drydock-cli 2.8.32__tar.gz → 2.8.34__tar.gz

drydock_cli-2.8.34/.auto_release.lock

@@ -0,0 +1 @@
+1883537

{drydock_cli-2.8.32 → drydock_cli-2.8.34}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.8.32
+Version: 2.8.34
 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-2.8.32 → drydock_cli-2.8.34}/drydock/core/agent_loop.py

@@ -838,9 +838,11 @@ class AgentLoop:
             STOP_NOW_WARN_AT = int(getattr(self, "_admiral_stop_now_warn_at",
                 int(os.environ.get("DRYDOCK_STOP_NOW_WARN_AT", 60))))
             STOP_NOW_TIME_SEC = int(os.environ.get("DRYDOCK_STOP_NOW_TIME_SEC", "0"))
+            TOOL_STOP_AFTER = int(os.environ.get("DRYDOCK_TOOL_STOP_AFTER", "0"))
             _prompt_start = time.perf_counter()
             _time_stop_injected = False
             _time_stop_escalated = False
+            _tool_stop_injected = False
             logger.warning("[TIMING] entering conversation while loop")
             while not should_break_loop:
                 # Drain any user messages typed while the agent was working.
@@ -925,6 +927,18 @@ class AgentLoop:
                         "guess."
                         + (f" {_stop_suffix}" if _stop_suffix else "")
                     )
+                if (TOOL_STOP_AFTER > 0
+                        and not _tool_stop_injected
+                        and tool_turns >= TOOL_STOP_AFTER):
+                    _tool_stop_injected = True
+                    _stop_suffix = os.environ.get("DRYDOCK_STOP_NOW_SUFFIX", "")
+                    self._inject_system_note(
+                        f"You have used {tool_turns} tool calls. "
+                        "You may NOT call any more tools. "
+                        "Your NEXT response must be plain text only — "
+                        "write your best answer right now."
+                        + (f" {_stop_suffix}" if _stop_suffix else "")
+                    )
                 if tool_turns == WRAP_UP_WARN_AT:
                     self._inject_system_note(
                         f"You have used {tool_turns} tool calls on this "
@@ -1028,7 +1042,9 @@ class AgentLoop:
                     elif ("context length" in error_str.lower()
                           or "maximum context" in error_str.lower()
                           or "400 bad request" in error_str.lower()
-                          or "status: 400" in error_str.lower()):
+                          or "status: 400" in error_str.lower()
+                          or "exceeds the available context" in error_str.lower()
+                          or "error code: 400" in error_str.lower()):
                         # Context limit or malformed request — aggressive recovery
                         # Step 0 (added 2026-05-09): if the error looks like a
                         # malformed tool call (most common 400 cause that ISN'T

{drydock_cli-2.8.32 → drydock_cli-2.8.34}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.8.32"
+version = "2.8.34"
 description = "Local-first CLI coding agent — tested with Gemma 4 26B via vLLM"
 readme = "README.md"
 requires-python = ">=3.12"

{drydock_cli-2.8.32 → drydock_cli-2.8.34}/scripts/hle_babysitter.sh

@@ -159,7 +159,9 @@ if [ -n "$GRAPHRAG_DB_OVERRIDE" ]; then
     DRYDOCK_WRAP_UP_WARN_AT=8 \
     DRYDOCK_STOP_NOW_WARN_AT=12 \
     DRYDOCK_STOP_NOW_TIME_SEC=240 \
-    DRYDOCK_THINKING_BUDGET_TOKENS=8000 \
+    DRYDOCK_THINKING_BUDGET_TOKENS=4000 \
+    DRYDOCK_TOOL_STOP_AFTER=3 \
+    DRYDOCK_STOP_NOW_SUFFIX="End your response with 'FINAL ANSWER: <your answer>' on its own line." \
     PYTHONUNBUFFERED=1 \
     nohup "$PY" -u "$DRYDOCK/scripts/hle_eval.py" \
         --source hle --limit "$LIMIT" --shuffle --seed "$SEED" --category "$CATEGORY" \
@@ -168,7 +170,9 @@ else
     DRYDOCK_WRAP_UP_WARN_AT=8 \
     DRYDOCK_STOP_NOW_WARN_AT=12 \
     DRYDOCK_STOP_NOW_TIME_SEC=240 \
-    DRYDOCK_THINKING_BUDGET_TOKENS=8000 \
+    DRYDOCK_THINKING_BUDGET_TOKENS=4000 \
+    DRYDOCK_TOOL_STOP_AFTER=3 \
+    DRYDOCK_STOP_NOW_SUFFIX="End your response with 'FINAL ANSWER: <your answer>' on its own line." \
     PYTHONUNBUFFERED=1 \
     nohup "$PY" -u "$DRYDOCK/scripts/hle_eval.py" \
         --source hle --limit "$LIMIT" --shuffle --seed "$SEED" --category "$CATEGORY" \

