drydock-cli 2.8.9__tar.gz → 2.8.11__tar.gz

drydock_cli-2.8.11/.auto_release.lock

@@ -0,0 +1 @@
+201506

{drydock_cli-2.8.9 → drydock_cli-2.8.11}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.8.9
+Version: 2.8.11
 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-2.8.9 → drydock_cli-2.8.11}/drydock/core/agent_loop.py

@@ -1080,7 +1080,8 @@ class AgentLoop:
                         and assistant_msg.tool_calls):
                     prev_tool_name = assistant_msg.tool_calls[-1].function.name if assistant_msg.tool_calls[-1].function else None
             _readonly_tools = {"read_file", "grep", "glob", "ls", "pwd",
-                               "ralph_repo_index", "retrieve", "search_files"}
+                               "ralph_repo_index", "ralph_file_summary",
+                               "retrieve", "search_files"}
             _write_tools = {"write_file", "search_replace"}
             _prev_was_read = prev_tool_name in _readonly_tools
             _prev_was_write = prev_tool_name in _write_tools
@@ -1127,6 +1128,27 @@ class AgentLoop:
                 and "file" in _prev_tool_result
                 and "changed" in _prev_tool_result
             )
+            # Detect bash that returned an error/traceback.
+            # These stalls fire as empty_after_tool:bash :: source=canned — the
+            # generic note doesn't tell the model what to DO with the error.
+            _prev_bash_had_error = (
+                prev_tool_name in ("bash", "run_command")
+                and not _prev_bash_nothing_to_commit
+                and not _prev_bash_commit_succeeded
+                and bool(_re.search(
+                    r"(Error|error|Traceback|FAILED|exit code [1-9]|command not found)",
+                    _prev_tool_result
+                ))
+            )
+            # Detect bash that returned non-empty output without an error.
+            # Model stalls instead of using the output to write or fix code.
+            _prev_bash_had_output = (
+                prev_tool_name in ("bash", "run_command")
+                and not _prev_bash_nothing_to_commit
+                and not _prev_bash_commit_succeeded
+                and not _prev_bash_had_error
+                and bool(_prev_tool_result.strip())
+            )
             if _stall_attempt == 0:
                 if _prev_write_path_error:
                     note = (
@@ -1142,6 +1164,13 @@ class AgentLoop:
                         "or grep(pattern='...') to search content. "
                         "Do NOT send an empty response."
                     )
+                elif prev_tool_name == "ralph_repo_index":
+                    note = (
+                        "You indexed the repository but produced no output. "
+                        "Now write a text answer to the user's question, "
+                        "or call read_file to inspect a specific file. "
+                        "Do NOT call ralph_repo_index again."
+                    )
                 elif _prev_was_read:
                     _tool_name_str = prev_tool_name or "read_file"
                     note = (
@@ -1167,6 +1196,18 @@ class AgentLoop:
                         "Respond with a short summary of what you changed and stop. "
                         "Do NOT run git add or git commit again."
                     )
+                elif _prev_bash_had_error:
+                    note = (
+                        "The command returned an error. Read the error message above, "
+                        "then fix the code with search_replace or write_file, or try a "
+                        "different command. Do NOT re-run the same failing command."
+                    )
+                elif _prev_bash_had_output:
+                    note = (
+                        "The command ran and returned output. Use that output now: "
+                        "write or update code files, fix any issues shown, or respond "
+                        "to the user. Do NOT re-run the same command."
+                    )
                 else:
                     note = (
                         "Continue working. Use a tool (write_file, "
@@ -1174,7 +1215,14 @@ class AgentLoop:
                         "your plan in text."
                     )
             elif _stall_attempt == 1:
-                if _prev_was_read:
+                if prev_tool_name == "ralph_repo_index":
+                    note = (
+                        "You sent an empty response after indexing the repository. "
+                        "Respond in TEXT now — answer the user's question directly, "
+                        "or state in one sentence why you cannot proceed. "
+                        "Do NOT call ralph_repo_index again."
+                    )
+                elif _prev_was_read:
                     _tool_name_str = prev_tool_name or "read_file"
                     note = (
                         f"You sent an empty response after calling {_tool_name_str}. "
@@ -1182,6 +1230,27 @@ class AgentLoop:
                         f"what you read — OR state in one sentence why you "
                         f"cannot proceed. Do NOT call {_tool_name_str} again."
