drydock-cli 2.7.42__tar.gz → 2.7.43__tar.gz

drydock_cli-2.7.43/.auto_release.lock

@@ -0,0 +1 @@
+2929491

{drydock_cli-2.7.42 → drydock_cli-2.7.43}/.gitignore

@@ -213,3 +213,10 @@ research/STOP
 # HLE eval outputs: per-run TUI logs, session symlinks, and copyrighted
 # HLE dataset content. Never commit.
 hle_results/
+
+# Untracked development clutter that shouldn't go to PyPI
+nohup.out
+.perf_baseline_done
+test/
+perf_results/baseline_*.json
+hle_results_iter/

{drydock_cli-2.7.42 → drydock_cli-2.7.43}/DEPLOYMENT.md

@@ -20,7 +20,60 @@ A single 24 GB GPU can run the model with slightly reduced context.
 The AWQ quant is what makes 16 GB cards viable — non-quantized
 Gemma 4 26B needs ≥40 GB.
 
-## vLLM (model server)
+## llama.cpp (recommended for tool-using workflows)
+
+Use this stack if drydock's TUI is your primary workflow. The `--jinja`
+flag is the loop-fix — without it, Gemma 4 enters infinite tool-call
+retry loops on multi-turn sessions (the 400 Bad Request loop fixed
+in v2.7.39, GH #14).
+
+```bash
+# 1. Download Unsloth GGUF (Q3_K_M is the article-recommended quant;
+#    UD-Q4_K_M is higher quality if you have ~17GB VRAM per slot)
+huggingface-cli download unsloth/gemma-4-26B-A4B-it-GGUF \
+    --include "gemma-4-26B-A4B-it-UD-Q3_K_M.gguf" \
+    --local-dir /path/to/models
+
+# 2. Build llama.cpp with CUDA, OR use the prebuilt Docker image
+#    ghcr.io/ggml-org/llama.cpp:server-cuda
+
+git clone https://github.com/ggerganov/llama.cpp.git
+cd llama.cpp
+cmake -B build -DGGML_CUDA=ON -DLLAMA_CURL=OFF
+cmake --build build --config Release -j8 --target llama-server
+
+# 3. Start the server (native binary)
+./build/bin/llama-server \
+    -m /path/to/models/gemma-4-26B-A4B-it-UD-Q3_K_M.gguf \
+    --host 0.0.0.0 --port 8000 \
+    -ngl 99 -c 32768 -np 1 \
+    --jinja \
+    -ctk q8_0 -ctv q8_0 \
+    --alias gemma4
+```
+
+**Don't omit** any of these flags — they're all load-bearing:
+
+| Flag | Why |
+|---|---|
+| `--jinja` | **The loop-fix.** Required for tool-using workflows. The GGUF's bundled chat template handles `tool` turns natively; without `--jinja`, tool results inject without proper turn markers and the model loops or returns empty assistants. |
+| `-ngl 99` | Offload all transformer layers to GPU. Anything less than full offload tanks decode speed on this model. |
+| `-c 32768` | 32K context. Fits in 16GB VRAM with q8 KV cache; raise to 65536 if you have headroom. |
+| `-ctk q8_0 -ctv q8_0` | Quantize K/V cache to q8 for longer contexts. f16 is ~2× slower per token at this size. |
+| `-np 1` | Single slot. Concurrent requests serialize — fine for one user, queue under load. |
+| `--alias gemma4` | What the API reports as the `model` field. Match this to your config.toml's `[[models]] name`. |
+
+**Critical client-side requirement:** temperature MUST be 1.0. Lower
+temperatures reinforce tool-call loops on Q3_K_M quants. The
+auto-detect path in v2.7.39+ (`local_detect.py`) bakes this in
+automatically when it sees a llama.cpp endpoint at first launch.
+
+## vLLM (alternative — higher decode throughput)
+
+Use this stack for batch eval / non-interactive workloads where you
+need ~70 tok/s decode (vs llama.cpp's ~15–17). The drydock-side
+empty-assistant filter (v2.7.39, GH #14) catches the loops vLLM is
+prone to without `--jinja`.
 
