drydock-cli 2.7.26__tar.gz → 2.7.28__tar.gz

drydock_cli-2.7.28/.auto_release.lock

@@ -0,0 +1 @@
+2010664

{drydock_cli-2.7.26 → drydock_cli-2.7.28}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.7.26
+Version: 2.7.28
 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

{drydock_cli-2.7.26 → drydock_cli-2.7.28}/drydock/admiral/detectors.py

@@ -64,11 +64,17 @@ def detect_struggle(messages: Sequence[LLMMessage], threshold: int = 20) -> Find
     """Fires when `threshold` tool calls have happened without any write.
 
     Reading + grepping forever without writing code is the classic
-    "stuck exploring" failure mode.
+    "stuck exploring" failure mode. Counter resets on each user turn so
+    test-only prompts (run tests, report results) don't accumulate across
+    turns and cause false positives.
     """
     calls_since_write = 0
     last_write_tool: str | None = None
     for m in messages:
+        if m.role == Role.user:
+            # Each new user prompt gives the model a fresh budget.
+            calls_since_write = 0
+            continue
         if m.role != Role.assistant or not m.tool_calls:
             continue
         for tc in m.tool_calls:

{drydock_cli-2.7.26 → drydock_cli-2.7.28}/drydock/core/agent_loop.py

@@ -2414,7 +2414,8 @@ class AgentLoop:
         # edit that already succeeded or that keeps failing with the same
         # "not found" error.  Nudge after just 2 identical attempts.
         recent_sr: list[str] = []
-        for msg in reversed(self.messages[-20:]):
+        recent_sr_files: list[str] = []
+        for msg in reversed(self.messages[-30:]):
             if msg.role == Role.assistant and msg.tool_calls:
                 for tc in msg.tool_calls:
                     if tc.function and tc.function.name == "search_replace":
@@ -2423,12 +2424,27 @@ class AgentLoop:
                             # Build a key from file_path + old_string (content block)
                             key = f"{args.get('file_path', '')}:{args.get('content', '')}"
                             recent_sr.append(key)
+                            recent_sr_files.append(args.get("file_path", ""))
                         except (json.JSONDecodeError, AttributeError):
                             pass
-                if len(recent_sr) >= 4:
+                if len(recent_sr) >= 6:
                     break
         if len(recent_sr) >= 2 and recent_sr[0] == recent_sr[1]:
             return "WARNING|search_replace"
+        # Detect when 5+ search_replace calls target the same file with
+        # varying search text (model adapts after HARD-STOP but still
+        # cannot find the right text). The per-file fail counter in
+        # search_replace.py escalates at count 3 — this adds a
+        # loop-detection layer that fires FORCE_STOP when the same file
+        # dominates 5 of the last 6 search_replace calls.
+        if len(recent_sr_files) >= 5:
+            from collections import Counter as _SRCounter
+            _sr_counts = _SRCounter(f for f in recent_sr_files if f)
+            if _sr_counts:
+                _top_sr_file, _top_sr_count = _sr_counts.most_common(1)[0]
+                if _top_sr_count >= 5:
+                    self._hot_tool_path = ("search_replace", _top_sr_file)
+                    return "FORCE_STOP"
 
         # Early check: write_file with _truncated args twice in a row for the
         # same path.  format.py already embeds the file content in the error,
@@ -2454,6 +2470,34 @@ class AgentLoop:
             self._hot_tool_path = ("write_file", recent_wf_truncated[0])
             return "FORCE_STOP"
 
+        # Early check: same bash command 5+ times across last 20 tool calls.
+        # Catches alternating bash/read_file exploration loops where neither
+        # the consecutive-N check nor the 9/12 path-dominance check fires
+        # (because bash and read_file alternate, keeping bash below 9/12).
+        _bash_cmds: list[str] = []
+        _total_tc = 0
+        for _msg in reversed(self.messages[-40:]):
+            if _msg.role == Role.assistant and _msg.tool_calls:
+                for _tc in _msg.tool_calls:
+                    _total_tc += 1
+                    if _tc.function and _tc.function.name == "bash":
+                        try:
+                            _a = json.loads(_tc.function.arguments or "{}")
+                            _cmd = _a.get("command", "")
+                            if _cmd:
+                                _bash_cmds.append(_cmd)
+                        except (json.JSONDecodeError, AttributeError):
+                            pass
+            if _total_tc >= 20:
+                break
+        if len(_bash_cmds) >= 5:
+            from collections import Counter as _Counter
+            _cmd_counts = _Counter(_bash_cmds)
+            _top_cmd, _top_count = _cmd_counts.most_common(1)[0]
+            if _top_count >= 5:
+                self._hot_tool_path = ("bash", _top_cmd)
+                return "FORCE_STOP"
+
         sigs: list[str] = []
         tool_names: list[str] = []
         lookback = 0

{drydock_cli-2.7.26 → drydock_cli-2.7.28}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.7.26"
+version = "2.7.28"
 description = "Local-first CLI coding agent — tested with Gemma 4 26B via vLLM"
 readme = "README.md"
 requires-python = ">=3.12"

{drydock_cli-2.7.26 → drydock_cli-2.7.28}/tests/test_admiral.py

@@ -64,6 +64,30 @@ def test_struggle_resets_on_write() -> None:
     assert detect_struggle(msgs) is None
 
 
+def test_struggle_resets_on_user_turn() -> None:
+    """Multiple test-only prompts must not accumulate into a false struggle."""
+    msgs: list[LLMMessage] = []
+    # 5 user prompts, each generating 8 bash calls (total 40 non-write calls)
+    for _ in range(5):
+        msgs.append(LLMMessage(role=Role.user, content="Test: run benchmarks"))
+        for _ in range(8):
+            msgs.append(_assistant("bash", '{"command":"python3 test.py"}'))
+            msgs.append(LLMMessage(role=Role.tool, content="ok"))
+        msgs.append(LLMMessage(role=Role.assistant, content="Tests passed."))
+    assert detect_struggle(msgs) is None
+
+
+def test_struggle_still_fires_within_single_turn() -> None:
+    """A single turn with 25+ non-write calls still fires struggle."""
+    msgs: list[LLMMessage] = [LLMMessage(role=Role.user, content="implement feature X")]
+    for i in range(25):
+        msgs.append(_assistant("read_file", f'{{"path":"f{i}.py"}}'))
+        msgs.append(LLMMessage(role=Role.tool, content="..."))
+    f = detect_struggle(msgs)
+    assert f is not None
+    assert "without writing" in f.directive
+
+
 def test_run_all_returns_multiple_findings() -> None:
     msgs: list[LLMMessage] = []
     for i in range(25):

