drydock-cli 2.10.2__tar.gz → 2.10.4__tar.gz

{drydock_cli-2.10.2 → drydock_cli-2.10.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.10.2
+Version: 2.10.4
 Summary: Local-first CLI coding agent — tested with Gemma 4 26B via llama.cpp
 Project-URL: Homepage, https://github.com/fbobe321/drydock
 Project-URL: Repository, https://github.com/fbobe321/drydock

{drydock_cli-2.10.2 → drydock_cli-2.10.4}/drydock/admiral/detectors_proposed.py

@@ -27,7 +27,6 @@ a clean "I have ..." / "The ... has been ..." summary.
 from __future__ import annotations
 
 from collections.abc import Sequence
-from typing import Optional
 
 from drydock.core.types import LLMMessage, Role
 from drydock.admiral.detectors import Finding, _tool_sig
@@ -78,7 +77,7 @@ def _is_hallucinated_tool_result(m: LLMMessage) -> bool:
 
 def detect_empty_after_tool(
     messages: Sequence[LLMMessage],
-) -> Optional[Finding]:
+) -> Finding | None:
     """Fires when the last assistant turn had no content AND no tool_calls,
     immediately following a tool result.
 
@@ -102,7 +101,7 @@ def detect_empty_after_tool(
     """
     # Walk backwards to find the last assistant message; look at the one
     # before it.
-    last_assistant_idx: Optional[int] = None
+    last_assistant_idx: int | None = None
     for i in range(len(messages) - 1, -1, -1):
         if messages[i].role is Role.assistant:
             last_assistant_idx = i
@@ -148,7 +147,7 @@ def detect_empty_after_tool(
 
 def detect_retry_after_error(
     messages: Sequence[LLMMessage],
-) -> Optional[Finding]:
+) -> Finding | None:
     """Fires when an identical tool call follows an errored tool result.
 
     In 400 sessions this fired 45 times (0.11/session). High-signal because

{drydock_cli-2.10.2 → drydock_cli-2.10.4}/drydock/admiral/history.py

@@ -6,7 +6,7 @@ no JSON, no rotation logic, just append-only timestamped lines.
 """
 from __future__ import annotations
 
-from datetime import datetime, timezone
+from datetime import datetime, UTC
 from pathlib import Path
 
 
@@ -17,7 +17,7 @@ def _log_path() -> Path:
 
 
 def append(event: str, detail: str) -> None:
-    ts = datetime.now(timezone.utc).isoformat(timespec="seconds")
+    ts = datetime.now(UTC).isoformat(timespec="seconds")
     line = f"{ts} | {event} | {detail}\n"
     with _log_path().open("a") as f:
         f.write(line)

{drydock_cli-2.10.2 → drydock_cli-2.10.4}/drydock/admiral/metrics.py

@@ -13,7 +13,7 @@ from __future__ import annotations
 import logging
 from collections.abc import Sequence
 from dataclasses import asdict, dataclass, field
-from datetime import datetime, timezone
+from datetime import datetime, UTC
 from typing import TYPE_CHECKING
 
 from drydock.admiral import persistence, task_classifier
@@ -75,7 +75,7 @@ def collect(agent_loop: AgentLoop, session_id: str, outcome: str = "unknown") ->
     total, writes = _count_tool_calls(msgs)
     tcpw = (total / writes) if writes else float(total)
     return SessionMetrics(
-        ts=datetime.now(timezone.utc).isoformat(timespec="seconds"),
+        ts=datetime.now(UTC).isoformat(timespec="seconds"),
         model=model_name,
         task_type=task_classifier.classify(msgs),
         session_id=session_id,

{drydock_cli-2.10.2 → drydock_cli-2.10.4}/drydock/admiral/opus_escalator.py

@@ -98,7 +98,7 @@ async def _try_claude_cli(prompt: str) -> str | None:
             stdout, _ = await asyncio.wait_for(
                 proc.communicate(), timeout=OPUS_TIMEOUT_SEC
             )
-        except asyncio.TimeoutError:
+        except TimeoutError:
             proc.kill()
             await proc.wait()
             return None

{drydock_cli-2.10.2 → drydock_cli-2.10.4}/drydock/admiral/persistence.py

@@ -17,7 +17,7 @@ from __future__ import annotations
 import json
 import logging
 import os
-from datetime import datetime, timezone
+from datetime import datetime, UTC
 from pathlib import Path
 from typing import Any
 