 Run via Docker:
 

{drydock_cli-2.7.42 → drydock_cli-2.7.43}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.7.42
+Version: 2.7.43
 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
@@ -73,30 +73,99 @@ Description-Content-Type: text/markdown
 DryDock is a TUI coding assistant designed to work with **local LLMs**. It provides a conversational interface to your codebase — explore, modify, build, and test projects through natural language and a powerful set of tools.
 
 > [!IMPORTANT]
-> DryDock is tested and optimized for **Gemma 4 26B-A4B** (26B MoE, 4B active parameters) served via vLLM. Other models and providers are supported (Mistral, OpenAI, Anthropic, Ollama) but are not as thoroughly tested. If you use a different model, expect to tune prompts and tool settings.
+> DryDock is tested and optimized for **Gemma 4 26B-A4B** (26B MoE, 4B active parameters). **Recommended serving stack: llama.cpp** with `--jinja` (the chat-template fix that prevents the tool-call loops Gemma 4 hits under other backends). vLLM is also documented below as a higher-throughput alternative for batch/eval workloads. Other models and providers are supported (Mistral, OpenAI, Anthropic, Ollama) but are not as thoroughly tested. If you use a different model, expect to tune prompts and tool settings.
 
 ## Tested Hardware + Model
 
 | Component | Spec |
 |-----------|------|
-| GPUs | 2x NVIDIA RTX 4060 Ti 16GB |
-| Model | [Gemma-4-26B-A4B-it-AWQ-4bit](https://huggingface.co/casperhansen/gemma-4-26b-a4b-it-AWQ-4bit) |
-| Serving | vLLM (Docker image `vllm/vllm-openai:gemma4`) |
-| Performance | ~70 tok/s, 131K context, 0% timeouts |
+| GPUs | 2× NVIDIA RTX 4060 Ti 16GB |
+| Model (llama.cpp, recommended) | [unsloth/gemma-4-26B-A4B-it-GGUF](https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF) — UD-Q3_K_M (12.7GB) or UD-Q4_K_M (16.9GB) |
+| Model (vLLM, alternative) | [casperhansen/gemma-4-26b-a4b-it-AWQ-4bit](https://huggingface.co/casperhansen/gemma-4-26b-a4b-it-AWQ-4bit) |
+| Performance | ~15–17 tok/s decode (llama.cpp Q3), ~70 tok/s decode (vLLM AWQ) |
 | Active params | 4B per token (MoE architecture — fast inference) |
 