{drydock_cli-2.8.32 → drydock_cli-2.8.34}/scripts/hle_eval.py

@@ -334,8 +334,9 @@ def _question_prompt(q: dict) -> str:
     detection race). Keeping this on one line until that's debugged.
     """
     return (
-        f"Answer this question. End your response with the literal string "
-        f"'FINAL ANSWER:' followed by your answer on the same line. "
+        f"Answer this question from your own knowledge. "
+        f"Use at most 1 tool lookup if truly needed, then commit to your answer. "
+        f"End your response with 'FINAL ANSWER:' followed by your answer on the same line. "
         f"QUESTION: {q['question']}"
     )
 
@@ -459,6 +460,7 @@ def run_one(q: dict, sk, run_dir: Path) -> dict:
            "DRYDOCK_WRAP_UP_WARN_AT": os.environ.get("DRYDOCK_WRAP_UP_WARN_AT", "8"),
            "DRYDOCK_STOP_NOW_WARN_AT": os.environ.get("DRYDOCK_STOP_NOW_WARN_AT", "12"),
            "DRYDOCK_STOP_NOW_TIME_SEC": os.environ.get("DRYDOCK_STOP_NOW_TIME_SEC", "300"),
+           "DRYDOCK_TOOL_STOP_AFTER": os.environ.get("DRYDOCK_TOOL_STOP_AFTER", "2"),
            "DRYDOCK_STOP_NOW_SUFFIX": "Write your best answer as 'FINAL ANSWER: <answer>' now."}
     start = time.time()
     # --dangerously-skip-permissions: HLE is batch eval against read-only

