drydock-cli 2.8.0__tar.gz → 2.8.2__tar.gz

drydock_cli-2.8.2/.auto_release.lock

@@ -0,0 +1 @@
+3550152

{drydock_cli-2.8.0 → drydock_cli-2.8.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.8.0
+Version: 2.8.2
 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-2.8.0 → drydock_cli-2.8.2}/drydock/core/agent_loop.py

@@ -1038,7 +1038,12 @@ class AgentLoop:
                 and "empty path" in _prev_tool_result
             )
             # Detect if prior tool was a hallucinated/suppressed tool call.
-            _prev_was_hallucinated = "does not exist — do not call it again" in _prev_tool_result
+            # Check against the live tool registry — the "does not exist" string
+            # is only in the system note, not the tool result, so string matching fails.
+            _prev_was_hallucinated = (
+                prev_tool_name is not None
+                and prev_tool_name not in self.tool_manager.available_tools
+            )
             # Detect successful write (no error keywords in result).
             _prev_write_success = (
                 _prev_was_write
@@ -1046,6 +1051,27 @@ class AgentLoop:
                 and "Error" not in _prev_tool_result
                 and "error" not in _prev_tool_result[:50]
             )
+            # Detect bash that returned "nothing to commit" / "working tree clean"
+            # → model is done; stall nudge should say so, not "continue working".
+            _prev_bash_nothing_to_commit = (
+                prev_tool_name in ("bash", "run_command")
+                and (
+                    "nothing to commit" in _prev_tool_result
+                    or "working tree clean" in _prev_tool_result
+                    or "nothing added to commit" in _prev_tool_result
+                )
+            )
+            # Detect bash that returned a successful git commit output.
+            # Signature: "[branch hash] message\n N file(s) changed".
+            # Without this, the model stalls after commit, gets "Continue working",
+            # then re-commits — wastes a round and adds a confusing duplicate commit.
+            import re as _re
+            _prev_bash_commit_succeeded = (
+                prev_tool_name in ("bash", "run_command")
+                and bool(_re.search(r"\[[\w/]+ [0-9a-f]{4,}\]", _prev_tool_result))
+                and "file" in _prev_tool_result
+                and "changed" in _prev_tool_result
+            )
             if _stall_attempt == 0:
                 if _prev_write_path_error:
                     note = (
@@ -1073,6 +1099,18 @@ class AgentLoop:
                         "write the next file in your plan, or run bash to test what "
                         "you have built so far. Do NOT re-read files you just wrote."
                     )
+                elif _prev_bash_nothing_to_commit:
+                    note = (
+                        "The git working tree is clean — your commit already succeeded. "
+                        "The task is COMPLETE. Respond with a short summary of what you did "
+                        "and stop. Do NOT run another git commit or git add."
+                    )
+                elif _prev_bash_commit_succeeded:
+                    note = (
+                        "Your git commit succeeded. The task is COMPLETE. "
+                        "Respond with a short summary of what you changed and stop. "
+                        "Do NOT run git add or git commit again."
+                    )
                 else:
                     note = (
                         "Continue working. Use a tool (write_file, "
@@ -2675,6 +2713,13 @@ class AgentLoop:
                             pass
             if _total_tc >= 20:
                 break
+        # Consecutive check: 3+ identical bash commands in a row → FORCE_STOP.
+        # The admiral nudges at 3 consecutive identical calls; without a hard
+        # stop here the model runs 2–4 more before the 5-total check below fires.
+        # (_bash_cmds is newest-first, so [0][1][2] = last 3 in order)
+        if len(_bash_cmds) >= 3 and _bash_cmds[0] == _bash_cmds[1] == _bash_cmds[2]:
+            self._hot_tool_path = ("bash", _bash_cmds[0])
+            return "FORCE_STOP"
         if len(_bash_cmds) >= 5:
             from collections import Counter as _Counter
             _cmd_counts = _Counter(_bash_cmds)

{drydock_cli-2.8.0 → drydock_cli-2.8.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.8.0"
+version = "2.8.2"
 description = "Local-first CLI coding agent — tested with Gemma 4 26B via vLLM"
 readme = "README.md"
 requires-python = ">=3.12"

