drydock-cli 2.8.37__tar.gz → 2.8.39__tar.gz

drydock_cli-2.8.39/.auto_release.lock

@@ -0,0 +1 @@
+2016173

{drydock_cli-2.8.37 → drydock_cli-2.8.39}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.8.37
+Version: 2.8.39
 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-2.8.37 → drydock_cli-2.8.39}/drydock/core/agent_loop.py

@@ -1051,7 +1051,10 @@ class AgentLoop:
                           or "400 bad request" 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()):
+                          or "error code: 400" in error_str.lower()
+                          or "500 internal server error" in error_str.lower()
+                          or "status: 500" in error_str.lower()
+                          or "llm backend error" 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
@@ -1127,11 +1130,31 @@ class AgentLoop:
                         except Exception:
                             pass
                         if dropped_bad_tool_call:
-                            error_text = (
-                                "Your last tool call was rejected by the "
-                                "server (likely malformed arguments). "
-                                "Try a simpler form, or use a different tool."
+                            # Detect JSON-truncation: llama.cpp returns this
+                            # when max_tokens is too low and the tool call
+                            # JSON gets cut off mid-string.
+                            _trunc = (
+                                "missing closing quote" in error_str.lower()
+                                or (
+                                    "parse error at" in error_str.lower()
+                                    and "column" in error_str.lower()
+                                )
                             )
+                            if _trunc:
+                                error_text = (
+                                    "Your write_file content was too large — "
+                                    "the server truncated the response mid-JSON "
+                                    "(hit max_tokens). Split the file into "
+                                    "smaller sections and write each with a "
+                                    "separate write_file call (aim for ≤50 "
+                                    "lines per call)."
+                                )
+                            else:
+                                error_text = (
+                                    "Your last tool call was rejected by the "
+                                    "server (likely malformed arguments). "
+                                    "Try a simpler form, or use a different tool."
+                                )
                         else:
                             error_text = (
                                 "Context compacted due to API error. "
@@ -1335,7 +1358,17 @@ class AgentLoop:
                 and bool(_prev_tool_result.strip())
             )
             if _stall_attempt == 0:
-                if _prev_write_path_error:
+                if _tool_stop_injected:
+                    _fa_suffix = os.environ.get(
+                        "DRYDOCK_STOP_NOW_SUFFIX",
+                        "End with 'FINAL ANSWER: <answer>'.",
+                    )
+                    note = (
+                        "STOP THINKING. Do NOT use any tools. "
+                        "Write your best answer as plain text RIGHT NOW. "
+                        + _fa_suffix
+                    )
+                elif _prev_write_path_error:
                     note = (
                         "Your write_file call failed because the path argument was empty. "
                         "Retry write_file RIGHT NOW with the correct path. "
@@ -1503,8 +1536,10 @@ class AgentLoop:
         # Detect repetitive text generation (Gemma 4 sometimes loops text within one response)
         if last_message.content and len(last_message.content) > 200:
             text = last_message.content
-            # Check if any sentence repeats 3+ times
-            sentences = [s.strip() for s in text.split('.') if len(s.strip()) > 30]
+            # Check if any sentence repeats 3+ times.
+            # Threshold 15: catches short repeated phrases like "(Wait, I'll call the tool."
+            # which split to 28-char fragments — previously filtered by the old > 30 threshold.
+            sentences = [s.strip() for s in text.split('.') if len(s.strip()) > 15]
             if sentences:
                 from collections import Counter
                 sentence_counts = Counter(sentences)

{drydock_cli-2.8.37 → drydock_cli-2.8.39}/drydock/core/llm/backend/generic.py

@@ -191,17 +191,27 @@ class OpenAIAdapter(APIAdapter):
         )
 
         # Enable thinking for models that support it (Gemma 4)
+        _budget_env = os.environ.get("DRYDOCK_THINKING_BUDGET_TOKENS")
         if thinking and thinking not in ("off", ""):
             payload["chat_template_kwargs"] = {"enable_thinking": True}
             # Cap thinking length (llama.cpp). Prevents runaway thinking stalls
             # on hard HLE questions (model can think >480s without this).
             # Set DRYDOCK_THINKING_BUDGET_TOKENS=8000 in HLE batch launches.
-            _budget = os.environ.get("DRYDOCK_THINKING_BUDGET_TOKENS")
-            if _budget:
+            if _budget_env:
                 try:
-                    payload["thinking_budget_tokens"] = int(_budget)
+                    # llama.cpp requires thinking_budget inside chat_template_kwargs,
+                    # not as a top-level field (top-level is silently ignored → INT32_MAX).
+                    payload["chat_template_kwargs"]["thinking_budget"] = int(_budget_env)
                 except ValueError:
                     pass
+        elif _budget_env:
+            # llama.cpp activates reasoning with budget=INT32_MAX by default even
+            # when thinking="off". Always cap it when the env var is set so that
+            # "off"-mode requests don't stall for 480s on hard questions.
+            try:
+                payload.setdefault("chat_template_kwargs", {})["thinking_budget"] = int(_budget_env)
+            except ValueError:
+                pass
 
