drydock-cli 2.8.58__tar.gz → 2.8.60__tar.gz

drydock_cli-2.8.60/.auto_release.lock

@@ -0,0 +1 @@
+2928182

{drydock_cli-2.8.58 → drydock_cli-2.8.60}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.8.58
+Version: 2.8.60
 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-2.8.58 → drydock_cli-2.8.60}/drydock/admiral/detectors.py

@@ -19,6 +19,7 @@ WRITE_TOOLS: frozenset[str] = frozenset({"write_file", "search_replace", "edit_f
 class Finding:
     code: str          # stable ID used for dedup, e.g. "loop:write_file:/p/x.py"
     directive: str     # text to inject into the conversation
+    force_stop: bool = False  # True → interventions.apply also sets FORCE_STOP on the loop
 
 
 def _tool_sig(tc) -> str:  # ToolCall — keep loose to avoid circular import
@@ -90,7 +91,8 @@ def detect_struggle(messages: Sequence[LLMMessage], threshold: int = 20) -> Find
     # Previously included the count, which caused 33 identical interventions
     # in a single session (each with a unique key that bypassed dedup).
     code = f"struggle:{last_write_tool or 'none'}"
-    if calls_since_write >= 30:
+    severe = calls_since_write >= 30
+    if severe:
         if last_write_tool is None:
             directive = (
                 f"Admiral: you have made {calls_since_write} tool calls without "
@@ -115,7 +117,7 @@ def detect_struggle(messages: Sequence[LLMMessage], threshold: int = 20) -> Find
             f"Admiral: you have made {calls_since_write} tool calls without "
             f"writing or editing any file. {hint}"
         )
-    return Finding(code=code, directive=directive)
+    return Finding(code=code, directive=directive, force_stop=severe)
 
 
 def run_all(messages: Sequence[LLMMessage]) -> list[Finding]:

{drydock_cli-2.8.58 → drydock_cli-2.8.60}/drydock/admiral/interventions.py

@@ -19,3 +19,9 @@ def apply(agent_loop: AgentLoop, finding: Finding) -> None:
     """Inject the finding's directive into the live conversation."""
     agent_loop._inject_system_note(finding.directive)
     history.append("intervention", f"{finding.code} :: {finding.directive[:160]}")
+    if finding.force_stop:
+        # Severe struggle (≥30 read-only calls) — advisory alone doesn't break
+        # the loop. Set FORCE_STOP so the next LLM call uses tool_choice=none,
+        # forcing the model to emit text rather than keep exploring.
+        agent_loop._loop_detected = True
+        agent_loop._loop_signal = "FORCE_STOP"

{drydock_cli-2.8.58 → drydock_cli-2.8.60}/drydock/admiral/worker.py

@@ -160,7 +160,7 @@ class AdmiralWorker:
                 "directive-source",
                 f"{finding.code} :: source={source}",
             )
-            finding_with_text = type(finding)(code=finding.code, directive=directive)
+            finding_with_text = type(finding)(code=finding.code, directive=directive, force_stop=finding.force_stop)
             interventions.apply(self.agent_loop, finding_with_text)
 
             # Watch for re-fires within the fail window to classify this

{drydock_cli-2.8.58 → drydock_cli-2.8.60}/drydock/core/agent_loop.py

@@ -1143,9 +1143,11 @@ 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 "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 "both backends failed" in error_str.lower()
                           or "500 internal server error" in error_str.lower()
                           or "status: 500" in error_str.lower()
                           or "error code: 500" in error_str.lower()
@@ -3374,7 +3376,8 @@ class AgentLoop:
                       'total_count: 0', 'retrieved 0 todos',
                       'no todos', '0 tasks', 'no tasks',
                       'no results', 'no matches', '0 matches',
-                      'no relevant information found'):
+                      'no relevant information found',
+                      '<tool_error>', 'tool_error'):
                 if p in s:
                     return True
             return False
@@ -3498,6 +3501,14 @@ class AgentLoop:
             len(sigs) >= REPEAT_WARNING_THRESHOLD
             and all(s == sigs[-1] for s in sigs[-REPEAT_WARNING_THRESHOLD:])
         ):
+            if last_tool in ("bash", "run_command"):
+                # Bash identical-call loops (e.g. repeated import checks) are
+                # never legitimate — escalate to FORCE_STOP (text-only for 1
+                # turn) so the model must summarise instead of looping.
+                # Temperature bump alone doesn't break these; the model needs
+                # to be unable to call bash.
+                self._hot_tool_path = None
+                return "FORCE_STOP"
             return f"WARNING|{last_tool}"
 
         # Check 2: Same tool called N+ times consecutively with different args

{drydock_cli-2.8.58 → drydock_cli-2.8.60}/drydock/core/tools/builtins/mcp_resources.py

@@ -152,7 +152,10 @@ class ReadMcpResource(
         servers = config.mcp_servers or []
 
         if not servers:
-            raise ToolError("No MCP servers configured. Add servers to ~/.drydock/config.toml")
+            raise ToolError(
+                "No MCP servers configured. "
+                "For local files use read_file; for shell commands use bash."
+            )
 
         from drydock.core.tools.mcp.tools import read_resource_http, read_resource_stdio
         from drydock.core.config import MCPStdio, MCPHttp
@@ -188,7 +191,8 @@ class ReadMcpResource(
                 last_error = e
                 continue
 
+        hint = " Use read_file for local paths instead." if args.uri.startswith("file://") else ""
         raise ToolError(
-            f"Could not read resource '{args.uri}' from any MCP server. "
-            f"Last error: {last_error}"
+            f"Could not read resource '{args.uri}' from any configured MCP server. "
+            f"Last error: {last_error}.{hint}"
         )

{drydock_cli-2.8.58 → drydock_cli-2.8.60}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.8.58"
+version = "2.8.60"
 description = "Local-first CLI coding agent — tested with Gemma 4 26B via vLLM"
 readme = "README.md"
 requires-python = ">=3.12"

{drydock_cli-2.8.58 → drydock_cli-2.8.60}/scripts/llm_balancer.py

@@ -22,6 +22,7 @@ import http.server
 import socket
 import sys
 import threading
+import urllib.error
 import urllib.request
 