{drydock_cli-2.8.0 → drydock_cli-2.8.2}/tests/test_loop_detection.py

@@ -183,19 +183,22 @@ class TestExactRepeatDetection:
         assert al._check_tool_call_repetition() is None
 
     def test_warning_at_threshold(self):
+        # Use write_file (not bash) — bash has a tighter 3-consecutive FORCE_STOP
+        # that fires before the generic WARNING_THRESHOLD=4.
         al = _make_agent()
-        args = '{"command":"ls -ltr"}'
+        args = '{"path":"foo.py","content":"x"}'
         for _ in range(REPEAT_WARNING_THRESHOLD):
-            _add_tool_call(al, "bash", args)
+            _add_tool_call(al, "write_file", args)
         result = al._check_tool_call_repetition()
         assert result is not None
         assert "WARNING" in result
 
     def test_below_threshold_returns_none(self):
+        # Use write_file with different paths — same path triggers Check 5 at 3
+        # (single-file rewrite), and bash has a 3-consecutive FORCE_STOP.
         al = _make_agent()
-        args = '{"command":"ls -ltr"}'
-        for _ in range(REPEAT_WARNING_THRESHOLD - 1):
-            _add_tool_call(al, "bash", args)
+        for i in range(REPEAT_WARNING_THRESHOLD - 1):
+            _add_tool_call(al, "write_file", f'{{"path":"file{i}.py","content":"x"}}')
         assert al._check_tool_call_repetition() is None
 
     def test_force_stop_at_threshold(self):
@@ -367,6 +370,24 @@ class TestBashExplorationLoop:
         assert result == "FORCE_STOP"
         assert getattr(al, "_hot_tool_path", None) == ("bash", cmd)
 
+    def test_3_consecutive_identical_bash_triggers(self):
+        """3 consecutive identical bash commands → FORCE_STOP (consecutive check)."""
+        al = _make_agent()
+        cmd = 'python3 -c "import re; match = re.match(r\'([^=!<>]+)(==|!=)\', \'age>25\')"'
+        for _ in range(3):
+            _add_tool_call(al, "bash", json.dumps({"command": cmd}))
+        result = al._check_tool_call_repetition()
+        assert result == "FORCE_STOP"
+
+    def test_2_consecutive_identical_bash_no_force_stop(self):
+        """2 identical bash commands should not fire FORCE_STOP."""
+        al = _make_agent()
+        cmd = "python3 app.py"
+        for _ in range(2):
+            _add_tool_call(al, "bash", json.dumps({"command": cmd}))
+        result = al._check_tool_call_repetition()
+        assert result != "FORCE_STOP"
+
     def test_5_identical_bash_triggers(self):
         """5 identical bash calls in a row → FORCE_STOP."""
         al = _make_agent()
@@ -383,8 +404,8 @@ class TestBashExplorationLoop:
         for _ in range(4):
             _add_tool_call(al, "bash", json.dumps({"command": cmd}))
         result = al._check_tool_call_repetition()
-        # Should not trigger the bash-exploration check
-        assert result != "FORCE_STOP"
+        # 4 consecutive should still trigger the new consecutive check
+        assert result == "FORCE_STOP"
 
     def test_varied_bash_no_trigger(self):
         """Different bash commands should not trigger."""