                     )
+                elif _prev_was_write:
+                    _tool_name_str = prev_tool_name or "write_file"
+                    note = (
+                        f"You sent an empty response after {_tool_name_str}. "
+                        "Write the NEXT file in your plan NOW with write_file, "
+                        "or run bash to test what you have built. "
+                        "Do NOT send another empty response."
+                    )
+                elif _prev_bash_had_error:
+                    note = (
+                        "Second empty response after a bash error. "
+                        "Fix the error NOW with search_replace or write_file. "
+                        "Do NOT re-run the same failing command."
+                    )
+                elif _prev_bash_had_output:
+                    note = (
+                        "Second empty response after bash output. "
+                        "Use the output now — write code with write_file, "
+                        "fix issues with search_replace, or respond to the user. "
+                        "Do NOT re-run the same command."
+                    )
                 else:
                     note = (
                         "You sent an empty response. Call a tool now "
@@ -1189,12 +1258,50 @@ class AgentLoop:
                         "OR explicitly say you are done with this task."
                     )
             else:
-                note = (
-                    "You have sent 3 empty responses in a row for "
-                    "this user request. Respond with either (a) a "
-                    "tool call to make progress, or (b) one "
-                    "sentence explaining why you cannot proceed."
-                )
+                if prev_tool_name == "ralph_repo_index":
+                    note = (
+                        "THIRD empty response after ralph_repo_index. "
+                        "Stop — write a text reply to the user RIGHT NOW. "
+                        "If you cannot answer, say 'I was unable to find that information' "
+                        "and stop. Do NOT call ralph_repo_index or any other tool."
+                    )
+                elif prev_tool_name in _readonly_tools:
+                    _tool_name_str = prev_tool_name or "read_file"
+                    note = (
+                        f"THIRD empty response after {_tool_name_str}. "
+                        "Stop reading — call write_file or search_replace NOW to apply "
+                        "the change, or respond in text explaining why you cannot proceed. "
+                        f"Do NOT call {_tool_name_str} again."
+                    )
+                elif _prev_was_write:
+                    _tool_name_str = prev_tool_name or "write_file"
+                    note = (
+                        f"THIRD empty response after {_tool_name_str}. "
+                        "Call bash NOW to run the tests or verify what you built, "
+                        "or write the next required file. "
+                        "Do NOT send another empty response."
+                    )
+                elif _prev_bash_had_error:
+                    note = (
+                        "THIRD empty response after a bash error. "
+                        "Fix the error NOW — call search_replace or write_file to "
+                        "correct the broken code, or respond in text with what went wrong. "
+                        "Do NOT run the same failing command again."
+                    )
+                elif _prev_bash_had_output:
+                    note = (
+                        "THIRD empty response after bash output. "
+                        "Act on the output NOW — write or fix code with write_file "
+                        "or search_replace, or respond in one sentence. "
+                        "Do NOT re-run the same command."
+                    )
+                else:
+                    note = (
+                        "You have sent 3 empty responses in a row for "
+                        "this user request. Respond with either (a) a "
+                        "tool call to make progress, or (b) one "
+                        "sentence explaining why you cannot proceed."
