drydock-cli 2.8.43__tar.gz → 2.8.44__tar.gz

drydock_cli-2.8.44/.auto_release.lock

@@ -0,0 +1 @@
+2170697

{drydock_cli-2.8.43 → drydock_cli-2.8.44}/CLAUDE.md

@@ -18,6 +18,15 @@ Drydock is a local CLI coding agent (fork of mistral-vibe, Apache 2.0).
   AWQ-4bit on vLLM is the historical alternate path and stays available
   via `/data3/Models/start_gemma4.sh` (the vLLM script — NOT the
   llama.cpp one). Speed: ~64 tok/s e2e on Q3.
+- **Vision (2026-05-16, GH #23):** `/data3/Models/mmproj-F16.gguf`
+  (1.2 GB, unsloth/gemma-4-26B-A4B-it-GGUF). Loaded via
+  `--mmproj /models/mmproj-F16.gguf` in `start_gemma4_llamacpp.sh`.
+  Container reports `projector: gemma4v` and accepts OpenAI-style
+  multimodal messages (`content: [{type:text}, {type:image_url}]`).
+  Verified with a base64 test image — model received 284 prompt tokens
+  for `text + 1×1 image`. Server-side ONLY; drydock TUI does not yet
+  have an image-attach affordance — the server-side capability is the
+  prereq for that work.
 - **The honest test:** `scripts/shakedown_interactive.py` — drives the real TUI
   via pexpect with a multi-step user conversation (24 steps per PRD: plan,
   build, test, add features, debug, review code, edge cases, README).

{drydock_cli-2.8.43 → drydock_cli-2.8.44}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.8.43
+Version: 2.8.44
 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-2.8.43 → drydock_cli-2.8.44}/drydock/core/agent_loop.py

@@ -729,6 +729,23 @@ class AgentLoop:
         if user_message.message_id is None:
             raise AgentLoopError("User message must have a message_id")
 
+        # Reset sticky error counters on every fresh user turn. Without
+        # this, once `_total_error_rounds` hits the 3-round stop ceiling
+        # (~45 API errors), it stays at 3 forever — every subsequent
+        # user message immediately re-trips the ceiling on its first
+        # API error and aborts. Users were stuck typing /clear (which
+        # wipes the whole session) just to recover. The user has
+        # manually intervened by typing again; they earn a fresh
+        # error budget. The bad messages may also have been
+        # dropped/compacted by the previous round's recovery path, so
+        # the new turn often succeeds where the prior round couldn't.
+        if getattr(self, "_total_error_rounds", 0) > 0:
+            logger.warning(
+                "[recovery] resetting _total_error_rounds=%d → 0 on new user turn",
+                self._total_error_rounds,
+            )
+            self._total_error_rounds = 0
+
         # Flush the user message to disk RIGHT NOW, before the LLM call.
         # Without this, messages.jsonl only updates after the model yields
         # — for silent/slow prompts the user message is invisible to any
@@ -1050,8 +1067,44 @@ class AgentLoop:
                             self._total_error_rounds = 0
                         self._total_error_rounds += 1
                         if self._total_error_rounds >= 3:
+                            # Hard-stop after 3 rounds of recovery attempts.
+                            # Drop the LAST user→assistant block so the next
+                            # user message doesn't immediately re-trigger the
+                            # same broken state. Counter resets on the next
+                            # _conversation_loop entry, so the user can just
+                            # type the next message and continue without
+                            # losing their entire context.
+                            try:
+                                # Find the last user message; truncate to it.
+                                # Keep everything up to AND INCLUDING that
+                                # message; drop the assistant garbage after.
+                                last_user_idx = -1
+                                for i in range(len(self.messages) - 1, -1, -1):
+                                    if self.messages[i].role == Role.user:
+                                        last_user_idx = i
+                                        break
+                                if last_user_idx >= 0 and last_user_idx < len(self.messages) - 1:
+                                    kept = list(self.messages[: last_user_idx + 1])
+                                    self.messages.reset(kept)
+                                    logger.warning(
+                                        "[recovery] hard-stop: dropped %d messages "
+                                        "after last user turn (idx=%d)",
+                                        len(self.messages) - last_user_idx - 1,
+                                        last_user_idx,
+                                    )
+                            except Exception as _drop_err:  # noqa: BLE001
+                                logger.warning(
+                                    "[recovery] hard-stop drop failed: %s",
+                                    _drop_err,
+                                )
                             yield AssistantEvent(
-                                content=f"\n\n[Stopping: {self._total_error_rounds * MAX_API_ERRORS}+ API errors. The model cannot process this request. Try /compact or /clear to free context.]\n",
+                                content=(
+                                    f"\n\n[Stopping after {self._total_error_rounds * MAX_API_ERRORS}+ "
+                                    f"API errors. Conversation rolled back to your "
+                                    f"last message — just type your next request "
+                                    f"to continue. (Use /compact if context is "
+                                    f"genuinely too long, /clear only to fully reset.)]\n"
+                                ),
                             )
                             return
 

