drydock-cli 2.7.23__tar.gz → 2.7.25__tar.gz

drydock_cli-2.7.25/.auto_release.lock

@@ -0,0 +1 @@
+691415

{drydock_cli-2.7.23 → drydock_cli-2.7.25}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.7.23
+Version: 2.7.25
 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-2.7.23 → drydock_cli-2.7.25}/drydock/core/agent_loop.py

@@ -1152,10 +1152,13 @@ class AgentLoop:
         # model a growing signal that nothing has changed.  Preserve last_result
         # so the message always shows the real bash/tool output, not a prior NOTE.
         self._tool_call_history[sig] = (count + 1, last_result)
+        # For read-only tools include the full cached content so the model
+        # has the data it needs and doesn't retry just to see more output.
+        result_preview = last_result if is_readonly else last_result[:200]
         return (
             f"NOTE: this exact call to `{tool_name}` has been made "
             f"{count} times this session with identical arguments. "
-            f"Last result: {last_result[:200]}\n\n"
+            f"Last result:\n{result_preview}\n\n"
             f"The result will not change on a {count + 1}th attempt. "
             f"Move on — call a DIFFERENT tool, use DIFFERENT arguments, "
             f"or end your turn with a text summary so the user can "
@@ -1227,7 +1230,12 @@ class AgentLoop:
             f"{tool_call.tool_name}:{args_str}".encode()
         ).hexdigest()
         count, _ = self._tool_call_history.get(sig, (0, ""))
-        self._tool_call_history[sig] = (count + 1, result_text[:500])
+        # Store more content for read-only tools so the NOTE advisory
+        # can include enough context for the model to act on it.
+        tool_name = tool_call.tool_name
+        is_readonly = tool_name in ("grep", "read_file", "glob", "ls")
+        store_limit = 2000 if is_readonly else 500
+        self._tool_call_history[sig] = (count + 1, result_text[:store_limit])
 
     async def _process_one_tool_call(
         self, tool_call: ResolvedToolCall
@@ -1523,6 +1531,13 @@ class AgentLoop:
                 self.format_handler.create_failed_tool_response_message(failed, error_msg)
             )
             self.stats.tool_calls_failed += 1
+            # Inject a [SYSTEM: ...] note (the format admiral uses) so the model
+            # is more likely to break out of the empty-response loop that often
+            # follows a suppressed hallucinated-tool call.
+            self._inject_system_note(
+                f"'{failed.tool_name}' does not exist — do NOT call it again. "
+                "Call glob, grep, or read_file NOW to make progress."
+            )
 
     async def _handle_tool_calls(
         self, resolved: ResolvedMessage

{drydock_cli-2.7.23 → drydock_cli-2.7.25}/drydock/core/llm/format.py

@@ -368,9 +368,11 @@ class APIToolFormatHandler:
                             tool_name=parsed_call.tool_name,
                             call_id=parsed_call.call_id,
                             error=(
-                                f"'{parsed_call.tool_name}' is not available. "
-                                f"Use only tools from your tool list: "
-                                f"{', '.join(sorted(active_tools.keys())[:8])}."
+                                f"'{parsed_call.tool_name}' does not exist — stop calling it. "
+                                f"Use glob/grep/read_file to explore the project instead. "
+                                f"Available tools: "
+                                f"{', '.join(sorted(active_tools.keys())[:8])}. "
+                                f"Call one of these NOW."
                             ),
                         )
                     )

{drydock_cli-2.7.23 → drydock_cli-2.7.25}/drydock/core/tools/builtins/search_replace.py

@@ -236,6 +236,30 @@ class SearchReplace(
                 # (would be catastrophic data loss). See CLAUDE.md learning
                 # #39: drydock nuked a 5171-char cli.py with a 16-line fragment.
                 raw_content = err_msg[len("NO_BLOCKS:"):]
+                # Guard: if the raw content contains SEARCH/REPLACE markers, it's
+                # a malformed block (e.g. missing >>>>>>> REPLACE closer). Writing
+                # it verbatim would corrupt the file with conflict markers, causing
+                # a retry loop. Return an error instead.
+                if "<<<<<<<" in raw_content or ">>>>>>>" in raw_content:
+                    yield SearchReplaceResult(
+                        file=args.file_path.strip() or "(unknown)",
+                        blocks_applied=0,
+                        lines_changed=0,
+                        warnings=[],
+                        content=(
+                            "PARSE ERROR: your search_replace content looks like a "
+                            "SEARCH/REPLACE block but it could not be parsed — likely "
+                            "the >>>>>>> REPLACE closer is missing or the markers are "
+                            "malformed. Fix the format:\n"
+                            "<<<<<<< SEARCH\n"
+                            "[exact text to find]\n"
+                            "=======\n"
+                            "[replacement text]\n"
+                            ">>>>>>> REPLACE\n"
+                            "Do NOT send conflict markers as literal file content."
+                        ),
+                    )
+                    return
                 file_path_str = args.file_path.strip()
                 if file_path_str and len(raw_content) > 20:
                     target = Path(file_path_str).expanduser()

{drydock_cli-2.7.23 → drydock_cli-2.7.25}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.7.23"
+version = "2.7.25"
 description = "Local-first CLI coding agent — tested with Gemma 4 26B via vLLM"
 readme = "README.md"
 requires-python = ">=3.12"

{drydock_cli-2.7.23 → drydock_cli-2.7.25}/tests/test_loop_detection.py

@@ -95,13 +95,19 @@ class TestCircuitBreaker:
             al._circuit_breaker_record(tc, "output ok")
         assert al._circuit_breaker_check(tc) is None
 
-    def test_successful_readonly_never_blocked(self):
-        """Read-only successful tools are never blocked."""
+    def test_readonly_note_includes_full_result(self):
+        """Read-only tools that repeat 5+ times get a NOTE that includes the
+        full cached result so the model can act on it without retrying."""
         al = _make_agent()
         tc = SimpleNamespace(tool_name="grep", args_dict={"pattern":"foo"})
+        content = "match.py:10:foo"
         for _ in range(10):
-            al._circuit_breaker_record(tc, "match.py:10:foo")
-        assert al._circuit_breaker_check(tc) is None
+            al._circuit_breaker_record(tc, content)
+        result = al._circuit_breaker_check(tc)
+        assert result is not None
+        # The full cached content must appear in the NOTE (not just 200 chars)
+        assert content in result
+        assert "NOTE:" in result
 
     def test_disabled_returns_none_even_for_failed(self):
         """CB disabled — returns None even for commands that would have been blocked."""
@@ -148,6 +154,19 @@ class TestCircuitBreaker:
         al._circuit_breaker_record(tc, "output ok")
         assert al._circuit_breaker_check(tc) is None
 