 BACKENDS = [
@@ -86,12 +87,34 @@ class ProxyHandler(http.server.BaseHTTPRequestHandler):
                 return
         except _CLIENT_DROP_ERRORS:
             return  # client gave up, no need to failover
+        except urllib.error.HTTPError as http_err:
+            # Backend returned a valid HTTP error (4xx/5xx). Forward the
+            # original response body so drydock can parse the detailed error
+            # message (e.g. "Failed to parse tool call arguments as JSON:
+            # missing closing quote"). Failovering to an identical backend
+            # would return the same error and strips the body, breaking
+            # drydock's bad-tool-call recovery.
+            try:
+                err_body = http_err.read()
+                self.send_response(http_err.code)
+                for key, val in http_err.headers.items():
+                    if key.lower() not in ('transfer-encoding', 'connection'):
+                        self.send_header(key, val)
+                self.end_headers()
+                self.wfile.write(err_body)
+            except _CLIENT_DROP_ERRORS:
+                pass
+            except Exception:
+                _safe_send_error(self, http_err.code, str(http_err))
+            return
         except Exception as e:
             primary_err = e
 
-        # Failover: try the next backend in rotation. (Was hardcoded to
-        # `1 - idx` which only works for 2-backend pools; with 3+ pool
-        # entries that became BACKENDS[-1] = same backend at idx 2 → loop.)
+        # Failover: try the next backend in rotation. Only reached for true
+        # network errors (connection refused, timeout) — not HTTP 4xx/5xx.
+        # (Was hardcoded to `1 - idx` which only works for 2-backend pools;
+        # with 3+ pool entries that became BACKENDS[-1] = same backend at
+        # idx 2 → loop.)
         other = BACKENDS[(idx + 1) % len(BACKENDS)]
         url = f"{other}{self.path}"
         req = urllib.request.Request(url, data=body, method='POST')
@@ -105,6 +128,19 @@ class ProxyHandler(http.server.BaseHTTPRequestHandler):
                         self.send_header(key, val)
                 self.end_headers()
                 self.wfile.write(data)
+        except urllib.error.HTTPError as http_err2:
+            try:
+                err_body = http_err2.read()
+                self.send_response(http_err2.code)
+                for key, val in http_err2.headers.items():
+                    if key.lower() not in ('transfer-encoding', 'connection'):
+                        self.send_header(key, val)
+                self.end_headers()
+                self.wfile.write(err_body)
+            except _CLIENT_DROP_ERRORS:
+                pass
+            except Exception:
+                _safe_send_error(self, http_err2.code, str(http_err2))
         except _CLIENT_DROP_ERRORS:
             return
         except Exception as e2:

{drydock_cli-2.8.58 → drydock_cli-2.8.60}/scripts/stress_shakedown.py

@@ -268,7 +268,14 @@ def _env_float(name: str, default: float) -> float:
 _ANSI_ESCAPE_RE = re.compile(r"\x1b\[[0-9;?]*[a-zA-Z]")
 _RAW_MARKDOWN_PATTERNS = [
     re.compile(r"\*\*\w[^*]{1,80}\*\*"),            # **bold words**
-    re.compile(r"(?m)^#{1,6}\s+\w"),                 # ##heading
+    # `^##\s+\w` — REQUIRE 2+ `#` for the heading pattern. The earlier
+    # `^#{1,6}` matched Python code comments like `# Instantiate GCS
+    # backend` inside code blocks the model emits, causing systematic
+    # false positives (raw_md=10+ on code-heavy prompts). Real markdown
+    # H1 is rare in conversational replies — the model uses **bold** or
+    # `## H2` for emphasis/sections — so requiring 2+ #s drops the
+    # false-positive rate without missing real leaks.
+    re.compile(r"(?m)^#{2,6}\s+\w"),                 # ##heading (2+ #s)
     re.compile(r"\[[^\]]{1,40}\]\([^)]{1,80}\)"),    # [link](url)
 ]
 