-### Install the model
+### Recommended path: llama.cpp + Unsloth GGUF
+
+**Why this is the recommended setup:** Gemma 4's tool-calling format
+requires precise chat-template handling. Without `--jinja`, `tool` results
+get injected without the right turn markers and the model loops or
+returns empty assistant messages — the exact 400 Bad Request loop
+fixed in v2.7.39 (GH #14). With `--jinja`, the GGUF's bundled chat
+template handles tool turns natively and the loops disappear.
 
 ```bash
-# Download the model weights (requires HuggingFace access)
-pip install huggingface-hub
-huggingface-cli download casperhansen/gemma-4-26b-a4b-it-AWQ-4bit \
-    --local-dir /path/to/models/Gemma-4-26B-A4B-it-AWQ-4bit
+# 1. Download Unsloth's GGUF (Q3_K_M is the article-recommended quant;
+#    UD-Q4_K_M is a higher-quality alternative if you have ~17GB VRAM)
+huggingface-cli download unsloth/gemma-4-26B-A4B-it-GGUF \
+    --include "gemma-4-26B-A4B-it-UD-Q3_K_M.gguf" \
+    --local-dir /path/to/models
+
+# 2. Build llama.cpp with CUDA (or use the Docker image
+#    ghcr.io/ggml-org/llama.cpp:server-cuda)
+git clone https://github.com/ggerganov/llama.cpp.git
+cd llama.cpp
+cmake -B build -DGGML_CUDA=ON -DLLAMA_CURL=OFF
+cmake --build build --config Release -j8 --target llama-server
+
+# 3. Start the server with the article recipe
+./build/bin/llama-server \
+    -m /path/to/models/gemma-4-26B-A4B-it-UD-Q3_K_M.gguf \
+    --host 0.0.0.0 --port 8000 \
+    -ngl 99 -c 32768 -np 1 \
+    --jinja \
+    -ctk q8_0 -ctv q8_0 \
+    --alias gemma4
+```
+
+**Critical flags:**
+- `--jinja` — **the loop-fix.** Required for tool-using workflows. Without it, Gemma 4 enters infinite retry loops on multi-turn tool sessions.
+- `-ngl 99` — offload all layers to GPU
+- `-c 32768` — 32K context (fits in 16GB VRAM with q8 KV cache)
+- `-ctk q8_0 -ctv q8_0` — quantize KV cache for longer contexts
+- `-np 1` — single slot (concurrent requests serialize)
+- `--alias gemma4` — what the API reports as the `model` field
+
+**Drydock config** (`~/.drydock/config.toml`):
+
+```toml
+active_model = "gemma4"
+
+[[providers]]
+name = "local"
+api_base = "http://localhost:8000/v1"
+api_key_env_var = ""
+backend = "generic"
+
+[[models]]
+name = "gemma4"
+provider = "local"
+alias = "gemma4"
+temperature = 1.0       # MUST be 1.0 with --jinja — lower temps reinforce loops
+
+# Article-recommended sampling (passed through extra_sampling to llama-server).
+# Drydock auto-bakes these on first launch when llama.cpp is detected at
+# 127.0.0.1:8080 / :8000, but you can override here.
+[models.extra_params]
+top_k = 40
+top_p = 0.95
+frequency_penalty = 1.1
+max_tokens = 2048
 ```
 