+    def test_read_file_note_stores_2000_chars(self):
+        """read_file NOTE stores up to 2000 chars of content so large files are
+        visible in the advisory — not truncated to 500 like bash results."""
+        al = _make_agent()
+        tc = SimpleNamespace(tool_name="read_file", args_dict={"path": "foo.py"})
+        long_content = "x" * 1800  # within 2000-char limit
+        for _ in range(6):
+            al._circuit_breaker_record(tc, long_content)
+        result = al._circuit_breaker_check(tc)
+        assert result is not None
+        # The full 1800-char content must be present in the NOTE
+        assert long_content in result
+
 
 # ============================================================================
 # Check 1: Exact Same Tool Call Repeated (REPEAT_WARNING_THRESHOLD)

drydock_cli-2.7.25/tests/tools/test_search_replace_malformed_block.py

@@ -0,0 +1,131 @@
+"""Regression test: search_replace must not write conflict markers to files.
+
+Observed in stress run 2026-04-30: when the model sends a truncated
+SEARCH/REPLACE block (e.g. missing >>>>>>> REPLACE closer), the regex fails
+to parse it, and the NO_BLOCKS fallback wrote the raw content — including
+'<<<<<<< SEARCH' markers — directly to the file. On the next turn the model
+found markers in the file and entered a retry loop trying to find original
+code that no longer existed.
+
+Fix: in NO_BLOCKS path, detect SEARCH/REPLACE marker characters in raw_content
+and return an error instead of writing the content to disk.
+"""
+from __future__ import annotations
+
+from pathlib import Path
+
+import pytest
+
+from drydock.core.tools.base import BaseToolState, InvokeContext
+from drydock.core.tools.builtins.search_replace import (
+    SearchReplace,
+    SearchReplaceArgs,
+    SearchReplaceConfig,
+    SearchReplaceResult,
+)
+
+
+@pytest.fixture
+def tool():
+    return SearchReplace(SearchReplaceConfig(), BaseToolState())
+
+
+@pytest.fixture
+def ctx() -> InvokeContext:
+    return InvokeContext(tool_call_id="tc_1", read_file_state={})
+
+
+async def _call(tool, args: SearchReplaceArgs, ctx: InvokeContext) -> SearchReplaceResult:
+    results = []
+    async for r in tool.run(args, ctx):
+        if isinstance(r, SearchReplaceResult):
+            results.append(r)
+    assert results, "Tool produced no output"
+    return results[-1]
+
+
+@pytest.mark.asyncio
+async def test_truncated_block_no_replace_closer(tmp_path: Path, tool, ctx):
+    """Malformed block missing >>>>>>> REPLACE must NOT be written to disk."""
+    target = tmp_path / "cli.py"
+    target.write_text("def main():\n    print('hello')\n")
+
+    # Simulate: model sends SEARCH/REPLACE without the closing >>>>>>> REPLACE
+    malformed = (
+        "<<<<<<< SEARCH\n"
+        "def main():\n"
+        "    print('hello')\n"
+        "=======\n"
+        "def main():\n"
+        "    print('world')\n"
+    )
+
+    ctx.read_file_state[str(target)] = {
+        "content": target.read_text(),
+        "timestamp": target.stat().st_mtime_ns,
+        "offset": 0,
+        "limit": None,
+    }
+    result = await _call(
+        tool,
+        SearchReplaceArgs(file_path=str(target), content=malformed),
+        ctx,
+    )
+
+    # Must not write anything — the malformed block is rejected
+    assert result.blocks_applied == 0
+    # The file must be unchanged regardless of the error path taken
+    assert target.read_text() == "def main():\n    print('hello')\n"
+
+
+@pytest.mark.asyncio
+async def test_content_with_only_open_marker(tmp_path: Path, tool, ctx):
+    """Content containing only '<<<<<<< SEARCH' (no equals/close) must not write."""
+    target = tmp_path / "server.py"
+    target.write_text("class Server:\n    pass\n")
+
+    ctx.read_file_state[str(target)] = {
+        "content": target.read_text(),
+        "timestamp": target.stat().st_mtime_ns,
+        "offset": 0,
+        "limit": None,
+    }
+    result = await _call(
+        tool,
+        SearchReplaceArgs(file_path=str(target), content="<<<<<<< SEARCH\nclass Server:\n    pass\n"),
+        ctx,
+    )
+
+    assert result.blocks_applied == 0
+    assert target.read_text() == "class Server:\n    pass\n"
+
+
+@pytest.mark.asyncio
+async def test_valid_block_still_works(tmp_path: Path, tool, ctx):
+    """A well-formed SEARCH/REPLACE block must still apply correctly."""
+    target = tmp_path / "app.py"
+    target.write_text("def greet():\n    return 'hello'\n")
+
+    ctx.read_file_state[str(target)] = {
+        "content": target.read_text(),
+        "timestamp": target.stat().st_mtime_ns,
+        "offset": 0,
+        "limit": None,
+    }
+    valid_block = (
+        "<<<<<<< SEARCH\n"
+        "def greet():\n"
+        "    return 'hello'\n"
+        "=======\n"
+        "def greet():\n"
+        "    return 'world'\n"
+        ">>>>>>> REPLACE\n"
+    )
+    result = await _call(
+        tool,
+        SearchReplaceArgs(file_path=str(target), content=valid_block),
+        ctx,
+    )
+
+    assert result.blocks_applied == 1
+    assert "world" in target.read_text()

{drydock_cli-2.7.23 → drydock_cli-2.7.25}/trip_log.md

@@ -3,6 +3,97 @@
 Autonomous Claude Code review ticks while the user is away. Each tick appended
 chronologically. Cron-driven every 30 min from `/data3/drydock/scripts/autonomous_review.sh`.
 