{drydock_cli-2.7.26 → drydock_cli-2.7.28}/tests/test_loop_detection.py

@@ -347,6 +347,96 @@ class TestAlternatingPattern:
         assert result is None or "FORCE_STOP" not in result
 
 
+# ============================================================================
+# Bash Exploration Loop (same command alternating with other tools)
+# ============================================================================
+
+class TestBashExplorationLoop:
+    """Same bash command 5+ times across 20 tool calls triggers FORCE_STOP."""
+
+    def test_alternating_bash_read_triggers(self):
+        """bash(same_cmd) + read_file(different) alternating 10x → FORCE_STOP."""
+        al = _make_agent()
+        cmd = "python3 -m tool_agent --list-tools"
+        for i in range(10):
+            if i % 2 == 0:
+                _add_tool_call(al, "bash", json.dumps({"command": cmd}))
+            else:
+                _add_tool_call(al, "read_file", json.dumps({"path": f"file_{i}.py"}))
+        result = al._check_tool_call_repetition()
+        assert result == "FORCE_STOP"
+        assert getattr(al, "_hot_tool_path", None) == ("bash", cmd)
+
+    def test_5_identical_bash_triggers(self):
+        """5 identical bash calls in a row → FORCE_STOP."""
+        al = _make_agent()
+        cmd = "python3 -m tool_agent --list-tools"
+        for _ in range(5):
+            _add_tool_call(al, "bash", json.dumps({"command": cmd}))
+        result = al._check_tool_call_repetition()
+        assert result == "FORCE_STOP"
+
+    def test_4_identical_bash_no_trigger(self):
+        """4 identical bash calls should not trigger (below threshold)."""
+        al = _make_agent()
+        cmd = "python3 app.py"
+        for _ in range(4):
+            _add_tool_call(al, "bash", json.dumps({"command": cmd}))
+        result = al._check_tool_call_repetition()
+        # Should not trigger the bash-exploration check
+        assert result != "FORCE_STOP"
+
+    def test_varied_bash_no_trigger(self):
+        """Different bash commands should not trigger."""
+        al = _make_agent()
+        cmds = ["ls -la", "cat README.md", "python3 app.py", "pwd", "echo hi"]
+        for cmd in cmds:
+            _add_tool_call(al, "bash", json.dumps({"command": cmd}))
+        result = al._check_tool_call_repetition()
+        assert result != "FORCE_STOP"
+
+
+# ============================================================================
+# Search/Replace File-Dominance Loop Detection
+# ============================================================================
+
+class TestSearchReplaceFileDominance:
+    """search_replace on same file 5+ times with varying content → FORCE_STOP."""
+
+    def test_five_sr_same_file_triggers(self):
+        """5 search_replace calls on same file (different search text) → FORCE_STOP."""
+        al = _make_agent()
+        for i in range(5):
+            _add_tool_call(
+                al, "search_replace",
+                json.dumps({"file_path": "tool_agent/cli.py", "content": f"text{i}"})
+            )
+        result = al._check_tool_call_repetition()
+        assert result == "FORCE_STOP"
+
+    def test_four_sr_same_file_no_trigger(self):
+        """4 search_replace on same file should not trigger."""
+        al = _make_agent()
+        for i in range(4):
+            _add_tool_call(
+                al, "search_replace",
+                json.dumps({"file_path": "tool_agent/cli.py", "content": f"text{i}"})
+            )
+        result = al._check_tool_call_repetition()
+        assert result != "FORCE_STOP"
+
+    def test_sr_different_files_no_trigger(self):
+        """search_replace spread across different files should not trigger."""
+        al = _make_agent()
+        for i in range(5):
+            _add_tool_call(
+                al, "search_replace",
+                json.dumps({"file_path": f"tool_agent/file{i}.py", "content": "text"})
+            )
+        result = al._check_tool_call_repetition()
+        assert result != "FORCE_STOP"
+
+
 # ============================================================================
 # Threshold Constants
 # ============================================================================

{drydock_cli-2.7.26 → drydock_cli-2.7.28}/trip_log.md

@@ -3,6 +3,82 @@
 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`.
 