{drydock_cli-2.8.0 → drydock_cli-2.8.2}/trip_log.md

@@ -1,5 +1,208 @@
 # Drydock Trip Log
 
+## 2026-05-07 16:47 UTC tick
+- Stress: PID 3209682 alive (>24h uptime); tool_agent prompts; idx 441/1658; done=223, skip=192 (44% SKIP rate — known TUI recycle timing issue, not a drydock bug)
+- Write rate: N/A (no progress file; session_20260507_162830 active, model reading/writing plugin files for "one-shot jobs" feature)
+- Admiral last 30 min: 4 new signals (3 harness:tool:hallucinated_name, 1 harness:bash:heredoc_loop) — all from autonomous_review.log trip_log misclassification, not live session events
+- vLLM 400s: 0; llamacpp-gemma4 "unhealthy" in docker ps but forwarding correctly; balancer PID 3175781 healthy on :8001 (1 model)
+- GH issues: 0 open
+- Dispatch queue: harness=57372, retrieval=74 (0 actionable — all current per consume_retrieval_queue); no steering queue
+- retrieval-drain: 0 projects ingested
+- Action this tick: no fix committed — all queued patterns (thinking_stall 35K, loop:bash_generic 11K, hallucinated_name 4.5K, search_replace:not_found_loop 4K, heredoc_loop 754, escape_loop 170, identical_blocks 46, tool_error_raised 25) addressed by commits in last 24h (v2.8.1 + fffaf7b); system fully healthy; no new actionable bug identified
+
+## 2026-05-07 16:01 UTC tick
+- Stress: PID 3209682 alive (1d+ uptime); stress_shakedown.py on tool_agent prompts, latest session session_20260507_155023 active and progressing (model doing search_replace, bash, write_file)
+- Write rate: N/A (no progress file)
+- Admiral last 30 min: harness:thinking_stall dominant (35810 total), harness:loop:bash_generic (11586), harness:tool:hallucinated_name (4500), harness:search_replace:not_found_loop (4107) — all pre-v2.8.1; queue entries are stale history, classifier is picking up trip_log text as signal (known issue)
+- vLLM 400s: 0; balancer PID 3175781 healthy on :8001; 0 JSONDecodeErrors in last 30m
+- GH issues: 0 open
+- Dispatch queue: harness=57218, retrieval=74 (0 ingested — all current per consume_retrieval_queue), steering=absent
+- Action this tick: no fix committed — all top patterns covered by commits in last 48h (fffaf7b, 8fd75fe, 7a119cc, a41f454); heredoc_loop and dedup_attempted entries are stale trip_log misclassifications, not live signals; system fully healthy
+
+## 2026-05-07 15:31 UTC tick
+- Stress: PID 3209682 alive (~28h uptime); no progress file; all queued patterns (thinking_stall, loop:bash_generic, hallucinated_name) addressed by v2.8.1 + fffaf7b
+- Write rate: N/A (progress file absent)
+- Admiral last 30 min: thinking_stall=440, loop:bash_generic=36, hallucinated_name=24 in last 500 dispatch entries — all from pre-fix sessions
+- vLLM 400s: 0; llamacpp-gemma4 healthy; balancer PID 3175781 healthy on :8001 (1 model)
+- GH issues: 0 open
+- Dispatch queue: harness=57065, retrieval=74 (0 ingested — all current), steering=absent
+- retrieval-drain: 0 projects ingested
+- Action this tick: no action — stall nudge code reviewed (all sub-patterns: hallucinated, read, write-success, nothing-to-commit, commit-succeeded covered); no new actionable patterns; system healthy
+
+## 2026-05-07 14:52 UTC tick
+- Stress: PID 3209682 alive; idx 416/1658; 184 SKIPs, 0 PASS/FAIL (harness TUI-recycle timing issue — session IS accepting prompts per messages.jsonl, 107 msgs in current session)
+- Write rate: n/a (no completed evaluations in this window)
+- Admiral last 30 min: 0 new fires; dispatch queue harness=56741 entries (all pre-fix, dominated by thinking_stall 35K)
+- vLLM 400s: 0; docker status "Up 2 days (unhealthy)" but responding; balancer OK on :8001 (PID 3175781)
+- GH issues: 0 open
+- Dispatch queue: harness=56741, retrieval=74 (0 actionable), steering=absent
+- retrieval-drain: 0 projects (all 74 already ingested recently)
+- Action this tick: no fix committed — all queued patterns (thinking_stall, loop:bash_generic, hallucinated_name, not_found_loop, heredoc_loop, escape_loop, identical_blocks, tool_error_raised) covered by commits in v2.8.0/v2.8.1; stress SKIP rate is a harness timing issue post-recycle, not a drydock bug
+
+## 2026-05-07 14:01 UTC tick
+- Stress: PID 3209682 alive (stress_shakedown on 403_tool_agent); balancer PID 3175781 healthy on :8001; gemma4 model responding
+- Write rate: n/a (no /tmp/stress_write_rate.txt this window)