{drydock_cli-2.8.32 → drydock_cli-2.8.34}/trip_log.md

@@ -1,5 +1,214 @@
 # Drydock Trip Log
 
+## 2026-05-15 10:34 UTC tick
+- Stress: restarted from step 1 (PID 1878801) — prior run had completed all 1658/1658 prompts; babysitter does not auto-recycle, so triggered manually
+- HLE lifetime: 32/365 = 8.8% (Math 11.4%, CS/AI 2.3%, Chem 0.0%, Bio/Med 16.7%, Eng 4.3%); active batch PID 1875152 (16m elapsed, on Q5/10 Biology/Medicine)
+- vLLM 400s: 0; model Q3_K_M on llamacpp-gemma4 healthy; balancer :8001 responding
+- GH issues: 0 open
+- Dispatch queue: harness=1284, retrieval=3, curiosity=1226 pending; retrieval drain: 0 actionable; consumed curiosity dd52a7abe7b2cfa1 (no_final_answer Math, addressed by a176ed7 DRYDOCK_TOOL_STOP_AFTER=3); harness:thinking_stall pattern addressed by same commit batch
+- Action this tick: restarted stress harness (was complete, not dead); consumed one curiosity item; no source fixes needed — healthy
+
+## 2026-05-15 10:00 UTC tick
+- Stress: COMPLETE at 1658/1658 (babysitter exits cleanly on completion, no restart needed)
+- HLE lifetime: 31/359 = 8.6% (Math 11.4%, CS/AI 2.3%, Chem 0.0%, Eng 4.3%, Bio/Med 16.7%); HLE pid 1864047 running 48m
+- vLLM 400s: 0; model Q3 on :8000, balancer on :8001 — healthy
+- GH issues: 0 open
+- Dispatch queue: harness=17 (thinking_stall, already addressed), retrieval=0 actionable, curiosity=1209 pending (191 hle_failure / 1018 unknown_term)
+- retrieval-drain: 0 projects (all recently ingested)
+- curiosity-queue: 1209 pending; top items are empty-prediction HLE failures already addressed by v2.8.33; no new action this tick
+- Action this tick: committed fix a176ed7 — DRYDOCK_TOOL_STOP_AFTER=3 was implemented (045169c) but never added to hle_babysitter.sh; last 5 runs show 54% empty:no_final_answer (26/48); fix wires the gate so the model is capped at 3 tool calls then forced to emit FINAL ANSWER text (addresses pattern harness:thinking_stall)
+
+## 2026-05-15 09:30 UTC tick
+- Stress: COMPLETE at 1658/1658 (babysitter idle, no action needed)
+- Write rate: N/A (stress done)
+- HLE lifetime: 30/353 = 8.5% (Math 11.4%, CS/AI 2.3%, Chem 0.0%, Eng 4.3%, Bio/Med 16.7%); HLE pid 1864047 running ~18m
+- vLLM 400s: 0 (llamacpp clean, 0 errors in last 30m)
+- GH issues: 0 open
+- Dispatch queue: harness=17 (thinking_stall, already addressed by 045169c), retrieval=0 actionable, curiosity=1199 pending (122 hle_failure / 1010 unknown_term)
+- retrieval-drain: 0 projects ingested (3 entries, all already ingested recently)
+- curiosity-queue: consumed 3 hle_failure items (5d4362861f, e9cd8146bf, 6ab35090c8) — all no_final_answer pattern after tool cap
+- Action this tick: committed fix 7e4f4ba (wire DRYDOCK_STOP_NOW_SUFFIX in hle_babysitter.sh so TOOL_STOP_AFTER and STOP_NOW nudges both append FINAL ANSWER format reminder; addresses curiosity hle_failure no_final_answer pattern)
+
+## 2026-05-15 09:00 UTC tick
+- Stress: COMPLETE at 1658/1658 (done=88, skip=53, recycle=42; babysitter correctly idle)
+- Write rate: N/A (stress done)
+- HLE lifetime: 29/349 = 8.3% (Math 11.0%, CS/AI 2.3%, Chem 0.0%, Eng 4.3%, Bio/Med 18.8%); current batch running 54m (PID 1851953)
+- Admiral last 30 min: N/A (no admiral log checked this tick)
+- vLLM 400s: 0 (llamacpp clean)
+- GH issues: 0 open
+- Dispatch queue: harness=17 (thinking_stall, already addressed), retrieval=3 (0 new ingests), curiosity=1186 pending
+- Action this tick: committed fix 045169c — DRYDOCK_TOOL_STOP_AFTER env var in agent_loop.py (caps tool calls per session; HLE sets=2 to stop Gemma 4 making 4-5 web_search calls before committing an answer); wired into hle_eval.py default=2; addresses pattern harness:thinking_stall. Consumed curiosity item 310f97edca6fab52 (empty-prediction Math, already addressed by ef07a22).
+
+## 2026-05-15 08:30 UTC tick
+- Stress: COMPLETE at 1658/1658 (done=88, skip=53, recycle=42; babysitter correctly not restarting)
+- Write rate: N/A (stress done)
+- HLE: 26/344 = 7.6% lifetime (28 runs); burndown daemon alive (PID 1851953, 24m elapsed); Math 9.7%, Bio/Med 18.8%, CS/AI 2.3%, Chem 0.0%, Eng 4.3%
+- vLLM 400s: 0; Q3_K_M on :8000 healthy; balancer on :8001 healthy
+- GH issues: 0 open
+- Dispatch queue: harness=17 (all thinking_stall circular noise, already addressed), retrieval=0 actionable (3 entries all recently ingested), curiosity=1177 pending after consuming 3 (992 unknown_term, 185 hle_failure)
