drydock-cli 3.0.106__tar.gz → 3.0.107__tar.gz

{drydock_cli-3.0.106/drydock_cli.egg-info → drydock_cli-3.0.107}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 3.0.106
+Version: 3.0.107
 Summary: Drydock — a local, provider-agnostic terminal coding agent for local LLMs
 Author: Frank Bobe III
 License-Expression: Apache-2.0

{drydock_cli-3.0.106 → drydock_cli-3.0.107}/drydock/agent.py

@@ -13,6 +13,20 @@ from typing import Generator
 # stall-loop while letting a long, productive plan run as far as it needs.
 PLAN_CONTINUE_CAP = 3
 
+# Over-think interrupt: when a call keeps stalling/over-thinking (repeated
+# StallRetry), re-issue in "decisive mode" — a forcing suffix + a hard token cap
+# so the model physically cannot burn thousands of reasoning tokens without
+# acting. The model that produced this loop (gemma) generates 5-15 min / 5k+
+# token "thinking" turns that stall_retry alone can't break (tokens flow); this
+# forces a short, single-action turn instead.
+_DECISIVE_MAX_TOKENS = 1500
+_DECISIVE_SUFFIX = (
+    "\n\nURGENT — you have spent too long reasoning without acting. STOP "
+    "analyzing. Your reply must be SHORT and take exactly ONE concrete action "
+    "right now: emit a single tool call (run a command or write a file) with "
+    "little or no explanation. Do NOT lay out a plan; act."
+)
+
 
 def _plan_has_unfinished(config: dict) -> bool:
     """True if the model laid out a `todo` plan that still has non-done items."""
@@ -160,6 +174,7 @@ def run(
         # Stream from LLM — with retry on context-length 400 error
         retries = 0
         stall_retries = 0
+        decisive = False  # over-think interrupt: force a short, single-action turn
         while retries < 2:
             try:
                 available = schemas()
@@ -167,7 +182,7 @@ def run(
                     available = [s for s in available if s.get("name") in allow]
                 for event in stream(
                     model=turn_config["model"],
-                    system=system_prompt,
+                    system=(system_prompt + _DECISIVE_SUFFIX) if decisive else system_prompt,
                     messages=state.messages,
                     tool_schemas=filter_tool_schemas(available, turn_config.get("model")),
                     config=turn_config,
@@ -180,17 +195,25 @@ def run(
                         assistant_turn = event
                 break  # success
             except StallRetry:
-                # The local server hung mid-generation (config stall_retry_secs).
-                # Re-issue the same request — a fresh generation usually isn't
-                # stalled. Bounded so a truly-dead server can't loop forever; on
-                # exhaustion end the turn cleanly. Stalls have their OWN budget,
-                # separate from the context-error `retries`.
+                # The call stalled or over-thought past stall_retry_secs. First
+                # retry re-issues as-is (handles a transient server hang cheaply);
+                # if it KEEPS happening it's over-thinking, so escalate to decisive
+                # mode — a forcing suffix + a hard token cap that make a long
+                # reasoning turn impossible. Bounded; on exhaustion end cleanly.
                 stall_retries += 1
                 if stall_retries > 3:
-                    yield TextChunk("\n[model server kept stalling — giving up on this step.]\n")
+                    yield TextChunk("\n[model kept stalling/over-thinking — giving up on this step.]\n")
                     assistant_turn = None
                     break
-                yield TextChunk("\n[model server stalled — retrying...]\n")
+                if stall_retries >= 2 and not decisive:
+                    decisive = True
+                    turn_config = dict(turn_config)
+                    cur = int(turn_config.get("max_tokens", 8192) or 8192)
+                    turn_config["max_tokens"] = min(cur, _DECISIVE_MAX_TOKENS)
+                    turn_config["reasoning_effort"] = "low"
+                    yield TextChunk("\n[taking too long — forcing a decisive, single-action step...]\n")
+                else:
+                    yield TextChunk("\n[model server stalled — retrying...]\n")
                 continue
             except Exception as e:
                 err = str(e)

