drydock-cli 2.7.6__tar.gz → 2.7.7__tar.gz

drydock_cli-2.7.7/.auto_release.lock

@@ -0,0 +1 @@
+3785750

drydock_cli-2.7.7/.claude/scheduled_tasks.lock

@@ -0,0 +1 @@
+{"sessionId":"514fb304-3a29-4466-985e-5e80a507b8cb","pid":3780811,"procStart":"363275523","acquiredAt":1777143665755}

{drydock_cli-2.7.6 → drydock_cli-2.7.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.7.6
+Version: 2.7.7
 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-2.7.6 → drydock_cli-2.7.7}/PRD.md

@@ -1,7 +1,7 @@
 # DryDock — Local CLI Coding Agent
 
 **Repository:** https://github.com/fbobe321/drydock
-**PyPI:** https://pypi.org/project/drydock-cli/ (v2.7.4)
+**PyPI:** https://pypi.org/project/drydock-cli/ (v2.7.6)
 **License:** Apache 2.0 (fork of [mistralai/mistral-vibe](https://github.com/mistralai/mistral-vibe))
 
 ## Vision
@@ -398,6 +398,54 @@ order:
 Auto-created if none exist. AGENTS.md is essential for devstral
 (it loops without one) but Gemma 4 works without it.
 
+## Recent fixes (April 25, 2026)
+
+### v2.7.6 (April 25, 09:18 CDT) — truncation JSON-validity fix
+**Root cause of the recurring vLLM 400 spiral.** Live capture from a
+2000-prompt stress run pinpointed it: `_truncate_old_tool_results` was
+shrinking old assistant `tool_calls.function.arguments` by appending
+`"\n[…truncated N bytes…]"` directly into the JSON-string field. vLLM's
+tool-call parser re-parses arguments as JSON; raw `\n` (0x0A) inside a
+JSON string is invalid → `JSONDecodeError: Invalid control character
+at: line 1 column 401` → HTTP 400 → drydock's "6 consecutive errors"
+banner → harness `/clear` → prompt lost. The pattern repeated every
+~30 prompts as truncation kicked in, costing dozens of prompts per
+batch.
+
+Fix in `agent_loop.py::_truncate_old_tool_results`: rebuild the
+truncated arguments as a small VALID JSON stub (`{"_truncated": true,
+"_original_bytes": N, "<id_field>": ...}`) preserving the most-useful
+identifying field (path / file_path / command / cmd / url / file)
+from the parsed original. 5 regression tests in
+`tests/test_truncate_args_valid_json.py`; 2 of them fail with the
+exact "char 400" error without the fix.
+
+**Stress measurement:** before fix, write rate 10% / vLLM 400s
+315 per 30 min. After fix, write rate sustained at ~44% / vLLM 400s
+0 per hour. **4.4× correctness improvement** on the same workload.
+
+### v2.7.6 (April 25) — search_replace APPEND fallback
+The 2026-04-25 stress run also exposed a search_replace REFUSED loop:
+Gemma 4 sent a single new function with no SEARCH/REPLACE markers,
+the safety check refused as "would shrink the file by 90%", model
+retried the same broken call, prompts dropped. Model intent in this
+shape is APPEND, not rewrite.
+
+Fix in `search_replace.py`: try appending the raw content first; for
+.py files only commit if the combined text still parses cleanly; if
+syntax breaks, return a directive REFUSED that explains the
+EOF-anchored SEARCH/REPLACE pattern. Tests in
+`tests/tools/test_search_replace_append_fallback.py`.
+
+### v2.7.5 (April 25, 07:25 CDT) — yesterday's GitHub issue batch
+Auto-release shipped the four fixes from 2026-04-24 plus the PRD
+refresh:
+- #13 control-char JSON sanitization in OpenAIAdapter
+- #11 fake `task(...)` text nuker in format.py
+- #12 wall-rescue at stream end in messages.py
+- #10 navygpt auto-disable in tool manager
+- PRD.md refresh for v2.7.4 (Admiral, Meta-Harness sections)
+
 ## Recent fixes (April 16–24, 2026)
 
 ### v2.7.4 + unreleased (April 24) — GitHub issue triage batch
@@ -749,7 +797,16 @@ Second claude-code pass produced two more wins:
 