@@ -76,7 +76,7 @@ def record_finding(code: str, session_id: str) -> dict[str, Any]:
     if session_id and session_id not in entry["sessions"]:
         entry["sessions"].append(session_id)
     entry["total_fires"] = int(entry["total_fires"]) + 1
-    entry["last_seen"] = datetime.now(timezone.utc).isoformat(timespec="seconds")
+    entry["last_seen"] = datetime.now(UTC).isoformat(timespec="seconds")
     save_state(state)
     return entry
 
@@ -118,7 +118,7 @@ def record_proposal_fingerprint(code: str, fingerprint: str, status: str) -> Non
     props.setdefault(code, []).append({
         "fingerprint": fingerprint,
         "status": status,
-        "ts": datetime.now(timezone.utc).isoformat(timespec="seconds"),
+        "ts": datetime.now(UTC).isoformat(timespec="seconds"),
     })
     save_state(state)
 

{drydock_cli-2.10.2 → drydock_cli-2.10.4}/drydock/admiral/stager.py

@@ -10,7 +10,7 @@ from __future__ import annotations
 import logging
 import subprocess
 from dataclasses import dataclass
-from datetime import datetime, timezone
+from datetime import datetime, UTC
 from pathlib import Path
 
 from drydock.admiral import history, persistence
@@ -50,7 +50,7 @@ def stage(proposal: Proposal, validation: ValidationResult) -> StagedProposal |
     if not validation.ok:
         raise ValueError("refusing to stage a proposal that didn't pass validation")
 
-    ts = datetime.now(timezone.utc).strftime("%Y%m%d-%H%M%S")
+    ts = datetime.now(UTC).strftime("%Y%m%d-%H%M%S")
     branch = f"admiral/{ts}-{_safe_code(proposal.code)}"
 
     # Invariant: we ALWAYS branch from main.

{drydock_cli-2.10.2 → drydock_cli-2.10.4}/drydock/cli/cli.py

@@ -23,8 +23,7 @@ from drydock.core.logger import logger
 from drydock.core.paths import HISTORY_FILE
 from drydock.core.programmatic import run_programmatic
 from drydock.core.session.session_loader import SessionLoader
-from drydock.core.types import EntrypointMetadata, LLMMessage, OutputFormat, Role
-from drydock.core.utils import ConversationLimitException
+from drydock.core.types import EntrypointMetadata, LLMMessage, Role
 from drydock.setup.onboarding import run_onboarding
 
 
@@ -254,6 +253,7 @@ def run_cli(args: argparse.Namespace) -> None:
     # adding more prints.
     import time as _t
     _t0 = _t.perf_counter()
+
     def _phase(name: str, start: float) -> float:
         now = _t.perf_counter()
         logger.warning("[TIMING:startup] %s=%.2fs", name, now - start)