+## 2026-04-30 14:04 UTC tick
+- Stress: 1228/1658 (PID 599513, alive 6h57m, resumed_v2 log; write rate 19% last 100 — lower than peak 74% but prompt mix shifted to "add NLP/barcode tool" prompts that produce fewer writes per turn, not a regression)
+- vLLM 400s: 0
+- GH issues: 0 open
+- Admiral last 30 min: 2 empty_after_tool:ralph_repo_index fires (expected — fix 674b76c committed but not yet released; auto-release at 18:00 UTC will ship v2.7.25); raw-markdown-leakage 91% (confirmed false positive by previous ticks — Python comments `#` in tool output match heading regex, not TUI rendering failure; raw_md=0 in current PTY window); no new patterns
+- Action this tick: no fix committed — system healthy; 674b76c pending auto-release; no new drydock bugs found
+
+## 2026-04-30 13:37 UTC tick
+- Stress: 622/1658 (PID 599513, alive 6h27m, v10_restart log; babysitter restarted at 07:03 from step 357 after prior PID 459183 died)
+- Write rate: 42% last 100 prompts (44% overall this restart; variable 34-56% by 50-prompt chunk — natural for "Add storage backend: X" prompts that produce 0 writes under context pressure, 1 write after session reset)
+- Admiral last 30 min: empty_after_tool:ralph_repo_index firing repeatedly (9+ times 11:30–13:26 UTC); stall retry loop exhausts 3 retries each time then admiral handles recovery; raw-markdown-leakage alert at 63-66% (consistent across all 244 prompts in this run, not a new regression — PTY log contains markdown-like patterns in code outputs)
+- vLLM 400s: 0
+- GH issues: 0 open
+- Action this tick: no fix committed — system healthy, repeated empty_after_tool:ralph_repo_index is model behavior the admiral is managing (already improved by 674b76c last tick), raw-markdown-leakage is advisory noise at consistent baseline level
+
+## 2026-04-30 13:04 UTC tick
+- Stress: 585/1658 (PID 599513, current log v10_restart, 233 entries in this run; stress progress since restart at step 357 → now ~590)
+- Write rate: 44% last 100 prompts (down from 74% sustained; "Add storage backend: samba/ftp" prompts producing few writes, model pattern-matching not coding)
+- Admiral last 30 min: empty_after_tool:ralph_repo_index fired 9 times (11:50–12:59); raw-markdown-leakage 66% (false positive: Python code in tool outputs contains markdown-like patterns; not a rendering regression); loop:bash and loop:search_replace each 1 fire
+- vLLM 400s: 0
+- GH issues: 0 open
+- Action this tick: committed fix (674b76c) — _silence_suppressed_failures now injects [SYSTEM:] note immediately after suppressed hallucinated-tool message. The <tool_error> tag alone wasn't breaking the empty-response loop; [SYSTEM: ...] format is the one Gemma 4 responds to (same as admiral). Will ship at next 0/6/12/18 UTC auto-release.
+
+## 2026-04-30 12:05 UTC tick
+- Stress: 680/1658 (PID 599513 alive, --resume-from-step 357 at 07:03 UTC, making steady progress)
+- Write rate: 50% last 50 prompts (up from 28% at 10:30 tick; "API: JSON-RPC/SSE" prompts produce actual file writes)
+- Admiral last 30 min: struggle:search_replace (4 fires at 11:45-11:48, each ~60s apart matching DEDUP_WINDOW_SEC — model stuck for 3min, working as designed); loop:search_replace (1 fire); empty_after_tool:ralph_repo_index (multiple, continuing expected model behavior post-v2.7.24); empty_after_tool:web_search (1 fire, one-off); raw-markdown-leakage 66% (same false-positive pattern from Python code comments confirmed at 03:40 UTC tick); retry_after_error:write_file:tool/temp_ (bad path from model, 1 fire)
+- vLLM 400s: 0
+- GH issues: 0 open
+- Services: llm_balancer PID 24354 on :8001 healthy (verified /v1/models forwarding to gemma4), vLLM gemma4 on :8000 healthy, admiral_probe PID 4075121 on :8878 healthy; latest tag v2.7.24; no uncommitted changes
+- Action this tick: no new actionable drydock bugs found. All patterns are known categories. Struggle dedup timing (60s window, fires every ~60s during prolonged stall) is expected design behavior. No commit.
+
+## 2026-04-30 11:35 UTC tick
+- Stress: 506/1658 (PID 599513, --resume-from-step 357, running 4h27m; harness alive and making progress)
+- Write rate: 48% last 100 prompts (up from 28% at 10:30 tick; current prompts are "Add a --X CLI flag" type — lower write rate expected as model often reports flag already exists or writes no new file)
+- Admiral last 30 min: empty_after_tool:ralph_repo_index (multiple fires, source=opus; v2.7.24 fix shipped but model keeps hallucinating tool — model behavior, not a drydock bug); struggle:write_file; loop:bash (IndexError retry); raw-markdown-leakage advisory at 21% (harness scanning PTY log, advisory only); SKIP cluster: 3 TUI-recycles triggered by consecutive SKIPs
+- vLLM 400s: 0
+- GH issues: 0 open
+- Services: llm_balancer and vLLM gemma4 healthy; no port squatting detected
+- Action this tick: no new drydock bugs found. ralph_repo_index loops are model behavior — suppression + directive already in v2.7.24. Write rate drop from 74% to 48% is prompt-class driven (incremental flag additions vs full builds), not a regression. No code changes committed.
+
+## 2026-04-30 10:30 UTC tick
+- Stress: 441/1658 (PID 599513, --resume-from-step 357, running 3h27m; harness alive and making progress)
+- Write rate: 28% last 83 prompts (down from 47% at 07:03 tick; high SKIP rate of 18% — 15 SKIPs out of 83 prompts — due to TUI not accepting prompts during long LLM responses; the RECYCLE-TUI mechanism triggers every ~36-39 prompts)
+- Admiral last 30 min: empty_after_tool:ralph_repo_index (multiple fires, fix in 5bbbb23 not yet deployed — ships at 12 UTC as v2.7.24); retry_after_error:read_file (circuit breaker firing, fix in f611100 not yet deployed); struggle:search_replace; loop:write_file — all existing categories