+- Action this tick: consumed 3 curiosity items (10be7315, f857714d, 2a1e7c12 — all empty:no_final_answer Math failures, addressed by ef07a22 "at most 1 tool lookup"). No code changes needed; system healthy.
+
+## 2026-05-15 08:00 UTC tick
+- Stress: COMPLETE at 1658/1658 (done=88, skip=53, recycle=42; babysitter correctly not restarting)
+- Write rate: N/A (stress done)
+- HLE: 25/340 = 7.4% lifetime (27 runs); burndown daemon alive; run_1778828063 post-ef07a22: 4/10 = 40% correct, 4 no_final_answer (down from 8/10 in run_1778823337 one hour earlier — fix confirmed working); Chemistry 0/39 = 0.0% (46% no_final_answer, 23% no_response — domain gap); Physics 0/12 = 0.0%
+- vLLM 400s: 0; Q3_K_M on :8000 healthy; balancer on :8001 healthy
+- GH issues: 0 open
+- Dispatch queue: harness=17 (all thinking_stall circular noise — classify_pulse reclassifies autonomous_review.log lines that contain "thinking_stall" text; not real signals); retrieval=0 actionable; curiosity=1170 pending (984 unknown_term, 186 hle_failure)
+- Action this tick: no code changes. System healthy. ef07a22 confirmed effective. Chemistry/Physics 0% is domain gap, not a drydock bug.
+
+## 2026-05-15 07:30 UTC tick
+- Stress: COMPLETE at 1658/1658 (done, babysitter not restarting — expected)
+- Write rate: N/A (stress done)
+- HLE: 25/336 = 7.4% lifetime (27 runs); burndown daemon alive (PID 1838832, Math batch, 36 min elapsed); Categories: Math 9.6%, Bio/Med 18.8%, CS/AI 2.3%, Chem 0.0%, Physics 0.0%
+- vLLM 400s: 0 llamacpp errors; Docker health shows "unhealthy" but API /v1/models and /props respond normally (false alarm from misconfigured healthcheck)
+- GH issues: 0 open
+- Dispatch queue: harness=830 (thinking_stall, all addressed by existing stall-nudge code in agent_loop), retrieval=0 actionable, curiosity=1157 pending (974 unknown_term, 183 hle_failure)
+- Action this tick: no code changes. Retrieval drain 0 actionable. Curiosity top 3 are blank-prediction Math failures (no_final_answer) — already targeted by ef07a22 "at most 1 tool lookup" prompt fix. Noted: HLE burndown daemon (PID 1838832) was started with DRYDOCK_THINKING_BUDGET_TOKENS=8000 (old value); script now says 4000 but the long-running process wasn't restarted. Minor — 8000 still caps thinking. Next babysitter batch at :45 will use 4000 via fresh launch.
+
+## 2026-05-15 06:00 UTC tick
+- Stress: COMPLETE at 1658/1658 (all prompts exhausted; babysitter correctly not restarting)
+- Write rate: N/A (stress done)
+- Admiral last 30 min: N/A
+- vLLM 400s: 0 (Q3_K_M llamacpp-gemma4 on :8000, balancer on :8001 PID 380535 healthy)
+- GH issues: 0 open
+- Dispatch queue: harness=17 (all thinking_stall, latest 04:10; pattern already addressed by v2.8.32 4000-token budget cap), retrieval=0 actionable (3 entries, all recently ingested), curiosity=1136 pending (177 hle_failure, 959 unknown_term)
+- HLE: 21/324 = 6.5% lifetime (26 runs); babysitter alive (PID 1824412, Math batch); recent 10 batches avg ~12%; dominant failure mode is now empty:no_final_answer (7/10 in run_1778818733): model makes 7 sequential tool calls (retrieve → web_search ×2 → web_fetch → web_search ×2 → web_fetch) over 482s, ignores all STOP_NOW injections at 240s/300s, exits only when QUESTION_TIMEOUT kills session; STOP_NOW hard-stop at 360s fires on turn 7 but returns without giving model a chance to write text — root cause identified, fix would require disabling tools on final LLM call (non-trivial)
+- Action: no code changes this tick — system healthy; retrieval drain: 0 new ingested; curiosity-queue: 1 item consumed would be premature (real issue is synthesis failure, not retrieval gap); no_final_answer root cause documented for user review on return
+
+## 2026-05-15 05:30 UTC tick
+- Stress: COMPLETE at 1658/1658 (run exhausted)
+- Write rate: N/A (stress done)
+- Admiral last 30 min: N/A
+- vLLM 400s: 0 (Q3_K_M llamacpp-gemma4 on :8000, balancer on :8001, both healthy)
+- GH issues: 0 open
+- Dispatch queue: harness=1283 (all historical; thinking_stall dominates), retrieval=3 (0 actionable, all already ingested), curiosity=1130 pending (175 hle_failure, 955 unknown_term)
