drydock-cli 2.7.30__tar.gz → 2.7.31__tar.gz

drydock_cli-2.7.31/.auto_release.lock

@@ -0,0 +1 @@
+2183712

{drydock_cli-2.7.30 → drydock_cli-2.7.31}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.7.30
+Version: 2.7.31
 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-2.7.30 → drydock_cli-2.7.31}/drydock/core/llm/format.py

@@ -353,6 +353,9 @@ class APIToolFormatHandler:
             # mini_db fresh build (called ralph_repo_index 5 times).
             "ralph_repo_index", "repo_index", "index_repo",
             "list_mcp_resources", "list_resources", "search_resources",
+            # Observed in stress run 2026-05-02: model hallucinates ralph_file_summary
+            # after receiving tool results, causing empty_after_tool stalls.
+            "ralph_file_summary", "file_summary", "repo_summary",
         }
 
         suppressed_failures = []

{drydock_cli-2.7.30 → drydock_cli-2.7.31}/drydock/core/tools/builtins/grep.py

@@ -4,6 +4,7 @@ import asyncio
 from collections.abc import AsyncGenerator
 from enum import StrEnum, auto
 from pathlib import Path
+import re as _re
 import shutil
 from typing import TYPE_CHECKING, ClassVar
 
@@ -131,6 +132,15 @@ class Grep(
         if not args.pattern.strip():
             raise ToolError("Empty search pattern provided.")
 
+        try:
+            _re.compile(args.pattern)
+        except _re.error as regex_err:
+            escaped = _re.escape(args.pattern)
+            raise ToolError(
+                f"Invalid regex pattern: {regex_err}\n"
+                f"To search for this as literal text, use pattern={escaped!r}"
+            ) from None
+
         path_obj = Path(args.path).expanduser()
         if not path_obj.is_absolute():
             path_obj = Path.cwd() / path_obj

{drydock_cli-2.7.30 → drydock_cli-2.7.31}/drydock/core/tools/builtins/search_replace.py

@@ -503,7 +503,23 @@ class SearchReplace(
             entry = state.get(fail_key, {"count": 0})
             entry["count"] += 1
             state[fail_key] = entry
-            if entry["count"] >= 2:
+            if entry["count"] == 1:
+                # First failure: embed file head so model sees actual content
+                # without waiting for a second retry cycle. Reduces
+                # retry_after_error:search_replace events significantly.
+                try:
+                    head = original_content[:1500]
+                    line_count = original_content.count("\n")
+                    error_message += (
+                        f"\n\n[HINT: search text not found in {file_path.name}. "
+                        f"File may have changed. Current file head "
+                        f"({line_count} lines total):\n"
+                        f"-----FILE HEAD-----\n{head}\n-----FILE HEAD END-----\n"
+                        f"Adjust your SEARCH text to match the actual content above.]"
+                    )
+                except Exception:
+                    pass
+            elif entry["count"] >= 2:
                 try:
                     line_count = original_content.count("\n")
                     count = entry["count"]

{drydock_cli-2.7.30 → drydock_cli-2.7.31}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.7.30"
+version = "2.7.31"
 description = "Local-first CLI coding agent — tested with Gemma 4 26B via vLLM"
 readme = "README.md"
 requires-python = ">=3.12"

{drydock_cli-2.7.30 → drydock_cli-2.7.31}/tests/test_hallucinated_tool_suppression.py

@@ -42,6 +42,9 @@ HALLUCINATED_TOOLS = [
     "list_mcp_resources",
     "list_resources",
     "search_resources",
+    "ralph_file_summary",
+    "file_summary",
+    "repo_summary",
 ]
 
 

{drydock_cli-2.7.30 → drydock_cli-2.7.31}/tests/tools/test_grep.py

@@ -109,6 +109,39 @@ async def test_fails_with_nonexistent_path(grep):
     assert "Path does not exist" in str(err.value)
 