+- Admiral last 30 min: harness:thinking_stall dominant in queue (56583 total entries); 3 commits shipped today (v2.8.1) cover all sub-cases: write success, hallucinated tools, nothing-to-commit, commit-succeeded
+- vLLM 400s: 0; balancer forwarding ok (balancer ok: gemma4)
+- GH issues: 0 open
+- Dispatch queue: harness=56583, retrieval=74 (all current, 0 ingested), steering=absent
+- retrieval-drain: 0 projects ingested (all 74 already current)
+- Action this tick: no action — healthy; all queued harness:thinking_stall patterns addressed by fffaf7b+8fd75fe+7a119cc (v2.8.1); next auto_release at 18:00 CDT will ship to PyPI
+
+## 2026-05-07 13:32 UTC tick
+- Stress: PID 3209682 alive; step 407/1658; skip rate 46% (183 skip, 213 done in this restart — elevated vs prior 34%, within known approval-modal variance)
+- Write rate: 54% done / 46% skip in current restart window
+- Admiral last 30 min: 188 thinking_stall, 16 loop:bash_generic, 8 hallucinated_name — all pre-fix sessions (fffaf7b + 8fd75fe + a41f454 + 7a119cc ship in v2.8.1, released today)
+- vLLM 400s: 0; llamacpp-gemma4 healthy; balancer PID 3175781 healthy on :8001
+- GH issues: 0 open
+- Dispatch queue: harness=55K+ (historical), retrieval=74 (0 actionable, all ingested), steering=absent
+- retrieval-drain: 0 projects ingested
+- Action this tick: no action — all queued patterns addressed by commits in v2.8.1; stress healthy and progressing; no new bugs found; no open GitHub issues
+
+## 2026-05-07 13:00 UTC tick
+- Stress: PID 3209682 alive, tool_agent prompts, resuming from step 18 (harness restarted per babysitter); ~400/1658 completed, high SKIP rate persists
+- Write rate: ~57% (207/360 at last read)
+- Admiral last 30 min: 0 new events (admiral probe last wrote at 11:58 UTC; admiral process not running); classify_pulse re-dispatching stale tail entries
+- vLLM 400s: 0 (llamacpp-gemma4 "unhealthy" in docker ps but responds correctly; balancer alive on :8001)
+- GH issues: 0 open
+- Dispatch queue: harness=56265, retrieval=74 (all 74 already ingested); dominant pattern harness:thinking_stall (34974 total, 282/hr last 2h)
+- Action this tick: no commit. 1 unreleased commit (fffaf7b: git-commit-success stall nudge) ships at next auto_release ~17:00 UTC. All active patterns addressed by v2.8.1 + fffaf7b. Retrieval drain: 0 actionable. No new actionable drydock bug found.
+
+## 2026-05-07 12:36 UTC tick
+- Stress: PID 3209682 alive at step 397/1658 (tool_agent); 174 skips so far; TUI recycling frequently (sessions created ~every 2min); new sessions exist but prompts aren't accepted — model server under load or TUI busy with concurrent HLE session; a session created at 12:34 shows HLE prompts, suggesting the SessionWatcher may be latching onto the wrong (HLE) session after recycle
+- Write rate: n/a (stress output observed at /tmp/stress_2000_v10_restart_1778067244.log)
+- Admiral last 30 min: dominant pattern harness:thinking_stall (continuous; 56K total queue entries); no new pattern types detected
+- vLLM 400s: 0 (llamacpp-gemma4 container "unhealthy" in docker ps but health endpoint returns ok; all balancer 400s come from both backends rejecting specific request formats, not a service outage)
+- GH issues: 0 open
+- Dispatch queue: harness=56106, retrieval=74 (0 newly ingested), steering=n/a
+- Action this tick: no commit. Stress is alive but SKIPping heavily; root cause appears to be TUI/SessionWatcher confusion when HLE runs concurrently. All top dispatch patterns addressed by recent commits (v2.8.1 + fffaf7b). No new actionable bug found. System alive.
+
+## 2026-05-07 12:04 UTC tick
+- Stress: PID 3209682 alive (running tool_agent prompts, --resume-from-step 18); balancer pid=3175781 healthy on :8001
+- Write rate: n/a (stress output not in reachable log; stress process running 36h, no timeout)
+- Admiral last 30 min: dominant dispatch pattern harness:thinking_stall (182 entries in last 30m); bash_generic 24 entries (all from May 6 evidence); tool:hallucinated_name 8 entries (evidence confirms fix 8fd75fe committed)
+- vLLM 400s: 0; GH issues: 0 open
+- Dispatch queue: harness=55941, retrieval=74, steering=n/a; today's totals: thinking_stall=2730, loop:bash_generic=438, tool:hallucinated_name=94
+- Retrieval drain: 0 actionable (all 74 entries already ingested)