+## 2026-05-01 12:34 UTC tick
+- Stress: 1020/1658 (PID 675181, alive 21h25m, log stress_2000_v10_restart_1777561483.log; write rate 10% last 100 prompts — all current prompts are "Test: unit/fuzz/property/golden" type; model runs bash tests and reports, not creating new files; overall 20% write rate for this restart run)
+- Write rate: 10% last 100 (expected for test-running block; 20% overall for this restart)
+- Skip rate: 109 total skips (babysitter confirmed skip=76 at 12:00 UTC); FORCE-RESET recovering normally
+- Admiral last 30 min: struggle:none firing every 60s (12:00-12:08 UTC) during "add storage backend" explore-without-write block (36-45 tool calls, model not responding to directives — known Gemma 4 limitation); also struggle:search_replace at 12:28-12:30; empty_after_tool:ralph_repo_index at 11:33 (one-off, model moved on); retry_after_error:write_file:truncated-history at 08:19 (model re-used truncated args, known pattern)
+- vLLM 400s: 0
+- GH issues: 0 open
+- Services: llm_balancer PID 713929 on :8001 healthy; vLLM gemma4 on :8000 healthy; admiral_probe PID 4075121 on :8878 healthy; harness RSS 7-8GB (harness TUI terminal log at 608MB, not a drydock leak); v2.7.27 is latest tag
+- Action this tick: no fix committed — system healthy; all observed admiral patterns are known (struggle:none is advisory-only limitation, retry_after_error handled by existing code, truncated-history errors expected from context compaction); no new actionable drydock bugs found
+
+## 2026-05-01 11:33 UTC tick
+- Stress: 990/1658 (PID 675181, alive 19h26m, log stress_2000_v10_restart_1777561483.log; 311 entries this restart; block is "Test: unit/integration" prompts)
+- Write rate: 17% last 100 prompts (expected — current block is "Test: X" prompts; model runs bash tests and reports, not writing files; matches 09:33 and 07:31 ticks)
+- Skip rate: 75/311 = 24% (down from 33% last tick; TUI wedge on API/Test server prompts, FORCE-RESET recovering normally; no orphan processes on ports)
+- Admiral last 30 min: 3 fires of loop:bash::k_anonymity_golden_test.py (10:28-10:30 UTC), 2 empty_after_tool:bash; no new retry_after_error:search_replace since 08:23 UTC
+- vLLM 400s: 0
+- GH issues: 0 open
+- Services: llm_balancer PID 713929 on :8001 healthy; vLLM gemma4 healthy; admiral_probe PID 4075121 on :8878 healthy; current version v2.7.26 (latest tag)
+- Action this tick: no fix committed — system healthy; write rate/skip rate consistent with prior ticks at "Test:" prompt block; search_replace LOOP-BREAKER and loop detection already in place for observed admiral patterns; no new actionable drydock bugs found
+
+## 2026-05-01 09:33 UTC tick
+- Stress: 982/1658 (PID 675181, alive 18h29m, writing to stress_2000_v10_restart_1777561483.log; 303 entries this run; block is "Test: regression/smoke/performance/memory" prompts)
+- Write rate: 17% last 100 prompts (expected — all current prompts are "Test: X" type; model runs bash tests and reports results, not writing files)
+- Skip rate: 101/303 = 33% (TUI wedge pattern consistent with prior ticks; FORCE-RESET recovering; no orphan servers on :8001/:8000/:8878)
+- Admiral last 30 min: 7 fires (all struggle:search_replace — model doing 20-21 tool calls without writing during a long search_replace session around 09:28-09:32 UTC; fixed by user-turn reset from 71cb046 for cross-turn accumulation but within-turn depth still fires normally)
+- vLLM 400s: 0
+- GH issues: 0 open
+- Services: llm_balancer PID 713929 on :8001 healthy; vLLM gemma4 up 7 days healthy; admiral_probe PID 4075121 on :8878 healthy; current version v2.7.26; 1 commit ahead of latest tag (ships at next 0/6/12/18 auto-release)
+- Action this tick: no fix committed — system healthy; write rate and skip rate consistent with prior ticks at same "Test:" prompt block; no new drydock bugs identified; harness progressing normally
+
+## 2026-05-01 07:31 UTC tick
+- Stress: 950/1658 (PID 675181, alive 16h27m, log stress_2000_v10_restart_1777561483.log; 271 entries this run)
+- Write rate: 21% last 100 prompts (expected — current block is "Test: fuzz/property/integration" prompts; model asks clarifying questions instead of writing, producing 0-writes responses; prompt-category effect, not regression)
+- Skip rate: 68/271 = 25% (TUI wedge on API/server prompts, FORCE-RESET recovering; no orphan servers on ports 50051/8765/8080 at this tick — clean)
+- Admiral last 30 min: 0 fires (model in text-response mode on Test: prompts, no tool-call loops observed)
+- vLLM 400s: 0
+- GH issues: 0 open
+- Services: llm_balancer PID 713929 on :8001 healthy; vLLM gemma4 up 7 days; 1 commit ahead of latest tag (fix(admiral): reset struggle counter on each user turn — ships at next 0/6/12/18 auto-release)
+- Action this tick: no fix committed — system healthy, no new drydock bugs identified; harness progressing; write rate and skip rate consistent with prior ticks at same prompt block
+
+## 2026-05-01 03:45 UTC tick
+- Stress: 680/1658 (PID 675181, alive, resumed from step 679; currently in "Add storage backend: X" / "API: JSON-RPC" prompt block)
+- Write rate: 32% last 100 prompts (low due to "Add storage backend: X" cluster — session already has implementations, model correctly returns 0 writes; prompt-category effect, not regression)