{drydock_cli-2.10.2 → drydock_cli-2.10.4}/drydock/cli/entrypoint.py

@@ -230,19 +230,50 @@ def main() -> None:
     # --local → set up local LLM provider without editing config
     if getattr(args, "local", None):
         os.environ["DRYDOCK_LOCAL_URL"] = args.local
-        # Detect model name from the server
-        try:
-            import httpx
-            resp = httpx.get(f"{args.local}/models", timeout=5)
-            if resp.status_code == 200:
-                models = resp.json().get("data", [])
-                if models:
-                    model_name = models[0]["id"]
-                    os.environ["DRYDOCK_LOCAL_MODEL"] = model_name
-                    rprint(f"[green]Using local model: {model_name} at {args.local}[/]")
-        except Exception:
-            os.environ["DRYDOCK_LOCAL_MODEL"] = "local"
-            rprint(f"[yellow]Using local server at {args.local} (couldn't detect model name)[/]")
+        # Respect an explicit env-var override (used by harness/CI).
+        # Otherwise probe /v1/models — retry once with a longer timeout
+        # because containerized tbench trials saw the 5s window expire on
+        # cold-start, falling back to model="local" which loses ALL
+        # Gemma 4 optimizations (slim prompt, tool disables, non-streaming).
+        if os.environ.get("DRYDOCK_LOCAL_MODEL", "").strip():
+            rprint(
+                f"[green]Using local model: "
+                f"{os.environ['DRYDOCK_LOCAL_MODEL']} at {args.local} "
+                f"(via DRYDOCK_LOCAL_MODEL env)[/]"
+            )
+        else:
+            model_name = None
+            try:
+                import httpx
+                for attempt in (1, 2):
+                    try:
+                        resp = httpx.get(f"{args.local}/models", timeout=15)
+                        if resp.status_code == 200:
+                            models = resp.json().get("data", [])
+                            if models and models[0].get("id"):
+                                model_name = models[0]["id"]
+                                break
+                    except Exception:
+                        if attempt == 2:
+                            raise
+            except Exception:
+                pass
+            if model_name:
+                os.environ["DRYDOCK_LOCAL_MODEL"] = model_name
+                rprint(f"[green]Using local model: {model_name} at {args.local}[/]")
+            else:
+                # Detection failed. Default to "gemma4" — drydock is
+                # optimized for Gemma 4 + llama.cpp per README, so this
+                # is the right fallback in the overwhelming common case.
+                # Loud warning lets the user override if they're running
+                # something else.
+                os.environ["DRYDOCK_LOCAL_MODEL"] = "gemma4"
+                rprint(
+                    f"[yellow]Couldn't detect model name at {args.local}; "
+                    f"defaulting to model=gemma4. Override with "
+                    f"DRYDOCK_LOCAL_MODEL=<name> if the server runs a "
+                    f"different model.[/]"
+                )
 
     is_interactive = args.prompt is None
     if is_interactive:

{drydock_cli-2.10.2 → drydock_cli-2.10.4}/drydock/cli/textual_ui/app.py

@@ -743,7 +743,6 @@ class DrydockApp(App):  # noqa: PLR0904
         else:
             # Clean up unreplaced placeholders
             skill_content = skill_content.replace("$ARGUMENTS", "")
-            import re
             for i in range(10):
                 skill_content = skill_content.replace(f"$ARGUMENTS[{i}]", "")
                 skill_content = skill_content.replace(f"${i}", "")

{drydock_cli-2.10.2 → drydock_cli-2.10.4}/drydock/cli/textual_ui/widgets/checkpoint_picker.py

@@ -23,7 +23,7 @@ from textual.widgets.option_list import Option
 from drydock.cli.textual_ui.widgets.no_markup_static import NoMarkupStatic
 
 if TYPE_CHECKING:
-    from drydock.core.checkpoint import Checkpoint, CheckpointStore
+    from drydock.core.checkpoint import Checkpoint
 
 
 _SECONDS_PER_MINUTE = 60
@@ -62,7 +62,7 @@ def _format_diff_stats(insertions: int, deletions: int,
     return " ".join(parts)
 
 
-def _build_checkpoint_option(cp: "Checkpoint",
+def _build_checkpoint_option(cp: Checkpoint,
                              diff_summary: str) -> Text:
     text = Text(no_wrap=True)
     text.append(f"#{cp.index:>3}  ", style="bold")
@@ -92,7 +92,7 @@ class CheckpointPickerApp(Container):
     class Cancelled(Message):
         """User hit Esc."""
 
-    def __init__(self, checkpoints: list["Checkpoint"],
+    def __init__(self, checkpoints: list[Checkpoint],
                  diff_summaries: dict[int, str], **kwargs: Any) -> None:
         super().__init__(id="checkpoint-picker", **kwargs)
         self._checkpoints = checkpoints

{drydock_cli-2.10.2 → drydock_cli-2.10.4}/drydock/cli/textual_ui/widgets/messages.py

@@ -380,7 +380,8 @@ class AssistantMessage(StreamingMessageBase):
                 self._content = rescued
                 await self._markdown.update("")
                 stream = self._ensure_stream()
-                await stream.write(rescued)
+                if stream is not None:
+                    await stream.write(rescued)
         await super().stop_stream()
 
     def compose(self) -> ComposeResult:
@@ -494,7 +495,8 @@ class ReasoningMessage(SpinnerMixin, StreamingMessageBase):
                     self._stream = None
                 await self._markdown.update("")
                 stream = self._ensure_stream()
-                await stream.write(self._content)
+                if stream is not None:
+                    await stream.write(self._content)
 
 
 class UserCommandMessage(Static):