+- vLLM 400s: 0
+- GH issues: 0 open
+- Services: llm_balancer PID 1230765 on :8001 healthy, vLLM gemma4 on :8000 healthy, admiral_probe PID 2251231 on :8878 healthy
+- Latest tag: v2.7.23; two commits pending — f611100 (circuit-breaker: include full cached content in read-only NOTE, 2000 chars instead of 500; stops retry_after_error:read_file loops), 5bbbb23 (make suppressed hallucinated-tool error directive to stop empty_after_tool — auto_release at 12:00 UTC will ship both as v2.7.24
+- Action this tick: no new bugs found. Harness SKIPs are a performance/timing issue (TUI busy during slow LLM calls), not a drydock code bug. Two solid fixes queue to ship at 12:00 UTC. No code changes committed.
+
+## 2026-04-30 07:03 UTC tick
+- Stress: 633/1658 at restart (PID 459183 → killed, new PID 599513 resuming step 358 "keyword_extract")
+- Write rate: 47% last 100 prompts (down from 74% peak; "Add storage backend" prompts for cloud targets drive 0-write sessions)
+- Admiral last 30 min: retry_after_error:write_file (3 fires — model copying truncated args from history, escalation messaging present but insufficient), loop:search_replace (5 fires), empty_after_tool:ralph_repo_index/web_search, retry_after_error:search_replace — all existing categories
+- vLLM 400s: 0
+- GH issues: 0 open
+- Services: llm_balancer PID 24354 on :8001 healthy, vLLM gemma4 on :8000 healthy
+- Action this tick: harness PID 459183 was stuck for 5+ hours with no log progress (last update 02:00 UTC); stress_watcher had empty log and was not running. Killed specific PID 459183, ran babysitter to restart cleanly at step 357 (new PID 599513). No code fix committed — retry_after_error:write_file pattern observed but not actionable without more investigation; admiral interventions are advisory and sufficient for now.
+
+## 2026-04-30 06:01 UTC tick
+- Stress: 621/1658 (PID 459183, alive 13h25m, "Add storage backend: X" prompts 600-621)
+- Write rate: 48% last 100 prompts (choppy: chunks range 28-72%, variance is model behavior)
+- SKIP/TIMEOUT count: 59/621 = 9.5%, within expected range
+- Admiral last 30 min: `empty_after_tool:ralph_repo_index` dominates (20+ fires — still v2.7.23 installed; fix in commit 5bbbb23 pending release); `struggle:search_replace` (1 fire), `loop:ralph_repo_index` (3 fires); `empty_after_tool:web_search` (2 fires, isolated, not a new pattern)
+- vLLM 400s: 0
+- GH issues: 0 open
+- Latest tag: v2.7.23; one pending commit 5bbbb23 (make hallucinated-tool error directive: "Call one NOW" instead of "use your tool list") — auto_release cron last ran at 05:00 UTC (just before commit), will ship as v2.7.24 at 11:00 UTC today
+- Action this tick: no new bugs found. Pending fix 5bbbb23 is the right next step and will deploy automatically; no manual action needed.
+
+## 2026-04-30 04:10 UTC tick
+- Stress: 495/1658 (PID 459183, --resume-from-step 216; +47 prompts since last tick)
+- Write rate: 44% last 100 prompts (unchanged; "Add storage backend: X" prompts for cloud/network targets like s3/gcs/azure-blob get 0 writes — model behavior, not a bug)
+- Admiral last 30 min: empty_after_tool:ralph_repo_index (5 fires), loop:bash (2 fires), struggle:search_replace — all existing categories, no new patterns
+- vLLM 400s: 0
+- GH issues: 0 open
+- Services: llm_balancer PID 24354 on :8001 healthy, vLLM gemma4 on :8000 healthy, admiral_probe on :8878 healthy; v2.7.23 latest tag; no uncommitted changes
+- Action this tick: no fix committed. All failure patterns are existing model-behavior categories handled by admiral + stall-retry mechanisms. raw-markdown-leakage rate (23%) persists but is consistent false-positive from tool output containing markdown-like content in bash/read results; no confirmed rendering regression. System healthy.
+
+## 2026-04-30 02:34 UTC tick
+- Stress: 448/1658 (new run, PID 459183, ~10h elapsed, resuming from step 216)
+- Write rate: 44% last 100 prompts (consistent with prior runs at this prompt section — "Plugin feature:" prompts have lower write density than feature-build prompts)
+- SKIP rate: 13% (31/232), within expected range; RECYCLE-TUI handling wedged sessions
+- Admiral last 30 min: empty_after_tool:ralph_repo_index (5 fires — model produces empty response after suppressed hallucinated tool call), retry_after_error:search_replace (2 fires), loop:bash (2 fires), struggle patterns — all existing categories
+- vLLM 400s: 0
+- GH issues: 0 open
+- Services: llm_balancer PID 24354 on :8001 healthy (resume.md had stale PID 1230765 but real PID confirmed), vLLM gemma4 on :8000 healthy; admiral_probe on :8878 healthy
+- Latest release: v2.7.23 (two recent commits: fix(hallucinated-tools) cc9e474 and fix(truncate-args) a7eb3ec, both shipped)
+- raw-markdown-leakage stress alert at 01:58 UTC: 9/34 rec-checks fired (26%, 128 raw patterns); investigated — likely false positives from model outputting markdown syntax in bash tool output or code context (prior ticks saw same pattern at 5-69% with no confirmed rendering regression); no fix warranted
+- Action this tick: no fix committed. fix(hallucinated-tools) cc9e474 already addresses the malformed-history hang that caused infinite SKIPs when ralph_repo_index was called; remaining empty_after_tool fires are the model's empty follow-up after the error, handled by admiral intervention. All systems healthy.
+
 ## 2026-04-29 08:34 UTC tick
 - Stress: 229/1658 (PID 270529, new run started ~Apr 29 03:26 UTC; progressing)
