drydock-cli 2.7.22__tar.gz → 2.7.23__tar.gz

drydock_cli-2.7.23/.auto_release.lock

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

{drydock_cli-2.7.22 → drydock_cli-2.7.23}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.7.22
+Version: 2.7.23
 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-2.7.22 → drydock_cli-2.7.23}/drydock/core/agent_loop.py

@@ -1032,7 +1032,7 @@ class AgentLoop:
         parsed = self.format_handler.parse_message(last_message)
         resolved = self.format_handler.resolve_tool_calls(parsed, self.tool_manager)
 
-        if not resolved.tool_calls and not resolved.failed_calls:
+        if not resolved.tool_calls and not resolved.failed_calls and not resolved.suppressed_failures:
             return
 
         async for event in self._handle_tool_calls(resolved):
@@ -1781,6 +1781,11 @@ class AgentLoop:
                 for tc in msg.tool_calls:
                     if not tc.function or not tc.function.arguments:
                         continue
+                    # search_replace args contain SEARCH/REPLACE blocks the model
+                    # NEEDS to see in history (to know what edits it already applied).
+                    # Only write_file carries full file content worth truncating.
+                    if tc.function.name == "search_replace":
+                        continue
                     args = tc.function.arguments
                     if len(args) <= SOFT_CAP_BYTES:
                         continue

{drydock_cli-2.7.22 → drydock_cli-2.7.23}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.7.22"
+version = "2.7.23"
 description = "Local-first CLI coding agent — tested with Gemma 4 26B via vLLM"
 readme = "README.md"
 requires-python = ">=3.12"

{drydock_cli-2.7.22 → drydock_cli-2.7.23}/tests/test_hallucinated_tool_suppression.py

@@ -85,3 +85,37 @@ def test_unknown_non_hallucinated_tool_goes_to_failed_calls():
     resolved = handler.resolve_tool_calls(parsed, tm)
     assert any(fc.tool_name == "nonexistent_random_tool" for fc in resolved.failed_calls)
     assert not resolved.suppressed_failures
+
+
+def test_suppressed_failures_not_bypassed_by_early_return():
+    """Regression for agent_loop early-return bug: when only suppressed_failures
+    exist (no real tool_calls, no failed_calls), the condition
+    `not tool_calls and not failed_calls` was True and _handle_tool_calls was
+    skipped, so _silence_suppressed_failures was never called. The tool result
+    message for the hallucinated tool was never added, leaving the conversation
+    malformed (assistant tool_call with no paired tool result).
+
+    Fix: early-return also checks `not suppressed_failures`.
+    This test verifies that a ResolvedMessage with only suppressed_failures
+    is NOT considered empty by the fixed condition.
+    """
+    handler = _make_handler()
+    tm = _make_tool_manager(["read_file"])
+    parsed = _parsed("ralph_repo_index")
+    resolved = handler.resolve_tool_calls(parsed, tm)
+
+    # Preconditions: only suppressed_failures, nothing else
+    assert not resolved.tool_calls
+    assert not resolved.failed_calls
+    assert resolved.suppressed_failures
+
+    # The fixed early-return condition must NOT consider this "no work to do"
+    should_skip_early = (
+        not resolved.tool_calls
+        and not resolved.failed_calls
+        and not resolved.suppressed_failures  # fixed: was missing this
+    )
+    assert not should_skip_early, (
+        "Bug regression: suppressed_failures were ignored by early-return, "
+        "leaving assistant tool_call with no paired tool result in history."
+    )