{drydock_cli-2.8.58 → drydock_cli-2.8.60}/trip_log.md

@@ -1,5 +1,283 @@
 # Drydock Trip Log
 
+## 2026-05-18 16:30 UTC tick
+- Stress: idx=1420+/1658 (PID 2666472, ~14.5h uptime) — healthy and progressing fast (~11 steps/min on Perf: prompts).
+- Write rate: 0 writes on Perf: prompt category (model giving text advice, not editing). 0 writes is expected for advisory prompts; force_stop at 30 tool calls handles any read-only loops.
+- Admiral last 30 min: struggle:none fired 4× (16:13, 16:14, 16:31, 16:32) all failed at 20-24 calls — working as designed; force_stop threshold is 30 (4b095b1).
+- vLLM 400s: balancer logged transient 500/400 cascade at start of 16:30 autonomous_review tick (romulus backend down), recovered automatically — llamacpp-gemma4 Q3_K_M healthy on :8000, balancer :8001 running with 1 backend.
+- GH issues: 0 open.
+- Dispatch queue: harness=7 (thinking_stall×4 informational, bash_generic×3 addressed by 83a2400+4b095b1), retrieval=0 actionable (all ingested), curiosity=2745 pending (hle_failure+unknown_term dominant; HLE babysitter paused by sentinel).
+- Action this tick: no code change. All queued patterns addressed by commits in v2.8.59. Balancer 500 cascade was transient/self-healed. System healthy.
+
+## 2026-05-18 15:30 UTC tick
+- Stress: PID 2666472, resumed-from-step 629, 13h+ uptime — healthy (tool_agent suite, active).
+- Write rate: N/A (stress harness running; model at ~64 tok/s Q3).
+- Admiral last 30 min: struggle:search_replace 3×advisory (20-25 calls, outcome "failed" — advisory only, FORCE_STOP wires at 30+); struggle:none 2×advisory, last outcome "unstuck" at 15:27 (4b095b1 FORCE_STOP wiring effective).
+- vLLM 400s: 2 minor errors in llama.cpp container (non-fatal); balancer :8001 and model :8000 both up.
+- GH issues: 0 open.
+- Dispatch queue: harness=7 (thinking_stall×4 + bash_generic×3, all addressed by commits in v2.8.59); retrieval=0 actionable; curiosity=2745 pending (hle_failure+unknown_term dominant; HLE babysitter paused by sentinel). 5 evidence_conflict items inspected — session-level noise (model claimed completion before file existed), no discrete code fix warranted.
+- Action this tick: no code change. All dispatch patterns addressed by v2.8.59 commits. FORCE_STOP wiring from 4b095b1 confirmed effective (struggle:none resolved "unstuck" at 15:27). System healthy.
+
+## 2026-05-18 15:00 UTC tick
+- Stress: 629+/2000 (resumed run), PID 2666472 uptime 12h49m, tool_agent suite actively writing storage backends; session logs show normal tool-call flow, no loops observed
+- Write rate: ~64 tok/s (Q3 model)
+- Admiral last 30 min: 0 new fires
+- vLLM 400s: 1 minor error in last 30 min
+- GH issues: 0 open
+- Dispatch queue: harness=7 (4 thinking_stall last at 2026-05-17T18:40, 3 bash_generic last at 2026-05-17T23:50 — both patterns addressed by recent commits 83a2400/4b095b1 in v2.8.59), retrieval=0 actionable, curiosity=2744 pending (HLE paused, not actionable)
+- Balancer: PID 2842264 correctly holds :8001, proxying to :8000
+- Sentinels: HLE babysitter + burndown + vLLM failover + watchdog all paused (user-set, not touched)
+- Action this tick: no fix needed — system healthy, all recent dispatch patterns addressed by v2.8.59 commits; retrieval drain ran (0 actionable); curiosity top-3 are HLE failures (HLE paused, skipped)
+
+## 2026-05-18 14:30 UTC tick
+- Stress: 1276+/1658, PID 2666472 (~12h); session at prompt 1276 actively writing tool_agent storage backends; last complete: done=757 skip=35 timeout=0 (skip rate 4.4%, down from 33% pre-May-17)
+- Write rate: ~62 tok/s (model), session at 1276 shows raw_md=2 on "Perf: coalesce parallel reads" prompt — advisory, not a bug
+- Admiral last 30 min: no interventions logged; empty_after_tool:bash at 18:11 UTC yesterday was resolved (unstuck)
+- vLLM 400s: 0; model at 62 tok/s, no actual errors in llamacpp-gemma4 logs (55 grep hits were false positives from "update:" lines)
+- GH issues: 0 open
+- Dispatch queue: harness=7 pending (thinking_stall×4, bash_generic×3) — all historical, addressed by 83a2400/4b095b1; retrieval=0 actionable; curiosity=4 evidence_conflict pending (model false-claim patterns in storage-backend PRDs, not directly fixable with a code change) + 2744 pending mostly hle_failure (paused)
+- Action this tick: no fix committed — system healthy, all patterns addressed, model/balancer running clean
+
+## 2026-05-18 14:00 UTC tick
+- Stress: 1263/1658 (76%), PID 2666472 (11h49m); done=630, skip=3, timeout=0, RSS=1640MB (dropped from 2954MB after session recycle)
+- Write rate: not computed this tick (harness stdout not captured to accessible log after 09:01 UTC)
+- Admiral last 30 min: 0 docker errors; 0 vLLM errors
+- vLLM 400s: 0
+- GH issues: 0 open
+- Dispatch queue: harness=7 (thinking_stall×4 last seen 2026-05-17T18:40, bash_generic×3 last seen 2026-05-17T23:50 — both addressed by 83a2400 / 4b095b1); retrieval=0 actionable; curiosity top-3 all hle_failure (paused per .pause_hle_babysitter sentinel)
+- Action this tick: no fix committed — system healthy. All active patterns addressed by recent commits. Balancer (PID 2842264) and model (Q3_K_M.gguf on :8000) confirmed healthy.
+
+## 2026-05-18 13:30 UTC tick
+- Stress: 1249/1658 (75%), PID 2666472 (running 16h+); 4 failures (3 SKIPs + 1 FORCE-RESET) out of 617 completed = 0.6% skip rate; current batch is Doc: prompts (1-message responses, fast throughput)
+- Write rate: 29/617 prompts with writes (4.7% — deflated by Doc/discussion tasks with no expected file output)
+- Admiral last 30 min: 0 docker errors; 0 vLLM 400s
+- vLLM 400s: 0
+- GH issues: 0 open
+- Dispatch queue: harness=7 (thinking_stall×4 last seen 2026-05-17T18:40, bash_generic×3 last seen 2026-05-17T23:50 — both addressed by recent commits); retrieval=0 actionable (3 entries all recently ingested); curiosity=2742 pending (dominated by unknown_term + hle_failure; HLE paused per operator sentinel)