 ## Roadmap
 
-### Currently in flight (April 24, 2026)
+### Currently in flight (April 25, 2026)
+- **2000-prompt stress test** running against v2.7.6 on
+  `/data3/drydock_test_projects/403_tool_agent` (PID 3713698, log
+  `/tmp/stress_2000_1777119799.log`). At 5+ hours in: 189/1658 prompts
+  done, write rate ~44% sustained, vLLM 400s eliminated, SKIPs 8%.
+  Babysitter cron will auto-restart on death.
+- **Resume sentinel** at `/data3/drydock/resume.md` — point a fresh
+  Claude session at it via `review the resume.md file and continue`.
+
+### Carried over from April 24, 2026
 - **GitHub issue triage** — 4 user-reported bugs (#10–#13) fixed this
   week, all with regression tests. Remaining open issues should be
   picked up on sight; `feedback_proactive_github_issues.md` codifies

{drydock_cli-2.7.6 → drydock_cli-2.7.7}/drydock/core/tools/builtins/search_replace.py

@@ -174,6 +174,9 @@ class SearchReplace(
                                     len(raw_content), target,
                                 )
                                 await self._write_file(target, combined)
+                                self.state.__dict__.setdefault(
+                                    "_sr_refused_raw_history", {}
+                                ).pop(str(target), None)
                                 yield SearchReplaceResult(
                                     file=str(target),
                                     blocks_applied=1,
@@ -186,11 +189,50 @@ class SearchReplace(
                                     content=raw_content[:100] + "...",
                                 )
                                 return
-                            raise ToolError(
+                            # Track consecutive REFUSED-raw failures per file
+                            # so we can escalate when the model retries the same
+                            # broken raw-content call. Pattern observed in
+                            # admiral logs: 14 retry_after_error:search_replace
+                            # fires per 6h, all with "REFUSED: the raw content".
+                            refused_state = self.state.__dict__.setdefault(
+                                "_sr_refused_raw_history", {}
+                            )
+                            refused_key = str(target)
+                            refused_entry = refused_state.get(
+                                refused_key, {"count": 0, "last_len": 0}
+                            )
+                            refused_entry["count"] += 1
+                            refused_entry["last_len"] = len(raw_content)
+                            refused_state[refused_key] = refused_entry
+                            base_msg = (
                                 f"REFUSED: the raw content you sent ({len(raw_content)} "
                                 f"chars) is much shorter than the existing {target.name} "
                                 f"({existing_size} chars), and appending it would break "
-                                f"the file's syntax. To add new code, use a SEARCH/REPLACE "
+                                f"the file's syntax."
+                            )
+                            if refused_entry["count"] >= 2:
+                                head = existing_text[:1500] if "existing_text" in locals() else target.read_text(errors="replace")[:1500]
+                                tail_src = existing_text if "existing_text" in locals() else target.read_text(errors="replace")
+                                tail = tail_src[-800:] if len(tail_src) > 2300 else ""
+                                tail_block = (
+                                    f"\n-----FILE TAIL (last 800 chars)-----\n{tail}\n"
+                                    if tail else ""
+                                )
+                                raise ToolError(
+                                    f"{base_msg}\n\n[LOOP-BREAKER: this is the "
+                                    f"#{refused_entry['count']} consecutive REFUSED on "
+                                    f"{target.name}. Stop sending the same raw content. "
+                                    f"Actual file content (first 1500 chars of "
+                                    f"{existing_size} bytes):\n"
+                                    f"-----FILE HEAD-----\n{head}\n-----FILE END HEAD-----"
+                                    f"{tail_block}\n"
+                                    f"Choose ONE: (a) call write_file with overwrite=True "
+                                    f"to replace the whole file with your raw content, OR "
+                                    f"(b) send a proper SEARCH/REPLACE block anchored on "
+                                    f"text from the file head/tail above.]"
+                                )
+                            raise ToolError(
+                                f"{base_msg} To add new code, use a SEARCH/REPLACE "
                                 f"block anchored on the last line of the file:\n"
                                 f"<<<<<<< SEARCH\n"
                                 f"[the file's actual final line, copied exactly]\n"
@@ -204,6 +246,9 @@ class SearchReplace(
                             "search_replace: no blocks — overwriting %s", target,
                         )
                         await self._write_file(target, raw_content)
+                        self.state.__dict__.setdefault(
+                            "_sr_refused_raw_history", {}
+                        ).pop(str(target), None)
                         yield SearchReplaceResult(
                             file=str(target),
                             blocks_applied=1,
@@ -340,6 +385,10 @@ class SearchReplace(
             # Success → reset the fail counter for this file
             state = self.state.__dict__.setdefault("_sr_fail_history", {})
             state.pop(str(file_path), None)
+            refused_state = self.state.__dict__.setdefault(
+                "_sr_refused_raw_history", {}
+            )
+            refused_state.pop(str(file_path), None)
 
         modified_content = block_result.content