 
+@pytest.mark.asyncio
+async def test_invalid_regex_unmatched_paren_raises_with_hint(grep, tmp_path):
+    # Pattern with unescaped ( triggers POSIX ERE error — model must get escaped form
+    (tmp_path / "test.py").write_text("def run(self):\n    pass\n")
+    with pytest.raises(ToolError) as err:
+        await collect_result(grep.run(GrepArgs(pattern="def run(self,")))
+
+    msg = str(err.value)
+    assert "Invalid regex" in msg
+    assert "literal text" in msg
+    # Escaped suggestion must contain escaped paren so model can retry without loop
+    assert r"\(" in msg
+
+
+@pytest.mark.asyncio
+async def test_invalid_regex_unclosed_bracket_raises_with_hint(grep, tmp_path):
+    (tmp_path / "test.py").write_text("items = [1, 2]\n")
+    with pytest.raises(ToolError) as err:
+        await collect_result(grep.run(GrepArgs(pattern="[abc")))
+
+    msg = str(err.value)
+    assert "Invalid regex" in msg
+    assert "literal text" in msg
+
+
+@pytest.mark.asyncio
+async def test_valid_escaped_regex_does_not_raise(grep, tmp_path):
+    (tmp_path / "test.py").write_text("def run(self):\n    pass\n")
+    # Properly escaped pattern must work without error
+    result = await collect_result(grep.run(GrepArgs(pattern=r"def run\(self\)")))
+    assert result.match_count >= 1
+
+
 @pytest.mark.asyncio
 async def test_searches_in_specific_path(grep, tmp_path):
     subdir = tmp_path / "subdir"

drydock_cli-2.7.31/tests/tools/test_search_replace_first_failure_hint.py

@@ -0,0 +1,132 @@
+"""Regression test: search_replace embeds file head on first failure.
+
+Admiral logs (2026-05-02) showed retry_after_error:search_replace events
+where the model retried a failed search_replace without re-reading the file.
+The file head was only embedded starting at count=2, so the model had no
+context on the first failure and would blindly retry the same text.
+
+Fix: embed a FILE HEAD hint on the first failure (count=1) so the model
+can immediately see the actual file content and adjust its search text
+without needing a second retry cycle.
+"""
+from __future__ import annotations
+
+import asyncio
+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:
+    # read_file_state=None bypasses Read-before-Write guard so tests can focus
+    # on the "search text not found" hint behavior specifically.
+    return InvokeContext(tool_call_id="tc_hint_1", read_file_state=None)
+
+
+async def _run(tool, ctx, tmp_path, file_content, search_text, replace_text="NEW"):
+    f = tmp_path / "target.py"
+    f.write_text(file_content)
+    args = SearchReplaceArgs(
+        file_path=str(f),
+        content=f"<<<<<<< SEARCH\n{search_text}\n=======\n{replace_text}\n>>>>>>> REPLACE",
+    )
+    last = None
+    async for event in tool.run(args, ctx):
+        last = event
+    return last
+
+
+def test_first_failure_embeds_file_head(tool, ctx, tmp_path):
+    """On first search_replace failure, result includes file content hint."""
+    content = "def foo():\n    pass\n\ndef bar():\n    return 42\n"
+    result = asyncio.run(_run(tool, ctx, tmp_path, content, search_text="def nonexistent_func():"))
+    assert isinstance(result, SearchReplaceResult)
+    assert result.blocks_applied == 0
+    assert "FILE HEAD" in result.content or "HINT" in result.content
+    assert "def foo():" in result.content
+
+
+def test_first_failure_hint_not_hard_stop(tool, ctx, tmp_path):
+    """First failure uses soft HINT language, not HARD-STOP prohibition."""
+    content = "x = 1\ny = 2\n"
+    result = asyncio.run(_run(tool, ctx, tmp_path, content, search_text="z = 99"))
+    assert isinstance(result, SearchReplaceResult)
+    assert "HARD-STOP" not in result.content
+    assert "HINT" in result.content or "FILE HEAD" in result.content
+
+
+def test_second_failure_escalates(tool, ctx, tmp_path):
+    """Second consecutive failure escalates to LOOP-BREAKER (existing behavior)."""
+    content = "alpha = 1\nbeta = 2\n"
+    f = tmp_path / "target.py"
+    f.write_text(content)
+    args = SearchReplaceArgs(
+        file_path=str(f),
+        content="<<<<<<< SEARCH\ngamma = 3\n=======\ndelta = 4\n>>>>>>> REPLACE",
+    )
+
+    async def run_twice():
+        async for _ in tool.run(args, ctx):
+            pass
+        last = None
+        async for event in tool.run(args, ctx):
+            last = event
+        return last
+
+    result = asyncio.run(run_twice())
+    assert isinstance(result, SearchReplaceResult)
+    assert "LOOP-BREAKER" in result.content or "consecutive" in result.content
+
+
+def test_success_resets_fail_counter(tool, ctx, tmp_path):
+    """After a successful edit, the fail counter resets for that file."""
+    content = "def hello():\n    pass\n"
+    f = tmp_path / "target.py"
+    f.write_text(content)
+
+    async def run_scenario():
+        # Fail once
+        fail_args = SearchReplaceArgs(
+            file_path=str(f),
+            content="<<<<<<< SEARCH\nno_such_function():\n=======\nreplaced\n>>>>>>> REPLACE",
+        )
+        async for _ in tool.run(fail_args, ctx):
+            pass
+
+        # Succeed
+        ok_args = SearchReplaceArgs(
+            file_path=str(f),
+            content="<<<<<<< SEARCH\ndef hello():\n=======\ndef goodbye():\n>>>>>>> REPLACE",
+        )
+        ok_result = None
+        async for event in tool.run(ok_args, ctx):
+            ok_result = event
+        assert ok_result.blocks_applied == 1
+
+        # Fail again — should be count=1 (reset), not count=2
+        fail2_args = SearchReplaceArgs(
+            file_path=str(f),
+            content="<<<<<<< SEARCH\nstill_no_such_function():\n=======\nreplaced\n>>>>>>> REPLACE",
+        )
+        fail2_result = None
+        async for event in tool.run(fail2_args, ctx):
+            fail2_result = event
+        return fail2_result
+
+    result = asyncio.run(run_scenario())
+    assert "HINT" in result.content or "FILE HEAD" in result.content
+    assert "LOOP-BREAKER" not in result.content