+- Admiral last 30 min: loop:bash::{fuser -k 8765/tcp} x11 (01:00–01:10 UTC, ~10-min port-kill spiral); loop:bash::{fuser -k 8000/tcp} x1; loop:search_replace conflict markers x1; empty_after_tool:task x1; empty_after_tool:read_file x1; retry_after_error:write_file truncated-history x2 (03:18, 03:30 — existing escalation logic in format.py handles these); struggle:none x3; loop:bash memory_benchmark x3; retry_after_error:search_replace x1
+- vLLM 400s: 0
+- GH issues: 0 open
+- Services: llm_balancer PID 713929 on :8001 healthy; vLLM gemma4 on :8000 healthy; admiral_probe PID 4075121 on :8878 healthy
+- Action this tick: no fix committed — all observed patterns are known types; write rate drop is prompt-category effect; no new drydock bugs identified; harness alive and progressing
+
+## 2026-05-01 04:52 UTC tick
+- Stress: 907/1658 (PID 675181, alive 14h25m, in "API surfaces" / "Test:" prompt block)
+- Write rate: 26% last 100 prompts (expected — current block is "Test: integration test / fuzz test" which produce 0 writes; first 100 prompts in this log also showed 20%)
+- Admiral top patterns overall: loop:read_file x302, loop:search_replace x234, loop:bash x181, struggle:none x92, retry_after_error:search_replace x75+71, empty_after_tool:ralph_repo_index x74 — all known, all handled; no new patterns
+- vLLM 400s: 0 (last 30m)
+- GH issues: 0 open
+- Services: llm_balancer PID 713929 on :8001 healthy (auto-restarted from PID 1230765); vLLM gemma4 on :8000 healthy (0 JSONDecodeErrors); stress PID 675181 alive; current version v2.7.26
+- Action this tick: no fix committed — system healthy, no actionable new bugs; write rate drop is prompt-category effect not regression; 32 total retry_after_error:write_file truncated-history fires across run (existing escalation logic handles each one); no drydock source changes needed
+
+## 2026-05-01 02:05 UTC tick
+- Stress: 866/1658 (PID 675181, alive 11h, writing to stress_2000_v10_restart_1777561483.log; 187 entries this run; 84 SKIPs = 45% SKIP rate; currently stuck at "Test:" block — integration/fuzz/property test prompts causing TUI wedge; FORCE-RESET not unsticking; rec-check shows raw_md=1 but no API-error banner)
+- Write rate: 24% this run (28/113 prompts with writes; "Test:" prompts tend to run bash rather than write files; also high SKIP rate artificially depresses metric)
+- Admiral last 30 min: 11x loop:bash::{fuser -k 8765/tcp} (model looping on port-kill for orphaned WebSocket server — cleared by this tick); 1x loop:search_replace conflict markers; 1x loop:bash::fuser -k 8000/tcp; 1x empty_after_tool:task; 1x empty_after_tool:read_file; 1x retry_after_error:bash
+- vLLM 400s: 0
+- GH issues: 0 open
+- Services: llm_balancer PID 713929 on :8001 healthy; vLLM gemma4 on :8000 healthy; RSS 3.3GB (babysitter threshold 4GB — not triggered); 11 orphan test server processes killed (PIDs 767439 on :8765 websocket_demo/server; 751293 751940 757659 758295 759238 764436 on :50051 gRPC servers)
+- Action this tick: killed 7 orphan test server processes squatting ports 8765 and 50051; these were causing 11-minute loop:bash spirals as the model tried fuser -k 8765 to reclaim the port; no drydock source bug fix committed (TUI wedge at Test: prompts is harness-side recovery gap, not drydock source bug; all other patterns are known)
+
+## 2026-05-01 00:35 UTC tick
+- Stress: 854/1658 (PID 675181, alive 9h25m, log stress_2000_v10_restart_1777561483.log; 174 prompts logged this run; harness in SKIP/FORCE-RESET cycle at current "API: SSE/gRPC" block — 50 total SKIPs and 19 FORCE-RESETs in log)
+- Write rate: 21% last 100 prompts (same API: block pattern — gRPC bidi-streaming, WebSocket, SSE prompts; model runs server+client bash tests rather than writing new files; expected depression for this prompt category)
+- Admiral last 30 min: empty_after_tool:bash firing every 15-20 min (model generates thinking-only response after long bash server/client commands; admiral intervenes each time; pattern is recurring but handled); struggle:none (2 fires); loop:search_replace and loop:bash (1 each); retry_after_error:write_file "truncated history" (1 — existing pattern, escalation logic in format.py handles it)
+- vLLM 400s: 0
+- GH issues: 0 open
+- Services: llm_balancer PID 713929 on :8001 healthy (confirmed cmd match); vLLM gemma4 on :8000 healthy; admiral_probe PID 4075121 on :8878 healthy; RSS 3.0GB
+- Action this tick: no fix committed — all observed patterns are known (empty_after_tool:bash, struggle:none, loop:search_replace); stress run progressing through hard API prompt block; no new drydock bugs identified; write rate drop is prompt-category effect, not regression
+
 ## 2026-04-30 17:55 UTC tick
 - Stress: 233/1658 in new run (PID 675181, --resume-from-step 679 but babysitter triggered a fresh wipe; log stress_2000_1777408317.log); 55 SKIPs (24%), 3 TIMEOUTs — elevated SKIP rate (was ~8%) is harness timing, not a drydock regression
 - Write rate: 19% last 100 prompts (expected for current prompt block — hash/cipher/math tools added via search_replace to existing files; metric only counts write_file, not search_replace edits)
@@ -12,6 +88,24 @@ chronologically. Cron-driven every 30 min from `/data3/drydock/scripts/autonomou
 - Services: llm_balancer PID 713929 on :8001 healthy; vLLM gemma4 on :8000 healthy; killed 5 orphaned gRPC server processes (PIDs 737458,738127,738759,744524,745299) that were squatting on port 50051 — stress run spawned these as background servers but never cleaned up; port 50051 now clear
 - Action this tick: killed orphaned gRPC servers; no drydock source fix committed (SKIP rate is harness timing, not a code bug; write rate drop is metric artifact from prompt type)
 