+- Action this tick: no new commit. All top-3 patterns (thinking_stall, loop:bash_generic, tool:hallucinated_name) are addressed by v2.8.1 or the pending unreleased fix fffaf7b (detect successful git commit; ships at next auto_release ~17:00 UTC). Post-v2.8.1 sessions (after 11:02 UTC) show 0 new bash_generic fires — fix is working. System healthy.
+
+## 2026-05-07 12:00 UTC tick
+- Stress: 383/1658, PID 3209682 alive (started May 06), done=207 skip=152 recycle=109 write_rate=57.7%; current run showing consecutive SKIPs (TUI prompt-acceptance issue — known approval-modal root cause per memory)
+- Write rate: 57.7% (207 done / 359 total); vLLM 0 JSONDecodeErrors; balancer pid=3175781 healthy on :8001
+- Admiral last 30 min: 83 fires today; dominant pattern: harness:thinking_stall (55780 queue entries). Secondary recurring pattern: retry_after_error:write_file:truncated_history (fires May 04-07, existing format.py detection active but model still retries)
+- vLLM 400s: 0; GH issues: 0 open (gh returned no output)
+- Dispatch queue: harness=55780, retrieval=74, steering=n/a
+- Retrieval drain: 0 actionable (all 74 entries already ingested recently)
+- Action this tick: no new commit — harness:thinking_stall already addressed by 4 commits this tick (fffaf7b, 8fd75fe, 7a119cc, a41f454). The truncated_history retry pattern warrants investigation but converting FailedToolCall to advisory result is nontrivial and deferred to next tick or user review.
+
+## 2026-05-07 11:30 UTC tick
+- Stress: tool_agent step 18 (PID 3209682, alive); vLLM container up 2d (unhealthy health-check but serving gemma4 OK); balancer on :8001 pid 3175781 healthy; GH issues: 0 open; auto_release last ran at 06:02 CDT shipping v2.8.1
+- Dispatch queue today: thinking_stall=2458, loop:bash_generic=402, hallucinated_name=82 (last are cron log false-positives, not real tool hallucinations); retrieval-drain: 0 actionable
+- Committed fix (fffaf7b): after successful git commit, model emits empty response and gets generic "Continue working" nudge, then re-commits; new `_prev_bash_commit_succeeded` detection returns task-done nudge instead, saving 1-2 wasted turns per commit operation (addresses pattern harness:thinking_stall)
+- Action this tick: committed fix for post-commit stall loop
+
+## 2026-05-07 10:31 UTC tick
+- Stress: resuming tool_agent from step 18 (PID 3209682, alive); vLLM 0 JSONDecodeErrors; balancer healthy on :8001 (pid 3175781); GH issues: 0 open
+- Dispatch queue: harness=55458, retrieval=74 (all consumed, 0 actionable), steering=0 (totals)
+- Pattern breakdown (all-time): thinking_stall=34281, loop:bash_generic=11430, hallucinated_name=4425, search_replace:not_found_loop=4107, bash:heredoc_loop=751 — all addressed by committed fixes; latest evidence for search_replace:not_found_loop and heredoc_loop is from 2026-05-06, pre-existing patterns
+- Today's stall patterns (thinking_stall after read_file, write_file, bash, ralph_repo_index hallucination): all covered by 8fd75fe, 7a119cc, 7243bff; pending deployment at ~11:00 UTC auto_release (v2.8.1)
+- retrieval-drain: 0 projects ingested (all already current)
+- Action this tick: no fix — system healthy; reviewed all active patterns, no new unhandled bugs found
+
+## 2026-05-07 10:10 UTC tick
+- Stress: 368/1658 (PID 3209682, ~24h uptime); skip rate ~41% (known approval-modal cause); FORCE_STOP count=0 (loop:bash_generic fix working)
+- vLLM 400s: 0; balancer: healthy on :8001 (pid 3175781); GH issues: 0 open
+- Dispatch queue: harness=55295, retrieval=74 (0 actionable), steering=0 (totals)
+- Today patterns: thinking_stall=2180, loop:bash_generic=366, hallucinated_name=70 — all addressed by this morning's 3 commits (8fd75fe, 7a119cc, a41f454); no new unhandled patterns today
+- retrieval-drain: 0 projects ingested (all already current)
+- Next auto_release at ~11:00 UTC will ship v2.8.1 with today's 3 commits to PyPI
+- Action this tick: no fix — system healthy; thoroughly reviewed all queued patterns including search_replace:not_found_loop (4107 entries, all pre-existing, already addressed by 444e4a5) and bash:heredoc_loop (751 entries, historical); no new actionable bugs found
+
+## 2026-05-07 09:42 UTC tick
+- Stress: 360/1658 (PID 3209682, ~22h uptime); skip rate ~42% (known approval-modal cause)