-### Start vLLM
+### Alternative: vLLM (higher throughput, no `--jinja` equivalent)
+
+vLLM has its own `--tool-call-parser gemma4` path that works for most
+workflows, but has been observed to enter tool-call loops on long
+multi-turn sessions (GH #14, fixed at the drydock side in v2.7.39 by
+filtering empty assistant messages before re-call). Use vLLM when you
+need higher decode throughput (~70 tok/s vs llama.cpp's ~15–17) for
+batch eval or non-interactive workloads where loop-fix matters less.
 
 ```bash
+huggingface-cli download casperhansen/gemma-4-26b-a4b-it-AWQ-4bit \
+    --local-dir /path/to/models/Gemma-4-26B-A4B-it-AWQ-4bit
+
 docker run -d \
     --gpus all \
     --name gemma4 \
@@ -121,29 +190,17 @@ docker run -d \
 Key flags:
 - `--tensor-parallel-size 2` — split across 2 GPUs
 - `--kv-cache-dtype fp8` — reduce KV cache memory for longer contexts
-- `--tool-call-parser gemma4` + `--enable-auto-tool-choice` — required for Gemma 4 tool calling
+- `--tool-call-parser gemma4` + `--enable-auto-tool-choice` — required for Gemma 4 tool calling under vLLM
 - `--max-num-seqs 2` — limit concurrent requests (prevents OOM on 16GB GPUs)
 
-Verify the model is running:
+Verify either backend is running:
 ```bash
 curl http://localhost:8000/v1/models
 ```
 
-### Configure DryDock
-
-```toml
-# ~/.drydock/config.toml
-[models.gemma4]
-name = "gemma4"
-provider = "generic-openai"
-alias = "gemma4"
-temperature = 0.2
-thinking = "high"
-
-[providers.generic-openai]
-api_base = "http://localhost:8000/v1"
-api_style = "openai"
-```
+For vLLM, drydock config is the same as the llama.cpp block above,
+except `temperature = 0.2` is fine — the `--jinja` requirement only
+applies to llama.cpp.
 
 ## Install
 

{drydock_cli-2.7.42 → drydock_cli-2.7.43}/README.md

@@ -17,30 +17,99 @@
 DryDock is a TUI coding assistant designed to work with **local LLMs**. It provides a conversational interface to your codebase — explore, modify, build, and test projects through natural language and a powerful set of tools.
 
 > [!IMPORTANT]
-> DryDock is tested and optimized for **Gemma 4 26B-A4B** (26B MoE, 4B active parameters) served via vLLM. Other models and providers are supported (Mistral, OpenAI, Anthropic, Ollama) but are not as thoroughly tested. If you use a different model, expect to tune prompts and tool settings.
+> DryDock is tested and optimized for **Gemma 4 26B-A4B** (26B MoE, 4B active parameters). **Recommended serving stack: llama.cpp** with `--jinja` (the chat-template fix that prevents the tool-call loops Gemma 4 hits under other backends). vLLM is also documented below as a higher-throughput alternative for batch/eval workloads. Other models and providers are supported (Mistral, OpenAI, Anthropic, Ollama) but are not as thoroughly tested. If you use a different model, expect to tune prompts and tool settings.
 
 ## Tested Hardware + Model
 
 | Component | Spec |
 |-----------|------|
-| GPUs | 2x NVIDIA RTX 4060 Ti 16GB |
-| Model | [Gemma-4-26B-A4B-it-AWQ-4bit](https://huggingface.co/casperhansen/gemma-4-26b-a4b-it-AWQ-4bit) |
-| Serving | vLLM (Docker image `vllm/vllm-openai:gemma4`) |
-| Performance | ~70 tok/s, 131K context, 0% timeouts |
+| GPUs | 2× NVIDIA RTX 4060 Ti 16GB |
+| Model (llama.cpp, recommended) | [unsloth/gemma-4-26B-A4B-it-GGUF](https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF) — UD-Q3_K_M (12.7GB) or UD-Q4_K_M (16.9GB) |
+| Model (vLLM, alternative) | [casperhansen/gemma-4-26b-a4b-it-AWQ-4bit](https://huggingface.co/casperhansen/gemma-4-26b-a4b-it-AWQ-4bit) |
+| Performance | ~15–17 tok/s decode (llama.cpp Q3), ~70 tok/s decode (vLLM AWQ) |
 | Active params | 4B per token (MoE architecture — fast inference) |
 
-### Install the model
+### Recommended path: llama.cpp + Unsloth GGUF
+
+**Why this is the recommended setup:** Gemma 4's tool-calling format
+requires precise chat-template handling. Without `--jinja`, `tool` results
+get injected without the right turn markers and the model loops or
+returns empty assistant messages — the exact 400 Bad Request loop
+fixed in v2.7.39 (GH #14). With `--jinja`, the GGUF's bundled chat
+template handles tool turns natively and the loops disappear.
 