         if enable_streaming:
             payload["stream"] = True

{drydock_cli-2.8.37 → drydock_cli-2.8.39}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.8.37"
+version = "2.8.39"
 description = "Local-first CLI coding agent — tested with Gemma 4 26B via vLLM"
 readme = "README.md"
 requires-python = ">=3.12"

{drydock_cli-2.8.37 → drydock_cli-2.8.39}/scripts/hle_burndown.sh

@@ -200,6 +200,8 @@ run_loop() {
             DRYDOCK_STOP_NOW_WARN_AT=12 \
             DRYDOCK_STOP_NOW_TIME_SEC=240 \
             DRYDOCK_THINKING_BUDGET_TOKENS=8000 \
+            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" \
@@ -210,6 +212,8 @@ run_loop() {
             DRYDOCK_STOP_NOW_WARN_AT=12 \
             DRYDOCK_STOP_NOW_TIME_SEC=240 \
             DRYDOCK_THINKING_BUDGET_TOKENS=8000 \
+            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" \

{drydock_cli-2.8.37 → drydock_cli-2.8.39}/scripts/hle_eval.py

@@ -422,6 +422,13 @@ def _extract_answer(messages: list[dict]) -> str:
         if m.get("role") != "assistant":
             continue
         content = _flat(m.get("content")).strip()
+        # Skip internal drydock error messages — they are not model answers.
+        if content.startswith("[Drydock:"):
+            continue
+        # llama.cpp --jinja stores thinking in reasoning_content; content may
+        # be empty even when the model wrote a FINAL ANSWER inside thinking.
+        if not content:
+            content = _flat(m.get("reasoning_content")).strip()
         if not content:
             continue
         for line in reversed(content.splitlines()):

{drydock_cli-2.8.37 → drydock_cli-2.8.39}/tests/test_user_issues.py

@@ -227,6 +227,51 @@ class TestWriteFileDedup:
         assert old != new  # Different = allow write
 
 
+class TestIssue20_500ErrorRecovery:
+    """#20: 500 Internal Server Error from llama.cpp backend triggers no-op
+    error path instead of context compaction, making recovery impossible."""
+
+    def test_500_error_matches_compaction_path(self):
+        """The condition that routes errors to aggressive compaction should
+        match 500 Internal Server Error strings from llama.cpp backends."""
+        # Reproduce the predicate extracted from agent_loop.py
+        error_str = "API error from local (model: navygpt): LLM backend error [local]\nstatus: 500 Internal Server Error reason: Internal Server Error"
+        cond = (
+            "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 "exceeds the available context" in error_str.lower()
+            or "error code: 400" in error_str.lower()
+            or "500 internal server error" in error_str.lower()
+            or "status: 500" in error_str.lower()
+            or "llm backend error" in error_str.lower()
+        )
+        assert cond, "500 error should trigger compaction path"
+
+    def test_generic_500_matches(self):
+        """Plain '500 Internal Server Error' string should also match."""
+        error_str = "status: 500 Internal Server Error"
+        cond = (
+            "500 internal server error" in error_str.lower()
+            or "status: 500" in error_str.lower()
+        )
+        assert cond
+
+    def test_non_500_error_does_not_match(self):
+        """Unrelated errors (e.g. timeout) should NOT match."""
+        error_str = "Connection timed out after 30s"
+        cond = (
+            "context length" in error_str.lower()
+            or "400 bad request" in error_str.lower()
+            or "status: 400" in error_str.lower()
+            or "500 internal server error" in error_str.lower()
+            or "status: 500" in error_str.lower()
+            or "llm backend error" in error_str.lower()
+        )
+        assert not cond
+
+
 class TestEscapeCharacters:
     """#70: Invalid escape characters causing API errors."""
 