{drydock_cli-2.8.43 → drydock_cli-2.8.44}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.8.43"
+version = "2.8.44"
 description = "Local-first CLI coding agent — tested with Gemma 4 26B via vLLM"
 readme = "README.md"
 requires-python = ">=3.12"

drydock_cli-2.8.44/tests/test_api_error_recovery.py

@@ -0,0 +1,149 @@
+"""Regression tests for the unrecoverable-API-error fix.
+
+The bug (reported 2026-05-16): once the agent loop hit
+`_total_error_rounds >= 3`, it stayed stuck — every subsequent user
+message tripped the ceiling on its first API error and aborted.
+The only escape was `/clear`, which wipes the entire session.
+
+The fix has two parts:
+  1. Reset `_total_error_rounds` to 0 at the top of each new
+     `_conversation_loop` (= new user turn).
+  2. On the hard-stop path, drop messages after the last user
+     turn so the broken state doesn't immediately re-trip the same
+     error on the next attempt.
+"""
+from __future__ import annotations
+
+import os
+from unittest.mock import MagicMock
+
+
+def _make_stub_loop():
+    """Minimal stub object that exercises just the reset semantics
+    without needing a full AgentLoop construction (which requires
+    a real backend, tools, config, etc.)."""
+    loop = MagicMock()
+    loop._total_error_rounds = 0
+    return loop
+
+
+def test_total_error_rounds_starts_at_zero():
+    """Baseline — fresh agent loops should have rounds=0."""
+    loop = _make_stub_loop()
+    assert loop._total_error_rounds == 0
+
+
+def test_reset_on_new_user_turn_clears_sticky_state():
+    """The hard-stop ceiling was sticky pre-fix. After the fix,
+    `_conversation_loop` resets `_total_error_rounds` at entry so
+    subsequent user messages get a fresh error budget."""
+    # Simulate the reset block from _conversation_loop:
+    loop = _make_stub_loop()
+    loop._total_error_rounds = 3  # hit the stop ceiling
+
+    # The fix logic — extracted to validate it in isolation:
+    if getattr(loop, "_total_error_rounds", 0) > 0:
+        loop._total_error_rounds = 0
+
+    assert loop._total_error_rounds == 0
+
+
+def test_drop_messages_after_last_user_turn_on_hard_stop():
+    """Hard-stop path should rewind history to the last user
+    message so the next turn doesn't re-trigger the same error.
+    Validates the slice/keep logic independent of AgentLoop."""
+
+    class FakeMsg:
+        def __init__(self, role, content=""):
+            self.role = role
+            self.content = content
+
+    class Role:
+        user = "user"
+        assistant = "assistant"
+        tool = "tool"
+
+    msgs = [
+        FakeMsg(Role.user, "first prompt"),
+        FakeMsg(Role.assistant, "first reply"),
+        FakeMsg(Role.user, "second prompt"),
+        FakeMsg(Role.assistant, "broken tool_call attempt"),
