drydock-cli 2.8.28__tar.gz → 2.8.29__tar.gz

drydock_cli-2.8.29/.auto_release.lock

@@ -0,0 +1 @@
+1502226

{drydock_cli-2.8.28 → drydock_cli-2.8.29}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.8.28
+Version: 2.8.29
 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-2.8.28 → drydock_cli-2.8.29}/drydock/core/agent_loop.py

@@ -809,7 +809,9 @@ class AgentLoop:
                 int(os.environ.get("DRYDOCK_WRAP_UP_WARN_AT", 30))))
             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"))
             _prompt_start = time.perf_counter()
+            _time_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.
@@ -846,6 +848,22 @@ class AgentLoop:
                         stopped_by_middleware=True,
                     )
                     return
+                # Time-based STOP_NOW: fires when wall-clock exceeds
+                # DRYDOCK_STOP_NOW_TIME_SEC regardless of turn count.
+                # Helps HLE batches where per-turn latency is 60-160s
+                # and the turn-based STOP_NOW fires too late or not at all.
+                if (STOP_NOW_TIME_SEC > 0
+                        and not _time_stop_injected
+                        and _elapsed > STOP_NOW_TIME_SEC):
+                    _time_stop_injected = True
+                    _stop_suffix = os.environ.get("DRYDOCK_STOP_NOW_SUFFIX", "")
+                    self._inject_system_note(
+                        f"Time limit reached: {int(_elapsed)}s elapsed on "
+                        "this single request. STOP NOW. Emit a final text "
+                        "response summarizing what you have or your best "
+                        "guess."
+                        + (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 "

{drydock_cli-2.8.28 → drydock_cli-2.8.29}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.8.28"
+version = "2.8.29"
 description = "Local-first CLI coding agent — tested with Gemma 4 26B via vLLM"
 readme = "README.md"
 requires-python = ">=3.12"

{drydock_cli-2.8.28 → drydock_cli-2.8.29}/scripts/hle_babysitter.sh

@@ -73,8 +73,12 @@ if [ -f "$PID_FILE" ] && kill -0 "$(cat "$PID_FILE")" 2>/dev/null; then
 fi
 
 # Stress harness owns the inference slot during its runs.
-if pgrep -fa "stress_.*\.py\b|stress_2000" >/dev/null 2>&1; then
-    log "skip — stress run in progress"
+# Use the PID file instead of pgrep — pgrep matches claude processes whose
+# --append-system-prompt text happens to contain "stress_2000", causing
+# false-positive skips every time autonomous_review runs.
+STRESS_PID_FILE=/tmp/stress_pid.txt
+if [ -f "$STRESS_PID_FILE" ] && kill -0 "$(cat "$STRESS_PID_FILE")" 2>/dev/null; then
+    log "skip — stress run in progress (pid $(cat "$STRESS_PID_FILE"))"
     exit 0
 fi
 

{drydock_cli-2.8.28 → drydock_cli-2.8.29}/scripts/hle_eval.py

@@ -276,6 +276,11 @@ def score_answer(q: dict, pred: str, outcome: dict | None = None) -> dict:
             return {"correct": True, "method": "exact", "verdict": "YES"}
         if fuzzy_score(pred, gold):
             return {"correct": True, "method": "fuzzy", "verdict": "YES"}
+        # Short pred (single letter/word) can't embed the right answer —
+        # skip the judge to avoid ERROR on an obviously wrong answer.
+        if len(normalize(pred)) <= 3:
+            return {"correct": False, "method": "exact", "verdict": "NO",
+                    "judge_reasoning": "short pred does not match gold"}
         # Fall through to judge for tricky MC where pred has explanation
     verdict, reasoning = judge_with_gemma(q["question"], gold, pred)
     return {
@@ -440,9 +445,15 @@ def run_one(q: dict, sk, run_dir: Path) -> dict:
            "COLUMNS": "120", "LINES": "30",
            # HLE sessions must commit quickly. Lower turn thresholds so the
            # model gets the wrap-up warning before the 480s wall-clock kills it.
-           # Default 30/60 is for coding tasks; 10/15 forces an answer sooner.
-           "DRYDOCK_WRAP_UP_WARN_AT": "10",
-           "DRYDOCK_STOP_NOW_WARN_AT": "15",
+           # Default 30/60 is for coding tasks; 8/12 forces an answer sooner.
+           # On engineering/math questions llama.cpp Q3 averages 60-160s/turn,
+           # so the turn-based STOP_NOW at 12 fires at 720-1920s — well past
+           # the 480s wall clock. STOP_NOW_TIME_SEC=300 fires at 5 minutes
+           # regardless of turn count, giving the model 3 minutes to respond.
+           # Respect env overrides (e.g. babysitter exports lower values).
+           "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_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-2.8.28 → drydock_cli-2.8.29}/trip_log.md