{drydock_cli-2.8.37 → drydock_cli-2.8.39}/trip_log.md

@@ -1,5 +1,181 @@
 # Drydock Trip Log
 
+## 2026-05-15 22:30 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- Write rate: N/A (stress paused)
+- HLE burndown: running (pid 2001562, 55 min elapsed); lifetime 9.2% (43/467 across 40 runs) — Math 11.5%, Bio/Med 16.7%, CS/AI 3.2%
+- vLLM 400s: 1 (minor, llama.cpp container otherwise clean)
+- GH issues: 1 open (#23 image-processing enhancement — requires mmproj download, deferred to user return)
+- Dispatch queue: harness=1291 total (thinking_stall pattern all addressed by prior commits), retrieval=3 (0 actionable, already ingested), curiosity=1627 pending
+- Retrieval-drain: 0 projects ingested (all already current)
+- Curiosity top 3: all hle_failure with empty predictions (Math/Bio) — genuine model non-responses; extraction already addressed by 8e92b1e/64a90b4
+- Action this tick: committed b0350cb fix(stall): targeted FINAL-ANSWER nudge when _tool_stop_injected stalls — unstaged agent_loop.py change (10-line addition) that injects "STOP THINKING, write FINAL ANSWER" when model stalls after hitting TOOL_STOP_AFTER limit, instead of the generic read-stall message
+
+## 2026-05-15 22:00 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- Write rate: N/A (stress paused)
+- HLE burndown: running (25 min elapsed this batch); lifetime 9.3% (43/463 across 40 runs) — Math 11.8%, Bio/Med 16.7%, CS/AI 3.2%
+- vLLM 400s: 0 (llama.cpp container clean)
+- GH issues: 1 open (#23 image-processing enhancement, not actionable this tick)
+- Dispatch queue: harness=18 thinking_stall (all addressed by prior commits), retrieval=0 actionable, curiosity=1619 pending (265 hle_failure + 1354 unknown_term)
+- Curiosity top items: hle_failure with empty-predicted on Math/Bio questions — genuine model non-responses, not extractable drydock bugs; prior 8e92b1e/64a90b4 commits covered the two structural extraction gaps
+- Action this tick: no action — system healthy, no actionable drydock bug; retrieval-drain: 0 projects ingested (all already current)
+
+## 2026-05-15 21:30 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- Write rate: N/A (stress paused)
+- HLE: 42/460 = 9.1% lifetime (39 runs); burndown daemon alive (PID in /tmp/hle_continuous.pid, engineering batch just completed 1/10=10%)
+- vLLM 400s: 0 (llamacpp-gemma4 "unhealthy" Docker healthcheck but /health endpoint returns ok, model responding fine)
+- GH issues: 1 open (#23 enhancement, image processing — not a drydock source bug)
+- Dispatch queue: harness=19 (all harness:thinking_stall, addressed by 64a90b4), retrieval=0 actionable, curiosity=1614 pending (264 hle_failure, 1350 unknown_term)
+- Action this tick: no fix committed. Engineering batch shows 4/10 questions hitting 481s timeout with `empty:no_response` (1 msg — model never produces content). The bc99b6b thinking-budget cap (DRYDOCK_THINKING_BUDGET_TOKENS=8000) is deployed in v2.8.38 and the burndown correctly passes the env var; unclear why stalls persist. Balancer passthrough confirmed correct. Next tick: verify llama.cpp is honoring chat_template_kwargs.thinking_budget by checking docker logs for "reasoning-budget:" lines on a live request.
+
+## 2026-05-15 21:00 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- Write rate: N/A (stress paused)
+- HLE: 41/456 = 9.0% lifetime (39 runs); burndown daemon alive (38:50 elapsed, engineering batch 19 in flight)
+- vLLM 400s: 0 (llamacpp-gemma4 build b9115 clean)
+- GH issues: 1 open (#23 enhancement, image processing — skipped, not a drydock source bug)
+- Dispatch queue: harness=19 (all thinking_stall, all addressed), retrieval=0 actionable, curiosity=1602 pending (262 hle_failure, 1340 unknown_term)