+                    )
             self._inject_system_note(note)
             logger.info(
                 "Empty-response stall (inline retry %d/%d, prev=%s)",

{drydock_cli-2.8.9 → drydock_cli-2.8.11}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.8.9"
+version = "2.8.11"
 description = "Local-first CLI coding agent — tested with Gemma 4 26B via vLLM"
 readme = "README.md"
 requires-python = ">=3.12"

{drydock_cli-2.8.9 → drydock_cli-2.8.11}/trip_log.md

@@ -1,5 +1,167 @@
 # Drydock Trip Log
 
+## 2026-05-10 11:12 UTC tick
+- Stress: PID 3875067 alive (1d 15h runtime), ~80% through tool_agent 2000-prompt suite
+- Write rate: degraded — 19 SKIPs in last 29 prompts prior to 10:18 UTC TUI recycle; recovering after recycle
+- Admiral last 30 min: 5 harness:thinking_stall fires (ralph_repo_index stalls 10:14-10:17), plus 3 TUI-recycle-requested stress-actions (09:52, 10:05, 10:18) and retry-spike alert (145% retry rate)
+- vLLM 400s: ~20 seen in balancer.log (context overflow bursts), resolved after TUI recycle at 10:18; balancer healthy now
+- GH issues: 0 open
+- Dispatch queue: harness=88511, retrieval=84 (all already consumed), steering=0
+- Action this tick: no fix — skip spike caused by context-overflow 400s from long sessions; AdmiralWorker recycled TUI at 10:18 and system is recovering; all harness:thinking_stall fixes (427a631, 0820952, 9f8d022) are in source awaiting next auto_release at 12:00 CDT; retrieval drain found nothing new
+
+## 2026-05-10 10:12 UTC tick
+- Stress: PID 3875067 alive (1d 14h runtime), sessions active at 09:29-09:30 UTC; stress run was at 1331/1658 at 09:00 UTC
+- Write rate: ~11 prompts/hr (steady)
+- Admiral last 30 min: harness:thinking_stall continuing to fire (88K queue total, all same pattern)
+- vLLM 400s: 0 (balancer healthy, 8000/8001 both OK)
+- GH issues: 0 open
+- Dispatch queue: harness=88057, retrieval=78 (all already consumed), steering=0
+- Action this tick: no action — all harness:thinking_stall patterns addressed by commits (427a631, 0820952, 9f8d022) shipped today; retrieval drain found nothing new to ingest; system healthy
+
+## 2026-05-10 09:12 UTC tick
+- Stress: 1331/1658 (PID 3875067 alive, 1d 13h runtime); done=250, skip=273, recycle=172; babysitter confirmed healthy through 09:00 UTC tick
+- Write rate: ~11 prompts/hr last hour (3 done, 8 skip, 5 recycle)
+- Admiral last 30 min: harness:thinking_stall dominant — 279 ralph_repo_index, 70 bash, 28 read_file/write_file in last hour; all stall sub-patterns covered by commits in current release queue (9f8d022 atm2 ralph/read, 0820952 write/sr atms 1-2, 427a631 bash atms 1-2), shipping at next auto_release ~11:00 UTC
+- vLLM 400s: 0; balancer responding on :8001 (1 model), vLLM on :8000 up
+- GH issues: 0 open
+- Dispatch queue: harness=87834, retrieval=75 (0 actionable — all consumed), steering=0
+- retrieval-drain: 0 projects ingested (all previously consumed)
+- Action this tick: no action — 3 unshipped stall fixes pending next auto_release (~11:00 UTC); system healthy; no new unaddressed patterns
+
+## 2026-05-10 08:52 UTC tick
+- Stress: 1320/1658 (PID 3875067 alive, 1d 13h runtime); done=247, skip=265, recycle=167; babysitter healthy, last tick 08:00 UTC
+- Write rate: ~14 prompts/hr last hour (8 done, 5 skip, 1 recycle)
+- Admiral last 30 min: harness:thinking_stall still dominant (87621 harness queue total); all recent stall types (bash, write_file, search_replace, ralph_repo_index, read_file) addressed by commits 427a631/0820952/9f8d022 shipped in v2.8.11+; no new unaddressed patterns observed
+- vLLM 400s: 0 (balancer on :8001 showing HTTP 400 502s — context overflow, not model crash); llamacpp-gemma4 container up and responding on :8000 and :8001
+- GH issues: 0 open
+- Dispatch queue: harness=87621, retrieval=74 (0 actionable — all consumed), steering=0
+- retrieval-drain: 0 projects ingested
+- Action this tick: no action — all queued patterns already addressed in 24h commits; stress run progressing steadily; system healthy
+
+## 2026-05-10 08:12 UTC tick
+- Stress: PID 3875067 alive (resume-from-step 807/1658 tool_agent); balancer OK on :8001; vLLM docker up (unhealthy status but responding, 0 400s in last 30m)
+- Write rate: N/A (mid-session)