{drydock_cli-3.0.106 → drydock_cli-3.0.107}/drydock/tuning.py

@@ -97,6 +97,10 @@ _DEFAULT_SYSTEM_PROMPT = (
     "RUN it and fix failures until it passes before you stop. Prefer acting — write "
     "files and run commands — over long planning; don't produce a big plan without "
     "taking a concrete step.\n"
+    "A script you write must run with ONLY what is already installed — prefer "
+    "the standard library or shelling out to the task's own tools (openssl, git, "
+    "sqlite3); do NOT depend on a package you just pip-installed unless the task "
+    "explicitly says to (the grader runs in a clean environment).\n"
     "You have VISION: when the user references an image file path (.png/.jpg/"
     ".jpeg/.gif/.webp/.bmp on disk — a screenshot, mockup, diagram, or photo), it "
     "is attached and you can SEE it; you can also call the `ViewImage` tool to "
@@ -133,6 +137,10 @@ _GEMMA_SYSTEM_PROMPT = (
     "RUN it and fix failures until it passes before you stop. Prefer acting — write "
     "files and run commands — over long planning; don't produce a big plan without "
     "taking a concrete step.\n"
+    "A script you write must run with ONLY what is already installed — prefer "
+    "the standard library or shelling out to the task's own tools (openssl, git, "
+    "sqlite3); do NOT depend on a package you just pip-installed unless the task "
+    "explicitly says to (the grader runs in a clean environment).\n"
     "You have VISION: when the user's message references an image file path "
     "(.png/.jpg/.jpeg/.gif/.webp/.bmp that exists on disk — a screenshot, mockup, "
     "diagram, or photo), that image is attached and you can SEE it. You can also "

{drydock_cli-3.0.106 → drydock_cli-3.0.107/drydock_cli.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 3.0.106
+Version: 3.0.107
 Summary: Drydock — a local, provider-agnostic terminal coding agent for local LLMs
 Author: Frank Bobe III
 License-Expression: Apache-2.0

{drydock_cli-3.0.106 → drydock_cli-3.0.107}/drydock_cli.egg-info/SOURCES.txt

@@ -87,6 +87,7 @@ tests/test_leaked_tool_call.py
 tests/test_loop_detect.py
 tests/test_mcp.py
 tests/test_oneshot_unreachable.py
+tests/test_overthink_interrupt.py
 tests/test_plan_autocontinue.py
 tests/test_poam.py
 tests/test_providers_unreachable.py

{drydock_cli-3.0.106 → drydock_cli-3.0.107}/pyproject.toml

@@ -7,7 +7,7 @@ build-backend = "setuptools.build_meta"
 # PyPI distribution name is drydock-cli (the established install name, continued
 # from the retired v2 fork); the import package + CLI command stay `drydock`.
 name = "drydock-cli"
-version = "3.0.106"
+version = "3.0.107"
 description = "Drydock — a local, provider-agnostic terminal coding agent for local LLMs"
 readme = "README.md"
 requires-python = ">=3.11"

drydock_cli-3.0.107/tests/test_overthink_interrupt.py

@@ -0,0 +1,47 @@
+"""Over-think interrupt: when a call keeps stalling/over-thinking (repeated
+StallRetry), the agent escalates to 'decisive mode' — a forcing suffix on the
+system prompt + a hard max_tokens cap — so the model can't burn a 5k-token
+reasoning turn without acting. First retry re-issues as-is; 2nd+ goes decisive."""
+from __future__ import annotations
+
+from drydock import agent
+from drydock.agent import AgentState, _DECISIVE_SUFFIX, _DECISIVE_MAX_TOKENS
+from drydock.providers import StallRetry, AssistantTurn
+
+
+def _drive(monkeypatch, stalls):
+    calls = []
+
+    def fake_stream(model, system, messages, tool_schemas, config):
+        calls.append({"system": system, "max_tokens": config.get("max_tokens"),
+                      "effort": config.get("reasoning_effort")})
+        if len(calls) <= stalls:
+            raise StallRetry(1)
+        yield AssistantTurn(text="done", tool_calls=[], input_tokens=10, output_tokens=5)
+
+    monkeypatch.setattr(agent, "stream", fake_stream)
+    state = AgentState()
+    cfg = {"model": "gemma4", "max_tokens": 8192, "stall_retry_secs": 1, "context_limit": 65536}
+    list(agent.run("do it", state, cfg, "SYS"))
+    return calls
+
+
+def test_first_retry_not_decisive_second_is(monkeypatch):
+    calls = _drive(monkeypatch, stalls=2)   # stall twice, succeed on 3rd
+    assert len(calls) == 3
+    assert _DECISIVE_SUFFIX not in calls[0]["system"]      # attempt 1: as-is
+    assert _DECISIVE_SUFFIX not in calls[1]["system"]      # attempt 2 (retry 1): as-is
+    assert _DECISIVE_SUFFIX in calls[2]["system"]          # attempt 3 (retry 2): decisive
+    assert calls[2]["max_tokens"] <= _DECISIVE_MAX_TOKENS   # hard token cap
+    assert calls[2]["effort"] == "low"
+
+
+def test_gives_up_after_bounded_retries(monkeypatch):
+    # stalls forever → bounded (>3) → ends cleanly, doesn't loop forever
+    calls = _drive(monkeypatch, stalls=99)
+    assert len(calls) <= 5   # 1 initial + up to ~3 retries then give up
+
+
+def test_no_stall_no_decisive(monkeypatch):
+    calls = _drive(monkeypatch, stalls=0)   # succeeds first try
+    assert len(calls) == 1 and _DECISIVE_SUFFIX not in calls[0]["system"]