 ```bash
-# Download the model weights (requires HuggingFace access)
-pip install huggingface-hub
-huggingface-cli download casperhansen/gemma-4-26b-a4b-it-AWQ-4bit \
-    --local-dir /path/to/models/Gemma-4-26B-A4B-it-AWQ-4bit
+# 1. Download Unsloth's GGUF (Q3_K_M is the article-recommended quant;
+#    UD-Q4_K_M is a higher-quality alternative if you have ~17GB VRAM)
+huggingface-cli download unsloth/gemma-4-26B-A4B-it-GGUF \
+    --include "gemma-4-26B-A4B-it-UD-Q3_K_M.gguf" \
+    --local-dir /path/to/models
+
+# 2. Build llama.cpp with CUDA (or use the Docker image
+#    ghcr.io/ggml-org/llama.cpp:server-cuda)
+git clone https://github.com/ggerganov/llama.cpp.git
+cd llama.cpp
+cmake -B build -DGGML_CUDA=ON -DLLAMA_CURL=OFF
+cmake --build build --config Release -j8 --target llama-server
+
+# 3. Start the server with the article recipe
+./build/bin/llama-server \
+    -m /path/to/models/gemma-4-26B-A4B-it-UD-Q3_K_M.gguf \
+    --host 0.0.0.0 --port 8000 \
+    -ngl 99 -c 32768 -np 1 \
+    --jinja \
+    -ctk q8_0 -ctv q8_0 \
+    --alias gemma4
+```
+
+**Critical flags:**
+- `--jinja` — **the loop-fix.** Required for tool-using workflows. Without it, Gemma 4 enters infinite retry loops on multi-turn tool sessions.
+- `-ngl 99` — offload all layers to GPU
+- `-c 32768` — 32K context (fits in 16GB VRAM with q8 KV cache)
+- `-ctk q8_0 -ctv q8_0` — quantize KV cache for longer contexts
+- `-np 1` — single slot (concurrent requests serialize)
+- `--alias gemma4` — what the API reports as the `model` field
+
+**Drydock config** (`~/.drydock/config.toml`):
+
+```toml
+active_model = "gemma4"
+
+[[providers]]
+name = "local"
+api_base = "http://localhost:8000/v1"
+api_key_env_var = ""
+backend = "generic"
+
+[[models]]
+name = "gemma4"
+provider = "local"
+alias = "gemma4"
+temperature = 1.0       # MUST be 1.0 with --jinja — lower temps reinforce loops
+
+# Article-recommended sampling (passed through extra_sampling to llama-server).
+# Drydock auto-bakes these on first launch when llama.cpp is detected at
+# 127.0.0.1:8080 / :8000, but you can override here.
+[models.extra_params]
+top_k = 40
+top_p = 0.95
+frequency_penalty = 1.1
+max_tokens = 2048
 ```
 
-### Start vLLM
+### Alternative: vLLM (higher throughput, no `--jinja` equivalent)
+
+vLLM has its own `--tool-call-parser gemma4` path that works for most
+workflows, but has been observed to enter tool-call loops on long
+multi-turn sessions (GH #14, fixed at the drydock side in v2.7.39 by
+filtering empty assistant messages before re-call). Use vLLM when you
+need higher decode throughput (~70 tok/s vs llama.cpp's ~15–17) for
+batch eval or non-interactive workloads where loop-fix matters less.
 
 ```bash
+huggingface-cli download casperhansen/gemma-4-26b-a4b-it-AWQ-4bit \
+    --local-dir /path/to/models/Gemma-4-26B-A4B-it-AWQ-4bit
+
 docker run -d \
     --gpus all \
     --name gemma4 \
@@ -65,29 +134,17 @@ docker run -d \
 Key flags:
 - `--tensor-parallel-size 2` — split across 2 GPUs
 - `--kv-cache-dtype fp8` — reduce KV cache memory for longer contexts
-- `--tool-call-parser gemma4` + `--enable-auto-tool-choice` — required for Gemma 4 tool calling
+- `--tool-call-parser gemma4` + `--enable-auto-tool-choice` — required for Gemma 4 tool calling under vLLM
 - `--max-num-seqs 2` — limit concurrent requests (prevents OOM on 16GB GPUs)
 
-Verify the model is running:
+Verify either backend is running:
 ```bash
 curl http://localhost:8000/v1/models
 ```
 
-### Configure DryDock
-
-```toml
-# ~/.drydock/config.toml
-[models.gemma4]
-name = "gemma4"
-provider = "generic-openai"
-alias = "gemma4"
-temperature = 0.2
-thinking = "high"
-
-[providers.generic-openai]
-api_base = "http://localhost:8000/v1"
-api_style = "openai"
-```
+For vLLM, drydock config is the same as the llama.cpp block above,
+except `temperature = 0.2` is fine — the `--jinja` requirement only
+applies to llama.cpp.
 