+## 2026-04-30 23:32 UTC tick
+- Stress: 843/1658 (PID 675181, alive 8h26m; same TUI session since 18:30 UTC — no new session dirs created in 5h, harness is reusing PID 675184 drydock process via /clear; harness RSS at 3.2GB and growing; 44/164 SKIPs this run = 27% SKIP rate in current log, elevated from prior ticks)
+- Write rate: 20% last 100 prompts (current "API:" block — GraphQL/REST/gRPC prompts; model gives text explanations, few file writes; known prompt-type effect)
+- Admiral last 30 min: not checked (budget limited)
+- vLLM 400s: 0
+- GH issues: 0 open
+- Services: llm_balancer PID 713929 on :8001 healthy; vLLM gemma4 on :8000 healthy; harness RSS 3.2GB (babysitter will warn but admiral threshold 4GB not reached)
+- Action this tick: no fix committed — harness alive and progressing through API prompt block; high SKIP rate (27%) is post-reset TUI timing, not a drydock code bug; no new GitHub issues; nothing actionable found
+
+## 2026-04-30 23:02 UTC tick
+- Stress: 838/1658 (PID 675181, alive 7h55m; 118 done, 40 SKIP, 0 TIMEOUT this run; babysitter restart at ~15:04 UTC, writing to stress_2000_v10_restart_1777561483.log)
+- Write rate: 20% last 100 prompts (current batch is "API:" prompts — REST GET/POST, rate limiters, GraphQL endpoints; model explores/tests servers; metric depressed by server-start loops)
+- Admiral last 30 min: loop:bash (pkill/restart rest_api_server pattern — orphan port squatting on model-spawned servers, known issue), loop:search_replace, struggle:none, empty_after_tool:bash/task — all known patterns, no new ones
+- vLLM 400s: 0
+- GH issues: 0 open
+- Services: llm_balancer PID 713929 on :8001 healthy, vLLM gemma4 on :8000 healthy; RSS 3018MB (babysitter warns >2GB; admiral actuator threshold 4GB, not reached)
+- Action this tick: no fix committed — infrastructure healthy, no new drydock bugs found; write rate drop is prompt-category effect (API server prompts harder for model)
+
 ## 2026-04-30 20:35 UTC tick
 - Stress: 812/1658 (PID 675181, alive 5h25m, writing to stress_2000_v10_restart_1777561483.log; 134 prompts this run, 107 done, 26 SKIP, 0 TIMEOUT; resume from step 679 by babysitter)
 - Write rate: 10% last 50 prompts (expected — currently in abstract "API:" prompt block: API versioning, rate limiters, REST/GraphQL endpoints; model explores 20-33 files before giving text responses; session reset at step 810 restored normal 4-write behavior for rate_limiter prompts after reset)
@@ -1150,3 +1244,180 @@ restarted, cron self-match bug fixed in this same session).
 - vLLM 400s: 0; llm_balancer PID 713929 on :8001 healthy (restarted by keepalive cron at ~19:00 UTC); vLLM gemma4 on :8000 healthy; TUI RSS 422MB, harness RSS 2753MB (below 4GB threshold)
 - GH issues: 0 open
 - Action this tick: no new drydock bugs found. Both recent fixes active in source (4e49bbe task continuation nudge, 92e5b3f successful_test_runs reset). Stress harness is alive and progressing; high skip rate is API-section timing behavior, not a drydock code bug. No uncommitted fixes.
+
+## 2026-05-01 00:01 UTC tick
+- Stress: 849/1658 (PID 675181 alive 8h56m, writing to stress_2000_v10_restart_1777561483.log, sessions resetting every 15 prompts through API block)
+- Write rate: 21% last 100 prompts (API gRPC/REST/GraphQL section — model generates long text explanations, expected low write rate)
+- SKIP rate: 63/170 = 37% in current session (elevated vs ~8% baseline; API prompts cause long model output that delays TUI readiness; FORCE-RESETs are recovering correctly)
+- Admiral last 30 min: struggle:none at 23:47–23:48 UTC (model exploring gRPC structure without writing), all known patterns, no new failure modes
+- vLLM 400s: 0; llm_balancer on :8001 healthy; vLLM gemma4 on :8000 healthy; no exceptions in drydock.log; GH issues: 0 open
+- Action this tick: no drydock bugs found; v2.7.26 deployed with both fixes (task-tool nudge + successful_test_runs reset); high SKIP rate is API-section harness timing, not a source regression; no fix committed
+
+## 2026-05-01 01:20 UTC tick
+- Stress: 857/1658 (PID 675181 alive, etime=9h55m, resumed from step 679; log /tmp/stress_2000_v10_restart_1777561483.log)
+- Write rate: 21% last 100 prompts (API/WebSocket/gRPC/GraphQL section — model generates long text explanations for network-server prompts; expected low write rate)
+- SKIP rate: 76/178 = 43% in current session (elevated — TUI stuck on 300s bash timeouts from WebSocket server tests; FORCE-RESET recovering correctly; harness progressing)
+- Admiral last 30 min: empty_after_tool:bash recurring (~8 fires since 22:19 UTC); loop:bash at 01:00 UTC (fuser -k 8765/tcp WebSocket server loop); all handled by existing canned interventions, sessions recover
+- vLLM 400s: 0; llm_balancer on :8001 healthy (PID 713929, different from resume.md PID 1230765 — was restarted by keepalive cron); vLLM gemma4 on :8000 healthy; GH issues: 0 open
+- Action this tick: investigated empty_after_tool:bash pattern (recurring ~8x in 3h); traced to _ensure_assistant_after_tools filler being detected by admiral detector at start of new user turns — not a real stall, just housekeeping filler; sessions recover without code change needed. No new drydock bugs found; v2.7.26 current (ships both task-tool nudge + successful_test_runs reset fixes). No commits.
+
+## 2026-05-01 02:31 UTC tick
+- Stress: 871/1658 (PID 675181 alive, etime=11h25m; step 871 "Test: concurrency" section, +1 write — past the API/gRPC/WebSocket SKIP cluster)
+- Write rate: 22% last 100 prompts (API/server section — expected; step 871 showed +1 write, transitioning back to test prompts)
+- SKIP rate: 64/192 = 33% in this restart run (elevated but consistent with prior ticks for API section; FORCE-RESETs recovering correctly)
+- Admiral last 30 min: 2 fires (empty_after_tool:read_file at 02:02, empty_after_tool:bash at 02:22) — all known patterns, handled by canned interventions
+- vLLM 400s: 0; llm_balancer PID 713929 on :8001 healthy; admiral_probe PID 4075121 on :8878 healthy; vLLM gemma4 on :8000 healthy; GH issues: 0 open
+- Action this tick: no new drydock bugs found; harness progressing normally through test-prompts section after API cluster; both v2.7.26 fixes active (task-tool nudge + successful_test_runs reset); no commits