+- Action this tick: no fix committed — system healthy. All harness patterns addressed. Retrieval drain: 0 new ingest. Curiosity top-3 all hle_failure (babysitter paused). Autonomous review log shows last two ticks also found nothing actionable; most recent fix was 79971a9 (balancer HTTP error forwarding) at 12:00 UTC.
+
+## 2026-05-18 13:00 UTC tick
+- Stress: 680/1658 (PID 2666472, running; last completed run: done=757 skip=35 timeout=0, write rate ~96%)
+- Write rate: ~96% (from last complete run)
+- Admiral last 30 min: 0 fires
+- vLLM 400s: 0
+- GH issues: 0 open
+- Dispatch queue: harness=7 (2 patterns: thinking_stall informational, bash_generic addressed by 83a2400 in last 24h), retrieval=0 actionable, curiosity=2739 pending (HLE paused)
+- Action this tick: no fix committed — system healthy. All harness dispatch patterns addressed by recent commits. Retrieval drain: 0 new ingest. Curiosity top-3 all hle_failure; HLE babysitter paused per operator sentinel.
+
+## 2026-05-18 12:30 UTC tick
+- Stress: 680/1658 (PID 2666472, 10h19m uptime, tool_agent prompts, ~3 SKIPs from API port-bind prompts resolved by FORCE-RESET)
+- Write rate: N/A (log format does not report rate directly)
+- Admiral last 30 min: 0 docker errors on llamacpp-gemma4; model on Q3_K_M.gguf; balancer healthy on :8001 (PID 2842259 from last tick fix)
+- vLLM 400s: 0 in last 30m
+- GH issues: 0 open
+- Dispatch queue: harness=1312 total (7 in last 24h: thinking_stall×4=ongoing noise, bash_generic×3=addressed by 83a2400), retrieval=3 entries all recently ingested, curiosity=2737 pending (HLE paused per sentinels; top items are hle_failure, not actionable)
+- Action this tick: no fix committed. System healthy. Recent commits (83a2400 bash_generic→FORCE_STOP, 4b095b1 admiral force_stop propagation, 79971a9 balancer HTTP error forwarding) are shipping well. Thinking_stall classifier still sees 4 events/24h but existing inline-retry fix is in place; rate is background noise, not an actionable regression. Retrieval drain ran: 0 new ingest (all 3 entries recently ingested). Curiosity top-3 all hle_failure; HLE is paused per operator directive.
+
+## 2026-05-18 12:00 UTC tick
+- Stress: 629+/~2000 (PID 2666472, 9h49m uptime, tool_agent prompts)
+- Write rate: N/A
+- Admiral last 30 min: llamacpp returning 500 "Failed to parse tool call arguments as JSON: missing closing quote" when search_replace args are too large — balancer was stripping the detailed error body, causing drydock's bad-tool-call recovery to miss it and loop on the same broken tool call
+- vLLM 400s: 6+ context-overflow 400s; 17 JSON-parse 500s (all from large search_replace blocks)
+- GH issues: 0 open
+- Dispatch queue: harness=7 (thinking_stall×4=informational, bash_generic×3=addressed by 83a2400), retrieval=3 entries but 0 actionable, curiosity=2732 pending (HLE paused)
+- Action this tick: committed fix(balancer): forward HTTP error responses instead of stripping body (79971a9). The balancer was catching urllib.error.HTTPError as a generic Exception, forwarding only "HTTP Error 500: Internal Server Error" in a 502 — stripping the JSON body with the actual error. Drydock's bad-tool-call drop never fired; same broken message was resent in a loop. Fix: catch HTTPError explicitly and forward original status+body. Balancer restarted (PID 2842259), verified forwarding on :8001.
+
+## 2026-05-18 10:30 UTC tick
+- Stress: 629+/~2000 (PID 2666472, 8h19m uptime, tool_agent prompts); sessions active (latest session_20260518_101002)
+- Write rate: N/A (no report in logs this window)
+- Admiral last 30 min: balancer returning 502 (both backends 400) from context overflow in stress sessions; recovery fix 64587fa already shipped
+- vLLM 400s: 6 errors in last 30 min (context overflow, expected under load)
+- GH issues: 0 open
+- Dispatch queue: harness=7 (thinking_stall×4=informational+working, bash_generic×3=addressed by 83a2400 today), retrieval=0 actionable (all up-to-date), curiosity=2724 pending (HLE paused — not actioning)
+- Action this tick: retrieval-drain: 0 actionable; no code committed — all active patterns addressed by v2.8.59 commits; harness:loop:bash_generic escalation fix (83a2400) shipped today covers the queued bash loop evidence; system healthy
+
+## 2026-05-18 10:00 UTC tick
+- Stress: 680/1658 (PID 2666472, 7h49m uptime); at session reset, WebSocket/SSE prompts
+- Write rate: 62 writes / 620 accepted = 10% (cumulative; API-discussion prompts deflate this)
+- Admiral last 30 min: empty_after_tool:bash nudge fired and unstuck (working)
+- vLLM 400s: 69 in last 30 min — all "request exceeds available context size (32768 tokens)"; recovery catching them but not always reducing below 32K limit (emergency compact stops at `len > 20` guard, misses sessions already at 6 msgs). Fix needed: lower guard from 20 to 4, add last-2 fallback. Also: startup script should use `-c 65536` (16 GB VRAM free; current 32K too small for drydock preamble + tools). See lines 1214-1225 agent_loop.py.
+- GH issues: 0 open
+- Dispatch queue: harness=14, retrieval=3 (all consumed), curiosity=2720 pending (HLE eval banned — not actioning)
+- Action this tick: investigated context overflow root cause; retrieval-drain: 0 projects ingested (all up-to-date); no code committed — budget exhausted mid-analysis. Fixes needed: (1) agent_loop.py line 1215 guard 20→4 + last-2 fallback, (2) start_gemma4_llamacpp.sh -c 32768→65536.
+
+## 2026-05-18 09:30 UTC tick
+- Stress: 1106/1658 (PID 2666472, 7h19m uptime, 0 skips/loops/FORCE_STOPs in log); running API rate-limiter prompts
+- Write rate: N/A (test/API prompts in current batch)
+- Admiral last 30 min: 0 fires (dispatch shows 7 harness signals in last 24h, all addressed by v2.8.59 commits)
+- vLLM 400s: 0 (llama.cpp Q3_K_M on :8000; balancer :8001 healthy)