{drydock_cli-2.7.30 → drydock_cli-2.7.31}/trip_log.md

@@ -717,3 +717,99 @@ restarted, cron self-match bug fixed in this same session).
 - Admiral last 2h: loop:bash ×8, struggle:none ×8, loop:search_replace ×4, struggle:write_file ×3, retry_after_error:search_replace ×3, retry_after_error:grep ×2, empty_after_tool:ralph_repo_index ×2 — all known model-behavior patterns; hallucinated ralph_repo_index is suppressed with system note per existing code; skip count stable at 121 (not growing since 03:00 UTC)
 - Latest release: v2.7.29; 2 unreleased commits (df5e0f5 read_file dedup cache, 96fa517 bash truncation notice) pending auto-ship as v2.7.30 at 06:00 UTC
 - Action this tick: no action — all services healthy, harness 83% complete, no new actionable drydock bugs found; stress should complete within ~3 hours at current rate (~18 prompts/hour in Perf: block)
+
+## 2026-05-02 05:03 UTC tick
+- Stress: 1398/1658 (84%) — PID 675181 alive 1d 13h 55m (RSS 1043 MB); active in Perf: block (coalesce-parallel-reads/warm-cache/evict-LRU); ~260 prompts remaining, ~4 hours at current rate
+- Write rate: 32% last 100 prompts — expected for Perf: text-advice block; model explains caching strategies rather than writing files
+- vLLM 400s: 0 — clean; vLLM gemma4 container healthy; llm_balancer PID 713929 on :8001 healthy and forwarding correctly
+- GH issues: 0 open
+- Latest release: v2.7.30 (auto-shipped at 06:00 UTC earlier today, 2 commits: read_file per-slot dedup cache + bash truncation notice)
+- Admiral last 500 entries: struggle:none ×82, struggle:write_file ×38, struggle:search_replace ×27, loop:bash ×14 — all expected for Perf: block; retry_after_error:read_file only ×1 (not actionable)
+- Action this tick: no action — all services healthy, harness 84% complete, no new actionable drydock bugs found
+
+## 2026-05-02 05:30 UTC tick
+- Stress: 1415/1658 (PID 675181, run 1 day 14h); progressing through "Perf:" block
+- Write rate: 26% (last 89 prompts) — expected low; "Perf: cache/lazy-load/memoize" prompts are advisory, model rarely writes files
+- Admiral last 30 min: retry_after_error:grep (Unmatched paren) fired 3 times since midnight; loop:grep fired twice — known pattern where model sends "def run(self," without escaping
+- vLLM 400s: 0
+- GH issues: 0 open
+- Latest release: v2.7.30
+- Action this tick: committed fix(grep) b79325f — _validate_args now runs re.compile() on pattern before subprocess invocation; on invalid regex raises ToolError with Python re.error message + re.escape() literal suggestion so model can immediately retry with correct pattern instead of looping. 3 regression tests added. Auto-release will ship v2.7.31 at next 0/6/12/18 UTC tick.
+
+## 2026-05-02 06:01 UTC tick
+- Stress: 1430/1658 on log stress_2000_v10_restart_1777561483 (PID 675181, running 1d 14h); the 680/1658 reading from earlier was from a stale log; harness babysitter relaunched and is near completion
+- Write rate: 24% last 100 prompts — mostly "Perf:" and "API:" prompts that legitimately generate 0 writes (model answers without coding); not a regression
+- Admiral last 30 min: loop:bash x2, loop:grep x1, empty_after_tool:ralph_repo_index x2, struggle:search_replace x2 — all known model-behavior patterns; no new failure types
+- vLLM 400s: 0
+- GH issues: 0 open
+- Latest release: v2.7.30; grep fix (b79325f) awaiting v2.7.31 at 06:00 CDT auto_release
+- Action this tick: no action — healthy; stress nearing completion, grep fix pending shipment
+
+## 2026-05-02 06:32 UTC tick
+- Stress: 1441/1658 (87%) — PID 675181 alive 1d 15h; active log stress_2000_v10_restart_1777561483 (last write 1 min ago); "Perf:" + "API:" block; ~217 prompts remain
+- Write rate: 23% last 89 prompts — expected for Perf: advisory block (model explains caching strategies, rarely writes files)