+- Write rate: N/A (stress run measures prompts, not writes directly)
+- Admiral last 30 min: 185 fires post-09:00Z; today breakdown: thinking_stall=2051, loop:bash_generic=348, hallucinated_name=64
+- vLLM 400s: 0; balancer: healthy on :8001 (pid 3175781)
+- GH issues: 0 open
+- Dispatch queue: harness=55142, retrieval=74, steering=0 (totals)
+- retrieval-drain: 0 projects ingested (all already current)
+- All top-3 today patterns (thinking_stall, loop:bash_generic, hallucinated_name) addressed by 3 commits from this morning (8fd75fe, 7a119cc, a41f454); not yet deployed — next auto_release ~11:00 UTC; no new unhandled patterns detected
+- Action this tick: no fix — healthy; queued patterns covered by pending commits, nothing new to address
+
+## 2026-05-07 09:13 UTC tick
+- Stress: 355/1658 (PID 3209682, step advancing); balancer PID 3175781 healthy :8001 (gemma4 OK); vLLM Up 46h (unhealthy health check but functional); GH issues: 0 open
+- Dispatch queue: harness=55004 (+140 since 09:00 tick), retrieval=74 (0 actionable — all already ingested); steering=0
+- retrieval-drain: 0 projects ingested (all current)
+- Stall debug log: all attempts=0 with valid tool_calls or content — no active stall retries; inline stall retry working
+- SKIP rate last 50 prompts: 56% (known approval-modal cause per memory, not a drydock source bug)
+- All 3 queued patterns (thinking_stall, loop:bash_generic, hallucinated_name) addressed by today's 3 commits (8fd75fe, 7a119cc, a41f454); next auto_release at ~11:00 UTC will push to PyPI
+- Action this tick: no fix — system healthy, all queued patterns already addressed by today's commits; no new patterns detected
+
+## 2026-05-07 09:00 UTC tick
+- Stress: 348/1658, PID 3209682 alive; balancer PID 3175781 healthy on :8001; vLLM 400s: 0; GH issues: N/A (gh cmd no output)
+- Dispatch queue: harness=54864 (all harness:thinking_stall), retrieval=74 (0 actionable); no steering queue
+- retrieval-drain: 0 projects ingested (all already current)
+- Action this tick: committed 8fd75fe — fix _prev_was_hallucinated in stall handler; it was checking for "does not exist — do not call it again" in the tool result, but that string only appears in the system note; result was always False so hallucinated-tool stalls (ralph_repo_index etc.) fell through to the generic nudge; fixed to check prev_tool_name against self.tool_manager.available_tools; 65 tests pass
+
+## 2026-05-07 08:00 UTC tick
+- Stress: 325/1658 (idx at babysitter report), PID 3209682 alive (etime 19h25m, done=178, skip=128, timeout=0, recycle=91); write rate 58%; balancer PID 3175781 healthy on :8001 (gemma4 forwarding OK)
+- vLLM 400s: 0; GH issues: 0 open
+- Dispatch queue: harness=54575 (top: thinking_stall 33536, loop:bash_generic 11322, hallucinated_name 4395 [false-positives from autonomous_review self-classification]); retrieval=74 (0 actionable); steering absent
+- Action this tick: no fix committed — harness:thinking_stall and harness:loop:bash_generic already addressed by two commits today (7a119cc, a41f454); hallucinated_name queue entries are autonomous_review.log re-classified by classifier (not live drydock instances); system healthy, all services up
+
+## 2026-05-07 06:01 UTC tick
+- Stress: PID 3209682 alive (step unknown, session active as of 06:00:39 UTC); balancer PID 3175781 healthy on :8001; vLLM up 43h (Docker unhealthy flag but functional, 0 JSONDecodeErrors last 30m)
+- vLLM 400s: 0; GH issues: 0 open
+- Dispatch queue: harness=54169 total (harness:thinking_stall=33193, harness:loop:bash_generic=11268, harness:tool:hallucinated_name=4386), retrieval=74 (0 actionable)
+- retrieval-drain: 0 projects ingested (all already current)
+- Action this tick: committed fix a41f454 — FORCE_STOP on 3 consecutive identical bash commands (addresses pattern harness:loop:bash_generic); admiral was nudging at 3 but drydock FORCE_STOP didn't fire until 5-total, leaving 2-4 extra identical calls; new consecutive-3 check closes that window; 9 bash tests pass, 20 smoke tests pass; auto_release will ship next 0/6/12/18 tick
+
+## 2026-05-07 05:31 UTC tick
+- Stress: PID 3209682 alive, step 680/1658 (tool_agent suite, running ~20h); skip rate ~8.5% (58 SKIPs/680 prompts); 22 FORCE-RESETs from consecutive SKIPs