+        FakeMsg(Role.tool, "error result"),
+        FakeMsg(Role.assistant, "another bad attempt"),
+    ]
+
+    # Find last user message index — matches the fix's loop
+    last_user_idx = -1
+    for i in range(len(msgs) - 1, -1, -1):
+        if msgs[i].role == Role.user:
+            last_user_idx = i
+            break
+
+    assert last_user_idx == 2
+    kept = msgs[: last_user_idx + 1]
+    assert len(kept) == 3
+    assert kept[-1].role == Role.user
+    assert kept[-1].content == "second prompt"
+
+
+def test_drop_is_safe_when_last_message_is_already_user():
+    """If the last message is already a user turn (nothing after),
+    the drop should be a no-op."""
+
+    class FakeMsg:
+        def __init__(self, role, content=""):
+            self.role = role
+            self.content = content
+
+    class Role:
+        user = "user"
+        assistant = "assistant"
+
+    msgs = [
+        FakeMsg(Role.user, "only message"),
+    ]
+    last_user_idx = -1
+    for i in range(len(msgs) - 1, -1, -1):
+        if msgs[i].role == Role.user:
+            last_user_idx = i
+            break
+
+    # The fix's guard: `last_user_idx < len(self.messages) - 1`
+    should_drop = last_user_idx >= 0 and last_user_idx < len(msgs) - 1
+    assert should_drop is False  # nothing to drop
+
+
+def test_reset_skipped_when_already_zero():
+    """The reset path logs when rounds > 0 and skips otherwise — no
+    side effects for healthy sessions."""
+    loop = _make_stub_loop()
+    loop._total_error_rounds = 0
+
+    triggered_reset = False
+    if getattr(loop, "_total_error_rounds", 0) > 0:
+        triggered_reset = True
+        loop._total_error_rounds = 0
+
+    assert triggered_reset is False
+    assert loop._total_error_rounds == 0
+
+
+# ── Integration via the real _conversation_loop entry ──────────────────
+
+def test_conversation_loop_resets_total_error_rounds_on_entry():
+    """The actual code path: `_conversation_loop` should reset the
+    counter to 0 before any LLM work happens. We patch AgentLoop's
+    dependencies to skip the LLM call and just verify the counter."""
+    # This is a partial-integration test — it imports the real method
+    # and exercises its early code path. We don't run a full LLM turn.
+    from drydock.core.agent_loop import AgentLoop  # noqa: F401
+
+    # Verify the source code contains the reset block at the
+    # `_conversation_loop` site. (A behavioural test would require
+    # mocking the whole agent loop; this string-check is the
+    # cheapest reliable regression.)
+    import inspect
+    src = inspect.getsource(AgentLoop._conversation_loop)
+    assert "_total_error_rounds" in src
+    assert "= 0" in src
+    assert "fresh user turn" in src or "Reset sticky error" in src