+- HLE: 20/320 = 6.2% lifetime (25 runs); burndown daemon alive (PID 1809726, 01:12+); recent 10 batches avg ~11%; 135 empty results (42% of all) — 92 are empty:no_final_answer (model calls retrieve+web_search 3-4 times but never emits FINAL ANSWER before 480s wall clock); top 3 curiosity items are all Math empty-prediction failures; autonomous_review 05:00 tick timed out (exit=124, no log output — likely investigated something expensive, no commit)
+- Action: no code changes — system healthy, no clear minimal fix for no_final_answer pattern within budget; retrieval drain: 0 new ingested; curiosity-queue: 3 pending hle_failure, no action this tick (all are empty-prediction, retrieval won't fix a synthesis failure)
+
+## 2026-05-15 05:00 UTC tick
+- Stress: COMPLETE at 1658/1658 (done=88, skip=53, recycle=42; run exhausted, babysitter correctly not restarting)
+- Write rate: N/A (stress done)
+- Admiral last 30 min: N/A
+- vLLM 400s: 0 (Q3_K_M llamacpp-gemma4 on :8000, balancer on :8001, both responsive)
+- GH issues: 0 open
+- Dispatch queue: harness=1283 total (all historical from completed stress run; dominant patterns: thinking_stall=829 addressed, loop:bash_generic=177, tool:hallucinated_name=133); retrieval=0 actionable (all ingested); curiosity=1122 pending (172 hle_failure, 950 unknown_term)
+- HLE: 20/316 = 6.3% lifetime (25 runs); burndown daemon alive (PID 1809726, 41+ min); recent batches: 0%, 20%, 0%, 0%, 20%, 10%, 10%, 10%, 20%, 20%; top curiosity items are all Math empty-response failures (empty:no_final_answer) — hard-stop fires before model outputs FINAL ANSWER line; 4000-token budget cap in effect; Chemistry 0/39, Physics 0/12 still at 0%
+- retrieval-drain: 0 new ingestions (3 queue entries, all already ingested)
+- curiosity-drain: no action this tick — top 3 items all Math empty-response failures already addressed by v2.8.32 budget cap
+- Action this tick: no code changes — system healthy; investigated empty:no_final_answer root cause (model makes tool calls then gets hard-stopped before emitting FINAL ANSWER; complex fix, deferred); stress run complete
+
+## 2026-05-15 04:00 UTC tick
+- Stress: COMPLETE at 1658/1658 (done=88, skip=53, recycle=42; babysitter reports dead each tick — expected, run exhausted)
+- Write rate: N/A (stress done)
+- Admiral last 30 min: N/A
+- vLLM 400s: 0 (Q3_K_M llamacpp-gemma4 on :8000, balancer on :8001, both responsive; Docker "unhealthy" label is a false alarm — health check hits port 8080 which was never configured, actual inference on :8000 is fine)
+- GH issues: 0 open
+- Dispatch queue: harness=15 (all self-referential thinking_stall false positives, already addressed), retrieval=0 actionable, curiosity=1099 pending (168 hle_failure, 931 unknown_term)
+- HLE: 19/308 = 6.2% lifetime; burndown daemon active (PID 1796996, 53m); current batch at Q8/10 (Math, braid groups); Chemistry 0/39 still 0%, CS/AI 1/43 2.3%; recent batches trending upward; 4000-token thinking budget (v2.8.32) reducing no_final_answer stalls
+- retrieval-drain: 0 new ingestions (all already current)
+- curiosity-drain: consumed `4d9503adb6b76691` (Math empty-prediction, thinking stall, addressed by v2.8.32)
+- Action this tick: no code changes — system healthy; stress run is complete; HLE burndown running normally
+
+## 2026-05-15 03:30 UTC tick
+- Stress: COMPLETE at 1658/1658 (all PRDs exhausted, babysitter will not restart — expected)
+- Write rate: N/A (stress done)
+- Admiral last 30 min: N/A
+- vLLM 400s: 0 (Q3_K_M on :8000, balancer on :8001, both healthy)
+- GH issues: 0 open
+- Dispatch queue: harness=14 (all self-referential thinking_stall false positives, already addressed), retrieval=0 actionable (all recently ingested), curiosity=1082 pending (top items all Math empty-prediction stalls already addressed by v2.8.32 thinking-budget 8000→4000 fix)
+- HLE: 17/304 = 5.6% lifetime; burndown daemon active (PID 1742185); current batch 0/3 in progress (first batch under 4000-token budget); Chemistry 0/39 and Physics 0/12 are PhD-level cheminformatics/QFT, no tool fix available; curiosity-drain: consumed `1d7e1b508cfe5aa6` (thinking stall, already fixed)
+- Action this tick: no code changes — system healthy; consumed 1 curiosity item (stall pattern addressed by v2.8.32)
+
+## 2026-05-15 03:00 UTC tick
+- Stress: dead (PID 340932 gone; run exhausted at 1658/1658, babysitter will not restart)
+- Write rate: N/A
+- Admiral last 30 min: N/A (stress done)
+- vLLM 400s: 0 (balancer/llamacpp healthy; prior 502s in log were stale)