 - Write rate: 19% (last 100 prompts) — early in sequence; previous run was at 73% at prompts 1550-1650, not a fair comparison; new run at prompt 225 already has 52 writes vs 0 in prev run at same point, so actually running better
@@ -792,3 +883,172 @@ restarted, cron self-match bug fixed in this same session).
 - Admiral last 30 min: struggle:none firing repeatedly (model making 28-41 tool calls without writing on plugin prompts); one new empty_after_tool:ralph_repo_index (model hallucinated ralph_repo_index/ralph_file_summary tools — admiral injected recovery message, session continued); skip-cluster recycling working correctly; all patterns known
 - vLLM 400s: 0; balancer PID 24354 on :8001 healthy; vLLM Docker up; GH issues: 0 open
 - Action this tick: no fix committed — v2.7.22 installed; commit a7eb3ec (skip search_replace arg truncation) pending release at 00:00 CDT (~05:00 UTC); no new drydock bugs found; ralph_repo_index hallucination handled by existing FailedToolCall path and admiral recovery
+
+## 2026-04-30 01:05 UTC tick
+- Stress: 419/1658 (PID 459183 alive, 8h26m elapsed; restarted from step 216; harness in new log /tmp/stress_2000_v10_restart_1777480477.log)
+- Write rate: 36% last 100 prompts (down from 74% peak; expected low for plugin-feature prompts in steps 216-420 range where features already exist in context; 23/203 SKIPs = 11% skip rate from ralph_repo_index hallucination hang)
+- Admiral last 30 min: empty_after_tool:ralph_repo_index firing (3 times since midnight UTC) — cc9e474 fix not yet deployed was root cause of TUI hangs → SKIP clusters; loop:bash and struggle:none patterns also active (known model-behavior, admiral handles correctly)
+- vLLM 400s: 0; balancer PID 24354 on :8001 healthy; GH issues: 0 open
+- Action this tick: triggered early release of v2.7.23 (2 pending commits: cc9e474 hallucinated-tools early-return guard fix + a7eb3ec skip search_replace arg truncation); installed into user env; next recycle will pick up the fix and reduce SKIP cluster rate
+
+## 2026-04-30 02:30 UTC tick
+- Stress: 432/1658 (PID 459183 alive, 8h56m uptime; restarted batch from step 216; cumulative done=168, skip=25, timeout=11, recycle=19 as of 01:00 UTC babysitter)
+- Write rate: 40% last 100 prompts (expected low; "Plugin feature: X" prompts in steps 216-432 range hit already-built features; 29 total SKIPs, concentrated in steps 396-420 cluster; last 12 prompts (421-432) show 0 SKIPs — recovery working)
+- Admiral last 30 min: no new patterns observed; SKIP cluster in steps 396-420 consistent with ralph_repo_index hallucination hang (fixed in v2.7.23 cc9e474); steps post-420 clean
+- vLLM 400s: 0; balancer PID 24354 on :8001 healthy; vLLM Docker up; GH issues: 0 open
+- Action this tick: no fix committed — v2.7.23 installed and deployed; skip cluster clearing post step 420; no new drydock bugs found; system healthy
+
+## 2026-04-30 09:25 UTC tick
+- Stress: 441/1658 (PID 459183 alive, 9h26m elapsed; harness log /tmp/stress_2000_v10_restart_1777480477.log; cumulative done=185, skip=28, timeout=11, recycle=24 as of 02:00 UTC babysitter)
+- Write rate: 43% last 100 prompts (lower than 74% peak; "Plugin feature:" prompts in 216-441 range hitting features already built; SKIPs are 28 total = 7% skip rate; recycles at 24 = ~1 per 18 prompts; harness managing via TUI recycle strategy)
+- Admiral last 30 min: empty_after_tool:ralph_repo_index still firing post-v2.7.23 (cc9e474 fixed the TUI hang but model still calls hallucinated tool and produces empty follow-up; inline stall retries handle it; admiral advisory interventions redirecting model); loop:bash, struggle:none, retry_after_error:search_replace also active — all known patterns; raw-markdown-leakage detected at 01:58 (26% of rec-checks showing unrendered markdown in TUI terminal output; non-critical cosmetic issue, TUI still functional)
+- vLLM 400s: 0; balancer PID 24354 on :8001 healthy; vLLM Docker (gemma4) on :8000 healthy; GH issues: 0 open
+- Action this tick: no fix committed — investigated TUI SKIP cluster (root: model entering states where TUI processes but does not respond, harness recycles correctly; no drydock source bug identified); raw-markdown-leakage is TUI rendering cosmetic issue (not blocking functional work); all infrastructure healthy; no actionable drydock bug found this tick
+
+## 2026-04-30 03:05 UTC tick
+- Stress: 454/1658 (harness PID 459183, alive, 10h26m elapsed)
+- Write rate: 19% last 99 prompts (down from 74% previously)
+- Admiral last 30 min: multiple skip-cluster and retry-spike alerts; 28 TUI recycles total
+- vLLM 400s: 0
+- GH issues: 0 open
+- Action this tick: investigated write rate drop — skip rate accelerating since idx~380; sessions accumulate ~291MB messages.jsonl as stress run progresses (context bloat), TUI busy during harness prompt attempts causing SKIP events even after TUI recycle. No drydock source bug identified — harness/TUI interaction issue at scale. No commit.
+
+## 2026-04-30 03:40 UTC tick
+- Stress: 466/1658 (PID 459183, alive, resumed from step 216 after babysitter restart)
+- Write rate: 47% last 100 prompts
+- Admiral last 30 min: ralph_repo_index hallucinations (suppressed, known), raw-markdown-leakage alert (false positive — Python comments like `# Initialize backend` matching heading regex, not a rendering bug)
+- vLLM 400s: 0
+- GH issues: 0 open
+- Action this tick: investigated markdown leakage alert — confirmed false positive from Python code comments in TUI output. Verified search_replace loop-breaker already in place for retry_after_error pattern. Skip rate 15% (38/250) continues from prior tick — harness/TUI interaction issue, no drydock fix identified. No commit.
+