drydock_cli-2.7.23/tests/test_search_replace_args_not_truncated.py

@@ -0,0 +1,181 @@
+"""Regression test: search_replace call args must NOT be truncated.
+
+Repro: _truncate_old_tool_results truncated ALL large tool-call arguments,
+including search_replace SEARCH/REPLACE blocks (~400-800 bytes). The model
+then saw {_truncated: true, file_path: "..."} in its own history and copied
+those truncated args as new call arguments. format.py detected this and
+returned a FailedToolCall, but the model retried with the same truncated
+args again and again, creating a retry_after_error:search_replace loop.
+
+Fix (2026-04-29): skip the arg-truncation loop for search_replace tool
+calls. Only write_file carries full file content worth truncating.
+"""
+from __future__ import annotations
+
+import json
+
+import pytest
+
+from drydock.core.types import (
+    FunctionCall,
+    LLMMessage,
+    Role,
+    ToolCall,
+)
+from tests.conftest import build_test_agent_loop
+
+
+def _make_tc(name: str, arguments: str, idx: int = 0) -> ToolCall:
+    return ToolCall(
+        id=f"tc_{idx}",
+        index=idx,
+        function=FunctionCall(name=name, arguments=arguments),
+    )
+
+
+def _build_history_with_search_replace(num_old: int) -> list[LLMMessage]:
+    """Build a conversation with `num_old` old search_replace messages
+    plus enough newer messages to exceed KEEP_RECENT."""
+    msgs: list[LLMMessage] = [
+        LLMMessage(role=Role.system, content="be helpful"),
+        LLMMessage(role=Role.user, content="please help"),
+    ]
+    # search_replace calls with realistic 600+ byte args (exceeds SOFT_CAP=500)
+    search_block = "    def run_tool(self, tool_name: str, args: dict) -> str:\n"
+    search_block += "        # This is the old implementation that needs to be replaced\n"
+    search_block += "        return ''\n"
+    replace_block = "    def run_tool(self, tool_name: str, args: dict) -> str:\n"
+    replace_block += "        plugin = self._load_plugin(tool_name)\n"
+    replace_block += "        if plugin is None:\n"
+    replace_block += "            raise ValueError(f'Unknown tool: {tool_name}')\n"
+    replace_block += "        return plugin.execute(args)\n"
+    sr_args = json.dumps({
+        "file_path": "/data3/drydock_test_projects/403_tool_agent/tool_agent/cli.py",
+        "content": (
+            f"<<<<<<< SEARCH\n{search_block}=======\n{replace_block}>>>>>>> REPLACE"
+        ),
+    })
+    assert len(sr_args) > 500, f"Test args must exceed SOFT_CAP_BYTES=500 to be meaningful, got {len(sr_args)}"
+
+    for i in range(num_old):
+        msgs.append(
+            LLMMessage(
+                role=Role.assistant,
+                content="",
+                tool_calls=[_make_tc("search_replace", sr_args, i)],
+            )
+        )
+        msgs.append(
+            LLMMessage(
+                role=Role.tool,
+                tool_call_id=f"tc_{i}",
+                name="search_replace",
+                content="Applied edit.",
+            )
+        )
+
+    # Add newer messages to exceed KEEP_RECENT (4) so truncation would trigger
+    # for write_file but should be skipped for search_replace.
+    for i in range(6):
+        msgs.append(
+            LLMMessage(
+                role=Role.assistant,
+                content="",
+                tool_calls=[_make_tc("read_file", '{"path":"/x"}', 100 + i)],
+            )
+        )
+        msgs.append(
+            LLMMessage(
+                role=Role.tool,
+                tool_call_id=f"tc_{100+i}",
+                name="read_file",
+                content="ok",
+            )
+        )
+    return msgs
+
+
+class TestSearchReplaceArgsNotTruncated:
+    def test_search_replace_args_preserved_after_truncation(self):
+        """search_replace call args must not be replaced with {_truncated: true} stub."""
+        loop = build_test_agent_loop()
+        loop.messages.reset(_build_history_with_search_replace(num_old=10))
+
+        loop._truncate_old_tool_results()
+
+        for m in loop.messages:
+            if m.role != Role.assistant or not m.tool_calls:
+                continue
+            for tc in m.tool_calls:
+                if tc.function.name != "search_replace":
+                    continue
+                args = tc.function.arguments or ""
+                parsed = json.loads(args)
+                assert not parsed.get("_truncated"), (
+                    f"search_replace args were truncated — model will copy the stub "
+                    f"and loop: {args[:120]}"
+                )
+                assert "content" in parsed, (
+                    "search_replace args lost the 'content' field after truncation"
+                )
+
+    def test_write_file_still_truncated(self):
+        """write_file args should still be truncated (they carry full file content)."""
+        msgs: list[LLMMessage] = [
+            LLMMessage(role=Role.system, content="be helpful"),
+            LLMMessage(role=Role.user, content="please help"),
+        ]
+        big_args = json.dumps({
+            "file_path": "/data3/test/foo.py",
+            "content": "x" * 5000,
+        })
+        for i in range(10):
+            msgs.append(
+                LLMMessage(
+                    role=Role.assistant,
+                    content="",
+                    tool_calls=[_make_tc("write_file", big_args, i)],
+                )
+            )
+            msgs.append(
+                LLMMessage(
+                    role=Role.tool,
+                    tool_call_id=f"tc_{i}",
+                    name="write_file",
+                    content=f"Wrote {i}",
+                )
+            )
+        for i in range(6):
+            msgs.append(
+                LLMMessage(
+                    role=Role.assistant,
+                    content="",
+                    tool_calls=[_make_tc("read_file", '{"path":"/x"}', 100 + i)],
+                )
+            )
+            msgs.append(
+                LLMMessage(
+                    role=Role.tool,
+                    tool_call_id=f"tc_{100+i}",
+                    name="read_file",
+                    content="ok",
+                )
+            )
+
+        loop = build_test_agent_loop()
+        loop.messages.reset(msgs)
+        loop._truncate_old_tool_results()
+
+        truncated_count = 0
+        for m in loop.messages:
+            if m.role != Role.assistant or not m.tool_calls:
+                continue
+            for tc in m.tool_calls:
+                if tc.function.name != "write_file":
+                    continue
+                args = tc.function.arguments or ""
+                parsed = json.loads(args)
+                if parsed.get("_truncated"):
+                    truncated_count += 1
+
+        assert truncated_count > 0, "write_file args should still be truncated"