+- Action this tick: no fix committed. Identified critical ongoing issue: engineering batch shows 3/5 questions hitting 481s QUESTION_TIMEOUT with only 1 message (model never responds). Docker logs show `reasoning-budget: activated, budget=2147483647` on ALL recent requests, confirming bc99b6b thinking-budget cap is NOT reaching llama.cpp despite correct payload construction. `chat_template_kwargs.thinking_budget=8000` is built correctly in code but llama.cpp b9115 appears to ignore it. Root cause unconfirmed — ran out of budget before deeper investigation. Next tick should verify if llama.cpp b9115 requires a different field path or if there's a proxy/stripping issue between drydock and the server.
+
+## 2026-05-15 20:30 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- Write rate: N/A (stress paused)
+- HLE: 41/452 = 9.1% lifetime (39 runs); burndown daemon alive (8:52 elapsed, engineering batch in flight)
+- vLLM 400s: 0 (llamacpp-gemma4 clean)
+- GH issues: 1 open (#23 enhancement, image processing — not actioned)
+- Dispatch queue: harness=19 (all thinking_stall, previously addressed), retrieval=0 actionable, curiosity=1591 pending (260 hle_failure, 1331 unknown_term)
+- Action this tick: investigated empty-prediction pattern in Humanities/Social Science batches (8/10 empty in run_1778867670, 19:07 UTC, after bc99b6b thinking-cap fix) — still ongoing post-fix; likely model-side, not drydock bug. No code committed. System healthy.
+
+## 2026-05-15 20:00 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- HLE: 40/448 = 8.9% lifetime (38 runs); burndown daemon alive (54 min elapsed, chemistry batch slot 18 in flight)
+- Write rate: N/A (stress paused)
+- vLLM 400s: 0 (llamacpp-gemma4 clean)
+- GH issues: 1 open (#23 enhancement: multimodal/image — not actionable)
+- Dispatch queue: harness=1290 total (stale thinking_stall, pattern covered); retrieval=0 actionable; curiosity=1888 total / 1578 pending
+- Action this tick: committed 64a90b4 — fix(hle): skip drydock error messages in _extract_answer. When MAX_STALL_RETRIES exhausted, agent_loop sets assistant content to "[Drydock: model returned an empty response...]"; _extract_answer was returning this error text as the prediction (causing judge to score NO on every case), wasting an API call per question. Now skips messages starting with "[Drydock:". Retrieval drain: 0 actionable (3 entries already ingested). Curiosity top-3 all hle_failure:empty — pattern partially addressed by 8e92b1e (reasoning_content fallback) and 64a90b4 (error-message skip); remaining empty cases are genuine timeouts.
+
+## 2026-05-15 19:00 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- HLE: 38/440 = 8.6% lifetime (37 runs); Math 11.4%, Bio/Med 16.7%, CS/AI 3.2%, Chem 0.0%; burndown daemon alive (01:05:50 elapsed)
+- Write rate: N/A (stress paused)
+- vLLM 400s: 0
+- GH issues: 1 open (#23 enhancement: multimodal/image processing — not actionable this tick)
+- Dispatch queue: harness=19 (stale thinking_stall, already covered by bc99b6b); retrieval=0 actionable; curiosity=1566 pending
+- Action this tick: committed 8e92b1e — fix _extract_answer to fall back to reasoning_content when content is empty. llama.cpp --jinja mode stores Gemma 4 thinking in reasoning_content; 44-78% of HLE questions were scoring as empty:no_final_answer because the function skipped those messages. Matches logic already in judge function. Addresses curiosity:b572406518cd9013.
+
+## 2026-05-15 18:30 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- HLE: 38/436 = 8.7% lifetime (37 runs); Math 11.4%, Bio/Med 16.7%, CS/AI 3.2%, Chem 0.0%; burndown daemon alive (35:49 elapsed)
+- Write rate: N/A (stress paused)
+- Admiral last 30 min: N/A
+- vLLM 400s: 0
+- GH issues: 1 open (#23 enhancement: image processing/multimodal); closed #21 and #22 (fixes already in v2.8.38)
+- Dispatch queue: harness=19 (all thinking_stall, entries from May 3 — stale, already addressed); retrieval=0 actionable; curiosity=1534 pending (top-3 are no_final_answer cases, covered by shipped TOOL_STOP_AFTER+STOP_NOW fixes)
+- Action this tick: closed GH #21 and #22 manually (commits had "closes" but no push to GitHub); no source fix needed — system healthy