{drydock_cli-2.10.2 → drydock_cli-2.10.4}/drydock/core/agent_loop.py

@@ -481,7 +481,7 @@ class AgentLoop:
             # Newest first, since later prompts usually supersede earlier
             # context. Number them so the model can refer back.
             numbered = "\n".join(
-                f"  ({i+1}) {text}"
+                f"  ({i + 1}) {text}"
                 for i, text in enumerate(reversed(injections))
             )
             note = (
@@ -1012,11 +1012,12 @@ class AgentLoop:
         self._last_post_edit_spec_fp = None
         # Clear any stale auto-goal state from a previous turn — the
         # user's new prompt may have nothing to do with the prior goal.
-        if (getattr(self, "goal", None) is not None
-                and getattr(self.goal, "active", False)):
+        _goal = getattr(self, "goal", None)
+        if _goal is not None and getattr(_goal, "active", False):
+            _cond = getattr(_goal, "condition", "") or ""
             logger.warning(
                 "[recovery] clearing stale goal on new user turn: %r",
-                self.goal.condition[:80] if self.goal.condition else "",
+                _cond[:80],
             )
             try:
                 self.clear_goal()
@@ -1347,7 +1348,7 @@ class AgentLoop:
                 if _elapsed > PER_PROMPT_BUDGET_SEC:
                     yield AssistantEvent(
                         content=(
-                            f"\n\n[Drydock: {int(_elapsed/60)} minutes "
+                            f"\n\n[Drydock: {int(_elapsed / 60)} minutes "
                             f"elapsed on this single prompt — over the "
                             f"{PER_PROMPT_BUDGET_SEC // 60}-min budget. "
                             "Stopping. Work done so far is on disk; "
@@ -1823,7 +1824,6 @@ class AgentLoop:
                     # message that didn't break loop because it had tool_calls)
                     self._consecutive_empty_turns = 0
                 else:
-                    streak = getattr(self, "_consecutive_empty_turns", 0)
                     # Count consecutive text-only assistant turns by walking
                     # backward through history.
                     text_only_streak = 0
@@ -2072,7 +2072,7 @@ class AgentLoop:
                                 f"⚠ PREMATURE EXIT BLOCK #{nudges + 1}/10. "
                                 f"The task explicitly names this verify "
                                 f"command: `{missing_verify_cmds[0]}`"
-                                f"{' (+%d more)' % (len(missing_verify_cmds)-1) if len(missing_verify_cmds) > 1 else ''}"
+                                f"{' (+%d more)' % (len(missing_verify_cmds) - 1) if len(missing_verify_cmds) > 1 else ''}"
                                 f". You have NOT run it. The external "
                                 f"verifier will use that exact command (or "
                                 f"equivalent) after you exit — run it NOW "
@@ -2165,9 +2165,8 @@ class AgentLoop:
                                 "go test", "python ", "python3 ",
                             )
                             if any(cmd.strip().startswith(p) for p in safe_prefixes):
+                                import subprocess as _sub
                                 try:
-                                    import subprocess as _sub
-                                    import shlex as _sx
                                     _proc = _sub.run(
                                         cmd, shell=True, capture_output=True,
                                         timeout=45, text=True,
@@ -2220,13 +2219,13 @@ class AgentLoop:
                 # system note and DON'T break — let the agent loop
                 # do another iteration. Cap at goal.max_iterations.
                 # Targets P1-S1 / P1-S2 partial-completion failures.
+                g = getattr(self, "goal", None)
                 if (should_break_loop
-                        and getattr(self, "goal", None) is not None
-                        and getattr(self.goal, "active", False)
+                        and g is not None
+                        and getattr(g, "active", False)
                         and os.environ.get(
                             "DRYDOCK_AUTO_GOAL", "1"
                         ).strip().lower() in ("1", "true", "yes")):
-                    g = self.goal
                     if g.iterations < g.max_iterations:
                         try:
                             ok, msg = self._verify_rename_goal(Path.cwd())
@@ -3135,7 +3134,7 @@ class AgentLoop:
             if len(last_result) > 800:
                 result_preview = (
                     last_result[:400]
-                    + f"\n  …[{len(last_result)-400} chars omitted on dedup re-display]"
+                    + f"\n  …[{len(last_result) - 400} chars omitted on dedup re-display]"
                 )
             else:
                 result_preview = last_result
@@ -3733,11 +3732,11 @@ class AgentLoop:
                     sr_args = tool_call.args_dict
                     sr_path = sr_args.get("file_path", sr_args.get("path", ""))
                     if sr_path and Path(sr_path).exists():
-                        with open(sr_path, "r", encoding="utf-8", errors="replace") as f:
+                        with open(sr_path, encoding="utf-8", errors="replace") as f:
                             lines = f.readlines()
                         # Show first 50 lines or the whole file if small
                         preview_lines = lines[:50]
-                        numbered = [f"{i+1}\t{line.rstrip()}" for i, line in enumerate(preview_lines)]
+                        numbered = [f"{i + 1}\t{line.rstrip()}" for i, line in enumerate(preview_lines)]
                         auto_read_content = (
                             f"\n\nAUTO-READ of {sr_path} (first {len(preview_lines)} lines):\n"
                             + "\n".join(numbered)
@@ -3886,7 +3885,7 @@ class AgentLoop:
         "Empty content provided",     # search_replace empty content
         "REFUSED: write blocked",     # write_file pre-write syntax gate (2026-05-31)
         "REFUSED: bash inplace file edit",       # bash sed -i redirect (2026-05-31)
-        "REFUSED: bash redirect-to-source-file", # bash > file redirect (2026-05-31)
+        "REFUSED: bash redirect-to-source-file",  # bash > file redirect (2026-05-31)
         # Pydantic schema rejections — model emitted `{}` for required-field tool.
         # Found in operator session 2026-05-31 14:24: model called bash({}) repeatedly
         # after a REFUSED sed -i, dispatcher rejected with this exact text, and the
@@ -3913,7 +3912,7 @@ class AgentLoop:
     )
 