+- Admiral last 30 min: harness:thinking_stall dominant (9890 today); harness:grep:unescaped_pattern 4014 today (model looping same grep call 5+ times, not invalid regex — advisory NOTE fires after 5 repeats); harness:loop:bash_generic 962; harness:tool:hallucinated_name 862
+- vLLM 400s: 0
+- GH issues: 0 open
+- Dispatch queue: harness=87403, retrieval=74 (0 actionable — all consumed), steering=0
+- retrieval-drain: 0 projects ingested
+- Action this tick: committed fix 427a631 — bash stall nudges at attempts 1 and 2 (addresses pattern harness:thinking_stall); bash error/output cases at attempts 1+2 fell through to generic else, now get specific fix/act-on-output nudges matching pattern used for write_file and read_file; auto_release will ship with prior 3 commits as v2.8.11
+
+## 2026-05-10 07:32 UTC tick
+- Stress: PID 3875067 alive (1d 12h runtime, resume-from-step 807/1658 tool_agent); vLLM 0 JSONDecodeErrors; balancer OK on :8001
+- Write rate: N/A (mid-session)
+- Admiral last 30 min: harness:thinking_stall dominant (57020 total in queue)
+- vLLM 400s: 0
+- GH issues: 0 open
+- Dispatch queue: harness=87171, retrieval=74 (all already consumed, 0 ingested)
+- Action this tick: no action — 2 commits unreleased since v2.8.10 (0820952 write_file/search_replace stall nudges at attempt 1+2, 9f8d022 ralph_repo_index+read_file at attempt 2); next auto_release will ship as v2.8.11; bash_generic (12896) and search_replace:not_found_loop (4798) remain high-volume but the advisory FORCE_STOP mechanism already handles them (loop detection removes bash from tools for 1 turn); no new actionable bug found this tick
+
+## 2026-05-10 07:11 UTC tick
+- Stress: PID 3875067 alive (1d 13h runtime, resume-from-step 807/1658); balancer healthy on :8001; vLLM 0 JSONDecodeErrors in last 30 min; docker gemma4 running
+- Write rate: N/A (mid-session)
+- Admiral last 30 min: harness:thinking_stall firing for ralph_repo_index/write_file/bash (all addressed by commits 0820952 + 9f8d022 pending next auto_release at ~11:00 UTC)
+- vLLM 400s: 0
+- GH issues: 0 open
+- Dispatch queue: harness=86968, retrieval=74 (0 actionable — all within 7-day re-ingest window), steering=0
+- retrieval-drain: 0 projects ingested (all already up to date)
+- Action this tick: no action — all top stall patterns addressed by today's commits (0820952 adds write_file/search_replace nudges at attempts 1+2; 9f8d022 adds ralph_repo_index/read_file at attempt 2); pending ship at next auto_release ~11:00 UTC; system healthy
+
+## 2026-05-10 06:32 UTC tick
+- Stress: PID 3875067 alive (1d 11h runtime, resume-from-step 807/1658); balancer on :8001 healthy; vLLM 0 JSONDecodeErrors
+- Write rate: N/A (session mid-flight)
+- Admiral last 30 min: harness:thinking_stall dominant; top stall tools from full queue: ralph_repo_index=30849, bash=9082, read_file=7136, write_file=5854, search_replace=2180
+- vLLM 400s: 0
+- GH issues: 0 open
+- Dispatch queue: harness=86765, retrieval=74 (0 actionable)
+- retrieval-drain: 0 projects ingested
+- Action this tick: committed fix — write_file and search_replace stalls at attempts 1 and 2 fell through to generic note (5854+2180 queue entries); added _prev_was_write branches at attempts 1 and 2 matching the ralph_repo_index/bash pattern; commit 0820952 (addresses pattern harness:thinking_stall)
+
+## 2026-05-10 06:02 UTC tick
+- Stress: PID 3875067 alive (resume-from-step 807/1658, 34h+ runtime); session_20260510_055041 active with 119 messages (57 assistant, 56 tool), messages.jsonl updated at 06:01 UTC — harness live and processing
+- Write rate: N/A (session mid-flight at tick time)
+- Admiral last 30 min: harness:thinking_stall=148, harness:loop:bash_generic=24, harness:grep:unescaped_pattern=3; ralph_repo_index dominant tool (102 stalls/30min), bash=15, read_file=9
+- vLLM 400s: 0 JSONDecodeErrors; balancer PID 3944578 healthy on :8001; simple completions and balancer respond correctly; balancer.log shows 502s for some context-overflow tool-call requests (expected — model stalls on long sessions); gemma4 docker logs clean
+- GH issues: 0 open
+- Dispatch queue: harness=86573 total (thinking_stall dominant, same pattern as prior ticks); retrieval=74 (0 actionable, all within 7-day window)
+- retrieval-drain: 0 projects ingested (all already ingested recently)