+
+## 2026-05-01 02:35 UTC tick
+- Stress: 861/1658 (PID 675181 alive, etime=10h25m; currently stuck at step 861 with SKIP retries; 57/182 SKIPs = 31% in this restart run)
+- Write rate: 21% last 100 prompts (API/rate-limiter section — long conceptual text responses, expected)
+- SKIP rate elevated: TUI (PID 675184) is processing turn 11 on remote vLLM (192.168.50.21:8000) — request has been in-flight ~56 min; remote host confirmed alive (0.6s latency, test inference returns in <15s); vLLM shows 1 running request now. Likely large context at step 861 causing very long generation on remote host. Local vLLM idle since 01:34 UTC.
+- vLLM 400s: 0; llm_balancer PID 713929 on :8001 round-robin healthy; both hosts up; orphan WebSocket server PID 767439 on :8765 present (from prior session, not impacting current run)
+- GH issues: 0 open
+- Action this tick: no drydock bugs found. Admiral last fires at 01:23 UTC (loop:bash fuser-k-8765, retry_after_error:bash — all known patterns, canned responses). Remote host processing long-running turn is benign — 300s socket timeout in balancer will recover if remote stalls. No commits.
+
+## 2026-05-01 03:02 UTC tick
+- Stress: 874/1658 (PID 675181 alive, etime=11h56m, writing to /tmp/stress_2000_v10_restart_1777561483.log; step 874 "Test: rollback test for I" with retry in progress)
+- Write rate: 23% last 100 prompts (test/API/concurrency section — model generates long text explanations, few file writes; expected low rate in this prompt cluster)
+- SKIP rate: 92/220 = 42% in current restart run (elevated but consistent with prior ticks for this section; FORCE-RESETs recovering correctly; harness progressing)
+- Admiral last 30 min: last fire at 01:48 UTC (empty_after_tool:task), quiet since — all known patterns, no new failure modes
+- vLLM 400s: 0; llm_balancer on :8001 healthy; vLLM gemma4 container up 7 days; no errors in today's drydock.log; GH issues: 0 open
+- Action this tick: no new drydock bugs found; v2.7.26 current; harness progressing normally through test-prompt cluster; no commits
+
+## 2026-05-01 03:36 UTC tick
+- Stress: 878/1658 (PID 675181 alive, etime=12h25m; log /tmp/stress_2000_v10_restart_1777561483.log)
+- Write rate: 25% last 100 prompts (test-writing prompt cluster — "Test: golden test for K" had TIMEOUT with 101 msgs, 15 writes)
+- SKIP rate: 67/196 = 34% in this restart run (consistent with prior ticks; 874 and 878 needed retries; 875-877 completed cleanly)
+- Admiral last 30 min: 2 fires (retry_after_error:write_file at 03:18 and 03:30 — truncated-history template; escalation logic in format.py already handles with file-content embed and escalating directives; no new failure modes)
+- vLLM 400s: 0; llm_balancer PID 713929 on :8001 healthy; vLLM gemma4 container up; raw_md=0 (TUI rendering clean); GH issues: 0 open
+- Action this tick: was reading wrong log file (/tmp/stress_2000_1777408317.log, stale); confirmed actual log via lsof; no drydock bugs found; v2.7.26 current; no commits
+
+## 2026-05-01 04:33 UTC tick
+- Stress: 890/1658 (PID 675181 alive, etime=13h26m, resuming from step 679 in /tmp/stress_2000_v10_restart_1777561483.log; 211 steps done since restart at ~15.7 steps/hr)
+- Write rate: 20% last 100 prompts (25% overall in current restart run — "Test: ..." and "API: ..." prompt cluster; 0-write results expected when model runs existing tests rather than writing new files)
+- SKIP rate: 67/211 = 32% (consistent with prior ticks; FORCE-RESETs recovering; harness progressing; elevated but not new pattern)
+- Admiral last fires: 04:05 UTC retry_after_error:write_file (truncated-history template, escalation logic in format.py handling), 04:19 UTC empty_after_tool:ralph_repo_index — all known patterns, no new failure modes
+- vLLM 400s: 0; llm_balancer PID 713929 on :8001 healthy (round-robin returns models list); vLLM gemma4 container up; GH issues: 0 open; v2.7.26 current
+- Action this tick: no new drydock bugs found; admiral shows same known patterns as prior ticks; all services healthy; no commits
+
+## 2026-05-01 05:21 UTC tick
+- Stress: 898/1658 (PID 675181 alive, etime=14h55m, resuming from step 679 in /tmp/stress_2000_v10_restart_1777561483.log; 219 steps done since restart)
+- Write rate: 23% last 100 prompts (Testing section — "Test: unit test", "Test: integration test" prompts; model runs existing tests rather than writing files; expected low write rate)
+- SKIP rate: 68/219 = 31% in this restart run (consistent with prior ticks; Test-section SKIP cluster around steps 863-869 recovered after session reset; steps 893-898 completing cleanly)
+- Admiral last 30 min: loop:search_replace at 04:46 and 04:53 UTC (<<<SEARCH loop on GrepTool and instruction block — both known patterns, canned interventions firing correctly); no new failure modes since 04:53 UTC
+- vLLM 400s: 0; llm_balancer PID 713929 on :8001 healthy (confirmed /data3/drydock/scripts/llm_balancer.py); admiral_probe PID 4075121 on :8878 healthy; vLLM gemma4 on :8000 healthy; GH issues: 0 open
+- Action this tick: no new drydock bugs found; all services healthy; retry_after_error:write_file (truncated-history template) pattern continues at ~3/hr during bad sessions — escalation logic in format.py already handles; no fix warranted. v2.7.26 current. No commits.
+
+## 2026-05-01 06:00 UTC tick
+- Stress: 918/1658 (PID 675181 alive, etime=14h55m, resuming from step 679; 239 steps done since last restart; current prompt cluster: "Test:" idempotency/rollback/snapshot/race/golden)
+- Write rate: 23% last 100 prompts (low but expected — "Test:" prompt block exercises existing code rather than writing new files)
+- vLLM 400s: 0; llm_balancer PID 713929 on :8001 healthy; vLLM gemma4 container up 7 days; GH issues: 0 open; v2.7.26 current
+- Action this tick: no new drydock bugs found; all services healthy; stress run progressing normally; no commits
+
+## 2026-05-01 06:35 UTC tick
+- Stress: 928/1658 (PID 675181, alive, 15h25m elapsed; currently in "Test:" prompt block — memory benchmark, concurrency, race condition prompts)
+- Write rate: 23% last 100 prompts (expected low — these are "Test:" prompts that run bash test files; no writes needed)