+- Write rate: N/A (no progress file); balancer PID 3175781 healthy on :8001; vLLM up 42h (Docker unhealthy flag but functional, 0 errors last 30m)
+- Admiral last 30 min: both harness:thinking_stall and harness:loop:bash_generic entries in queue (54043 total — historical accumulation)
+- vLLM 400s: 0 (last 30m)
+- GH issues: 0 open
+- Dispatch queue: harness=54043, retrieval=74 (0 actionable, all ingested)
+- retrieval-drain: 0 projects ingested (all 74 already current)
+- Action this tick: no action — harness:thinking_stall addressed by 7243bff (v2.8.0, shipped today); harness:loop:bash_generic has admiral nudge coverage at threshold 3, drydock nudge at 8, no drydock-level fix indicated; system healthy
+
+## 2026-05-07 05:02 UTC tick
+- Stress: PID 3209682 alive, step ~303/1658, skip rate 37% (approval modal — known), 171 accepted / 111 skipped of last 300 prompts
+- Write rate: ~60% (consistent with prior ticks); 9 writes observed in prompt 302 segment
+- Admiral last 30 min: patterns present (thinking_stall + loop:bash_generic) but all from pre-v2.8.0 sessions
+- vLLM 400s: 0; balancer :8001 healthy, vLLM :8000 healthy
+- GH issues: 0 new (poll_issues.log last entry 2026-05-06 22:00 UTC — no new issues)
+- Dispatch queue: harness=53917 (all historical), retrieval=74 (0 actionable, all ingested within 7-day window)
+- retrieval-drain: 0 projects ingested
+- Action this tick: no action — harness:thinking_stall already addressed in v2.8.0 (7243bff); harness:loop:bash_generic admiral-handled with no new source improvement identified; everything healthy
+
+## 2026-05-07 04:30 UTC tick
+- Stress: PID 3209682 alive (started 06:34 CDT May 6, ~22h); latest session dir created 04:27 UTC — actively cycling; ~300/1658 estimated
+- Write rate: ~60% (consistent with prior ticks)
+- Admiral last 30 min: harness:thinking_stall + harness:loop:bash_generic both firing (both addressed by commits 7243bff and 5fd307e respectively)
+- vLLM 400s: 0
+- GH issues: 0 open
+- Dispatch queue: harness=53,791 (top: thinking_stall, bash_generic — covered); retrieval=74 (0 actionable, all ingested)
+- retrieval-drain: 0 projects ingested (all current)
+- Action this tick: no action — system healthy; all queued patterns covered by recent commits; harness:thinking_stall addressed in last 24h (7243bff); bash loop detection and python3 -c SyntaxError handlers already in place
+
+## 2026-05-07 04:02 UTC tick
+- Stress: 292/1658 (PID 3209682, elapsed ~21.5h); done=168, skip=110 (~40% skip rate); drydock child PID 3402157 active (started 03:54 UTC, PPID confirmed = stress harness); balancer PID 3175781 on :8001 healthy; vLLM 400s=0
+- Write rate: ~60% (168/278)
+- Admiral last 30 min: 0 logged fires (admiral_probe.log empty — probe may not have fired since last cron; classify_pulse fires every 10m)
+- GH issues: 0 open
+- Dispatch queue: harness=53665, retrieval=74 (drain: 0 ingested, all already current)
+- Action this tick: no fix — stress run is alive and progressing; harness:thinking_stall addressed by 7243bff (<24h); harness:loop:bash_generic (confidence 0.6) has in-loop FORCE_STOP at 5 identical calls and admiral nudge at 3 consecutive, no additional fix warranted; retrieval drain found 0 actionable projects
+
+## 2026-05-07 03:57 UTC tick
+- Stress: 287/1658 (PID 3209682 alive); done=166, skip=102 (~38% skip rate — approval modal, known); balancer PID 3175781 on :8001 healthy; vLLM 400s=0
+- Write rate: ~62% (166/268)
+- Admiral last 30 min: 40 signals (34 thinking_stall, 6 bash_generic) from classify_pulse
+- GH issues: 0 open
+- Dispatch queue: harness=53539, retrieval=74 (0 actionable, drain=0 ingested)
+- Action this tick: no fix — all queued patterns addressed by recent commits (7243bff thinking_stall targeted nudge, 255eb4b hallucinated-tool-aware nudge); bash_generic loop still advisory-only (admiral nudge firing per queue); no new actionable bugs found
+
 ## 2026-05-07 03:03 UTC tick
 - Stress: 280/1658 (PID 3209682 alive, 15h27m elapsed); done=164, skip=97 (~37% skip rate — approval modal, known); balancer OK, vLLM 400s=0
 - Write rate: ~63% (164 done of 261 non-skip)
@@ -2679,3 +2882,42 @@ restarted, cron self-match bug fixed in this same session).
 - Dispatch queue: harness=53047 (historical pre-fix), retrieval=74 (0 actionable, all ingested)
 - retrieval-drain: 0 projects ingested
