drydock-cli 2.8.57__tar.gz → 2.8.59__tar.gz

drydock_cli-2.8.59/.auto_release.lock

@@ -0,0 +1 @@
+2716549

{drydock_cli-2.8.57 → drydock_cli-2.8.59}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.8.57
+Version: 2.8.59
 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-2.8.57 → drydock_cli-2.8.59}/drydock/core/agent_loop.py

@@ -1143,11 +1143,14 @@ 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()
                           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
@@ -3373,7 +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
@@ -3497,6 +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-2.8.57 → drydock_cli-2.8.59}/drydock/core/session/session_logger.py

@@ -53,6 +53,12 @@ class SessionLogger:
         self.save_dir.mkdir(parents=True, exist_ok=True)
         self.session_dir = self.save_folder
         self.session_metadata = self._initialize_session_metadata()
+        # Publish current session dir at init too (not just on reset)
+        # so external watchers can find the very first session.
+        try:
+            self._publish_current_session()
+        except Exception:  # noqa: BLE001
+            pass
 
     @property
     def save_folder(self) -> Path:
@@ -327,16 +333,15 @@ class SessionLogger:
         Materializes the new session_dir on disk immediately so external
         watchers (drydock-using tools, dashboards, the stress harness)
         can discover the new session BEFORE the first message gets
-        persisted. Previously the dir only appeared on the first
-        persist_messages call — a race for any watcher polling between
-        /clear and the next user message.
-
-        Also touches an empty messages.jsonl marker file. Watchers
-        commonly use the existence of messages.jsonl to identify a
-        "real" session dir (vs an in-progress one). Without the marker,
-        a watcher polling right after /clear would fall back to the
-        PREVIOUS session's dir as the most recent "real" one, and
-        miss messages routed to the new session.
+        persisted.
+
+        Also publishes the new session_dir path to
+        ~/.drydock/current_session.txt — any external tool that wants to
+        know "which drydock session is active right now" can read this
+        single file instead of doing a heuristic mtime-sorted search of
+        all ~3k+ historical session dirs. Solves the post-/clear race
+        where a watcher polling between reset_session and the OS
+        filesystem flush would cache the OLD dir as "newest".
         """
         if not self.enabled:
             return
@@ -351,6 +356,8 @@ class SessionLogger:
             # discoverable session dir immediately, not after the first
             # message lands.
             (self.session_dir / MESSAGES_FILENAME).touch(exist_ok=True)
+            # Publish the active session_dir for external watchers.
+            self._publish_current_session()
             self.log_event({
                 "event": "session_reset",
                 "session_id": session_id,
@@ -358,6 +365,24 @@ class SessionLogger:
         except Exception:  # noqa: BLE001 — never block on disk
             pass
 
+    def _publish_current_session(self) -> None:
+        """Write the active session_dir path to ~/.drydock/current_session.txt
+        so external watchers don't have to mtime-scan the whole history.
+
+        Atomic via write-then-rename. Fail-silent — publishing is an
+        observability convenience, never block the session on it.
+        """
+        if self.session_dir is None:
+            return
+        try:
+            pub_path = Path.home() / ".drydock" / "current_session.txt"
+            pub_path.parent.mkdir(parents=True, exist_ok=True)
+            tmp_path = pub_path.with_suffix(".txt.tmp")
+            tmp_path.write_text(str(self.session_dir) + "\n")
+            tmp_path.replace(pub_path)
+        except Exception:  # noqa: BLE001
+            pass
+
     def resume_existing_session(self, session_id: str, session_dir: Path) -> None:
         if not self.enabled:
             return

{drydock_cli-2.8.57 → drydock_cli-2.8.59}/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-2.8.57 → drydock_cli-2.8.59}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.8.57"
+version = "2.8.59"
 description = "Local-first CLI coding agent — tested with Gemma 4 26B via vLLM"
 readme = "README.md"
 requires-python = ">=3.12"

{drydock_cli-2.8.57 → drydock_cli-2.8.59}/scripts/shakedown_interactive.py

@@ -1386,9 +1386,27 @@ class SessionWatcher:
     def find_session(self) -> Path | None:
         if self.session_dir is not None:
             return self.session_dir
-        # meta.json is only written at session EXIT (CLAUDE.md learning #37).
-        # For active sessions we fall back to: new dir (mtime > since) that
-        # has messages.jsonl. meta.json match is preferred when available.
+        # Fast path (added 2026-05-17 with v2.8.58): drydock now
+        # publishes the active session_dir to ~/.drydock/current_session.txt
+        # at session init and on every reset_session. This avoids the
+        # post-/clear race where mtime-scan returned the stale OLD dir
+        # because reset_session's mkdir hadn't flushed yet.
+        pub_path = Path.home() / ".drydock" / "current_session.txt"
+        try:
+            published = pub_path.read_text().strip()
+            if published:
+                p = Path(published)
+                # Only accept the published path if it's recent enough —
+                # protects against a leftover file from a previous run.
+                if p.is_dir() and p.stat().st_mtime >= self.since - 5:
+                    self.session_dir = p
+                    return p
+        except (FileNotFoundError, OSError):
+            pass
+        # Fallback: meta.json is only written at session EXIT (CLAUDE.md
+        # learning #37). For active sessions we fall back to: new dir
+        # (mtime > since) that has messages.jsonl. meta.json match is
+        # preferred when available.
         fallback: Path | None = None
         for entry in sorted(SESSION_ROOT.iterdir(), reverse=True):
             try:

{drydock_cli-2.8.57 → drydock_cli-2.8.59}/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-2.8.57 → drydock_cli-2.8.59}/trip_log.md

@@ -1,5 +1,138 @@
 # Drydock Trip Log
 
+## 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)
+- llamacpp 400s: 0
+- GH issues: 0 open
+- Dispatch queue: harness=9 in last 24h (all thinking_stall, existing nudge resolves them per log evidence), retrieval=0 actionable, curiosity=2655 pending (all hle_failure/unknown_term; HLE paused per directive)
+- Model: Q3_K_M on llamacpp-gemma4 (:8000), balancer on :8001 (PID 380535) — both healthy
+- Action this tick: no action — system healthy, no new bugs, retrieval-drain: 0 ingested
+
+## 2026-05-17 20:30 UTC tick
+- Stress: 279/1658 (PID 2501045, ~3h36m elapsed, tool_agent suite, 195 done / 241 steps = ~81% write rate, 0 timeouts, 3 recycles)
+- Write rate: ~81% (done/attempted since restart at step 38)
+- Admiral last 30 min: 1019 total findings; recent-window check skipped (schema traversal error)
+- llamacpp 400s: 0 blocking 400s; 6 minor slot-management lines in last 30m (normal)
+- GH issues: 0 open
+- Dispatch queue: harness=1309 total (9 in last 24h, all thinking_stall — existing nudge logic covers), retrieval=0 actionable, curiosity=2655 pending (409 hle_failure paused, 2245 unknown_term noise)
+- Action this tick: no action — previous tick (20:00 UTC) already committed 332057d; system healthy; no new bugs found; retrieval-drain: 0 ingested (all already recent)
+
+## 2026-05-17 20:00 UTC tick
+- Stress: 238/1658 (PID 2501045, ~3h04m elapsed, tool_agent suite, running cleanly with 3 TUI recycles)
+- Write rate: N/A (no session write-rate metric available for this run)
+- Admiral last 30 min: N/A
+- vLLM 400s: 51 errors in last 60m from docker logs (mix of slot updates and JSON parse errors, none blocking)
+- GH issues: 0 open
+- Dispatch queue: harness=9 (all thinking_stall, already fixed), retrieval=0 actionable, curiosity=588 pending (473 unknown_term noise, 114 hle_failure paused, 1 evidence_conflict)
+- Action this tick: committed fix 332057d — added "error code: 500" to the 500-error recovery branch in agent_loop.py; llama.cpp returns "Error code: 500 - {..." not "500 internal server error", so the tool-call drop + truncation advice path was silently bypassed for all JSON-parse-error 500s; the bad tool call was re-sent unchanged on every retry; auto_release will ship at next CDT tick
+
+## 2026-05-17 19:30 UTC tick
+- Stress: PID 2501045, ~02:35 elapsed, tool_agent suite step ~250+/1658, running cleanly
+- Write rate: N/A
+- Admiral last 30 min: 3 harness:thinking_stall fires at 18:20, 1 at 18:40 (all informational — nudge-inject path working as designed)
+- llamacpp 400s: 6 minor errors in container logs (none JSONDecode/400-class)
+- GH issues: 0 open
+- Dispatch queue: harness=1309 (9 in last 24h, all thinking_stall resolved), retrieval=0 actionable (drain: 0 ingested), curiosity=2646 pending (top: hle_failure banned, unknown_term noise, evidence_conflict stale)
+- Action this tick: no action — system healthy; model Q3_K_M + balancer :8001 both up; stress advancing; HLE paused per directive; no actionable bugs found
+
 ## 2026-05-17 19:00 UTC tick
 - Stress: 192/1658 (PID 2501045, 02:04 elapsed, tool_agent suite, 154 done / 0 skip / 0 timeout since restart at 16:56 UTC)
 - Write rate: low (~2/92 steps show writes — normal for informational prompts on an existing package)
@@ -7519,3 +7652,40 @@ 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.

{drydock_cli-2.8.57 → drydock_cli-2.8.59}/uv.lock

@@ -326,7 +326,7 @@ wheels = [
 
 [[package]]
 name = "drydock-cli"
-version = "2.8.49"
+version = "2.8.58"
 source = { editable = "." }
 dependencies = [
     { name = "agent-client-protocol" },

drydock_cli-2.8.57/.auto_release.lock

@@ -1 +0,0 @@
-2543523