+- vLLM 400s: 0 — gemma4 container up 8 days; llm_balancer PID 713929 on :8001 healthy
+- GH issues: 0 open
+- Latest release: v2.7.30; grep validation fix (b79325f) unreleased, awaiting v2.7.31 at next auto_release (12:00 UTC)
+- Admiral last 30 min: loop:bash x4 (model looping concurrent.futures benchmark), retry_after_error:grep x2 (unmatched paren — fixed by b79325f not yet shipped), loop:grep x2 — all known model-behavior patterns
+- Action this tick: no action — all services healthy, stress 87% complete, grep fix pending shipment
+
+## 2026-05-02 07:02 UTC tick
+- Stress: 1447/1658 (87%) — PID 675181 alive 1d 15h; active log stress_2000_v10_restart_1777561483; 162 SKIPs (TUI slow during Perf/API block), 329 retries
+- Write rate: 22% last 89 prompts — expected low; "Perf: cache/memoize/lazy-load/batch-writes" prompts are advisory-only, model explains without writing files
+- vLLM 400s: 0; llm_balancer PID 713929 on :8001 healthy; gemma4 container up 8+ days
+- GH issues: 0 open
+- Latest release: v2.7.30; grep validation fix (b79325f) unreleased, pending v2.7.31 at 12:00 UTC auto_release
+- Admiral last 30 min: loop:bash x6 (model stuck on concurrent.futures benchmark, same command every ~50s 06:28–06:33 UTC) — model-behavior, not a drydock bug; retry_after_error:grep x3 (unescaped paren — will be resolved by b79325f post-shipment)
+- Action this tick: no action — all services healthy, stress 87% complete, no new actionable drydock bugs
+
+## 2026-05-02 07:32 UTC tick
+- Stress: 1466/1658 (88%) — PID 675181 alive; active log stress_2000_v10_restart_1777561483; in "Perf:" + "API:" block; ~192 prompts remaining
+- Write rate: 20% last 89 prompts — expected for Perf:/API: advisory block (model explains strategies without writing files)
+- vLLM 400s: 0 — gemma4 container healthy; llm_balancer PID 713929 on :8001 healthy
+- GH issues: 0 open
+- Latest release: v2.7.30; grep validation fix (b79325f) + search_replace first-failure hint (516d0c6) both unreleased, will auto-ship as v2.7.31 at next 0/6/12/18 UTC tick
+- Action this tick: committed fix(search_replace) 516d0c6 — file head now embedded on first search-not-found failure (count=1) so model can immediately see actual content and adjust, instead of blindly retrying; previously file head was only shown at count=2. 4 regression tests added. Also fixed test file to use correct SearchReplaceArgs (file_path/content fields) and async tool.run() API — the test was written with wrong constructor args (file/diff, InvokeContext(cwd=)) that never matched the actual codebase.
+
+## 2026-05-02 08:30 UTC tick
+- Stress: 1473/1658 (88.8%), PID 675181 healthy (1d 17h elapsed, resumed from step 679)
+- Write rate: 21% last 90 prompts — expected for "Integrate: X" advisory block (Slack/Discord/Datadog/etc., model explains integration without writing files)
+- Admiral last 30 min: 0 fires (no new admiral patterns observed)
+- vLLM 400s: 0 — gemma4 container healthy; llm_balancer PID 713929 on :8001 healthy
+- GH issues: 0 open
+- Action this tick: no fix committed — 08:00 UTC tick already committed 516d0c6 and b79325f; system healthy; skip rate ~17% (134/781) is stable, caused by TUI log hitting 1.24GB slowing PTY tail checks, expected to self-resolve as run finishes in ~2h
+
+## 2026-05-02 09:02 UTC tick
+- Stress: 1482/1658 (89.4%) — PID 675181 alive (1d 17h elapsed); active log stress_2000_v10_restart_1777561483; 139 SKIPs total, 52 TUI recycles
+- Write rate: 23% last 90 prompts — expected for "Integrate: X" advisory block (Slack/Datadog/Vercel/OCI/etc., model explains integrations without writing files)
+- vLLM 400s: 0 — gemma4 container healthy; llm_balancer PID 713929 on :8001 healthy (1d 14h uptime)