+- GH issues: 0 open
+- Dispatch queue: harness=7 (thinking_stall×4 + bash_generic×3, both addressed by 83a2400/1191431), retrieval=0 actionable, curiosity=311 pending (276 unknown_term + 31 hle_failure + 4 evidence_conflict; babysitter paused per sentinel)
+- Action this tick: no fix committed — all active dispatch patterns addressed by recent commits; retrieval drain 0 actionable; curiosity top 3 are HLE failures (babysitter paused); system healthy
+
+## 2026-05-18 08:30 UTC tick
+- Stress: ~680/1658 (PID 2666472, 6h20m uptime); 55 skips/3 timeouts by step 675 (~8% skip rate); 3 consecutive SKIPs at 677-679 triggered FORCE-RESET, step 680 in progress. Log size 469MB suggests TUI busy with JSON-RPC/SSE prompts.
+- Write rate: 62 total writes across 675 prompts (~9%)
+- Admiral last 30 min: 0 new fires (last fire 2026-05-17T23:50, harness:loop:bash_generic, addressed by 83a2400)
+- vLLM 400s: 0 (llama.cpp Q3_K_M on :8000, 44h uptime; balancer on :8001 healthy)
+- GH issues: 0 open
+- Dispatch queue: harness=7 (all addressed by v2.8.59 commits), retrieval=0 actionable, curiosity=2706 pending (dominated by unknown_term + hle_failure; babysitter paused)
+- Action this tick: no code change — all active patterns addressed by previous ticks; retrieval drain 0 actionable; system healthy
+
+## 2026-05-18 08:00 UTC tick
+- Stress: 1021/1658 (PID 2666472, 5h49m uptime, done=391/392 processed = 99.7% success, rss=1992MB growing), 0 skip/timeout
+- Write rate: 0 writes (test/regression/smoke prompts, no file writes expected)
+- Admiral last 30 min: 0 fires
+- vLLM 400s: 0 (llama.cpp Q3_K_M on :8000, balancer :8001 healthy)
+- GH issues: 0 open
+- Dispatch queue: harness=7 signals in 24h (thinking_stall×4 pre-fix, bash_generic×3 addressed by 83a2400), retrieval=0 actionable, curiosity=2705 pending (2292 unknown_term + 409 hle_failure paused + 4 evidence_conflict)
+- Action this tick: no fix committed — all dispatch patterns addressed by recent commits (83a2400 FORCE_STOP escalation, 1191431 tool-error-as-empty-result, 64587fa balancer-wrapped 400 recovery); retrieval drain 0 actionable; curiosity top 3 all hle_failure (babysitter paused per sentinel); system healthy
+
+## 2026-05-18 07:30 UTC tick
+- Stress: ~1015/1658 est. (PID 2666472, 5h19m uptime, was 978 at 07:00, ~74/hr rate), 0 skip/timeout
+- Write rate: N/A (tool_agent test-writing prompts)
+- Admiral last 30 min: 0 fires
+- vLLM 400s: 0 (llama.cpp Q3_K_M :8000, balancer :8001 — both healthy, container up 43h)
+- GH issues: 0 open
+- Dispatch queue: harness=8 (thinking_stall informational, bash_generic addressed by 83a2400), retrieval=3 entries all already ingested, curiosity=2705 pending (2292 unknown_term, 409 hle_failure paused, 4 evidence_conflict)
+- Action this tick: no fix committed — all dispatch patterns addressed; retrieval drain 0 actionable; curiosity top 3 are HLE failures (babysitter paused per operator sentinel); autonomous_review.sh running concurrently; system healthy
+
+## 2026-05-18 07:00 UTC tick
+- Stress: 978/1658, PID 2666472 (4h49m uptime), done=348, skip=0, timeout=0, recycle=0, RSS=1543MB
+- Write rate: N/A (test-writing prompts — few file writes expected)
+- Admiral last 30 min: 0 fires
+- vLLM 400s: 0 (llama.cpp Q3_K_M on :8000, balancer on :8001 — both healthy)
+- GH issues: 0 open
+- Dispatch queue: harness=8 (5 thinking_stall pre-fix, 3 bash_generic addressed by 83a2400 in v2.8.59), retrieval=3 entries all already ingested, curiosity=2704 pending (2291 unknown_term, 409 hle_failure paused, 4 evidence_conflict not actionable)
+- Action this tick: no fix committed — all dispatch patterns already addressed by v2.8.59; retrieval drain 0 actionable; curiosity top items are HLE failures (babysitter paused per operator directive); system healthy
+
+## 2026-05-18 06:30 UTC tick
+- Stress: 680/1658, PID 2666472 (4h20m uptime), --resume-from-step 629 (babysitter restarted with new resume point); 3 recent SKIPs on SSE server/client and JSON-RPC server prompts (network-server tasks binding orphan ports), then recovered
+- Write rate: N/A (API/network-server tasks)
+- Admiral last 30 min: 0 fires
+- vLLM 400s: 0 (llama.cpp Q3_K_M on :8000, balancer pid=380535 on :8001 — both healthy)
+- GH issues: 0 open
+- Dispatch queue: harness=17 thinking_stall (old, pre-fix, latest 2026-05-17T18:40), harness=180 bash_generic (addressed by 83a2400 in v2.8.59); retrieval=0 actionable; curiosity=2703 pending (unknown_term dominant, 4 evidence_conflict, HLE failures; HLE paused)
+- Action this tick: no fix committed — all dispatch patterns already addressed by v2.8.59 commits; retrieval drain 0 actionable; 4 evidence_conflict items are false-claim stress artifacts (model said "I implemented X" without verifying), not immediately actionable; system healthy
+
+## 2026-05-18 06:00 UTC tick
+- Stress: 929/1658, PID 2666472 (3h49m uptime), done=300, skip=0 this run — healthy progress (~74/hr)
+- Write rate: low (current PRDs appear to be knowledge-answer tasks, 0-1 writes/session)
+- Admiral last 30 min: 0 new harness fires; dispatch queue has 5 thinking_stall (latest 2026-05-17T18:40, all old, existing fix handles them) + 3 bash_generic (addressed 83a2400 in v2.8.59)
+- vLLM 400s: 0 (llama.cpp on :8000, no context overflow errors this run)
+- GH issues: 0 open
+- Dispatch queue: harness=8 (no new patterns), retrieval=0 actionable, curiosity=2703 pending (unknown_term dominant, hle_failure × 409 — HLE paused per sentinel)
+- Action this tick: no action — all dispatch patterns already addressed by recent commits; stress run healthy; model+balancer healthy
+
+## 2026-05-18 05:30 UTC tick
+- Stress: ~867+/1658, PID 2666472 (3h20m uptime), 0 skip/timeout this run — skip rate effectively 0% (excellent, vs 30% before v2.8.51 fixes)
+- Admiral last 30 min: 0 new harness fires; dispatch shows thinking_stall×6 + bash_generic×3 (all addressed by commits 83a2400/1191431 shipped in v2.8.59)