{drydock_cli-2.7.22 → drydock_cli-2.7.23}/trip_log.md

@@ -770,3 +770,25 @@ restarted, cron self-match bug fixed in this same session).
 - vLLM 400s: 0
 - GH issues: 0 open
 - Action this tick: committed fix (39515a1) — tool-aware escalation message for truncated args; search_replace now gets SEARCH/REPLACE recovery guidance instead of write_file advice; 3 new regression tests all pass; ships as v2.7.22 at next 0/6/12/18 auto-release tick
+
+## 2026-04-29 23:15 UTC tick
+- Stress: 384/1658 (current restart batch from step 216; harness alive, PID 459183, 6h25m uptime)
+- Write rate: 40% last 100 prompts (down from 74% peak — traced to new bug)
+- Admiral last 30 min: ~12 retry_after_error:search_replace fires + SKIP clusters
+- vLLM 400s: 0
+- GH issues: 0 open
+- Action this tick: committed fix `a7eb3ec` — skip search_replace arg truncation in `_truncate_old_tool_results`. The 500-byte SOFT_CAP was hitting search_replace SEARCH/REPLACE blocks (600-800 bytes), replacing them with `{_truncated:true, file_path:...}` stubs. Model copied the stubs as new call args, triggering persistent retry_after_error loops. Only write_file (full file content) warrants truncation. Regression test added. Auto-release will ship at next 0/6/12/18 UTC tick.
+
+## 2026-04-29 23:33 UTC tick
+- Stress: 398/1658 (current restart batch from step 216, PID 459183 alive 6h58m; previous run died at step 422/1658 at 16:34 UTC and was restarted by babysitter)
+- Write rate: 35% last 100 prompts (expected low for API-tool plugin prompts in steps 216-400 range — many features already implemented; peak 74% was during initial build phase)
+- Admiral last 30 min: 2 interventions (struggle:none at 23:04 UTC source=opus, empty_after_tool:write_file at 23:28 UTC source=opus); skip-cluster stress-alert at 23:21; tui-recycle-requested at 23:28 — all known patterns, FORCE-RESET and recycle cycling correctly
+- vLLM 400s: 0; balancer PID 24354 on :8001 healthy; vLLM Docker up 5 days; GH issues: 0 open
+- Action this tick: no fix committed — system healthy; pending commit a7eb3ec (skip search_replace arg truncation) ships as v2.7.23 at 00:00 UTC auto-release (committed at 23:15 UTC, after the 18:00 UTC v2.7.22 release window)
+
+## 2026-04-30 00:05 UTC tick
+- Stress: 406/1658 (PID 459183 alive, 7h25m elapsed; cumulative done=160, skip=16, timeout=11, recycle=15)
+- Write rate: 46% last 30 prompts (35% last 100; low expected for plugin-feature prompts where features already present in context)
+- 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

{drydock_cli-2.7.22 → drydock_cli-2.7.23}/uv.lock

@@ -326,7 +326,7 @@ wheels = [
 
 [[package]]
 name = "drydock-cli"
-version = "2.7.21"
+version = "2.7.22"
 source = { editable = "." }
 dependencies = [
     { name = "agent-client-protocol" },

drydock_cli-2.7.22/.auto_release.lock

@@ -1 +0,0 @@
-522842