+- GH issues: 0 open
+- Admiral last 30 min: skip-cluster alerts (10-13 skips/34 prompts), tui-recycle-requested x5 between 08:08-09:00 UTC — all caused by large TUI log (1.28GB) slowing PTY detection; model-behavior patterns (loop:bash, retry_after_error:search_replace) but no new drydock bug class
+- Latest release: v2.7.30; two unreleased fixes pending (b79325f grep validation + 516d0c6 search_replace first-failure hint) — will auto-ship as v2.7.31 at 11:00 UTC
+- Action this tick: no fix committed — skip cluster is harness/log-size issue not drydock bug; all services healthy; ~176 prompts remaining (~10h at current pace)
+
+## 2026-05-02 09:32 UTC tick
+- Stress: 1498/1658 (90.3%) — PID 675181 alive (1d 18h elapsed, resumed from step 679); 139 SKIPs, 52 TUI recycles; ~160 prompts remaining
+- Write rate: 30% last 100 prompts (18% overall) — expected: "Integrate: X" block (Slack/Datadog/Opsgenie/Honeycomb/PagerDuty, model explains integrations without writing files)
+- vLLM 400s: 0 — gemma4 container healthy; llm_balancer PID 713929 on :8001 healthy
+- GH issues: 0 open
+- Admiral last 30 min: skip-cluster alerts (9-13 skips/34 prompts) + tui-recycle-requested x5; empty_after_tool:ralph_file_summary observed — model hallucinating non-existent summary tool causing stalls
+- Action this tick: committed fix(hallucinated-tools) bf9f0e2 — added ralph_file_summary, file_summary, repo_summary to _IGNORE_TOOLS so they get suppressed with redirect instead of visible FailedToolCall error; 15 regression tests pass; will ship as v2.7.31 at 12:00 UTC auto_release
+
+## 2026-05-02 10:02 UTC tick
+- Stress: 1508/1658 (91.0%) — PID 675181 alive (1d 18h elapsed, resumed from step 679); active log stress_2000_v10_restart_1777561483; ~178 SKIPs (21%), consistent with prior ticks
+- Write rate: 34% last 100 prompts — expected variation in "Integrate: X" block (Docker/OCI/Singularity/CI/monitoring prompts; model explains integration strategy without writing files)
+- vLLM 400s: 0 — gemma4 container up 8 days; llm_balancer PID 713929 on :8001 healthy; no JSONDecodeErrors last 30 min
+- GH issues: 0 open
+- Admiral last 30 min: no banner=True, raw_md=0 — no write loops or raw-markdown dumps; skip cluster continuing from prior ticks (TUI log ~1.31GB, PTY tail slow)
+- Unreleased commits since v2.7.30: 3 (bf9f0e2 ralph_file_summary suppression, 516d0c6 search_replace first-failure hint, b79325f grep validation) — will auto-ship as v2.7.31 at 12:00 UTC
+- Action this tick: no fix committed — no new drydock bug class observed; all services healthy; ~150 prompts remaining, run likely completes before 18:00 UTC
+
+## 2026-05-02 10:31 UTC tick
+- Stress: 1528/1658 (92% complete); harness PID 675181 alive, running since 1d19h; 130 prompts remain, likely finishes this hour
+- Write rate: 35% last 100 (lower than 74% peak; "Integrate: X" prompts in this range yield 0 writes — expected model behavior, not a regression)
+- Skip rate: 143/1528 = 9.4% (stable, within prior range; all "TUI did not accept after 3 retries")
+- vLLM 400s: 0 — gemma4 container healthy; balancer PID 713929 on :8001 unchanged
+- GH issues: 0 open
+- Unreleased commits since v2.7.30: 3 (bf9f0e2 hallucinated-tools, 516d0c6 search_replace hint, b79325f grep validation) — auto-ships as v2.7.31 at 12:00 UTC
+- Action this tick: no fix committed — no new drydock bug class observed; all services healthy; run close to completion

{drydock_cli-2.7.30 → drydock_cli-2.7.31}/uv.lock

@@ -326,7 +326,7 @@ wheels = [
 
 [[package]]
 name = "drydock-cli"
-version = "2.7.29"
+version = "2.7.30"
 source = { editable = "." }
 dependencies = [
     { name = "agent-client-protocol" },

drydock_cli-2.7.30/.auto_release.lock

@@ -1 +0,0 @@
-2123904