+- Action this tick: no fix committed — today's v2.8.10 release (commits 9f8d022, a6a448e, ee354e5, cd6b56c) already added specific stall nudges for ralph_repo_index at all 3 attempt levels, bash error/output cases, and ralph_file_summary; all top queue patterns are addressed; no new unaddressed patterns found; system healthy
+
+## 2026-05-10 03:32 UTC tick
+- Stress: PID 3875067 alive (resume-from-step 807/1658, 32h+ runtime); new TUI child PID 115006 just spawned at 22:29 CDT; harness actively recycling sessions (12146 total sessions)
+- Write rate: strained — admiral reports 91% retry rate and 14+ SKIP clusters in last 35 prompts; TUI recycle mechanism firing correctly
+- Admiral last 30 min: high skip/retry pattern; last fires at 03:29 UTC (bootstrap after recycle); dominant patterns: thinking_stall, loop:bash_generic
+- vLLM 400s: balancer showing 4905 total 502 errors (context overflow), balancer healthy on :8001 forwarding to gemma4 OK
+- GH issues: 0 open
+- Dispatch queue: harness=85708 (thinking_stall dominant, bash_generic secondary), retrieval=74 (0 actionable, all within 7-day window)
+- retrieval-drain: 0 projects ingested (all up to date)
+- Action this tick: no fix committed — ee354e5+cd6b56c (bash stall nudges + ralph_file_summary to readonly-tools) already address all top dispatch queue patterns and are pending next auto_release at 00:00 CDT (06:00 UTC); no new unaddressed patterns found
+
+## 2026-05-10 02:47 UTC tick
+- Stress: PID 3875067 alive (resume-from-step 807, running since May 8, 45+ hr runtime); recent sessions cycling through "Doc:" single-letter prompts; harness advancing
+- Write rate: N/A (stress sessions in Doc prompt segment)
+- Admiral last 30 min: harness:thinking_stall dominant (6,962 in last 24h); harness:grep:unescaped_pattern 3,981 (fix b3922fa in v2.8.6 deployed, queue entries are historical); loop:bash_generic 660
+- vLLM 400s: 0; llm_balancer PID 3944578 healthy on :8001 (gemma4 OK)
+- GH issues: 0 open
+- Dispatch queue: harness=85,384 (956 classifier batches), retrieval=74 (0 actionable, all within 7-day window)
+- Retrieval-drain: 0 projects ingested (74 entries all already ingested recently)
+- Action this tick: no fix committed — all top patterns (thinking_stall, grep:unescaped_pattern, heredoc_loop) addressed by commits ee354e5/cd6b56c/c8a5f7f/dd07921/b3922fa in last 24h; search_replace:not_found_loop at 114/24h, HARD-STOP logic already handles it; system healthy
+
+## 2026-05-10 01:32 UTC tick
+- Stress: PID 3875067 alive; current cycle at ~680/1658 (stress_2000 log tail); harness proceeding through API prompt segment; watcher active
+- Write rate: N/A (log-based estimate not available this tick)
+- Admiral last 30 min: harness:thinking_stall dominant; harness:loop:bash_generic secondary; all patterns have existing nudge coverage
+- vLLM 400s: 0; llm_balancer PID 3944578 healthy on :8001; no port squatters
+- GH issues: 0 open
+- Dispatch queue: harness=85,052, retrieval=74 (0 actionable; all within 7-day re-ingest window); steering=0
+- Retrieval-drain: 0 projects ingested (74 entries all already ingested recently)
+- Action this tick: no action — system healthy; recent commits cd6b56c + c8a5f7f already address dominant harness:thinking_stall pattern; harness:loop:bash_generic covered by existing FORCE_STOP + admiral nudges; no new actionable drydock bugs found
+
+## 2026-05-10 01:30 UTC tick
+- Stress: 1197/1658 (done=166, skip=223, recycle=134); PID 3875067 alive (30h runtime); high skip rate in "Doc:" prompt segment (single-letter shortcuts confuse model, causes bash loops → TUI approval modal stall → harness skips); harness is recycling TUI correctly and advancing
+- Write rate: all recent steps skipped (Doc segment, model inherently loops on ambiguous prompts)
+- Admiral last 30 min: harness:thinking_stall (55,104 cumulative); harness:loop:bash_generic (12,574); harness:tool:hallucinated_name (6,425); harness:search_replace:not_found_loop (4,798)
+- vLLM 400s: 0; llm_balancer PID 3944578 healthy on :8001; vLLM model count=1
+- GH issues: 0 open
+- Dispatch queue: harness=84,888, retrieval=74 (0 actionable); retrieval-drain: 0 projects ingested
+- Action this tick: no action — system healthy; recent commits cd6b56c + c8a5f7f address dominant thinking_stall pattern; bash_generic covered by existing FORCE_STOP + admiral; no new actionable drydock bugs found