{drydock_cli-2.8.43 → drydock_cli-2.8.44}/trip_log.md

@@ -1,5 +1,53 @@
 # Drydock Trip Log
 
+## 2026-05-16 12:00 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- Write rate: N/A (stress paused)
+- HLE burndown: running (PID 2157358, slot 33 "other" category); lifetime 12.1% (72/594): Math 17.1%, Bio/Med 14.3%, CS/AI 2.4%, Chemistry 6.1%, Engineering 6.1%, Physics 3.1%
+- Recent batches declining (60%→0%) as burndown cycles into hard CS/AI questions — not a regression; CS/AI 2.4% is baseline for this model on symbolic/graph-theory problems; earlier Math 6/10 batch confirms Math loop is healthy
+- vLLM 400s: 0 (llamacpp-gemma4 healthy, reasoning-budget=12000 confirmed in container config)
+- GH issues: 1 open (#23 image-processing multimodal enhancement — not actionable this tick)
+- Dispatch queue: harness=8 (all thinking_stall, fully addressed by prior commits), retrieval=0 actionable, curiosity=1950 pending (316 hle_failure, 1635 unknown_term)
+- Retrieval drain: 0 projects ingested (all already ingested recently)
+- Curiosity: consumed 30cceb0645e3c41e (JS timeout hle_failure — pre-fix thinking stall, addressed by reasoning-budget=12000 + STOP_NOW fixes)
+- Action this tick: no code changes — all dispatch patterns addressed, CS/AI 0% is model capability not a drydock bug
+
+## 2026-05-16 12:30 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- Write rate: N/A (stress paused)
+- HLE burndown: running (PID 2157358, 43m elapsed); lifetime 12.3% (74/600): Math 17.1%, Bio/Med 14.3%, CS/AI 2.4%, Chemistry 6.1%, Other 18.2%
+- vLLM 400s: 0 (llamacpp-gemma4 healthy, Q3_K_M, reasoning-budget=12000)
+- GH issues: 1 open (#23 image-processing enhancement — not actionable)
+- Dispatch queue: harness=8 (all thinking_stall, fully addressed), retrieval=0 actionable, curiosity=1971 pending (319 hle_failure, 1652 unknown_term)
+- Retrieval drain: 0 projects ingested (all already ingested recently)
+- Curiosity: consumed 21fc1bf2f3913021 (directed multigraph cycle cover — model used tools but no FINAL ANSWER; addressed by 23f3530 force tool_choice=none + b0350cb stall nudge)
+- Action this tick: no code changes — system nominal; all harness:thinking_stall patterns addressed by prior 24h commits; HLE score holding steady at 12.3%
+
+## 2026-05-16 11:30 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- Write rate: N/A (stress paused)
+- HLE burndown: alive (PID 2144966, 52 min), lifetime 71/589 = 12.1% (Math 50/292 = 17.1%, CS/AI 2/81 = 2.5%)
+- Recent batch trend: scores dropped 60%→0% over last 8 batches — NOT a regression. Category rotation has been hitting CS/AI (10:40 UTC 0/10, 10:46 UTC 0/8 ongoing) and Biology (10%). CS/AI is historically 2.5% for this model; both recent CS/AI batches are consistent with baseline. Earlier 6/10 Math batch confirms Math is healthy at ~17%.
+- vLLM (llamacpp-gemma4): healthy, Q3_K_M, reasoning-budget=12000 confirmed active
+- Balancer: healthy on :8001
+- Autonomous review: started 11:30 UTC (prior tick ended 11:05 UTC exit=0)
+- GH issues: 1 open (#23 image processing enhancement, non-critical)
+- Dispatch queue: harness=12 (thinking_stall, all addressed per evidence), retrieval=0 actionable (3 entries, already ingested), curiosity=1945 pending (1632 unknown_term, 314 hle_failure)
+- Action this tick: consumed curiosity d287eae28286f71b (Make 24 no-FINAL-ANSWER case, addressed by 23f3530 force-tool_choice-none fix). No code change — system healthy, HLE running, all loops nominal.
+
+## 2026-05-16 11:00 UTC tick
+- Stress: paused (.pause_stress sentinel active)
+- Write rate: N/A (stress paused)
+- HLE burndown: alive (PID 2144966, 19h+ elapsed), lifetime 71/584 = 12.2% (Math 50/292 = 17.1%, CS/AI 2/76 = 2.6%)
+- Recent batch trend: score drop from ~30% (Math) to 0% (CS/AI) is category-rotation driven — model historically 2.6% on CS/AI vs 17.1% Math; empty:no_response on two coding questions is thinking-budget exhaustion (12000-token budget × ~50 tok/s ≈ 240s, plus prompt overhead)
+- vLLM (llamacpp-gemma4): healthy, Q3_K_M, reasoning-budget=12000
+- Balancer: healthy on :8001
+- Autonomous review: started 11:00 UTC (prior run ended 10:35 UTC)
+- GH issues: 1 open (#23 image processing enhancement, non-critical)
+- Dispatch queue: harness=12 (thinking_stall, all addressed per prior commits), retrieval=0 actionable, curiosity=1925 pending
+- Retrieval drain: 0 ingested (3 entries, all recently ingested)
+- Action this tick: consumed 3 stale curiosity items (chemistry no_final_answer from pre-v2.8.42 era, before judge fix e3a3a9d and stall fixes b0350cb/23f3530 — no code change needed)
+
 ## 2026-05-16 10:30 UTC tick
 - Stress: paused (.pause_stress sentinel active)
 - Write rate: N/A (stress paused)

drydock_cli-2.8.43/.auto_release.lock

@@ -1 +0,0 @@
-2148271