+- GH issues: none checked (budget tight)
+- Dispatch queue: harness=15 (all self-referential thinking_stall), curiosity=1070 pending
+- HLE: 17/300 = 5.7% lifetime; latest Math batch 1/10 = 10%; all failures empty:no_final_answer (482s = 480s outer deadline, NOT 360s hard-stop)
+- Action this tick: committed fix 1e741c7 — lower thinking budget 8000→4000 tokens in hle_babysitter.sh (addresses curiosity:3cd829328e46a225). Root cause: 8000-tok thinking = up to 125s per LLM call; last call starts at ~320s, 480s outer deadline fires before hard-stop at 360s can trigger. 4000 tokens caps call at ~63s, giving hard-stop room to fire. Retrieval drain: 0 new (all already ingested).
+
+## 2026-05-15 02:30 UTC tick
+- Stress: 1658/1658 COMPLETE (final: done=88, skip=53, recycle=42; babysitter reporting dead each tick — run is exhausted, not broken)
+- Write rate: N/A (run finished)
+- HLE burndown: PID 1782298 alive 42m, slot=3 Math batch in progress; aggregate 23 runs, 297 questions, 17 correct = 5.7%; Math 10/148=6.8%, Bio/Med 3/16=18.8%, Chemistry 0/39=0.0%, CS/AI 1/43=2.3%, Engineering 1/23=4.3%; recent batches (last 4): 10%, 10%, 16.7%, 20.0% — upward trend on Math
+- vLLM 400s: 0 (llamacpp-gemma4 Q3_K_M :8000 healthy, balancer :8001 healthy)
+- GH issues: 0 open
+- Dispatch queue: harness=18 (all thinking_stall, all from autonomous_review self-referential logging, pattern already addressed in agent_loop.py), retrieval=0 actionable, curiosity=1058 pending (161 hle_failure, 897 unknown_term)
+- retrieval-drain: 0 new ingestions (all already current)
+- curiosity-queue: top 3 items are all hle_failure with empty predictions (stall cases, not retrieval gaps) — no GraphRAG ingest would help; no action
+- Investigated Chemistry 0/39: confirmed same root-cause pattern (Böttcher Molecular Complexity, peptidomimetic foldamers, SMILES design) — PhD-level cheminformatics beyond any tool capability; hard-stop and loop fixes from v2.8.31 were shipped after these batches; awaiting Chemistry batch in v2.8.32+ to see if stall rate improves
+- Action this tick: no commit — system healthy, burndown active, stress complete
+
+## 2026-05-15 02:00 UTC tick
+- Stress: 1658/1658 COMPLETE (stress run fully finished — all prompts exhausted, babysitter confirming dead each tick).
+- HLE: 5.5% lifetime (16/293); burndown daemon pid=1742185 alive (9h+), current batch pid=1782298 alive (12m+, Math category); 23 batches run to date.
+- Model/balancer: Q3_K_M on :8000, llm_balancer on :8001 — both healthy; 0 errors in last 30m.
+- Admiral last 30 min: 15 harness:thinking_stall (all self-referential classifier false positives from autonomous_review.log text).
+- vLLM 400s: 0.
+- GH issues: 0 open.
+- Dispatch queue: harness=15 (all self-referential, not actionable), retrieval=0 actionable, curiosity=1050 pending (159 hle_failure + 891 unknown_term).
+- Retrieval drain: 0 actionable (all recently ingested).
+- Curiosity: consumed cd5b64179bc51be0 (Math empty prediction = thinking stall, covered by existing hard-stop). 288 total consumed.
+- Action this tick: no code committed — system healthy. Stress run is complete (done=88, skip=53, timeout=0, recycle=42 across 1658 prompts); HLE burndown continues autonomously.
+
+## 2026-05-15 01:30 UTC tick
+- Stress: 1658/1658 COMPLETE (PID 340932 gone — run finished, babysitter will restart next tick if needed).
+- HLE: 5.6% lifetime (16/288); burndown batch pid=1769413 alive (54m, Math). Recent completed batch (pid=1755519, Math): 1/10 correct. Babysitter log shows most batches skipping (prior eval still alive >60min) — normal given 60-160s per question.
+- Model/balancer: Q3_K_M on :8000, llm_balancer on :8001 — both up; 0 errors in last 30m.
+- Admiral last 30 min: 14 harness:thinking_stall entries (all self-referential — classifier reading autonomous_review log text, not new drydock sessions).
+- vLLM 400s: 0 (llama.cpp backend).
+- GH issues: 0 open.
+- Dispatch queue: harness=14 (all self-referential), retrieval=0 actionable, curiosity=1037 pending (154 hle_failure + 883 unknown_term).
+- Retrieval drain: 0 actionable (all recently ingested).
+- Curiosity top 3: all "no answer extracted" Math failures (empty predictions from thinking stalls — the last LLM call of the session starts ~t=410s, takes 70s mid-generation, harness kills at 480s before content is written; hard-stop at +120s fires after harness kill in this race). Pattern noted; deeper fix (per-LLM-call timeout or watcher.refresh after loop exit) deferred to user review.