+
+## 2026-05-10 01:05 UTC tick
+- Stress: active; PID 3875067 alive (29h runtime, etimes=104563, resume-from-step 807); latest session session_20260510_002909_d671f1bd started 19:30 local; harness cycling normally
+- Write rate: N/A (not exposed in live log)
+- Admiral last 30 min: harness:thinking_stall dominant (54,976 cumulative); harness:loop:bash_generic (12,544); harness:tool:hallucinated_name (6,422); harness:search_replace:not_found_loop (4,798); harness:grep:unescaped_pattern (3,969)
+- vLLM 400s: 0; llm_balancer PID 3944578 healthy on :8001 (confirmed legit); docker serving gemma4 normally
+- GH issues: 0 open
+- Dispatch queue: harness=84,724, retrieval=74 (0 actionable, all within 7-day window); retrieval-drain: 0 projects ingested
+- Action this tick: no action — system healthy; 1 unreleased commit cd6b56c (ralph_file_summary stall nudge) pending next auto_release; all major patterns (thinking_stall, grep:unescaped_pattern, search_replace:not_found_loop, hallucinated_name) addressed by commits in last 48h; bash_generic (12K) handled by existing FORCE_STOP + admiral, inherent model limitation
+
+## 2026-05-10 00:04 UTC tick
+- Stress: 1189/1658 (stress_2000_v10, log updated 1 min ago); PID 3875067 (stress_shakedown.py, 28.5h runtime from step 807) also alive
+- Write rate: not recalculated (doc-heavy segment ongoing — architecture diagrams, release notes, changelog entries driving high SKIP rate)
+- Admiral last 30 min: harness:thinking_stall dominant (54K cumulative), harness:loop:bash_generic (12K), harness:tool:hallucinated_name (6.4K), harness:search_replace:not_found_loop (4.8K); all covered by existing fixes
+- vLLM 400s: 0; llm_balancer PID 3944578 confirmed healthy on :8001 (not an orphan); no :8001 squatting
+- GH issues: 0 open
+- Dispatch queue: harness=84547, retrieval=74 (0 actionable, all within 7-day window); retrieval-drain: 0 projects ingested
+- Action this tick: no action — all patterns addressed by commits from last 24h (cd6b56c, c8a5f7f, b3922fa, dd07921); ralph_repo_index is top stall trigger (29K queue) but already in _readonly_tools since prior fix; system healthy
+
+## 2026-05-09 23:33 UTC tick
+- Stress: 1183/1658; PID 3875067 alive; done=165, skip=205, recycle=125 (babysitter shows step 1183 active)
+- Write rate: ~44.6% (165/370 non-timeout prompts written)
+- Admiral last 30 min: 0 fires (classify_pulse log empty for this window)
+- vLLM 400s: 0; llm_balancer PID 3944578 healthy on :8001 (confirmed legit)
+- GH issues: 0 open
+- Dispatch queue: harness=84359, retrieval=74 (0 actionable); retrieval-drain: 0 ingested (all within 7-day window)
+- Action this tick: no action — all queued patterns (harness:thinking_stall, harness:loop:bash_generic) addressed by recent commits cd6b56c + c8a5f7f + 2cfa973; TUI showing "Queued: 4 pending" indicating multi-prompt queue buildup during doc-heavy segment, not a drydock source bug; system healthy
+
+## 2026-05-09 23:03 UTC tick
+- Stress: 1178+/1658; PID 3875067 alive (27.5h runtime); 165 done, 205 skip, 125 recycles since step 807 resume; doc-heavy segment (architecture diagrams, release notes) causing high SKIP rate from prompt-acceptance retries
+- Write rate: ~31% overall (stable)
+- Admiral last 30 min: harness:thinking_stall dominant (86/100 recent queue entries), harness:loop:bash_generic (12/100), hallucinated_name (1), grep:unescaped_pattern (1)
+- vLLM 400s: 0; balancer PID 3944578 healthy on :8001 (confirmed llm_balancer.py); docker status unknown (CLI unavailable)
+- GH issues: 0 open
+- Dispatch queue: harness=84167, retrieval=74 (0 actionable, all within 7-day window); retrieval-drain: 0 projects ingested
+- Action this tick: committed fix cd6b56c — added ralph_file_summary to _readonly_tools set in stall handler (addresses pattern harness:thinking_stall); stalls after ralph_file_summary were getting generic "Continue working" nudge instead of the specific "you called {tool} but produced no output" message
+
 ## 2026-05-09 23:30 UTC tick
 - Stress: 1173/1658; PID 3875067 alive (27h runtime); progressing from 1130 in prior tick — ~43 steps/1.5h
 - Write rate: ~31% (stable, unchanged across multiple ticks)
@@ -3941,3 +4103,67 @@ restarted, cron self-match bug fixed in this same session).