+
+## 2026-05-15 18:00 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- HLE: 38/432 = 8.8% lifetime (37 runs); Math 11.4%, Bio/Med 16.7%, CS/AI 3.2%, Chem 0.0%; burndown daemon alive (05:54 elapsed at tick time)
+- Write rate: N/A (stress paused)
+- Admiral last 30 min: N/A
+- vLLM 400s: 0
+- GH issues: 2 open (#21 #22 — fixed by recent commits, not yet pushed)
+- Dispatch queue: harness=19 (all thinking_stall), retrieval=0 actionable, curiosity=1515 pending (645 in 24h)
+- Action this tick: investigated empty:no_response pattern in Physics HLE batch (0/10, 480s timeouts) — model likely exhausts thinking budget on hard QFT/gauge-theory questions before producing FINAL ANSWER; thinking_budget cap (bc99b6b, 8000 tok) is in place but stall debug shows retries working normally for other sessions; no commit — deeper diagnosis deferred. Retrieval drain: 0 actionable. Autonomous_review 17:30 tick failed (budget exceeded).
+
+## 2026-05-15 17:00 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- HLE: 38/422 = 9.0% lifetime (36 runs); Math 11.4%, Bio/Med 16.7%, CS/AI 3.2%, Chem 0.0%; burndown daemon alive (11:07 elapsed)
+- vLLM 400s: 0; llamacpp-gemma4 healthy (Q3_K_M on :8000); balancer :8001 up
+- GH issues: 2 open (#21 still showing open despite close commit a338858; #22 looping bug filed by RedHat8 user — "(Wait, I'll call the tool.)" repeated 10+ times on large C++-to-Python port request)
+- Dispatch queue: harness=19 (thinking_stall, addressed), retrieval=0 actionable, curiosity=1483 pending (top hle_failure: empty-prediction pre-thinking-budget-fix, consumed 2 stale items)
+- retrieval-drain: 0 ingestions
+- Action this tick: committed 50db330 fix(loop): lower repetitive-text sentence threshold 30→15 (closes #22) — "(Wait, I'll call the tool." fragments are 28 chars, previously filtered by >30 guard; now caught at >15, truncation fires after 3 repeats
+
+## 2026-05-15 16:30 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- HLE: 38/418 = 9.1% lifetime (35 runs); Math 11.4%, Bio/Med 16.7%, CS/AI 3.2%, Chem 0.0%; burndown daemon alive (43:51 elapsed)
+- vLLM 400s: 0; llamacpp-gemma4 healthy (Q3_K_M on :8000); balancer :8001 up
+- GH issues: 1 open (#21 — user on RedHat8/navygpt reports 500 JSON-parse error when write_file content exceeds max_tokens; also notes temperature 0.02 vs 1 in config.toml — not reproduced locally)
+- Dispatch queue: harness=19 (all thinking_stall, addressed by bc99b6b), retrieval=0 actionable, curiosity=611 signals / top hle_failure items = empty-prediction failures (no FINAL ANSWER output)
+- retrieval-drain: 0 ingestions (all recently consumed)
+- curiosity-drain: no action — top 3 hle_failure items have empty predictions traceable to thinking stall / FINAL ANSWER format; root cause already addressed
+- Action this tick: committed a338858 fix(recovery): detect JSON-truncation 500 (closes #21) — when llama.cpp returns HTTP 500 with 'missing closing quote' or 'parse error at column N', now injects targeted message telling model to split write_file into ≤50-line chunks instead of the generic 'malformed arguments' message
+
+## 2026-05-15 16:00 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- HLE: 38/414 = 9.2% lifetime (35 runs); Math 11.4%, Bio/Med 16.7%, CS/AI 3.2%, Chem 0.0%; burndown daemon PID 1938107 alive (14h elapsed)
+- vLLM 400s: 0 in last 30min; llamacpp-gemma4 healthy (Q3_K_M); balancer :8001 up (pid 380535, llm_balancer.py confirmed)
+- GH issues: 1 open (#21 — user on RedHat8/navygpt reports 500 JSON-parse error + temperature 0.02 instead of 1 from config.toml; JSON-parse 500 is covered by c174552 compaction routing; temperature 0.02 bug not reproduced locally — no code path produces 0.02, may be misread of 0.2 default or extra_params temperature collision; no fix this tick, needs repro)