     def _scrub_validation_error_call(
-        self, failed_tool_call: "ResolvedToolCall"
+        self, failed_tool_call: ResolvedToolCall
     ) -> None:
         """Strip the args of a validation-error tool_call from history.
 
@@ -4652,10 +4651,10 @@ class AgentLoop:
         )
         if not passed:
             note += (
-                f"\nThe test suite went RED after your edit. Read the failure "
-                f"above, fix the cause BEFORE declaring done or moving to a "
-                f"different file. Re-running pytest manually if you need a "
-                f"fresh view.\n"
+                "\nThe test suite went RED after your edit. Read the failure "
+                "above, fix the cause BEFORE declaring done or moving to a "
+                "different file. Re-running pytest manually if you need a "
+                "fresh view.\n"
             )
         self._inject_system_note(note)
         logger.warning(
@@ -5133,23 +5132,6 @@ class AgentLoop:
                 )
                 if not is_stub:
                     continue
-                # Extract path hint from any of the known fields.
-                path_hint = ""
-                for k in ("path", "file_path", "command", "cmd",
-                          "url", "file"):
-                    v = parsed.get(k)
-                    if isinstance(v, str) and v:
-                        path_hint = (
-                            v if len(v) <= 200 else v[:200] + "…"
-                        )
-                        break
-                # Try to recover an approximate original size.
-                orig_bytes = parsed.get("_original_bytes")
-                size_str = (
-                    f"original {int(orig_bytes)} bytes"
-                    if isinstance(orig_bytes, (int, float))
-                    else "compacted"
-                )
                 # 2026-05-29: upgrade legacy AND post-2026-05-24 stubs
                 # (which used the marker key `__drydock_compacted_args__`)
                 # to the new empty-args format `{}`. The marker-bearing
@@ -5158,10 +5140,9 @@ class AgentLoop:
                 # fake arg, triggering format.py's placeholder-loop
                 # recovery dance. Empty args produce a clean pydantic
                 # field-required error instead. Path hint and size
-                # info from `note` is dropped — debugging value was
-                # offset by the copy-bait cost.
+                # info dropped — debugging value was offset by the
+                # copy-bait cost.
                 tc.function.arguments = "{}"
-                _ = size_str  # retain for future logging if useful
                 try:
                     self.stats.legacy_stubs_upgraded += 1
                 except Exception:
@@ -6158,7 +6139,7 @@ class AgentLoop:
                         # Truncate args for readability
                         if len(args_str) > 120:
                             args_str = args_str[:120] + "..."
-                        recent.append(f"  {count+1}. {name}({args_str})")
+                        recent.append(f"  {count + 1}. {name}({args_str})")
                         count += 1
                         if count >= 8:
                             break
@@ -6166,11 +6147,11 @@ class AgentLoop:
                 # Add result summary to last entry
                 result = str(msg.content or "")[:150]
                 if "error" in result.lower() or "Error" in result:
-                    recent[-1] += f" → ERROR"
+                    recent[-1] += " → ERROR"
                 elif "not found" in result.lower():
-                    recent[-1] += f" → NOT FOUND"
+                    recent[-1] += " → NOT FOUND"
                 else:
-                    recent[-1] += f" → ok"
+                    recent[-1] += " → ok"
 
         if len(recent) < 3:
             return None
@@ -6592,7 +6573,6 @@ class AgentLoop:
         ):
             return
 