@@ -1,5 +1,90 @@
 # Drydock Trip Log
 
+## 2026-05-14 10:30 UTC tick
+- Stress: 1658/1658 complete (full suite finished, stress harness PID 340932 dead as expected)
+- Write rate: n/a (stress run complete)
+- vLLM 400s: 0 (llamacpp-gemma4 healthy on :8000; balancer :8001 responding)
+- GH issues: 0 open
+- Dispatch queue: harness=3 in 24h (thinking_stall, already addressed), retrieval=0 actionable, curiosity=764 pending (99 hle_failure + 665 unknown_term)
+- HLE: 5.0% lifetime (12 runs, 201 total, 10 correct); batches at 08:45 and 09:45 were silently skipped by babysitter bug
+- Action this tick: committed cb73ae2 fix(hle_babysitter) — pgrep -fa "stress_2000" false-positively matched autonomous_review's claude process (whose --append-system-prompt contains "stress_2000" in the text), blocking all HLE ticks while autonomous_review ran; replaced pgrep with PID-file kill -0 check; curiosity-queue: 764 pending, top items are empty-prediction PhD-level math — no actionable prompt fix
+
+## 2026-05-14 10:00 UTC tick
+- Stress: 1658/1658 complete (done=88, skip=53, timeout=0, recycle=42) — full suite finished; babysitter confirmed no-restart correct (idx>=total)
+- Write rate: n/a (stress run complete)
+- vLLM 400s: 0 (llamacpp-gemma4 Up 10h, healthy on :8000; docker shows "unhealthy" because old container was started with healthcheck pointing to :8080 instead of :8000 — cosmetic, model works; balancer :8001 PID=380535 confirmed llm_balancer.py)
+- GH issues: 0 open
+- Dispatch queue: harness=3 in 24h (thinking_stall, already addressed in prior commits), retrieval=0 actionable (all ingested), curiosity=764 pending (99 hle_failure + 665 unknown_term)
+- HLE: 5.0% lifetime (12 runs, 201 total, 10 correct); Math 2.9% (3/102), Bio/Med 18.8% (3/16); no active batch running at tick time
+- Action this tick: no code commit — top curiosity items are empty-prediction PhD-level math (p-adic algebraic extensions, matrix-indexed-by-subsets, 5D hypercube game theory) — capability gap not addressable by prompt rules or corpus ingest; retrieval-drain: 0 ingested (already done); stress babysitter verified complete state
+
+## 2026-05-14 09:30 UTC tick
+- Stress: 1658/1658 complete (done=88, skip=53, timeout=0, recycle=42) — full run finished, babysitter holding (correct: no restart when idx>=total)
+- Write rate: n/a (stress run complete)
+- vLLM 400s: 0 (llamacpp-gemma4 clean; :8000 and :8001 both live)
+- GH issues: 0 open
+- Dispatch queue: harness=3 (thinking_stall, already addressed), retrieval=3 entries all already ingested, curiosity=764 pending (99 hle_failure + 665 unknown_term)
+- HLE: 5.0% lifetime (12 runs, 201 total, 10 correct); Math 2.9% (3/102), Bio/Med 18.8% (3/16); no active batch (next fires at :45)
+- Action this tick: no action — system healthy; top curiosity items are empty-prediction PhD-level math (p-adic extensions, matrix subset indexing, 5D hypercube games) — capability gap, not fixable by prompt rules; autonomous_review running in parallel (started 09:30 UTC); retrieval-drain: 0 projects ingested (all already done)
+
+## 2026-05-14 09:00 UTC tick
+- Stress: 1658/1658 complete (done=88, skip=53, timeout=0, recycle=42) — full run finished, harness not restarted
+- Write rate: n/a (stress run complete)
+- Admiral last 30 min: n/a
+- vLLM 400s: 0
+- GH issues: 0 open
+- Dispatch queue: harness=3 new (thinking_stall, already addressed in agent_loop.py), retrieval=0 ingested (all recent), curiosity=765 pending (100 hle_failure + 665 unknown_term)
+- HLE: 5.0% lifetime (12 runs, 201 total, 10 correct); Math 2.9% (3/102), Bio/Med 18.8% (3/16); no active batch (PID dead, next :45 tick)
+- Action this tick: consumed curiosity:cf2cdaa2615aa0f8 (Conway/prime/GoL HLE stall — deep math, empty prediction, prompt rules already address pattern; no new fix); retrieval drain: 0 projects ingested (all already done); autonomous_review started new run at 09:00 UTC
+
+## 2026-05-14 08:30 UTC tick
+- Stress: 1658/1658 complete (done=88, skip=53, recycle=42) — suite exhausted, not a failure