+- Dispatch queue: harness=18 (all thinking_stall, latest 12:10 UTC, addressed by bc99b6b), retrieval=0 actionable, curiosity=1458 pending (230 hle_failure, 1228 unknown_term)
+- retrieval-drain: 0 ingestions (3 entries all recently consumed)
+- curiosity-drain: 1458 pending, no action this tick — top hle_failure items are empty-prediction failures traceable to thinking stalls + missing FINAL ANSWER format; root causes addressed by bc99b6b + f8f4f06; autonomous_review cron just started at 16:00:01 UTC
+- Action this tick: no fix committed — all queued patterns addressed by today's earlier commits; system healthy
+
+## 2026-05-15 15:30 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- HLE: 37/408 = 9.1% lifetime (34 runs); Math 11.4%, Bio/Med 16.7%, CS/AI 3.2%, Chem 0.0%, Eng 4.3%; burndown daemon PID 1926757 alive (slot 14, category=other)
+- vLLM 400s: 0; llamacpp-gemma4 healthy (Q3_K_M); balancer :8001 up
+- GH issues: 0 open (closed #20 this tick — already fixed in c174552, shipping in next auto-release; added close comment)
+- Dispatch queue: harness=19 (all thinking_stall, no new fires since 39ddef9), retrieval=0 actionable, curiosity=1436 pending (226 hle_failure, 1210 unknown_term)
+- retrieval-drain: 0 ingestions (all recently consumed)
+- curiosity-drain: 1436 pending, no action this tick — top items are empty-prediction failures; root cause traced to same thinking_budget gap below
+- Action this tick: committed fix bc99b6b (addresses pattern harness:thinking_stall) — docker logs showed budget=2147483647 on ~80% of requests even after 39ddef9, because the budget cap was only applied when thinking != "off"; llama.cpp activates reasoning with INT32_MAX by default on ALL requests; fix: always include thinking_budget in chat_template_kwargs when DRYDOCK_THINKING_BUDGET_TOKENS is set, regardless of thinking mode.
+
+## 2026-05-15 15:00 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- HLE: 36/403 = 8.9% lifetime (34 runs); Math 11.4%, Bio/Med 16.7%, CS/AI 3.2%, Chem 0.0%, Eng 4.3%; burndown daemon PID 1926757 alive (slot 14, category=other, batch in progress)
+- vLLM 400s: 0; llamacpp-gemma4 healthy (Q3_K_M, budget=8000 confirmed in docker logs); balancer :8001 up
+- GH issues: 1 open (#20 — already fixed in c174552, ships at next auto-release tick)
+- Dispatch queue: harness=19 in 24h (all harness:thinking_stall, last fire 12:10 UTC documenting the 39ddef9 fix, no new fires post-fix); retrieval=0 actionable; curiosity=1409 pending (222 hle_failure, 1187 unknown_term)
+- retrieval-drain: 0 ingestions (3 entries, all recently consumed)
+- curiosity-drain: 1409 pending, no action this tick — top items are empty-prediction failures on extremely hard questions (MIT Mystery Hunt, Go puzzles); thinking budget IS working per docker logs (budget=8000, forced-end on exhaust), these are model-capability failures on questions above the model's ceiling
+- Action this tick: no fix committed — thinking_budget fix (39ddef9) confirmed operational; thinking_stall dispatch pattern has no new fires since fix shipped; system healthy
+
+## 2026-05-15 14:30 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- HLE: 35/397 = 8.8% lifetime (33 runs); Math 11.4%, Bio/Med 16.7%, CS/AI 3.2%, Chem 0.0%, Eng 4.3%; burndown daemon PID 1917195 alive (category=other, 40+ min elapsed — stalling on 9-msg sessions)
+- vLLM 400s: 0; llamacpp-gemma4 healthy (Q3_K_M); balancer :8001 up
+- GH issues: 1 open (#20 "API error with tool call" — already fixed in c174552, pending next auto-release)
+- Dispatch queue: harness=20 in 24h (all thinking_stall, addressed by 39ddef9), retrieval=0 actionable, curiosity=1380 pending (217 hle_failure, 1163 unknown_term)
+- retrieval-drain: 0 ingestions (3 entries, all recently consumed)
+- curiosity-drain: 1380 pending, no action this tick (top items are pre-fix no_final_answer failures, will clear naturally)
+- Action this tick: committed fix f8f4f06 — hle_burndown.sh was missing DRYDOCK_TOOL_STOP_AFTER=3 and DRYDOCK_STOP_NOW_SUFFIX that hle_babysitter.sh had; burndown batches were spinning 9+ tool turns into 480s timeouts instead of capping at 3 tool calls