 - Action this tick: no action — all queued patterns already addressed by 255eb4b (v2.7.51); stress healthy; no new bugs found
+
+## 2026-05-07 07:01 UTC tick
+- Stress: PID 3209682 alive (~20h); currently at step 325/1658; SKIPs continue (approval-modal issue, known)
+- Write rate: N/A (progress file absent)
+- Admiral last 2h: harness:thinking_stall=434, harness:loop:bash_generic=72, harness:tool:hallucinated_name=12 (all pre-fix sessions; recent commits 7a119cc + a41f454 not yet deployed)
+- vLLM 400s: 0; llamacpp-gemma4 healthy (Up 44h, unhealthy flag but functional); balancer PID 3175781 healthy on :8001
+- GH issues: 0 open
+- Dispatch queue (24h): harness:thinking_stall=9060, harness:loop:bash_generic=591, harness:tool:hallucinated_name=60, harness:search_replace:not_found_loop=4
+- retrieval-drain: 0 projects ingested (all 74 already current)
+- Action this tick: no action — all queued patterns already addressed by commits this morning (7a119cc, a41f454, 255eb4b); stress healthy; no new GitHub issues; web_search stall (36 entries) traced to single session at 2026-05-06T20:01 UTC, not a spreading pattern; auto_release ships at 11:00 UTC
+
+## 2026-05-07 06:35 UTC tick
+- Stress: PID 3209682 alive (~19h); step count not readable (progress file absent)
+- Write rate: N/A
+- Admiral last 2h: harness:thinking_stall=420, harness:loop:bash_generic=72, harness:tool:hallucinated_name=12
+- vLLM 400s: 0; llamacpp-gemma4 healthy; balancer PID 3175781 healthy on :8001
+- GH issues: 0 open
+- Dispatch queue: harness=54295 (historical), retrieval=74 (0 actionable, all ingested), steering=absent
+- retrieval-drain: 0 projects ingested
+- Action this tick: committed fix(stall): targeted nudge when bash returns 'nothing to commit' (7a119cc, addresses pattern harness:thinking_stall). Root cause: after successful git commit, bash returns "working tree clean" and model stalls; generic "continue working" nudge caused retry of identical git commit, triggering bash_generic FORCE_STOP loop. Fix adds _prev_bash_nothing_to_commit detection and injects "task complete" nudge. Also fixed two test_loop_detection.py tests broken by a41f454's bash-specific 3-consecutive FORCE_STOP (tests used bash with same command; switched to write_file with distinct paths to isolate generic threshold). 45 tests pass.
+
+## 2026-05-07 09:32 UTC tick
+- Stress: PID 3209682 alive (~22h); stress_shakedown.py tool_agent --resume-from-step 18; no progress file
+- Write rate: N/A (progress file absent)
+- Admiral last 30 min: 123 thinking_stall, 18 loop:bash_generic, 3 hallucinated_name — all pre-fix sessions (7a119cc + a41f454 not yet in deployed wheel)
+- vLLM 400s: 0; llamacpp-gemma4 healthy; balancer PID 3175781 healthy on :8001
+- GH issues: 0 open
+- Dispatch queue (24h): harness:thinking_stall=8908, harness:loop:bash_generic=591, harness:tool:hallucinated_name=60
+- retrieval-drain: 0 projects ingested (all 74 already current)
+- Action this tick: no action — all queued patterns already addressed by 7a119cc + a41f454 (v2.8.0); stress healthy; signal counts expected to fall once auto_release ships next run
+
+## 2026-05-07 15:10 UTC tick
+- Stress: PID 3209682 alive at idx=421/1658; done=215, skip=187, recycle=127 (slow but progressing ~10 prompts/hr)
+- Write rate: n/a (progress file absent)
+- Admiral last 30 min: harness:thinking_stall dominant (56K queue total); all sub-patterns covered by v2.8.1 + fffaf7b
+- vLLM 400s: 0; balancer PID 3175781 healthy on :8001 (1 model returned)
+- GH issues: 0 open
+- Dispatch queue: harness=56902, retrieval=74 (0 ingested — all 74 already current per consume_retrieval_queue)
+- Action this tick: no fix committed. All top patterns (thinking_stall, loop:bash_generic, hallucinated_name, search_replace:not_found_loop) are addressed by recent commits; escape_loop (170 entries, all 2026-05-04) predates the FORCE_STOP fix in a41f454. Heredoc loop (751 entries) predates the targeted cat-heredoc notice in bash.py. No new actionable patterns found. System healthy.

{drydock_cli-2.8.0 → drydock_cli-2.8.2}/uv.lock

@@ -326,7 +326,7 @@ wheels = [
 
 [[package]]
 name = "drydock-cli"
-version = "2.7.52"
+version = "2.8.0"
 source = { editable = "." }
 dependencies = [
     { name = "agent-client-protocol" },

drydock_cli-2.8.0/.auto_release.lock

@@ -1 +0,0 @@
-3393464