+- Action this tick: no code committed — system healthy, investigated empty-prediction root cause (race between hard-stop check and 480s harness kill), no actionable queue items.
+
+## 2026-05-15 00:30 UTC tick
+- Stress: 1658/1658 COMPLETE (harness not restarted — no remaining prompts).
+- Write rate: N/A.
+- HLE: 5.4% lifetime (15/280); burndown daemon pid=1742185 alive (2h28m), current batch pid=1755519 (1h12m, Math category); babysitter coexisting correctly.
+- Chemistry: 0/39 (0.0%); Physics: 0/12 (0.0%) — both dominated by empty predictions from thinking stalls (481s timeout, 1 msg). SMILES design questions structurally hard for generalist model. No fixable prompt rule identified.
+- Admiral last 30 min: 12 harness:thinking_stall entries (all self-referential — classifier picking up autonomous_review log text, not new drydock sessions).
+- vLLM 400s: N/A (llama.cpp backend, 400s are rare).
+- GH issues: 0 open.
+- Dispatch queue: harness=12 (all thinking_stall, previously addressed), retrieval=0 actionable, curiosity=1007 pending (148 hle_failure + 859 unknown_term).
+- Retrieval drain: 0 actionable (all recently ingested).
+- Curiosity drain: consumed item 4288f930bec646ff (Math empty prediction, Gold=16, thinking stall — covered by existing hard-stop + 8k token cap).
+- Action this tick: no code committed — system healthy, no actionable bugs found.
+
+## 2026-05-15 00:00 UTC tick
+- Stress: 1658/1658 COMPLETE — all stress prompts done; harness not restarted (expected).
+- Write rate: N/A (stress completed).
+- Admiral last 30 min: 0 new fires (autonomous_review ran at 23:31 UTC, exit=0).
+- vLLM 400s: 0 (llama.cpp Q3 model healthy, balancer forwarding).
+- GH issues: 0 open.
+- Dispatch queue: harness=11 (all thinking_stall, previously addressed), retrieval=0 actionable, curiosity=999 pending (dominated by Math empty-prediction thinking stalls, already fixed by hard-stop + 8k cap).
+- HLE: 15/276 = 5.4% overall (21 runs). Math 8/127=6.3%, CS/AI 1/43=2.3%, Chemistry 0/39=0.0%, Engineering 1/23=4.3%, Bio/Med 3/16=18.8%. Babysitter running (PID 1755519, 42 min). Burndown daemon running.
+- Action this tick: stress run confirmed complete (1658/1658); retrieval drain 0 actionable; consumed curiosity item ffa559ac292e6fe7 (Math blank prediction, thinking stall — addressed by existing hard-stop + 8k cap); no new bugs found.
+
+## 2026-05-14 23:30 UTC tick
+- Stress: PID 340932 dead; babysitter hourly cron will restart at 00:00 UTC.
+- HLE: 5.1% overall (14/272, 21 runs). Math 7/123=5.7%, CS/AI 1/43=2.3%, Chemistry 0/39=0.0%, Engineering 1/23=4.3%, Bio/Med 3/16=18.8%. Current batch (PID 1755519, hle_eval.py math) alive at 13 min — burndown log shows Q1 NO (empty:no_final_answer, 481s, Riemannian manifold), Q2 YES (judge, 302s, dB addition). Hard math still hitting the no_final_answer wall; llama.cpp ignores thinking_budget_tokens so model runs uncapped until hard-stop fires.
+- Write rate: N/A (stress not running).
+- Admiral last 30 min: not checked (autonomous_review started at 23:30 UTC).
+- vLLM 400s: 0 (llamacpp-gemma4 logs clean).
+- GH issues: not checked (budget conservation).
+- Dispatch queue: harness=10 (thinking_stall, all addressed), curiosity=995 pending (143 hle_failure + 852 unknown_term).
+- Action this tick: consumed curiosity:12cb56df66dfe2d2 (Math empty prediction, thinking stall, addressed by existing hard-stop + 8000-token cap); retrieval drain 0 actionable; no code committed.
+
+## 2026-05-14 23:00 UTC tick
+- Stress: COMPLETE (1658/1658, done=88, skip=53). Stress babysitter correctly not restarting — run finished.
+- HLE: 5.2% overall (14/268 across 20 runs). Math 7/119=5.9%, CS/AI 1/43=2.3%, Chemistry 0/39=0.0%, Engineering 1/23=4.3%, Bio/Med 3/16=18.8%. Burndown batch pid=1741798 still alive (~60 min into 10 Math questions — slow but progressing, log confirms Q7 YES at 355s). `no_response, 1 msg, 482s` failures still occurring on hardest Math despite DRYDOCK_THINKING_BUDGET_TOKENS=8000; likely llama.cpp ignores `thinking_budget_tokens` parameter, leaving the model uncapped. Chemistry flat at 0% — `no_final_answer` pattern across all 39 questions, organic synthesis and SMILES prediction beyond current symbolic tools.