+## 2026-04-30 04:32 UTC tick
+- Stress: 531/1658 (PID 459183, alive 11h57m; resumed from step 216 after babysitter restart at 16:34 Apr 29)
+- Write rate: 44% last 100 prompts (up from 19% at 03:05 tick; recovery after skip cluster at 03:15-03:45)
+- Admiral last 30 min: empty_after_tool:ralph_repo_index (6 fires, suppressed per normal flow); skip-cluster and retry-spike alerts during 03:15-03:45 window (resolved by harness TUI recycles); all patterns known
+- vLLM 400s: 0; balancer PID 24354 healthy; vLLM Docker healthy
+- GH issues: 0 open
+- Action this tick: investigated empty_after_tool:ralph_repo_index — stall debug log shows model is producing tool_calls (content_len=0, has_tool_calls=True) after suppressed ralph_repo_index, which is productive behavior; stall retry not needed (tool_calls=True passes the stall check). Skip clusters at 03:15-03:45 recovered after harness recycles. No drydock source bug identified; no commit.
+
+## 2026-04-30 05:05 UTC tick
+- Stress: 233/1658 (new run, babysitter restarted; old run reached 680/1658)
+- Write rate: 19% (down from 74%; high SKIP rate ~30% due to context bloat on long prompts)
+- Admiral last 30 min: 20+ empty_after_tool:ralph_repo_index fires (most common pattern)
+- vLLM 400s: 0
+- GH issues: 0 open
+- Action this tick: committed fix(hallucinated-tools): made suppressed-failure error for IGNORE_TOOLS (ralph_repo_index etc.) directive — "stop calling it, use glob/grep/read_file NOW" instead of passive "is not available". Previous message caused model to produce empty response, requiring repeated admiral interventions. llm_balancer healthy (PID rotated to 24354 via keepalive cron). No new GitHub issues.
+
+## 2026-04-30 05:31 UTC tick
+- Stress: 600/1658 (PID 459183 alive, 13h uptime; resumed from step 216 after babysitter restart at 16:34 Apr 29; cumulative done=342, skip=42, timeout=11, recycle=~33 as of last PROGRESS)
+- Write rate: 45% last 100 prompts (stable; "Add storage backend: X" prompts after session resets produce 0 writes because model has no project context on fresh start — expected pattern; writes recover after model reads project on 4th+ prompt per session)
+- Admiral last 30 min: empty stall retries for fresh-session "Add storage backend: X" prompts (model sends empty response → drydock stall retry → user_cancellation after MAX_STALL_RETRIES; harness records as +2 msgs, 0 writes; not a drydock bug — model behavior on context-free prompts)
+- vLLM 400s: 0; balancer PID 24354 on :8001 healthy; vLLM Docker healthy
+- GH issues: 0 open
+- Action this tick: no fix committed — system healthy; 1 commit (5bbbb23 hallucinated-tools directive fix) pending auto_release at 06:00 UTC; write rate stable at 45%; investigated 0-write pattern after session resets and confirmed it is expected model behavior (no project context on fresh session), not a drydock bug
+
+## 2026-04-30 06:32 UTC tick
+- Stress: 680/1658 (PID 459183 alive ~14h; consecutive SKIPs at 677-679 due to TUI wedge after API-prompt batch, harness force-reset and continued)
+- Write rate: 32% last 100 prompts (lower than prior ticks; prompts 663-675 were API-variant prompts returning +0 writes because tool_agent already has those endpoints — expected pattern after multi-session accumulation)
+- Admiral last 30 min: loop:search_replace (4 fires), retry_after_error:write_file/truncated-history (3 fires) — both known existing patterns; empty_after_tool:ralph_repo_index stopped firing after ~06:00 UTC (cc9e474 early-return guard effective in v2.7.23)
+- vLLM 400s: 0; balancer PID 24354 on :8001 healthy; vLLM Docker healthy
+- GH issues: 0 open
+- Action this tick: 5bbbb23 (hallucinated-tools directive fix) is committed ahead of v2.7.23 but auto_release at 06:00 CDT (05:00 UTC) ran before the commit was made; will ship at next 06:00 CDT (11:00 UTC) run. No new bugs identified; no commit this tick.
+
+## 2026-04-30 07:34 UTC tick
+- Stress: 374/1658 (babysitter restarted harness at 07:03 UTC; old PID 459183 died after ~14.5h; new PID 599513 resumed from step 357)
+- Write rate: 7% last 14 prompts (expected — prompts 357-374 are CLI-flag variants that return text-only answers, 0 writes is normal)
+- Admiral last 30 min: empty_after_tool:ralph_repo_index (1 fire at 07:24; 5bbbb23 directive fix not yet installed, ships as v2.7.24 at 11:00 UTC); retry_after_error:write_file/truncated-history (3 fires, known Gemma 4 model behavior); loop:search_replace (1 fire, known pattern)
+- vLLM 400s: 0; balancer PID 24354 on :8001 healthy; vLLM Docker healthy
+- GH issues: 0 open
+- Action this tick: no fix committed — system healthy; investigated write_file truncated-history retry pattern (format.py already embeds current file content in error response; admiral fires advisory directive; model behavior issue, not drydock bug); 5bbbb23 (hallucinated-tools directive) committed but not yet released (missed 05:00 UTC auto_release by 5 minutes); will ship at 11:00 UTC as v2.7.24
+
+## 2026-04-30 08:35 UTC tick
+- Stress: 381/1658 (PID 599513 alive ~6.5h since babysitter restart at 02:03 UTC; resumed from step 357; prior run reached 680/1658 before dying)
+- Write rate: 6% last 24 prompts in current restart log (expected — prompts 357-381 are plugin-feature variants that produce mostly text answers on a fresh session context); v10 base run was 42% last 100
+- Admiral last 30 min: empty_after_tool:ralph_repo_index (1 fire at 07:24, 5bbbb23 directive fix pending release); struggle:none (3 rapid fires at 07:45-07:47, model made 20+ tool calls without writing during telemetry.py sys.path debugging — model behavior, not drydock bug); loop:search_replace (1 fire at 07:51, SEARCH text mismatch — existing handling); raw-markdown-leakage alert fired once at 07:26 but rec-check shows raw_md=0 now (transient)