+- Write rate: n/a (stress run complete)
+- Admiral last 30 min: n/a
+- vLLM 400s: 0
+- GH issues: 0 open
+- Dispatch queue: harness=1267 (all old thinking_stall entries, already fixed), retrieval=0 actionable, curiosity=765 pending (101 hle_failure + 664 unknown_term)
+- HLE: 5.0% lifetime (12 runs, 201 total, 10 correct); Math 2.9% (3/102), Bio/Med 18.8% (3/16); PID 1401018 dead, auto-restart at :45
+- Action this tick: retrieval-drain: 0 ingests; consumed curiosity:6f5be8e925adedfe (empty-prediction hle_failure — covered by time-based STOP_NOW fix f80c8ca, shipping at 11:00 UTC); no new drydock bugs found; autonomous_review running since 08:30
+
+## 2026-05-14 07:00 UTC tick
+- Stress: PID 340932 dead; babysitter invoked (no output, likely dormant — no PRDs left to run)
+- Write rate: n/a (no stress run)
+- Admiral last 30 min: n/a
+- vLLM 400s: 1 minor error in llamacpp logs
+- GH issues: 0 open
+- Dispatch queue: harness=10, retrieval=0 actionable (all ingested), curiosity=737 pending (98 hle_failure + 639 unknown_term)
+- HLE: 4.6% lifetime (12 runs, 194 total, 9 correct); current Engineering batch in progress; prior batches showed `no_final_answer, 480s, 21 msgs` — STOP_NOW at tool-turn 15 fired too late (15 turns * ~35s = 525s > 480s wall clock)
+- Action this tick: committed fix ff12396 — lowered hle_eval.py STOP_NOW threshold 15→12 tool turns (WRAP_UP 10→8) so nudge fires at ~360s, leaving 120s buffer before wall clock; also fixed babysitter env-override being silently clobbered by hle_eval.py's hardcoded dict literal. Retrieval drain: 0 projects ingested (all recent). Curiosity: consumed item 86652a02e46cefbd (hle_failure/no_final_answer timing root cause addressed by this fix).
+
+## 2026-05-14 06:00 UTC tick
+- Stress: 1658/1658 complete (babysitter dormant; no stress PID in /tmp/stress_pid.txt)
+- HLE: 4.7% lifetime (11 runs, 191 total, 9 correct); 04:45 batch (CS/AI) final: 0/10 — 4x empty:no_response (480s, 1 msg each), 4x empty:no_final_answer, 1 wrong answer, 1 judge ERROR
+- vLLM 400s: 0 (llamacpp-gemma4 Q3 healthy; Docker health=unhealthy is false alarm — health check uses port 8080 but container is on 8000)
+- GH issues: 0 open
+- Dispatch queue: harness=2 (thinking_stall × 2, previously addressed), retrieval=0 actionable, curiosity=717 pending (95 hle_failure + 622 unknown_term)
+- Retrieval drain: 0 projects ingested (all already current)
+- Action this tick: no fix committed. The 04:45 batch ran AFTER 85d1a70 (WRAP_UP=10/STOP_NOW=15 fix) but still shows 4 no_response failures — root cause is model generating thinking-only responses on its first turn (msg_count=1 after 480s), which the STOP_NOW nudge cannot reach because it only fires after N tool calls. Top curiosity items (finite groups, BVP, spy-notebook) are hard math failures with empty predictions, consistent with no_response pattern; not retrieval gaps. No new actionable fix identified this tick without deeper investigation of thinking suppression for first-turn HLE calls.
+
+## 2026-05-14 05:30 UTC tick
+- Stress: 1658/1658 complete (run finished; babysitter correctly dormant; not a death)
+- HLE: 4.8% lifetime (11 runs, 187 total, 9 correct); batch PID 1353401 running (45 min elapsed)
+- vLLM 400s: 0 (llamacpp-gemma4 Q3 healthy)
+- GH issues: 0 open
+- Dispatch queue: harness=2 (thinking_stall, already addressed), retrieval=0 actionable, curiosity=706 pending (91 hle_failure + 615 unknown_term)
+- Retrieval drain: 0 projects ingested (all already current)
+- Action this tick: no fix committed — top curiosity items are hard math questions (group theory, ODE, combinatorics) where retrieval/prompt changes cannot help; unknown_term noise is assembly opcodes (MUL/DIV/JUMP/CALL/RET) from stress PRD; system healthy
+
+## 2026-05-14 05:00 UTC tick
+- Stress: 1658/1658 complete (babysitter dormant; next restart at top of hour by cron)