+- vLLM 400s: 0 new JSONDecodeErrors; llamacpp-gemma4 container up 41h, healthy
+- GH issues: 0 open
+- Dispatch queue: harness=9 (no new entries since 23:50 May 17), retrieval=0 actionable (fully consumed), curiosity=2701 pending (HLE failures + unknown_terms; HLE paused per operator)
+- Action this tick: no fix committed — system healthy; retrieval-drain 0 ingested; curiosity items are HLE-failures (paused directive); bash_generic pattern already addressed by 83a2400 in prior tick
+
+## 2026-05-18 05:00 UTC tick
+- Stress: 867/1658, PID 2666472 (2h50m uptime), progressing steadily with no SKIPs or timeouts in recent steps
+- Write rate: 0 file writes on last 5 prompts (test-type prompts answered with text, expected)
+- Admiral last 30 min: curiosity queue active (unknown_term signals at 04:51 UTC), no new harness fires since 23:50 yesterday
+- vLLM 400s: 6 minor errors in docker logs (not JSONDecodeErrors; llama.cpp Q3_K_M healthy on :8000)
+- GH issues: 0 open
+- Dispatch queue: harness=9 (thinking_stall×6 + bash_generic×3, all addressed by commits 83a2400/1191431), retrieval=0 actionable (fully consumed), curiosity=2701 pending (HLE failures + unknown_terms; babysitter paused per operator directive)
+- Action this tick: no fix committed — system healthy; retrieval-drain 0 ingested; top curiosity items are HLE-related (banned per CLAUDE.md directive); balancer confirmed on :8001 (PID 380535)
+
+## 2026-05-18 04:00 UTC tick
+- Stress: 793/1658, 164 done, 0 skip, 0 timeout (PID 2666472, 1h49m uptime, RSS=618MB)
+- Write rate: normal (sessions completing, no stalls)
+- Admiral last 30 min: 0 new fires (harness:bash_generic last at 2026-05-17T23:50, thinking_stall at 18:40)
+- vLLM 400s: 0 new (llama.cpp on :8000 healthy, Q3_K_M confirmed)
+- GH issues: 0 open
+- Dispatch queue: harness=9 total (6 thinking_stall + 3 bash_generic, all addressed by existing commits), retrieval=0 actionable (all already ingested), curiosity=2690 pending (mostly unknown_term + hle_failure; HLE paused per operator)
+- Action this tick: no fix — system healthy; retrieval-drain 0 actionable; bash_generic loop was isolated incident (one import-check loop in tool_agent session, existing FORCE_STOP detection handled it, model resumed after mute which is expected advisory-only behavior)
+
+## 2026-05-18 03:00 UTC tick
+- Stress: PID 2666472 alive (49+ min, cycling through PRDs normally — API: OpenAPI generation at tick time)
+- Write rate: sessions completing (session_025001 ran 126 messages on tool_agent/MinIO PRD)
+- Admiral last 30 min: 0 new fires (bash_generic last at 23:50, thinking_stall last at 18:40)
+- vLLM 400s: 6 context-overflow errors (33388 > 32768 tokens; emergency compaction handled; session continued)
+- GH issues: 0 open
+- Dispatch queue: harness=9 (6 thinking_stall + 3 bash_generic, both patterns already addressed by commits 1191431 / source), retrieval=0 actionable, curiosity=2674 pending (all hle_failure/unknown_term; HLE babysitter paused per operator directive)
+- Action this tick: no fix — system healthy; 2 post-v2.8.58 fixes (1191431, 77275c2) in source pending next auto-release at ~05:00 UTC; retrieval-drain 0 actionable; curiosity queue all HLE-pattern (paused)
+
+## 2026-05-18 02:30 UTC tick
+- Stress: PID 2666472 (babysitter restarted, now at step 629/1658, tool_agent PRD, healthy)
+- Write rate: N/A (session cycling normally)
+- Admiral last 30 min: thinking_stall last at 18:40, bash_generic 3x at 23:50 (addressed by 1191431)
+- vLLM 400s: 0 context errors in last 30m (1bcd5d1 search_replace truncation fix holding)
+- GH issues: 0 open
+- Dispatch queue: harness=9 (all addressed), retrieval=0 actionable, curiosity=2665 pending (409 hle_failure paused, 2254 unknown_term noise)
+- Action this tick: no action — system healthy; all dispatch patterns addressed by recent commits; curiosity queue is HLE/unknown_term noise (babysitter paused per operator directive); retrieval-drain: 0 projects ingested
+
+## 2026-05-18 02:00 UTC tick
+- Stress: PID 2501045 alive (9h+), at prompt 620/1658 (storage backend cycle, +0 writes normal for context-less follow-on prompts)
+- Write rate: 0 writes/prompt (expected post-session-reset for storage backend prompts without prior context)
+- Admiral last 30 min: thinking_stall fired once, outcome=unstuck (working correctly)
+- llamacpp 400s: 5 context overflow errors (33668 > 32768 token limit) — multiple concurrent sessions each failing once; emergency compaction truncates and recovers; no drydock bug
+- GH issues: 0 open
+- Dispatch queue: harness=10 (thinking_stall + bash_generic), retrieval=0 actionable, curiosity=2665 pending (all hle_failure/unknown_term — babysitter paused)
+- Sentinels: hle_babysitter, hle_burndown, hle_burndown_keepalive, vllm_failover, watchdog all paused
+- Action this tick: no fix committed — last tick (01:30) shipped 1191431 (bash_generic loop-break via tool-error detection); retrieval drain: 0 actionable; curiosity top 3 all HLE-failure with babysitter paused — no action taken; model+balancer healthy
+
+## 2026-05-18 01:30 UTC tick
+- Stress: PID 2501045 alive (8h34m on tool_agent PRD step 38+, sessions cycling every ~15 min — healthy)
+- Write rate: N/A
+- Admiral last 30 min: 3 bash_generic fires at 23:50 (model repeating import-check bash call; addressed by commit 1191431 last tick)
+- vLLM 400s: 0 (llamacpp-gemma4 Q3_K_M on :8000, balancer on :8001 both healthy)
+- GH issues: 0 open
+- Dispatch queue: harness=1312 total (17 thinking_stall + 3 bash_generic in last 20; bash_generic addressed; no new patterns since 23:50), retrieval=3 (all stale, already ingested), curiosity=2665 pending (409 hle_failure, 2254 unknown_term — HLE paused per directive)