-        from drydock.core.prompts import SystemPrompt
         try:
             # SystemPrompt enum value names lowercased = file stems.
             # gemma4_math isn't in the enum — load by path.
@@ -7196,6 +7176,43 @@ class AgentLoop:
             logger.warning("[AUTO-RETRIEVE] setup failed: %s", e, exc_info=True)
             return
 
+        # 2026-06-08: first-call lazy bootstrap. Inside fresh containers
+        # (tbench trials, fresh pip-install drydock-cli) the home DB
+        # doesn't exist, so AUTO-RETRIEVE silently no-ops and the
+        # bundled cookbook never gets a chance. Detect that case and
+        # ingest the bundled cookbook into ~/.drydock/graphrag.sqlite
+        # one time. Opt-out via DRYDOCK_AUTO_BOOTSTRAP_GRAPHRAG=0.
+        if not getattr(self, "_graphrag_bootstrap_attempted", False):
+            self._graphrag_bootstrap_attempted = True
+            if os.environ.get(
+                "DRYDOCK_AUTO_BOOTSTRAP_GRAPHRAG", "1"
+            ).strip().lower() not in ("0", "false", "no"):
+                try:
+                    _home_db_check = Path.home() / ".drydock" / "graphrag.sqlite"
+                    if not _home_db_check.is_file():
+                        import drydock as _dd
+                        _bundle = (
+                            Path(_dd.__file__).parent
+                            / "data" / "cookbook"
+                        )
+                        if _bundle.is_dir():
+                            _home_db_check.parent.mkdir(parents=True, exist_ok=True)
+                            logger.warning(
+                                "[AUTO-RETRIEVE] bootstrap: ingesting bundled "
+                                "cookbook %s into %s",
+                                _bundle, _home_db_check,
+                            )
+                            _bootstrap_idx = Index(str(_home_db_check))
+                            _counts = _bootstrap_idx.ingest_path(_bundle)
+                            logger.warning(
+                                "[AUTO-RETRIEVE] bootstrap done: %d files / "
+                                "%d chunks",
+                                _counts.get("files", 0),
+                                _counts.get("chunks", 0),
+                            )
+                except Exception as e:
+                    logger.warning("[AUTO-RETRIEVE] bootstrap failed: %s", e)
+
         # Skip auto-retrieve for file-producing tasks (terminal-bench /
         # harbor wraps everything with "This task has multiple steps...").
         # Retrieve adds 2K tokens of (usually irrelevant) cookbook content
@@ -7219,13 +7236,60 @@ class AgentLoop:
             # tool calls (write_file/bash) not a text-only retrieval response.
             or ("__init__.py" in msg_text and "Requirements:" in msg_text)
         )
-        if is_file_task and os.environ.get(
+        # 2026-06-08: file-task skip used to be unconditional unless
+        # DRYDOCK_AUTO_RETRIEVE_FORCE_ON. That blocked cookbook chunks
+        # (path-tracing, MIPS, FEAL, etc.) from helping tbench multi-step
+        # prompts — exactly the cases where domain cookbook IS the
+        # missing piece. New behavior: do a probe retrieve first; if the
+        # top hit scores >= AUTO_RETRIEVE_BYPASS_SCORE (25.0 by default,
+        # well above the 8.0 quality floor and the typical "incidental
+        # match" range), inject it regardless of file-task heuristic.
+        # Generic noise (binary_search code injected into chess-best-move)
+        # rarely scores that high — the chess prompt's top hit was ~12.
+        _bypass = False
+        if is_file_task:
+            try:
+                _bypass_thresh = float(os.environ.get(
+                    "DRYDOCK_AUTO_RETRIEVE_BYPASS_SCORE", "25.0"
+                ))
+            except ValueError:
+                _bypass_thresh = 25.0
+            # Cheap probe — short query, just enough to read the top score.
+            try:
+                from drydock.graphrag import Index as _Probe
+                _probe_query = (user_msg or "").strip()[:300]
+                if _probe_query:
+                    _env_db = os.environ.get("DRYDOCK_GRAPHRAG_DB")
+                    _home_db = str(Path.home() / ".drydock" / "graphrag.sqlite")
+                    _project_db = Path.cwd() / ".drydock" / "graphrag.sqlite"
+                    _probe_db = _env_db or (
+                        str(_project_db) if _project_db.is_file() else _home_db
+                    )
+                    if Path(_probe_db).is_file():
+                        _r = _Probe(_probe_db).retrieve(
+                            _probe_query, symbol_limit=0, text_limit=1
+                        )
+                        _top = _r.text[0] if _r.text else None
+                        _top_score = getattr(_top, 'score', 0.0) if _top else 0.0
+                        if _top_score >= _bypass_thresh:
+                            _bypass = True
+                            logger.warning(
+                                "[AUTO-RETRIEVE] file-task SKIP bypassed by "
+                                "strong cookbook hit (score=%.1f >= %.1f)",
+                                _top_score, _bypass_thresh,
+                            )
+            except Exception as e:
+                logger.warning("[AUTO-RETRIEVE] probe failed: %s", e)
+
+        if is_file_task and not _bypass and os.environ.get(
             "DRYDOCK_AUTO_RETRIEVE_FORCE_ON", ""
         ).strip().lower() not in ("1", "true", "yes"):
             logger.warning(
                 "[AUTO-RETRIEVE] skipped — file-production task detected "
-                "(harbor wrap / multiple /app/ or /tmp/ paths). Override "
-                "with DRYDOCK_AUTO_RETRIEVE_FORCE_ON=1."
+                "(harbor wrap / multiple /app/ or /tmp/ paths) and probe "
+                "found no strong cookbook hit. Override with "
+                "DRYDOCK_AUTO_RETRIEVE_FORCE_ON=1 or lower "
+                "DRYDOCK_AUTO_RETRIEVE_BYPASS_SCORE."
             )
             return
 