+- HLE: 4.9% lifetime (11 runs, 183 total, 9 correct). 04:45 batch (Computer Science/AI) still in progress — 2/10 visible: 1 correct(B/E mix-up), 1 empty:no_response (480s, 1 msg — JS deobfuscation puzzle, model produced no content after thinking tokens)
+- vLLM 400s: 0 (llamacpp-gemma4 Q3 healthy)
+- GH issues: 0 open
+- Dispatch queue: harness=2 (thinking_stall, addressed by 85d1a70), retrieval=0 actionable, curiosity=688 pending (87 hle_failure + 601 unknown_term)
+- Retrieval drain: 0 projects ingested (all already current)
+- Autonomous review: 04:00Z tick hit budget; 04:30Z tick committed 85d1a70 (HLE turn limit fix); 05:00Z tick started concurrently with this cron tick
+- Action this tick: no fix committed — system healthy, AR covering curiosity drain, no new actionable patterns found
+
 ## 2026-05-14 04:30 UTC tick
 - Stress: 1658/1658 complete (run finished, babysitter correctly dormant)
 - HLE: 5.0% lifetime (10 runs, 181 total, 9 correct). Chemistry batch 024501 shows 0/10: all `empty:no_response` or `empty:no_final_answer` at 480s wall-clock timeout. Root cause: model used 11-60 tool calls searching web but STOP_NOW warning (default turn 60) never fired within 480s budget.
@@ -5800,3 +5885,23 @@ restarted, cron self-match bug fixed in this same session).
 - retrieval-drain: 0 new ingestions
 - curiosity-queue: 3 pending hle_failure items, pure knowledge gaps, no actionable fix this tick
 - Action this tick: committed fix for hle_eval._extract_answer — now handles \boxed{}, $...$, markdown emphasis on FINAL ANSWER, multipart content lists, trailing code fences (a74b484). Pre-existing unstaged changes, not authored this tick.
+
+## 2026-05-14 06:35 UTC tick
+- Stress: 1658/1658 (completed full suite — babysitter correctly dormant; PID 340932 dead, was expected completion not failure)
+- Write rate: 88/1658 done, 53 skip
+- Admiral last 30 min: N/A (stress complete, no active sessions)
+- vLLM 400s: 0 (llamacpp-gemma4 healthy, Q3_K_M, balancer :8001 healthy)
+- GH issues: 0 open
+- Dispatch queue: harness=2 (thinking_stall, already addressed), retrieval=3 (0 actionable, all ingested), curiosity=717 pending (top 3: hle_failure — group theory/PDE/combinatorics, pure knowledge gaps)
+- retrieval-drain: 0 new ingestions
+- curiosity-queue: 717 pending; top items are genuinely hard math (no_response pattern = thinking stall, no_final_answer pattern = hits 480s wall before STOP_NOW turn-limit fires at turn 15); identified fix: lower DRYDOCK_STOP_NOW_WARN_AT from 15→8 in hle_eval.py so FINAL ANSWER nudge fires at ~turn 8 (~480s) rather than turn 15 (~750s). HLE latest batch: 0/10 CS/AI, 4 no_response (thinking stall), 3 no_final_answer (timeout before nudge), 1 judge error. Overall HLE: 9/191 = 4.7%.
+- Action this tick: no commit — restarted stress babysitter (found PID 340932 dead but babysitter was correctly treating it as completed run, not failure); identified STOP_NOW timing bug in HLE eval for next tick to fix.
+
+## 2026-05-14 07:30 UTC tick
+- Stress: watchdog paused (no stress idx available)
+- Write rate: N/A (stress paused)
+- Admiral last 30 min: N/A
+- vLLM 400s: 0 (llama.cpp Q3 healthy, balancer on :8001 serving)
+- GH issues: 0 open
+- Dispatch queue: harness=1267, retrieval=3, steering=N/A, curiosity=1016 pending (totals)
+- Action this tick: committed feat(agent_loop): time-based STOP_NOW for slow-turn HLE sessions (f80c8ca). Engineering HLE batch (07:30 UTC): 4/7 questions hit 480s wall clock with empty predictions — per-turn latency 60-160s means turn-based STOP_NOW at 12 never fires before timeout. Added DRYDOCK_STOP_NOW_TIME_SEC env var (fires at wall-clock threshold regardless of turns); hle_eval.py sets it to 300s so model gets a FINAL ANSWER nudge at 5 min, 3 min before wall clock. Retrieval drain: 0 ingested (all already current). Consumed curiosity item 1bcaa8f5aa40216a (cellular automaton — computation gap, not retrieval gap). HLE babysitter running (PID 1401018, 45+ min, engineering category). Lifetime HLE: 9/197 = 4.6%.

drydock_cli-2.8.28/.auto_release.lock

@@ -1 +0,0 @@
-1359290