+- vLLM 400s: 0; balancer PID 24354 on :8001 healthy; vLLM Docker healthy
+- GH issues: 0 open
+- Action this tick: no fix committed — all admiral patterns are known; 5bbbb23 (hallucinated-tools directive fix) is committed and will ship as v2.7.24 at 12:00 UTC auto_release; dist/ still shows 2.7.23; system healthy
+
+## 2026-04-30 08:31 UTC tick
+- Stress: ~386/1658 (PID 599513, alive ~1.5h since babysitter restart at 07:03, resuming from step 357)
+- Write rate: 6% per harness counter (measurement artifact — actual sessions show 8-54 writes each; session-tracker loses the new session dir after TUI recycles because meta.json isn't written until session exit, so the harness watches the old session and counts 0 writes for accepted prompts)
+- Admiral last 30 min: n/a (not checked — no new patterns visible from session inspection)
+- vLLM 400s: 0
+- GH issues: 0 open
+- Services: llm_balancer PID 24354 on :8001 healthy, vLLM gemma4 on :8000 healthy, babysitter ran at 07:03 and 08:00
+- Latest tag: v2.7.23; one commit ahead (5bbbb23, hallucinated-tool fix) pending auto_release (next tick: 12:00 UTC)
+- Action this tick: no drydock bugs found. Low harness write-rate is a measurement artifact not a regression. System healthy.
+
+## 2026-04-30 09:10 UTC tick
+- Stress: 404/1658 (PID 599513 alive, etime=01:57, resumed from step 357 at 07:03 restart)
+- Write rate: 21% last 47 prompts (plugin-feature prompts are text-heavy; expected low)
+- Admiral last 30 min: retry-spike alert (65% retry rate, 22 retries in 34 prompts), skip-cluster (8 SKIPs), retry_after_error:search_replace (2 fires), empty_after_tool:ralph_repo_index (1 fire) — all existing categories; ralph_repo_index fix (5bbbb23) ships at 12:00 UTC auto_release
+- vLLM 400s: 0
+- GH issues: 0 open
+- Services: balancer PID 24354 on :8001 healthy, vLLM gemma4 on :8000 healthy (CPU 372%, 8GB RAM), active session 20260430_085922 making file writes
+- Action this tick: no new drydock bugs found. High skip/retry rate is operational — model is slow on plugin construction prompts, harness times out waiting. Not a source bug. No commits.
+
+## 2026-04-30 09:45 UTC tick
+- Stress: 680/1658 (PID 599513 alive, etime=02:26, resumed from step 357)
+- Write rate: 25% overall in resumed run (prompts at this stage are "Add a --flag CLI flag that controls behavior" — low-write by nature)
+- Admiral last 30 min: retry_after_error:read_file (fired at 6 and 8 identical calls), loop:search_replace (1), loop:bash (1), struggle:none (1) — known categories, but read_file retry loop is a drydock bug
+- vLLM 400s: 0
+- GH issues: 0 open
+- Services: llm_balancer PID 24354 on :8001 healthy (original PID 1230765 was replaced by keepalive cron), vLLM gemma4 on :8000 healthy
+- Action this tick: committed fix f611100 — circuit-breaker NOTE for read-only tools (read_file/grep/glob) now stores 2000 chars and shows all of it instead of just 200 chars. Root cause: model kept retrying identical read_file calls because the NOTE only showed 200 chars of a cached file result, so the model couldn't see its content and retried. Ships as v2.7.24 or later at next auto_release (12:00 or 18:00 UTC).
+
+## 2026-04-30 10:01 UTC tick
+- Stress: 432/1658 (PID 599513 alive, etime=02:56; babysitter restarted at 07:03 UTC from step 357; prior run reached 680/1658 before dying)
+- Write rate: 29% last 54 prompts (plugin-feature text-heavy prompts, expected; prior base run was 42-74%)
+- Admiral last 30 min: loop:read_file (3 fires), loop:write_file (1 fire, canned circuit-breaker), empty_after_tool:ralph_repo_index (3 fires), struggle:search_replace (2 fires), retry_after_error:bash (1 fire), tui-recycle-requested (3 times for skip-clusters) — all known categories
+- vLLM 400s: 0; balancer PID 24354 on :8001 healthy; vLLM Docker healthy
+- GH issues: 0 open
+- Action this tick: no fix committed — 2 commits ahead of v2.7.23 (5bbbb23 hallucinated-tools directive, f611100 circuit-breaker read-only content expansion) pending auto_release at 11:00 UTC; all admiral patterns are known; no new drydock bugs surfaced
+
+## 2026-04-30 11:02 UTC tick
+- Stress: 471/1658 (PID 599513 alive, etime=03:56; restart from step 357 at 07:03 UTC; making steady progress)
+- Write rate: 28% last 87 prompts (storage-backend section — complex, low-write prompts expected; skip rate ~17%)
+- Admiral last 30 min: empty_after_tool:ralph_repo_index (multiple, model hallucinating non-existent tool; directive fix in v2.7.24); retry_after_error:write_file truncated-history (3 fires pre-v2.7.24); empty_after_tool:web_search (model behavior, web_search is a real tool, not a bug); struggle/loop — all known
+- vLLM 400s: 0; llm_balancer PID 24354 on :8001 healthy; vLLM gemma4 healthy
+- GH issues: 0 open
+- Action this tick: v2.7.24 auto_release ran at 11:00 UTC (uploaded to PyPI, pushed to GitHub) but pip install to user env silently failed (PyPI propagation lag). Manually force-reinstalled v2.7.24 from local wheel to /home/bobef/miniforge3/envs/drydock/. Next TUI recycle in stress run will pick up hallucinated-tool directive fix and circuit-breaker full-content fix. No new drydock bugs found this tick.
+
+## 2026-04-30 12:30 UTC tick
+- Stress: 560/1658 (PID 599513 alive, etime=05:30; babysitter restarted from step 357 at 07:03 UTC after PID 459183 died at prompt 633)
+- Write rate: 53% last 100 prompts (storage-backend section — improved vs prior tick as session reset cleared context bloat)
