 \
+  -e NCCL_P2P_DISABLE=1 \
+  vllm/vllm-openai:v0.26.0 \
+  --model /models/gemma-4-31B-it-qat-w4a16-ct \
+  --served-model-name gemma4 \
+  --tensor-parallel-size 2 \
+  --max-model-len 131072 \
+  --max-num-seqs 2 \
+  --gpu-memory-utilization 0.97 \
+  --kv-cache-dtype fp8 \
+  --tool-call-parser gemma4 \
+  --enable-auto-tool-choice
+```
+
+Point drydock at either with
+`--provider vllm --base-url http://<host>:8000/v1 --model gemma4`.
+
+**vLLM-on-Gemma-4 notes** (drydock handles these for you as of **3.1.7**):
+
+- **`skip_special_tokens: false`** is sent on every request — without it, a turn
+  truncated at `max_tokens` (tool-call / reasoning tails) comes back with *empty*
+  content, which looks like a refusal. If `finish_reason == "length"`, retry with
+  a larger `max_tokens` (≥ 2000 for reasoning-heavy calls).
+- **Reasoning is off by default.** Enable per-request via
+  `chat_template_kwargs.enable_thinking`; the vLLM reasoning parser is disabled,
+  so `<|channel>thought …` markers arrive inside `content` (stripped client-side).
+- **`--max-num-seqs 2`** caps concurrency at 2 — the cost of fitting 128K context
+  into the VRAM. Use llama.cpp if you need more concurrent sessions than context.
+- Tool-calling is the standard OpenAI format; `/props` is llama.cpp-only (vLLM
+  404s it — drydock falls back to `/v1/models` for the context length).
+
 ## Using it
 
 Type a task and press **Enter**. Drydock reads/writes/edits files and runs

{drydock_cli-3.1.6 → drydock_cli-3.1.8}/drydock/providers.py

@@ -584,6 +584,17 @@ def stream(
     if config.get("temperature") is not None:
         kwargs["temperature"] = config["temperature"]
 
+    # vLLM-Gemma quirk: a turn truncated at max_tokens (esp. tool-call / reasoning
+    # turns, whose tail is special tokens) comes back with EMPTY content unless we
+    # send skip_special_tokens=false — and a checker-scored harness reads an empty
+    # turn as a FAILURE. Always send it for vLLM (harmless / ignored on llama.cpp),
+    # and floor an already-configured-but-tiny budget so reasoning turns don't
+    # truncate to nothing. We never impose a cap where none was set.
+    if provider == "vllm":
+        kwargs.setdefault("extra_body", {})["skip_special_tokens"] = False
+        if kwargs.get("max_tokens") and kwargs["max_tokens"] < 2000:
+            kwargs["max_tokens"] = 2000
+
     # Adaptive reasoning budget (harmony/gpt-oss style models accept this;
     # endpoints without the knob ignore the extra field). Set per turn by the
     # agent loop: high for planning, low for routine continuation.