 - GH issues: 0 open
 - Dispatch queue: harness=83168, retrieval=74 (0 actionable); retrieval-drain: 0 projects ingested (74 entries all within 7-day re-ingest window)
 - Action this tick: no action — all top patterns have existing fixes; dd07921+c8a5f7f pending auto_release at 00:00 CDT (05:00 UTC); search_replace:not_found_loop (4798 total) already has file-head embedding fix in source; bash:escape_loop (170 total) already has sed/echo escape detection in bash.py; system healthy
+
+## 2026-05-10 02:30 UTC tick
+- Stress: 1218/1658 (actively running, PID 3875067)
+- Write rate: N/A (no report-every line visible in current log slice)
+- Admiral last 30 min: harness:thinking_stall dominant (55364 total queue), harness:loop:bash_generic (12634), harness:tool:hallucinated_name (6431), harness:search_replace:not_found_loop (4798)
+- vLLM 400s: 0; balancer OK on :8001; docker gemma4 healthy
+- GH issues: 0 open
+- Dispatch queue: harness=55364+, retrieval=74 (0 actionable, all within re-ingest window), steering=0
+- Action this tick: committed ee354e5 fix(stall): add specific bash stall nudges for error and output cases (addresses pattern harness:thinking_stall) — empty_after_tool:bash::source=canned was falling through to generic 'Continue working'; new _prev_bash_had_error and _prev_bash_had_output branches give the model concrete direction. Auto_release will ship at next 0/6/12/18 CDT tick.
+
+## 2026-05-10 04:03 UTC tick
+- Stress: 1230/1658 (wedged — harness alive PID 3875067, TUI child 110424 alive since 22:03 CDT, log stale 5+ hours since 22:01 CDT)
+- Write rate: 0 new writes observed (183 accepted / 235 skipped total at last progress report, prompt 1225)
+- Admiral last 30 min: not measurable (admiral_history.log not updated)
+- vLLM 400s: 0
+- GH issues: gh unavailable this tick
+- Dispatch queue: harness=85546 entries (2 patterns: thinking_stall 42 fires addressed in 24h commits, bash_generic 8 fires advisory), retrieval=74 entries (all already ingested), steering=0
+- retrieval-drain: 0 projects ingested (all up to date)
+- Action this tick: no fix committed — harness is alive but wedged (TUI child spawned at 22:03 created session_20260510_030327 with only 3 messages then stalled; babysitter cannot restart a live harness; watcher PIDs 3875068+110078 are running but recycle signals haven't unblocked it). harness:thinking_stall pattern was already addressed in today's commits (ee354e5, cd6b56c); harness:loop:bash_generic is low-confidence (0.6) with admiral nudge already firing — no source fix warranted.