+
+## 2026-05-15 13:30 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- HLE: 34/387 = 8.8% lifetime (32 runs); Math 11.4%, Bio/Med 16.7%, CS/AI 3.4%, Chem 0.0%, Eng 4.3%; burndown daemon PID 1905338 alive (slot=12 Computer Science/AI batch, ~47min elapsed, still running)
+- vLLM 400s: 5 (minor); llamacpp-gemma4 healthy (Q3_K_M confirmed, budget=8000 in docker logs — 39ddef9 fix verified working); balancer :8001 up (PID 380535)
+- GH issues: 0 open
+- Dispatch queue: harness=21 in 24h (all thinking_stall, addressed by 39ddef9), retrieval=0 actionable, curiosity=1349 pending (135 hle_failure, 1140 unknown_term in 24h); dispatch report 578 signals in 24h window
+- retrieval-drain: 0 ingestions (3 entries, all recently consumed)
+- curiosity-drain: consumed item 4bd8d88fca523801 (Math empty prediction 3.9942, root cause = thinking budget unlimited; addressed by 39ddef9 chat_template_kwargs fix; total consumed: 305)
+- Action this tick: no code changes needed; 39ddef9 thinking_budget fix is current work; current batch (slot=12) predates fix and still shows 480s stalls, next batch will benefit
+
+## 2026-05-15 13:00 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- HLE: 34/384 = 8.9% lifetime (32 runs); Math 11.4%, Bio/Med 16.7%, CS/AI 3.6%, Chem 0.0%, Physics 0.0%; burndown daemon PID 1905338 alive (slot=12, Computer Science batch in progress); 135/197 failures are method=empty (thinking stalls)
+- vLLM 400s: 0; llamacpp-gemma4 healthy (Q3_K_M confirmed); balancer :8001 up; thinking budget CONFIRMED working — docker logs show "reasoning-budget: activated, budget=8000 tokens" (was INT32_MAX before 39ddef9)
+- GH issues: 0 open
+- Dispatch queue: harness=1289 total (22 in 24h window, all thinking_stall — ADDRESSED by 39ddef9), retrieval=3 (0 actionable), curiosity=1636 total (135 hle_failure:no_final_answer, 419 unknown_term in 24h window)
+- retrieval-drain: 0 new ingestions (3 entries all already consumed recently)
+- curiosity-drain: consumed item 1b7d0cc4aaf08f52 (Math blank prediction 6*sqrt(2); root cause = thinking stall, addressed by 39ddef9 chat_template_kwargs fix; total consumed: 304)
+- Action this tick: no new commit — 39ddef9 thinking_budget fix confirmed live in source tree (miniconda3 Python uses source dev path); current batch slot=12 uses fixed code; next auto-release at ~17:00 UTC will ship fix to PyPI as v2.8.38; system healthy
+
+## 2026-05-15 12:30 UTC tick
+- Stress: paused (.pause_stress sentinel)
+- Write rate: N/A (stress paused)
+- HLE: 33/378 = 8.7% lifetime (31 runs); burndown daemon alive (PID 1896475, ~34 min, running Computer Science/AI batch)
+- vLLM 400s: 0 errors; llamacpp-gemma4 healthy (Q3_K_M, model_path confirmed)
+- GH issues: 0 open
+- Dispatch queue: harness=22 (thinking_stall, root cause fixed by 39ddef9 this tick), retrieval=0 actionable, curiosity=1297 pending (1093 unknown_term, 204 hle_failure)
+- retrieval-drain: 0 new ingestions (all 3 entries already ingested)
+- curiosity-drain: consumed item 8d95bf4cc3fc45d3 (Math blank prediction pi/2-1/4; root cause = thinking stall, fixed by 39ddef9)
+- Action this tick: no new commit — 39ddef9 (thinking_budget into chat_template_kwargs for llama.cpp) already committed and live in source tree; current burndown batch is pre-fix but next batch will use correct 4000-token budget cap; system healthy, no new actionable patterns
+
 ## 2026-05-15 11:30 UTC tick
 - Stress: alive PID 1878801, 55m elapsed (restarted this cycle)
 - HLE lifetime: 33/371 = 8.9% (Math 11.4%, CS/AI 2.3%, Chem 0.0%, Bio/Med 16.7%, Eng 4.3%); latest batch 2/10 = 20% Biology/Medicine; burndown daemon PID 1742185 alive
@@ -6399,3 +6575,34 @@ restarted, cron self-match bug fixed in this same session).
 - 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
+
+## 2026-05-15 12:00 UTC tick