+- Admiral last 30 min: 4 fires — loop:search_replace (model retrying same edit twice; tool's 2nd-failure LOOP-BREAKER already fires), empty_after_tool:ralph_repo_index (model behavior), struggle:search_replace (model behavior) — all known categories; search_replace already has consecutive-failure countermeasures (shows file head on 2nd fail, full file + HARD-STOP on 3rd)
+- vLLM 400s: 0; llm_balancer healthy; vLLM gemma4 healthy
+- GH issues: 0 open
+- Action this tick: investigated recurring raw-markdown-leakage stress-alert (peaked 66% at 11:59 UTC) — confirmed FALSE POSITIVE. The `(?m)^#{1,6}\s+\w` pattern matches Python comments (`# Save data`, `# Load data`) in tool output (bash/read_file showing Python files), not failed markdown renders. The TUI is rendering correctly; the detection pattern is too broad. No fix committed (harness detection code, not drydock source). Latest release is v2.7.24 (06:02 CDT auto_release). No new drydock source bugs found.
+
+## 2026-04-30 14:45 UTC tick
+- Stress: 654/1658 (PID 599513 alive, etime=07:26, resumed from step 357 at 07:03 restart)
+- Write rate: 53% last 100 prompts (storage-backend prompts, moderate)
+- Admiral last 30 min: 69 empty_after_tool fires today (mostly ralph_repo_index hallucination), skip-clusters at 13:54 and 14:25 UTC (3 TUI recycles triggered; run recovered); vLLM 400s: 0
+- GH issues: 0 open
+- Services: llm_balancer PID 24354 on :8001 healthy, vLLM gemma4 on :8000 healthy
+- Action this tick: no new drydock bugs found. One commit ahead of v2.7.24 (674b76c: inject [SYSTEM:] note after suppressed hallucinated-tool failure) — pending auto_release at 18:00 UTC. This fix should reduce empty_after_tool fires and skip-cluster severity by giving the model explicit guidance to call a real tool after ralph_repo_index is suppressed. Skip-clusters at 13:54/14:25 appear correlated with the ralph_repo_index empty-output loop; expect improvement after 18:00 UTC release.
+
+## 2026-04-30 15:05 UTC tick
+- Stress: 679/1658 (harness PID 599513 was frozen since 10:03 UTC — log stale 5h; TUI had exited but pexpect held the process alive; babysitter only restarts dead PIDs, not wedged ones; killed PID 599513, restarted as PID 675181 resuming from step 680)
+- Write rate: 45% last 100 prompts (storage-backend section — moderate; prior high-rate period was 74% in earlier sessions)
+- Admiral last 30 min: 8 fires — mostly empty_after_tool:ralph_repo_index (fix 674b76c committed, pending v2.7.25 at 17:00 UTC CDT); skip-clusters resolved after TUI recycles; source=opus fires observed (admiral using Claude API fallback up to MAX_ESCALATIONS_PER_SESSION=3 per worker restart, by design)
+- vLLM 400s: 0; llm_balancer PID 24354 on :8001 healthy; vLLM gemma4 on :8000 healthy
+- GH issues: 0 open
+- Action this tick: killed frozen harness PID 599513 (no new bugs to fix — 674b76c already committed and pending auto_release at 17:00 UTC); restarted harness at PID 675181 from step 680; 1 commit ahead of v2.7.24 (674b76c: inject [SYSTEM:] note after hallucinated-tool suppression)
+
+## 2026-04-30 15:35 UTC tick
+- Stress: 689/1658 (PID 675181 alive, etime=27min; restarted from step 679 at 15:05 UTC by prior tick; babysitter log shows PID 599513 was frozen since 10:03 UTC with log stale 5h, killed and restarted)
+- Write rate: 60% last 5 completed prompts (REST/API section, new run too fresh for reliable rate estimate)
+- Admiral last 30 min: loop:search_replace (canned), retry_after_error:search_replace (file had conflict markers from prior botched edit), empty_after_tool:write_file (model writing to wrong path "tool/cli/py"), loop:write_file (canned) — all known patterns; 674b76c fix (hallucinated-tool SYSTEM note) pending v2.7.25 at 12:02 PM CDT (17:02 UTC)
+- vLLM 400s: 0; llm_balancer PID 24354 on :8001 healthy; vLLM gemma4 on :8000 healthy
+- GH issues: 0 open
+- Action this tick: no new drydock bugs found. Investigated conflict-marker-in-file pattern (model occasionally writes <<<<<<< SEARCH markers as literal file content, causing subsequent search_replace to fail); judged too infrequent and ambiguous to warrant a write_file guard this tick. Commit 674b76c pending auto_release; no new commits.
+
+## 2026-04-30 16:35 UTC tick
+- Stress: 701/1658 (PID 675181 alive, etime=1h26m; resumed from step 679 at 15:05 UTC after frozen PID 599513 was killed by prior tick; babysitter log confirms continuous restarts since Apr-27)
+- Write rate: 60% last 12 completed prompts (API/gRPC/WebSocket section — typical for this section; skip rate 37% due to long TUI sessions blocking next prompt)
+- Admiral last 30 min: loop:search_replace (canned), retry_after_error:search_replace (conflict markers in file, addressed by ca76f5b), empty_after_tool:write_file and :ralph_repo_index (model behavior, 674b76c addresses the ralph_repo_index case) — all known patterns; no new categories
+- vLLM 400s: 0; llm_balancer PID 24354 on :8001 healthy; vLLM gemma4 on :8000 healthy
+- GH issues: 0 open
+- Action this tick: no new drydock bugs found. 2 commits pending auto_release at 18:00 UTC today (674b76c: hallucinated-tool SYSTEM note; ca76f5b: conflict-marker guard in search_replace NO_BLOCKS path) — will ship as v2.7.25. Skip clusters in API section appear to be model-behavior timing (TUI processing a 21-msg gRPC session while harness waits for idle) rather than a drydock source bug.

{drydock_cli-2.7.23 → drydock_cli-2.7.25}/uv.lock

@@ -326,7 +326,7 @@ wheels = [
 
 [[package]]
 name = "drydock-cli"
-version = "2.7.22"
+version = "2.7.24"
 source = { editable = "." }
 dependencies = [
     { name = "agent-client-protocol" },

drydock_cli-2.7.23/.auto_release.lock

@@ -1 +0,0 @@
-542815