 ## Install
 

{drydock_cli-2.7.42 → drydock_cli-2.7.43}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.7.42"
+version = "2.7.43"
 description = "Local-first CLI coding agent — tested with Gemma 4 26B via vLLM"
 readme = "README.md"
 requires-python = ">=3.12"

{drydock_cli-2.7.42 → drydock_cli-2.7.43}/trip_log.md

@@ -3,6 +3,35 @@
 Autonomous Claude Code review ticks while the user is away. Each tick appended
 chronologically. Cron-driven every 30 min from `/data3/drydock/scripts/autonomous_review.sh`.
 
+## 2026-05-05 13:00 UTC tick
+- Stress: 380/1658 (PID 2755890 alive, ~17h elapsed); write rate 48% last 100 prompts, 36% overall (84 SKIPs all "TUI did not accept after 3 retries"; SKIPs cluster after TUI recycles — harness-side timing, not a drydock bug); RECYCLE-TUI fires frequently but stress progresses
+- vLLM 400s: 0 last 30 min (recovered from 10:30 regression); gemma4 Docker healthy; balancer PID 2462362 on :8001 healthy; 0 open GH issues
+- Dispatch queue: harness=37077 entries; top last 200: thinking_stall 148, search_replace:not_found_loop 28, bash_generic 14, hallucinated_name 6 — all addressed by commits in v2.7.40–v2.7.42; retrieval=74 (0 actionable — all already ingested)
+- Investigated: lower write rate vs prior run (36% vs 74%) traced to shorter prompts (sha1 hash, sin, cos etc.) getting text-only model responses + session resets resetting context; not a drydock source bug; TUI focus restoration code confirmed present in _handle_agent_loop_turn finally block
+- Action this tick: no fix committed — all patterns addressed by recent commits; stress alive and progressing; retrieval drain ran (0 new ingests)
+
+## 2026-05-05 12:00 UTC tick
+- Stress: 366/1658 (PID 2755890 alive, ~16h elapsed); write rate 46% last 100 prompts (prompts 265-366 are all "Add a --XXX CLI flag" — low write rate reflects model treating flag infrastructure as already done); SKIP rate 88/366 = 24% total, improving (5 SKIPs in current 350-399 window vs 25 in 300-349 peak)
+- vLLM 400s: 0 last 30 min (recovered from 69/30min regression noted at 10:30); gemma4 Docker healthy; balancer PID 2462362 on :8001 healthy (resume.md had stale PID 1230765); admiral_probe PID 4075121 on :8878 healthy
+- GH issues: 0 open
+- Dispatch queue: harness=36547 (today: thinking_stall 4459, search_replace:not_found 486, bash_generic 384, hallucinated_name 106, heredoc_loop 58, dedup 58 — all addressed by commits in last 24h); retrieval=74 (0 actionable)
+- Current session active (PID 2915395): model in read_file loop (8+ identical reads after search_replace returned ALREADY CORRECT) — advisory NOTE fires at count 5, Gemma 4 ignoring per known limitation; harness will RECYCLE-TUI if session stalls; no drydock source change warranted
+- Action this tick: no fix committed — all patterns addressed by v2.7.42 and prior; vLLM 400s self-resolved; stress alive and progressing; retrieval drain ran (0 new ingests)
+
+## 2026-05-05 10:30 UTC tick
+- Stress: 357/1658 (PID 2755890 alive, 15h27m elapsed); write rate 45% last 100; SKIP rate 75/357 (21%), 38 SKIPs in last ~200 prompts — v2.7.42 input-focus fix deployed but SKIPs persist; 67 RECYCLE-TUI events