+- Admiral last 30 min: struggle:none x18, struggle:search_replace x12, empty_after_tool:bash x7 — bulk of struggle:none was false positives from test prompts accumulating tool calls across turns
+- vLLM 400s: 0
+- GH issues: 0 open
+- Action this tick: committed fix(admiral): reset struggle counter on each user turn — detect_struggle was accumulating calls_since_write across all messages, so 5 test prompts × 8 bash calls = 40 non-write calls triggered struggle:none every minute even when model was working correctly. Counter now resets on each user message. 2 regression tests added.
+
+## 2026-05-01 07:02 UTC tick
+- Stress: 938/1658 (PID 675181, alive 15:57 elapsed; TUI PID 675184 active; log /tmp/stress_shakedown_1777561484.tui.log growing normally — timezone CDT vs UTC caused false "log stalled" alarm)
+- Write rate: 22% last 100 prompts (expected: current prompt block is "Test: unit test / integration test / fuzz test / smoke test" vague prompts; model makes 20-31 tool calls reading files but rarely writes; admiral fires ~1/min "no writes" on these)
+- Admiral last 30 min: struggle:none firing repeatedly (~1/min) on "Test:" prompts — model explores deeply (20-31 tool calls) without writing test files; not a drydock code bug, model behavioral on vague test requests
+- vLLM 400s: 0
+- GH issues: 0 open
+- Action this tick: no fix committed — harness healthy, no actionable drydock bug found; write rate drop 74%→22% is prompt-category effect (test prompts vs build prompts), not regression
+
+## 2026-05-01 08:01 UTC tick
+- Stress: 960/1658 (PID 675181 alive, 16h55m elapsed; log /tmp/stress_2000_v10_restart_1777561483.log updating normally; prompt 960 "Test: golden test for K" in-flight)
+- Write rate: 20% last 100 prompts (expected — current block is "Test:" prompts, model responds with clarifying text or bash-only exploration; no file writes needed)
+- Admiral last 30 min: struggle:none fired repeatedly from 07:00-07:19 UTC on previous session (22-24 tool calls w/o write); session reset at 07:19 cleared it; new session quiet since
+- vLLM 400s: 0
+- GH issues: 0 open
+- Action this tick: no new drydock bugs found; all services healthy; no commits
+
+## 2026-05-01 08:31 UTC tick
+- Stress: 966/1658 (PID 675181 alive, 17h25m elapsed; log /tmp/stress_2000_v10_restart_1777561483.log; currently at "Test: smoke test for CLI C" prompt cluster)
+- Write rate: 23% last 100 prompts (expected — current block is "Test:" prompts; model reads and runs bash rather than writing files; 4 TIMEOUT events from integration/property test prompts with 58-84 msgs each)
+- Admiral last 30 min: single retry_after_error:search_replace:file at 08:09 UTC (model retried a failed search_replace on test_integ path; canned intervention fired correctly); no struggle:none false positives in current session (admiral counter reset fix 71cb046 working as intended)
+- vLLM 400s: 0; llm_balancer PID 713929 on :8001 healthy; vLLM gemma4 container up; GH issues: 0 open; v2.7.26 current
+- Action this tick: no new drydock bugs found; all services healthy; no commits
+
+## 2026-05-01 08:30 UTC tick
+- Stress: 977/1658 (PID 675181 alive, 18h elapsed; log /tmp/stress_2000_v10_restart_1777561483.log; session reset at 975, now on "Test:" prompt cluster)
+- Write rate: 19% last 100 prompts (expected — entire range is "Test: unit/integration/fuzz/property/regression test" prompts; model runs bash to explore, rarely writes files; this matches prompt-type not a regression)
+- Admiral last 30 min: 1 retry_after_error:write_file (truncated history, model retried same call once then self-corrected); 2 retry_after_error:search_replace; 1 loop:search_replace (BaseTool class replacement); 107 interventions total today, all normal patterns
+- vLLM 400s: 0; GH issues: 0 open; 1 pending commit (71cb046 fix(admiral): reset struggle counter per user turn) ships at next 12:00 UTC auto-release as v2.7.27
+- Action this tick: no new drydock bugs found; stress healthy and progressing; no commits
+
+## 2026-05-01 10:01 UTC tick
+- Stress: 984/1658 (PID 675181 alive, 19h elapsed; babysitter reporting done=229 skip=72 at 10:00 UTC tick)
+- Write rate: 17% last 100 prompts (slight dip from 19% prior; same "Test:" prompt cluster, expected low-write range)
+- Admiral last 30 min: 6x struggle:search_replace (model making 20-28 read-only calls before attempting edits; count reaching 28 in one session), 2x empty_after_tool:bash; all canned interventions; no new unknown patterns
+- vLLM 400s: 0; llm_balancer healthy (PID 713929 on :8001, forwarding confirmed); GH issues: 0 open
+- TUI session session_20260501_093430_40e5bc06 stuck in search_replace struggle loop for ~30min; harness FORCE-RESET (ESC+/clear) cycling; session last modified 09:58 UTC so still alive; harness will auto-recover
+- Action this tick: no actionable drydock bugs found; all failure patterns are known model-behavior (struggle:search_replace, empty_after_tool); no commits
+
+## 2026-05-01 11:01 UTC tick
+- Stress: 998/1658 (60% complete), PID 675181 alive ~20h, log /tmp/stress_2000_v10_restart_1777561483.log, progressing on "Test:" prompt cluster
+- Write rate: 17% last 100 prompts (expected — test prompts where model reads/runs bash rather than writing files)
+- Admiral last 30 min: loop:bash (k_anonymity_golden_test.py repeated 3x, 10:28-10:30), struggle:none (22 tool calls w/o write, 7 interventions 10:54-11:01); all canned interventions firing; patterns are known model-behavior
+- vLLM 400s: 0; llm_balancer PID 713929 on :8001 healthy; vLLM gemma4 container up; GH issues: 0 open
+- Action this tick: no new drydock bugs found; all services healthy; no commits
+
+## 2026-05-01 11:30 UTC tick
+- Stress: 1006/1658 (61% complete), PID 675181 alive 20h+ (--resume-from-step 679 active restart log /tmp/stress_2000_v10_restart_1777561483.log)
+- Write rate: 14% last 100 prompts (expected — model in "Test:" prompt block; running bash for test prompts, no file writes)
+- Admiral last 30 min: 20x struggle:none (model making bash calls for test prompts without writing, each turn 30+ tool calls), 1x empty_after_tool:ralph_repo_index (hallucinated tool, already handled); all known patterns, no new types