+
+## 2026-05-10 04:30 UTC tick
+- Stress: 1253/1658 prompts (idx), done=200, skip=245, timeout=0, recycle=148; PID 3875067 alive (1d 8h runtime)
+- Write rate: not directly measured this tick
+- Admiral last 30 min: harness:thinking_stall dominant (55,902 total queue entries); ralph_repo_index accounts for 30,357 (54% of all stalls)
+- vLLM 400s: 0 in last 30 min
+- GH issues: 0 open
+- Dispatch queue: harness=68686 entries (multiple patterns), retrieval=74 (all already ingested), steering=0
+- Action this tick: committed fix for ralph_repo_index stall nudge (addresses pattern harness:thinking_stall) — prev generic nudge said "make changes" which is wrong after repo-indexing; new nudge tells model to write TEXT answer or call read_file; retrieval-drain: 0 projects ingested (all already ingested recently)
+
+## 2026-05-10 04:31 UTC tick
+- Stress: PID 3875067 alive (33h runtime, resume-from 807/1658); current prompt "Perf: batch DB writes"; TUI child PID 125661 alive (128s), session_20260510_042832 active with 33 messages, last updated 04:30 UTC
+- Write rate: not directly measured; sessions cycling normally (12152 total)
+- Admiral last 30 min: harness:thinking_stall (137 fires, all ralph_repo_index/bash), harness:loop:bash_generic (24), harness:grep:unescaped_pattern (3)
+- vLLM 400s: 0; balancer PID 3944578 healthy on :8001 (confirmed llm_balancer.py); docker gemma4 serving normally
+- GH issues: 0 open (gh returned no results)
+- Dispatch queue: harness=86034 (top: thinking_stall 56039, bash_generic 12758, hallucinated_name 6431), retrieval=74 (0 actionable — all within 7-day window), steering=0
+- retrieval-drain: 0 projects ingested (all already up to date)
+- Action this tick: no action — all recent patterns addressed by a6a448e/ee354e5/cd6b56c (thinking_stall nudges for ralph_repo_index/bash); auto_release ships at ~05:00 UTC; system healthy
+
+## 2026-05-10 05:03 UTC tick
+- Stress: 807+/1658 (PID 3875067 alive, tool_agent run)
+- vLLM 400s: 0; balancer healthy on :8001; docker gemma4 serving normally
+- GH issues: 0 open
+- Dispatch queue: harness=86219, retrieval=74 (0 actionable), steering=0
+- retrieval-drain: 0 projects ingested (all within 7-day window)
+- Action this tick: committed fix for ralph_repo_index:canned and read_file:canned stalls (56+53 fires in 3h at attempt 2 were falling through to generic note; added tool-specific nudges at _stall_attempt==2). Addresses pattern harness:thinking_stall. resume.md not present; worked from dispatch queue directly.
+
+## 2026-05-10 05:35 UTC tick
+- Stress: PID 3875067 alive (1d 10h runtime, resume-from-step 807/1658); most recent session session_20260510_051933 (05:19 UTC); harness cycling normally
+- vLLM 400s: 0; balancer PID 3944578 healthy on :8001 (confirmed llm_balancer.py); docker gemma4 serving normally
+- GH issues: 0 open
+- Dispatch queue: harness=86398 (top: thinking_stall 56349, bash_generic 12806, hallucinated_name 6431, search_replace:not_found_loop 4798, grep:unescaped_pattern 3999), retrieval=74 (0 actionable — all within 7-day re-ingest window), steering=0
+- retrieval-drain: 0 projects ingested (all already up to date)
+- Action this tick: no action — all top patterns addressed by commits in last 24h (9f8d022/a6a448e/ee354e5/cd6b56c fix thinking_stall for ralph_repo_index+bash at all 3 stall attempts; b3922fa fixes grep:unescaped_pattern; dd07921 fixes heredoc_loop); bash_generic and hallucinated_name are model-behavior with advisory circuit-breaker already catching repeats; system healthy awaiting next auto_release at 06:00 UTC to ship v2.8.10
+
+## 2026-05-10 11:12 UTC tick
+- Stress: PID 3875067 alive, resumed at step 807; current step ~1343/1658
+- Write rate: high skip rate (~81%: 280/345 prompts skipped in current run segment); each SKIP takes 3×120s retries before giving up, so run is progressing slowly but has advanced ~536 steps since resuming
+- Admiral last 30 min: 10+ thinking_stall fires (ralph_repo_index, write_file, bash patterns)
+- vLLM 400s: 0; balancer PID 3944578 healthy on :8001; gemma4 docker serving normally
+- GH issues: 0 open
+- Dispatch queue: harness=88280 (top: thinking_stall 58836, bash_generic 12978, hallucinated_name 6431, search_replace:not_found_loop 4798, grep:unescaped_pattern 4038), retrieval=81 (all within 7-day window, 0 actionable)
+- retrieval-drain: 0 projects ingested (all up to date)
+- Action this tick: no action — all top patterns addressed by recent commits (427a631/0820952/9f8d022/a6a448e/ee354e5/cd6b56c cover thinking_stall at all 3 retry levels; dd07921 covers heredoc_loop); high skip rate is pexpect harness behavior under model stall conditions, not a new drydock bug; system healthy

{drydock_cli-2.8.9 → drydock_cli-2.8.11}/uv.lock

@@ -326,7 +326,7 @@ wheels = [
 
 [[package]]
 name = "drydock-cli"
-version = "2.8.7"
+version = "2.8.10"
 source = { editable = "." }
 dependencies = [
     { name = "agent-client-protocol" },

drydock_cli-2.8.9/.auto_release.lock

@@ -1 +0,0 @@
-65938