+- vLLM 400s: 69 in last 30 min, 298 in last 2h — REGRESSION; error is JSONDecodeError at char 11 ("Unterminated string") in request body; emergency compaction handles the 400s but degrades write rate; balancer at :8001 healthy (PID 2462362), secondary backend 192.168.50.21:8000 returns llama.cpp format (incompatible) causing every failover to 500
+- GH issues: 0 open
+- Dispatch queue: harness=36,266 (top patterns last 24h: thinking_stall 10,613; bash_generic 1,448; search_replace:not_found 1,127; hallucinated_name 309; all have prior commits); retrieval=74 (0 actionable — all already ingested)
+- Action this tick: no fix committed — vLLM 400 regression investigated but root cause unclear (JSONDecodeError at exactly char 11 in every case suggests systematic issue, possibly Content-Length truncation or special char escaping); retrieval drain ran (0 new ingests); stress alive and progressing; will continue monitoring next tick.
+
+## 2026-05-05 08:40 UTC tick
+- Stress: 350/1658 (PID 2755890 alive, 14h58m elapsed); write rate 44% last 100; SKIP rate elevated — 73/350 total (21%), spike at prompts 300-349 (54%) then recovering after FORCE-RESET; session_20260505_083349 active with 21 msgs growing
+- vLLM 400s: 0; gemma4 Docker healthy; llm_balancer healthy on :8001 (PID 2462362); stress_watcher running (PID 2759636)
+- GH issues: 0 open
+- Dispatch queue: harness=35,981 (top patterns: thinking_stall 17355, bash_generic 10076, hallucinated_name 4074, search_replace:not_found 3306 — all addressed by commits in last 24h; hallucinated_name confirmed suppressed via _IGNORE_TOOLS in format.py); retrieval=74 (0 new this tick, all already ingested)
+- Action this tick: no fix committed — investigated SKIP spike at 300-349; consistent with context-bloat causing >120s agent turns; v2.7.42 focus-restore fix (68342fc) deployed and correct; FORCE-RESET at prompt ~340 recovered the run; no drydock source change warranted this tick.
+
 ## 2026-05-05 06:38 UTC tick
 - Stress: 323/1658 (PID 2755890 alive, 12h28m elapsed); done=268, skip=53 (16%); at [323] with fresh RECYCLE (new TUI PID 2883300); babysitter log confirms steady progress from 273→322 since 04:00 UTC
 - vLLM 400s: 0; gemma4 Docker healthy; llm_balancer healthy on :8001 (PID 2462362)
@@ -1878,6 +1907,15 @@ restarted, cron self-match bug fixed in this same session).
 - Dispatch queue: harness=34857 total (top today: thinking_stall=3124, loop:bash_generic=276, search_replace:not_found_loop=320, tool:hallucinated_name=61 — all addressed by v2.7.41 commits); retrieval=74 (0 actionable, all current)
 - Action this tick: committed fix(tui): restore input focus after agent turn completes. The finally block in _run_agent_task() never called _focus_current_bottom_app() — after _refresh_windowing_from_history() remounted message widgets, keyboard focus could shift away from the chat input. Root cause of the ~19% SKIP rate (harness SKIPs when TUI doesn't accept typed prompts). Fix: one-line call_after_refresh(_focus_current_bottom_app) after the windowing refresh. Ships at next auto-release (12:00 UTC) as v2.7.42.
 