+- Stress: paused (.pause_stress sentinel)
+- Write rate: N/A (stress paused)
+- HLE: 33/372 = 8.9% lifetime (31 runs); burndown daemon alive (PID 1896475, ~4h elapsed)
+- vLLM 400s: 0 critical; llamacpp-gemma4 shows "unhealthy" Docker health status (healthcheck hitting wrong port 8080 vs actual 8000) — API responds normally, false alarm
+- GH issues: 0 open
+- Dispatch queue: harness=21 (thinking_stall), retrieval=0 actionable, curiosity=1263 pending (1063 unknown_term, 200 hle_failure)
+- Action this tick: fixed real bug — thinking_budget_tokens was being sent as a top-level payload field which llama.cpp silently ignores (logs showed budget=2147483647=INT32_MAX on every HLE call). Correct path is chat_template_kwargs.thinking_budget — verified empirically. This was causing all 480s+ thinking stalls despite DRYDOCK_THINKING_BUDGET_TOKENS=4000/8000 being set. Committed 39ddef9; addresses pattern harness:thinking_stall. Consumed curiosity item 7e62485e82c6849c (top hle_failure blank-prediction, root cause same). Retrieval drain: 0 actionable.
+
+## 2026-05-15 14:00 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- HLE: 35/393 = 8.9% lifetime (33 runs); Math 11.4%, Bio/Med 16.7%, CS/AI 3.2%, Chem 0.0%, Eng 4.3%; burndown daemon PID 1917195 alive (running "other" category batch, ~9m elapsed)
+- vLLM 400s: n/a (llamacpp); balancer :8001 up; model Q3_K_M confirmed
+- GH issues: 1 new (issue #20 filed 13:56 UTC — user on Red Hat 8 / navygpt hitting "Stopping: 15+ API errors" after search_replace "file path required" error)
+- Dispatch queue: harness=21 (all thinking_stall, addressed by 39ddef9), retrieval=0 actionable, curiosity=1365 pending (1152 unknown_term, 213 hle_failure)
+- retrieval-drain: 0 ingestions (3 entries, all recently consumed)
+- curiosity-drain: no action — top 3 are empty-predicted hle_failure items (symptoms of thinking_stall fixed by 39ddef9, not a separate issue)
+- Action this tick: fixed issue #20 — commit c174552. Root cause: llama.cpp backends return HTTP 500 (not 400) for context overflow; the error handler only applied aggressive compaction+truncation for 400 errors. 500s hit the fallback path that injected an error note WITHOUT truncating context, making each retry larger and recovery impossible. Fix: route "500 Internal Server Error", "status: 500", "LLM backend error" through the same compaction path as 400 errors. 3 regression tests added. Issue commented with fix details.
+
+## 2026-05-15 17:30 UTC tick
+- Stress: paused (`.pause_stress` sentinel active)
+- Write rate: N/A (stress paused)
+- HLE daemon: alive (PID 1949327, 41 min elapsed at tick start); current batch is Physics, 6/10 done, showing `no_response/no_final_answer` at 480s — expected for pre-v2.8.38 sessions where thinking="off" mode had no budget cap (bc99b6b). Future batches post-v2.8.38 have the cap applied to all modes.
+- HLE score: 38/427 = 8.9% lifetime (Math 11.4%, Bio 16.7%, CS 3.2%, Physics 3.6%, Chemistry 0.0%)
+- vLLM 400s: 0 (llama.cpp container clean)
+- GH issues: 2 open (#21, #22 — fixes committed 50db330+a338858, pending 4 AM rsync push to GitHub for auto-close)
+- Dispatch queue: harness=1290 (all addressed by bc99b6b), retrieval=3 (0 actionable), steering=N/A, curiosity=1492 pending
+- retrieval-drain: 0 ingestions (all recently consumed)
+- curiosity-drain: top items are pre-thinking-budget-fix empty-predicted failures — not actionable as code changes
+- Action this tick: no fix committed — system healthy, previous ticks (15:30, 16:30, 17:00) shipped bc99b6b+a338858+50db330 addressing all queued patterns; no new dispatch signals or GH issues since 17:00 tick

{drydock_cli-2.8.37 → drydock_cli-2.8.39}/uv.lock

@@ -326,7 +326,7 @@ wheels = [
 
 [[package]]
 name = "drydock-cli"
-version = "2.8.34"
+version = "2.8.38"
 source = { editable = "." }
 dependencies = [
     { name = "agent-client-protocol" },

drydock_cli-2.8.37/.auto_release.lock

@@ -1 +0,0 @@
-1897081