+- vLLM 400s: 0; llm_balancer PID 713929 on :8001 healthy (verified via /v1/models); vLLM gemma4 container up; GH issues: 0 open
+- Action this tick: no actionable drydock bugs found; system healthy; no commits
+
+## 2026-05-01 13:03 UTC tick
+- Stress: 1027/1658 (62% complete), PID 675181 alive ~27h, writing to /tmp/stress_2000_v10_restart_1777561483.log (stress_log_path.txt is stale — still points to old Apr 26 log, not a drydock bug)
+- Write rate: 10% last 100 prompts (expected — harness in "Test:" prompt cluster; model runs bash to verify code rather than writing files)
+- Admiral last 2h: 19 interventions, all known patterns (struggle:none, struggle:search_replace, retry_after_error:search_replace, empty_after_tool:bash); no new unknown types
+- vLLM 400s: 0; llm_balancer PID 713929 on :8001 healthy (verified /v1/models); vLLM gemma4 container up; GH issues: 0 open
+- Action this tick: no actionable drydock bugs found; all services healthy; no commits
+
+## 2026-05-01 14:30 UTC tick
+- Stress: 1041/1658 (62.8%), PID 675181 alive ~23h, idx at 1041 per babysitter; 280 done, 78 skip, 4 timeout
+- Write rate: 10% last 100 (expected — harness deep in "Test:" prompt cluster; these prompts run tests not build files); 18% overall in current log
+- Admiral last 30 min: 29 interventions, all struggle:none (model repeatedly making 30-42 bash calls without writing on test prompts; loop_detection nudges but doesn't stop — expected behavior); 1x loop:bash at 13:56 UTC (python3 race_condition_test.py repeated 3x)
+- vLLM 400s: 0; llm_balancer PID 713929 on :8001 healthy; vLLM gemma4 container up; GH issues: 0 open
+- Action this tick: committed fix(loop-detection): new bash exploration loop check catches same bash command 5+ times in last 20 tool calls (alternating bash/read_file pattern). 4 regression tests pass; 40 total tests pass. Will ship in next auto-release.
+
+## 2026-05-01 14:31 UTC tick
+- Stress: 1051/1658 (63.4%), PID 675181 alive ~23.5h, log /tmp/stress_2000_v10_restart_1777561483.log; harness in "Test:" prompt cluster (Test: concurrency/race-condition/idempotency/rollback/snapshot prompts)
+- Write rate: 9% last 100 prompts (expected — model runs bash to execute existing test code rather than write new files; same test-prompt behavior as prior ticks)
+- Admiral last 5 min: continuous struggle:none starting 14:12 UTC; model at 36 bash calls on single session with 0 writes; canned interventions firing every ~1 min; no new unknown patterns
+- vLLM 400s: 0; llm_balancer PID 713929 on :8001 healthy; vLLM gemma4 container up 7 days; GH issues: 0 open
+- Action this tick: no new drydock bugs; 1 pending commit (cc0c467 loop-detection bash/read pattern) will ship at 18:00 UTC auto-release as v2.7.28; all services healthy
+
+## 2026-05-01 15:20 UTC tick
+- Stress: 1057/1658 (PID 675181, alive 23h55m, log stress_2000_v10_restart_1777561483.log)
+- Write rate: 11% last 100 prompts — current block is test-analysis prompts ("Test: snapshot/race-condition/idempotency for X"); model correctly runs existing tests and summarizes without writing; expected for this prompt type
+- SKIPs: 78 total (7.4% skip rate); concentrated in steps 800-899 (42 SKIPs); FORCE-RESET handling working; skip rate improved to 3/50 in steps 1000-1049
+- Admiral last 2h: 53 struggle:none (model exploring 35+ calls before analyzing test prompts — expected), 11 struggle:write_file (search_replace retry loops), 1 retry_after_error:bash; all known patterns, no new unrecognized patterns
+- vLLM 400s: 0
+- GH issues: 0 open
+- cc0c467 unreleased (bash-exploration loop detection); auto_release will ship as v2.7.28 at 18:00 UTC
+- Notable observation: "progressive grep funnel" pattern in session — model runs same grep 7-8 times with incremental | grep -v additions; not caught by cc0c467 (requires exact match); minor inefficiency, not blocking
+- Action this tick: no action — healthy. Low write rate is prompt-type artifact, not a drydock regression. Stress progressing normally.
+
+## 2026-05-01 15:32 UTC tick
+- Stress: 1066/1658 (64.3%), PID 675181 alive ~24h30m; log /tmp/stress_2000_v10_restart_1777561483.log; active prompt "Doc: changelog entry for E"; harness self-restored after brief v8/v9 internal restarts (false alarm — main process never died)
+- Write rate: 11% last 100 prompts (expected — in "Doc:" prompt block; model responds to documentation prompts with text or raw_md, not file writes)
+- Admiral last 30 min: continuous struggle:write_file (15:06–15:25 UTC), 59 tool calls without write, same directive firing every 60s — model ignoring advisory nudges; known model behavior per CLAUDE.md #2, not a new drydock bug
+- vLLM 400s: 0; llm_balancer PID 713929 on :8001 healthy (verified /v1/models); vLLM gemma4 container up; GH issues: 0 open
+- Action this tick: no actionable drydock bugs found; 1 unreleased commit (cc0c467 bash/read loop detection) will ship as v2.7.28 at 18:00 UTC auto-release; all services healthy; no commits
+
+## 2026-05-01 16:33 UTC tick
+- Stress: 1083/1658 (65.3% complete), PID 675181 alive 25h+, log /tmp/stress_2000_v10_restart_1777561483.log; harness in "Doc:" prompt block (Doc: changelog/release-notes/architecture-diagram/README prompts)
+- Write rate: 8% last 100 prompts (expected — model responds with text to documentation prompts; after session reset at step 1080 writes resumed at +5 and +7 per prompt; prompt-type artifact not regression)
+- Admiral last hour: 0 new interventions since 15:25 UTC; admiral probe confirmed alive (PID 4075121, 4+ days uptime; resume.md had stale PID 2251231); current sessions not triggering patterns
+- vLLM 400s: 0; llm_balancer PID 713929 on :8001 healthy; vLLM gemma4 container up; GH issues: 0 open
+- 2 unreleased commits (8ddf09d search_replace-dominates-file FORCE_STOP, cc0c467 bash/read alternating loop) will ship as v2.7.28 at 17:00 UTC auto_release
+- Action this tick: no new drydock bugs found; all services healthy; no commits

{drydock_cli-2.7.26 → drydock_cli-2.7.28}/uv.lock

@@ -326,7 +326,7 @@ wheels = [
 
 [[package]]
 name = "drydock-cli"
-version = "2.7.25"
+version = "2.7.27"
 source = { editable = "." }
 dependencies = [
     { name = "agent-client-protocol" },

drydock_cli-2.7.26/.auto_release.lock

@@ -1 +0,0 @@
-749372