+- Write rate: N/A (stress complete)
+- Admiral last 30 min: 0 fires (autonomous_review just started at 23:00)
+- vLLM 400s: 2 (docker health "unhealthy" but /health endpoint returns ok — likely stale health check)
+- GH issues: 0 open
+- Dispatch queue: harness=1273, retrieval=3, steering=0, curiosity=988 pending (totals)
+- Action this tick: retrieval-drain: 0 actionable (already ingested); curiosity item cf34579c9a29f293 consumed (empty Math prediction — classified as thinking-stall not corpus gap); no new bug committed — system healthy, all known patterns already addressed by prior ticks.
+
 ## 2026-05-14 22:00 UTC tick
 - Stress: COMPLETED — 1658/1658 prompts done (done=88, skip=53, recycle=42). Babysitter intentionally not restarting.
 - HLE: 4.6% overall (12/261 correct across 18 runs). Chemistry 0/39=0.0%, Math 5/112=4.5%, CS/AI 1/43=2.3%, Bio/Med 3/16=18.8%, Engineering 1/23=4.3%.
@@ -6153,3 +6362,32 @@ restarted, cron self-match bug fixed in this same session).
 - curiosity-queue: 3 top items are generic Math hle_failure (no answer extracted, gold=24/20/3) — insufficient context to act on without reading full session; skipping
 - Investigated Chemistry 0/39: session logs confirm model looped on web_search for obscure terms (Böttcher Molecular Complexity) — root cause was the missing empty-result string in _looks_empty, fixed in 7aa9507 (v2.8.31). Chemistry batches in v2.8.31+ should improve.
 - Action this tick: no commit — system healthy, v2.8.31 deployed with all hard-stop + loop fixes, burndown daemon running
+
+## 2026-05-15 04:35 UTC tick
+- Stress: n/a (watchdog paused)
+- Write rate: n/a
+- Admiral last 30 min: not checked (stress paused)
+- vLLM 400s: 108 context-overflow errors in last 30 min (llamacpp -c 32768; requests hitting 40–41K tokens on HLE sessions with 15+ messages)
+- GH issues: 0 open
+- Dispatch queue: harness=17 (all thinking_stall, pattern addressed by v2.8.32 budget cap), retrieval=0 actionable, curiosity=1105 pending (170 hle_failure, 935 unknown_term)
+- retrieval-drain: 0 new ingestions
+- Action this tick: committed fix 69ef332 — added "exceeds the available context" and "error code: 400" to emergency-compaction trigger in agent_loop.py so HLE sessions overflow recovery fires correctly under llamacpp (was silently missing the vLLM-only patterns).
+
+## 2026-05-15 06:30 UTC tick
+- Stress: 1658/1658 (completed, harness correctly not restarted)
+- Write rate: N/A (stress run done)
+- HLE: 21/328 = 6.4% lifetime (26 runs); recent batches 10-20%; by method: 135 empty, 99 empty:no_final_answer, 28 empty:no_response
+- vLLM 400s: container shows "unhealthy" (3603 failing health checks) but /health returns {"status":"ok"} — health check misconfigured, model functional
+- GH issues: 0 open
+- Dispatch queue: harness=17 (thinking_stall, all self-referential), retrieval=0 actionable, curiosity=1145 pending (966 unknown_term, 179 hle_failure)
+- Action this tick: committed fix ef07a22 — cap tool lookups in HLE question prompt to 1 max. Root cause: model making 3-4 web_search calls (60-160s each) on Q3_K_M, exhausting 480s wall-clock before 300s STOP_NOW fires; 99/328 questions (30%) failing as empty:no_final_answer. Prompt now instructs "answer from own knowledge, at most 1 tool lookup." No other actionable harness patterns; retrieval drain found nothing new.
+
+## 2026-05-15 07:00 UTC tick
+- Stress: COMPLETE at 1658/1658 (run finished; babysitter not restarting — expected)
+- Write rate: N/A (stress run done)
+- Admiral last 30 min: N/A (stress done)
+- vLLM 400s: 6 minor errors in container logs; llamacpp shows "unhealthy" Docker health status but API (/v1/models, /props) responds normally — false alarm from misconfigured healthcheck
+- GH issues: 0 open
+- HLE: 22/331 = 6.6% lifetime (27 runs); burndown daemon alive (PID 1838832, 6h+ elapsed, running Math batch)
+- Dispatch queue: harness=16 (thinking_stall, all self-referential false positives already addressed), retrieval=0 actionable, curiosity=1154 pending (972 unknown_term, 182 hle_failure)
+- Action this tick: consumed curiosity item bd848692c98c1362 (Math blank-prediction thinking stall, already addressed by v2.8.32 4000-token cap); retrieval drain 0 actionable; no code changes needed — system healthy

drydock_cli-2.8.32/.auto_release.lock

@@ -1 +0,0 @@
-1751768