+- Action this tick: no action — system healthy; retrieval-drain: 0 actionable; top curiosity items are hle_failure (babysitter paused) and unknown_term noise; bash_generic pattern already addressed by 1191431; autonomous_review started at 01:30 UTC
+
+## 2026-05-18 01:00 UTC tick
+- Stress: 562/1658 (tool_agent storage backends, PID 2501045, 9h elapsed, healthy)
+- Write rate: N/A (storage-backend add prompts produce mostly text responses)
+- Admiral last 30 min: N/A
+- vLLM 400s: 6 log lines in 30m (normal llamacpp slot-checkpoint output, not errors)
+- GH issues: 0 open
+- Dispatch queue: harness=11 (8 thinking_stall informational, 3 bash_generic from 23:48 loop — all handled by existing mechanisms; intervention on bash_generic at 23:48 failed but FORCE_STOP at 5 runs covers it), retrieval=0 actionable, curiosity=494 pending (88 hle_failure paused, 405 unknown_term noise)
+- Action this tick: no action — system healthy. 00:35 UTC autonomous_review shipped `_looks_empty` tool_error match fix (v2.8.58, already released). Bash_generic loop pattern reviewed; existing 5-run repetition loop-breaker in bash.py covers import-check loops. No new actionable bugs.
+
+## 2026-05-18 00:00 UTC tick
+- Stress: 486/1658 (tool_agent storage backend prompts, PID 2501045, 12h elapsed)
+- Write rate: 14% (7/50 last steps had write_file calls — storage backend adds expected to be low)
+- Admiral last 30 min: 22 fires (23:xx UTC window)
+- vLLM 400s: 0 (llamacpp-gemma4 Q3_K_M clean)
+- GH issues: 0 open
+- Dispatch queue: harness=11 (3 new loop:bash_generic at 23:48 from single session, outcome=failed then FORCE_STOP took over; 8 stale thinking_stall), retrieval=0 actionable, curiosity=2664 pending (HLE babysitter paused per sentinel)
+- Action this tick: no action — system healthy. loop:bash_generic pattern examined; existing 3-consecutive-identical FORCE_STOP in agent_loop.py is catching it correctly. HLE curiosity items not actionable (babysitter paused). Retrieval drain: 0 ingested.
+
+## 2026-05-17 23:30 UTC tick
+- Stress: PID 2501045 running (tool_agent PRD, 6h34m elapsed, healthy); new sessions created at 23:13, 23:15, 23:24 UTC — progressing normally
+- Write rate: N/A (no session log linked to stress pid)
+- Admiral last 30 min: N/A
+- vLLM 400s: 0 (44 grep hits are informational slot-update lines, not errors)
+- GH issues: 0 open
+- Dispatch queue: harness=8 (thinking_stall, stale — already addressed), retrieval=0 actionable (3 queue entries, all recently ingested), curiosity=2664 pending (409 hle_failure paused per operator directive, 2254 unknown_term noise)
+- Action this tick: no action — system healthy; autonomous_review.sh started its 23:30 tick; top curiosity items are hle_failure (HLE babysitter paused per CLAUDE.md 2026-05-17 directive) and unknown_term noise not worth acting on
+
+## 2026-05-17 22:30 UTC tick
+- Stress: PID 2501045 running 5h37m, stress_shakedown.py on tool_agent PRD (--resume-from-step 38), 181 sessions today; current session 48ba8dd6 writing plugin files (secret_scanning_tool.py, pii_redaction_tool.py) — healthy
+- Write rate: N/A (tool log shows writes in current session, exploration in prior)
+- Admiral last 30 min: N/A (HLE paused)
+- vLLM 400s: ~102 log lines in 6h (normal for llamacpp, not error class)
+- GH issues: 0 open
+- Dispatch queue: harness=8 (all thinking_stall, no new patterns), retrieval=0 actionable (all already ingested), curiosity=2664 pending (409 hle_failure w/ HLE paused, 2254 unknown_term noise — top items are math capability gaps, not prompt or infra bugs)
+- Action this tick: no code change — system stable; all queued harness patterns addressed by recent commits; autonomous_review started at 22:30; curiosity items are HLE capability gaps not actionable while babysitter paused
+
+## 2026-05-17 22:00 UTC tick
+- Stress: PID 2501045 running 5h+, tool_agent suite (step 38 resume), session 214835 active at 106 messages — healthy
+- vLLM 400s: 0 (model Q3_K_M, llamacpp-gemma4 healthy 33h, 62 tok/s)
+- GH issues: 0 open
+- Dispatch queue: harness=1309 (thinking_stall only, admiral nudge self-resolves per log), retrieval=3 drained (1 project ingested: 403_tool_agent), curiosity=2660 pending (HLE loops paused per directive — not actionable)
+- Action this tick: retrieval-drain: 1 project ingested; no source fix needed — system healthy
+
+## 2026-05-17 21:30 UTC tick
+- Stress: PID 2501045 running 4h34m, tool_agent stress_shakedown.py resuming from step 38 — active and healthy
+- llamacpp 400s: 0
+- GH issues: 0 open
+- Dispatch queue: harness=9 (all thinking_stall, admiral nudge resolving — latest log shows "unstuck"), retrieval=0 actionable, curiosity=2659 pending (2249 unknown_term, 409 hle_failure, 1 evidence_conflict; HLE paused per directive)
+- Model: Q3_K_M on llamacpp-gemma4 (:8000), balancer on :8001 — both healthy, 0 errors
+- Action this tick: no action — system fully healthy, no queued patterns unaddressed by recent commits
+
 ## 2026-05-17 21:00 UTC tick
 - Stress: ~680/1658 (PID 2501045, ~4h05m elapsed, tool_agent suite; consecutive SKIPs on API-shape prompts normal, FORCE-RESET recovering)
 - Write rate: tracking (no aggregated figure — recent prompts are API-heavy, SKIP-prone)
@@ -7555,3 +7833,67 @@ restarted, cron self-match bug fixed in this same session).
 - GH issues: 0 open
 - Dispatch queue: harness=7 (all stale thinking_stall, resolved), retrieval=0 actionable, curiosity=2548 pending (409 hle_failure paused, 2139 unknown_term noise)
 - Action this tick: committed fix 1bcd5d1 — removed search_replace exclusion from _truncate_old_tool_results(); old search_replace args (SEARCH/REPLACE blocks with full file content) were accumulating indefinitely past KEEP_RECENT=4 and pushing context over the 32768-token llama.cpp limit in long sessions with 20+ edits; truncated to JSON stub same as write_file args. auto_release will ship at next CDT tick.
+
+## 2026-05-17 23:00 UTC tick
+- Stress: PID 2501045 (tool_agent, step 38+, 6h+ elapsed, healthy)