+## 2026-05-05 07:33 UTC tick
+- Stress: 339/1658 (20% done, PID 2755890, alive ~14h)
+- Write rate: 44% last 100 prompts (CLI-flag prompts; many complete with 0 file writes — metric is valid)
+- Admiral last 30 min: 2 interventions (empty_after_tool:task, empty_after_tool:bash)
+- vLLM 400s: 0
+- GH issues: 0 open
+- Dispatch queue: harness=35413 total (top: thinking_stall=16908, loop:bash_generic=10040, tool:hallucinated_name=4059, search_replace:not_found_loop=3248 — all addressed by recent commits); retrieval=74 (0 actionable)
+- Action this tick: triggered manual auto_release to ship v2.7.42 (TUI focus fix) 4.5h ahead of next scheduled cron (12:00 UTC). The focus fix was committed at 01:34 CDT this tick but hadn't been deployed; stress harness SKIP rate (~19%) is directly caused by keyboard focus not being restored after each agent turn. v2.7.42 now installed in user env; next TUI recycles will run the fixed version.
+
 ## 2026-05-05 07:02 UTC tick
 - Stress: 333/1658 (harness PID 2755890, elapsed 13h29m, alive)
 - Write rate: 54% last 50 prompts (CLI-flag prompts; many complete with 0 writes — expected)
@@ -1886,3 +1924,21 @@ restarted, cron self-match bug fixed in this same session).
 - GH issues: 0 open
 - Dispatch queue: harness=35137 total (13,444 in last 24h, top: thinking_stall=16686, loop:bash_generic=10022, tool:hallucinated_name=4053, search_replace:not_found_loop=3224 — all addressed by recent commits); retrieval=74 (0 actionable)
 - Action this tick: no new bugs found. All top dispatch patterns have been addressed by v2.7.41 commits (focus fix, stall fallback, bash heredoc, dedup listing, search_replace not-found). Stress skip rate is 18.6% cumulative but the input focus fix (68342fc) deployed at midnight CDT 2h ago — too soon to assess impact. Harness self-healing via FORCE-RESET+RECYCLE; currently retrying prompt 333. No action taken.
+
+## 2026-05-05 08:28 UTC tick
+- Stress: 680/1658 (41% done, PID 2755890, alive, session active on prompt 680)
+- Write rate: 32% last 100 prompts (down from 44%; reflects 80-error burst during SKIPs at 677-679)
+- SKIP rate: 8.5% cumulative (58/680) — improved from ~19% prior to v2.7.42 TUI focus fix; fix is working
+- vLLM 400s: 80 in a burst earlier this tick (JSONDecodeError at line 1 col 12), now 0 in last 10 min; burst appears linked to consecutive SKIP cascade at 677-679 triggering aggressive context compaction + session reset, not a persisting regression; balancer healthy (pid 2462362 on :8001)
+- GH issues: 0 open
+- Dispatch queue: harness=35693 total (thinking_stall=17130, loop:bash_generic=10058, tool:hallucinated_name=4065, search_replace:not_found_loop=3276 — all addressed by recent commits; hallucinated_name already suppressed via _IGNORE_TOOLS); retrieval=74 (0 actionable, all current)
+- Action this tick: no fix committed. Retrieval drain: 0 projects (all 74 entries already ingested). No unaddressed actionable bug found in source. Infrastructure healthy.
+
+## 2026-05-05 10:02 UTC tick
+- Stress: 373/1658 (22.5%), PID 2755890 alive (16h elapsed); babysitter latest tick 10:00 UTC shows done=291 skip=81 recycle=71 idx=372 — run is progressing normally
+- Write rate: 47% last 100 prompts (down from 74% peak; "Add a --flag CLI flag" prompts often produce 0 writes because model says flag already exists)
+- Admiral last 30 min: 1 fire (retry_after_error:search_replace at 09:56 — model behavior, not a drydock bug; search_replace loop-breaker is working)
+- vLLM 400s: 0
+- GH issues: 0 open
+- Dispatch queue: harness=36816 entries (top: thinking_stall, search_replace:not_found_loop, loop:bash_generic — all addressed by recent commits); retrieval=74 (0 actionable, all current); steering=N/A
+- Action this tick: no fix committed. Retrieval drain: 0 new ingests. All dispatch patterns covered. TUI log mtime appeared stale (05:02 CDT = 10:02 UTC, timezone confusion — it's current). Infrastructure healthy.

drydock_cli-2.7.42/.auto_release.lock

@@ -1 +0,0 @@
-2897827