@@ -7301,7 +7365,6 @@ class AgentLoop:
             db_chain.append(fallback_db)
 
         good_hits: list = []
-        text_hits: list = []
         used_db: str | None = None
         for db in db_chain:
             if not Path(db).is_file():
@@ -7336,7 +7399,6 @@ class AgentLoop:
                 db, len(hits), len(gh), QUALITY_THRESHOLD,
             )
             if gh:
-                text_hits = hits
                 good_hits = gh
                 used_db = db
                 break

{drydock_cli-2.10.2 → drydock_cli-2.10.4}/drydock/core/auto_solve.py

@@ -77,6 +77,7 @@ def _emit_telemetry(record: dict) -> None:
     except OSError:
         pass
 
+
 if TYPE_CHECKING:
     from drydock.core.constraint_extract import ExtractResult
 
@@ -95,7 +96,7 @@ _DEFAULT_BOUNDS = {
 }
 
 
-def _build_args(extr: "ExtractResult") -> dict | None:
+def _build_args(extr: ExtractResult) -> dict | None:
     """Build the SolveArgs dict for this ExtractResult, or None if the
     predicate isn't one Z3 can decide. The dict is what the model
     *would* have written when calling the tool — used both for the
@@ -249,7 +250,7 @@ def _run_solve_sync(args_dict: dict) -> dict | None:
         # bounds the whole orchestration including streaming setup.
         wall_ms = int(args_dict.get("timeout_ms", 15000)) + 5000
         result = asyncio.run(asyncio.wait_for(_run(), timeout=wall_ms / 1000))
-    except asyncio.TimeoutError:
+    except TimeoutError:
         logger.warning("[AUTO-SOLVE] hard timeout exceeded")
         return None
     except RuntimeError as e:
@@ -267,7 +268,7 @@ def _run_solve_sync(args_dict: dict) -> dict | None:
 
 
 def _format_result(args_dict: dict, result_dict: dict,
-                   extr: "ExtractResult") -> str:
+                   extr: ExtractResult) -> str:
     """Render the Solve result the way the real Solve tool prints it,
     matching what the model sees from a normal tool call. Adds a
     one-line summary at the bottom so the model can extract the