+- Write rate: N/A (no session log for current run)
+- Admiral last 30 min: N/A
+- vLLM 400s: 0 (llamacpp-gemma4 Q3_K_M, no errors in last 30m)
+- GH issues: 0 open
+- Dispatch queue: harness=8 (all thinking_stall, stale — already addressed in agent_loop.py), retrieval=0 actionable, curiosity=2664 pending (409 hle_failure, 2254 unknown_term — HLE babysitter paused per operator directive)
+- Action this tick: no action — system healthy; retrieval-drain: 0 projects ingested; thinking_stall dispatch pattern already handled in source; top curiosity items are hle_failure (paused) and unknown_term noise; autonomous_review confirmed healthy at 22:30 UTC same findings
+
+## 2026-05-18 00:30 UTC tick
+- Stress: PID 2501045 alive (7h+ on tool_agent PRD, step 38+, sessions cycling normally)
+- Write rate: session dirs updating ~every 10-15 min
+- Admiral last 30 min: 3 bash_generic fires at 23:50, thinking_stall last at 18:40
+- vLLM 400s: 0 (model on Q3_K_M at :8000, balancer at :8001 healthy)
+- GH issues: 0 open
+- Dispatch queue: harness=11, retrieval=0 actionable, curiosity=2664 pending
+- Sentinels: hle_babysitter, hle_burndown, hle_burndown_keepalive, vllm_failover, watchdog all paused
+- Action this tick: committed fix — _looks_empty now matches <tool_error> so Check 0 loop detection fires after 3 repeated failing tool calls (was 8 via Check 1); also improved read_mcp_resource error messages to suggest read_file/bash when file:// URI fails with no MCP servers configured. Addresses pattern harness:loop:bash_generic. Retrieval drain: 0 actionable (already ingested). All 72 loop detection tests pass.
+
+## 2026-05-18 03:30 UTC tick
+- Stress: 749/1658
+- Write rate: N/A (current prompts are API knowledge questions with 0 writes expected)
+- Admiral last 30 min: 0 fires
+- vLLM 400s: 47 in 30 min (all context-overflow: 34K–35K token requests against 32K ctx)
+- GH issues: 0 open
+- Dispatch queue: harness=9 pending (thinking_stall×6 pre-fix, bash_generic×3 addressed by 1191431), retrieval=0 actionable, curiosity=2681 pending (mostly unknown_term/hle_failure, HLE paused)
+- Action this tick: committed fix 64587fa — context-overflow recovery was silently failing because the llm_balancer wraps backend 400s as "HTTP Error 400: Bad Request" (with colon), but the recovery condition checked for "400 bad request" (no colon). Added "400: bad request" and "both backends failed" to the elif. All other services healthy.
+
+## 2026-05-18 04:30 UTC tick
+- Stress: PID 2666472 alive (2h19m uptime)
+- Write rate: N/A (stress harness running)
+- Admiral last 30 min: 3 fires (harness:loop:bash_generic, 2026-05-17T23:50)
+- vLLM 400s: N/A (llama.cpp on :8000, balancer on :8001 — both healthy)
+- GH issues: no output (gh command returned empty)
+- Dispatch queue: harness=9, retrieval=0 actionable, curiosity=2694 pending (unknown_term dominant)
+- Action this tick: committed fix — escalate bash identical-call WARNING to FORCE_STOP (addresses pattern harness:loop:bash_generic). Model was running `python3 -c "from tool_agent.storage.factory import get_storag..."` 3+ times after import check passed; WARNING path only bumped temperature which the model ignored. 4+ identical bash calls now force text-only for 1 turn. All 72 loop-detection tests pass.
+
+## 2026-05-18 09:00 UTC tick
+- Stress: PID 2666472, tool_agent suite, resume-from-step 629, 6h49m uptime — healthy.
+- Write rate: N/A (stress harness running).
+- Admiral last 30 min: 0 new fires; last entry 2026-05-17T18:20 (empty_after_tool:bash, canned).
+- vLLM 400s: 0 (llama.cpp Q3_K_M on :8000; balancer :8001 both up).
+- GH issues: 0 open.
+- Dispatch queue: harness=7 in 24h (all addressed by 83a2400/1191431/64587fa shipped yesterday); retrieval=0 actionable; curiosity=2710 pending (hle_failure+unknown_term dominant, HLE babysitter paused by sentinel).
+- Action this tick: no code change. All harness patterns addressed by prior commits. HLE babysitter paused by operator sentinel — curiosity top items are HLE failures, not actioned. System healthy.
+
+## 2026-05-18 11:30 UTC tick
+- Stress: PID 2666472, 9h19m uptime — healthy.
+- Write rate: N/A (stress harness running).
+- Admiral last 30 min: 0 new fires.
+- vLLM 400s: 0 (llama.cpp Q3_K_M on :8000; balancer :8001 healthy).
+- GH issues: 0 open.
+- Dispatch queue: harness=7 (thinking_stall×4 informational, bash_generic×3 addressed by 83a2400), retrieval=0 actionable (all up-to-date), curiosity=2729 pending (hle_failure+unknown_term dominant; HLE babysitter paused by sentinel).
+- Action this tick: committed feat(admiral) 4b095b1 — propagate force_stop flag through struggle detector to interventions; severe struggle (>=30 read-only calls) now sets FORCE_STOP on agent loop, not just advisory nudge. All 72 loop-detection tests pass. Retrieval drain: 0 actionable.
+
+## 2026-05-18 16:00 UTC tick
+- Stress: idx=1021/1658 (PID 2666472, running ~13h50m).
+- Write rate: n/a (no stress metrics file).
+- Admiral last 30 min: 0 new fires (previous tick at 15:30 confirmed struggle:none resolved "unstuck").
+- vLLM 400s: 0 (llamacpp-gemma4 Q3_K_M on :8000; balancer :8001 healthy, 1 model).
+- GH issues: 0 open.
+- Dispatch queue: harness=7 (thinking_stall×4 informational, bash_generic×3 all addressed by 83a2400+4b095b1 in last 24h), retrieval=0 actionable, curiosity=3 pending (hle_failure×2 + unknown_term×1; HLE babysitter paused by sentinel — not actionable).
+- Action this tick: no action — all queued patterns addressed by commits already shipped in v2.8.59. System healthy.

{drydock_cli-2.8.58 → drydock_cli-2.8.60}/uv.lock

@@ -326,7 +326,7 @@ wheels = [
 
 [[package]]
 name = "drydock-cli"
-version = "2.8.49"
+version = "2.8.59"
 source = { editable = "." }
 dependencies = [
     { name = "agent-client-protocol" },

drydock_cli-2.8.58/.auto_release.lock

@@ -1 +0,0 @@
-2574982