drydock-cli 2.6.38__tar.gz → 2.6.40__tar.gz

drydock_cli-2.6.40/.auto_release.lock

@@ -0,0 +1 @@
+844174

{drydock_cli-2.6.38 → drydock_cli-2.6.40}/CLAUDE.md

@@ -483,6 +483,50 @@ v3 is a from-scratch rewrite using nano-claude-code as foundation. 4 core files,
     are a better signal for whether drydock can handle real user work.
     Easy PRDs were ceiling-ed out — all passing means you can't tell
     what's broken.
+31. **Harness bugs faked all previous pass rates.** (2026-04-12) The
+    shakedown harness typed prompts via PTY but never verified the TUI
+    actually accepted them. Result: 25 out of 28 prompts were silently
+    dropped per session — the TUI only saw the first 3-5 prompts, and
+    the harness counted assistant/tool messages from that ongoing work
+    as "condition met." Every "24/24 passed" was a lie. Fix: wait for
+    a NEW user-role message to appear in the session log after typing,
+    retry up to 3x if it doesn't. Also wait for TUI to go idle (6s no
+    new messages) before typing the next prompt.
+32. **`pkg_works` via `--help` was the exact bug CLAUDE.md warns
+    against.** (2026-04-12) I reintroduced `python3 -m pkg --help` as
+    the pass criterion. Every PRD "passed." Actually only 3/11 worked
+    — the rest had broken imports, missing features, fake outputs.
+    Replace with `functional_tests.sh` per PRD that runs real feature
+    commands and checks real outputs (e.g., calculator must compute
+    23*47=1081, lang_interp must run let x=42; print(x*2) and output 84).
+33. **Have the TUI run the tests, not the harness.** Externally
+    running `functional_tests.sh` tests the CODE, not drydock's
+    ability to test-and-fix. Real test: TUI runs tests via bash tool,
+    observes failures, reads failing files, applies fixes, iterates.
+    Harness only observes the session log for 'RESULT: X passed, Y failed'.
+34. **RALPH loop works much better than pipelined prompts.** (2026-04-12)
+    `shakedown_interactive.py` with 24-30 prompts racing the TUI never
+    landed properly. `ralph_loop.py` with ONE prompt → wait for
+    completion → run tests → send failures back → iterate works great.
+    Got 10/11 packages to 100% test pass via this approach in one night.
+35. **Auto-rollback on test regression is essential.** Gemma 4 26B with
+    only 4B active params regresses passing tests when trying to fix
+    failing ones. Snapshot the package dir before each fix iteration;
+    if score goes down, restore. Even with "don't break passing tests"
+    in the prompt, the model will destroy working code.
+36. **Drydock takes 3-5 MINUTES to create a session dir after spawn.**
+    (2026-04-12) When GPU is busy from a prior run, drydock may take
+    up to 4 minutes before the session directory is even created. My
+    180s wait was timing out with 1 second to spare. Bumped to 300s.
+37. **meta.json is only written at session EXIT.** (2026-04-12) Can't
+    match sessions by working_directory during an active session.
+    `find_session()` must match on directory creation time relative
+    to harness start, not meta.json content. messages.jsonl is also
+    inconsistently flushed during the session.
+38. **NEVER use broad pkill/kill matching 'drydock'.** (2026-04-12) I
+    killed the user's active TUI session twice by doing
+    `ps aux | grep drydock | kill`. Only kill by tracked PIDs from
+    my own background tasks. User's TUI is also a drydock process.
 
 ## Testing
 
@@ -592,12 +636,22 @@ bash scripts/shakedown_suite.sh
   - doc_qa: 22/24, 693s, pkg YES (2 steps timeout)
   - prompt_optimizer: 24/24, 431s, pkg YES
 - Autonomous mega-prompt: 4/5 pass, 1 flaky (eval_harness thinking stall)
-  - tool_agent: 155s, pkg YES
-  - stock_screener: 207s, pkg YES
-  - eval_harness: 669s timeout, pkg NO (model stalled on complex evaluator code)
-  - doc_qa: 174s, pkg YES
-  - prompt_optimizer: 210s, pkg YES
 - Key fix: autonomous prompts now write cli.py FIRST + pkg_works condition
+- **NOTE: These were fake pass rates. The harness was dropping 25/28 prompts
+  silently. Real results via functional_tests.sh (below) showed only 3/11
+  packages actually worked.**
+
+*Functional test results (2026-04-12, post harness fixes, v2.6.38):*
+- Ladder: 5-min, 15-min, 30-min, 60-min tiers (complexity, not duration)
+- Session started: 29/49 tests passing (59%) via real functional_tests.sh
+- Session ended: **49/52 tests passing (94%)** via ralph_loop.py
+  - doc_qa: 5/5, prompt_optimizer: 4/4, tool_agent: 5/5
+  - stock_screener: 4/4, eval_harness: 4/4, mini_db: 4/7
+  - site_gen: 5/5, lang_interp: 5/5, pkg_manager: 4/4
+  - web_frame: 4/4, build_sys: 5/5
+- Only mini_db not at 100%: WHERE filter, UPDATE, DELETE don't work
+- Key tooling: scripts/ralph_loop.py replaces shakedown_interactive.py
+  as the primary test harness. One big prompt, observe, fix, iterate.
 
 *Small suite (older):*
 - 4/4 (roman_converter, prime_tool, todo_list, codec) with functional

{drydock_cli-2.6.38 → drydock_cli-2.6.40}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.6.38
+Version: 2.6.40
 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-2.6.38 → drydock_cli-2.6.40}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.6.38"
+version = "2.6.40"
 description = "Local-first CLI coding agent — tested with Gemma 4 26B via vLLM"
 readme = "README.md"
 requires-python = ">=3.12"

{drydock_cli-2.6.38 → drydock_cli-2.6.40}/scripts/ralph_loop.py

@@ -46,22 +46,45 @@ class SessionWatcher:
         self.messages: list[dict] = []
 
     def find_session(self) -> Path | None:
+        """Find the drydock session directory for our cwd.
+
+        Problem: drydock only writes meta.json at session EXIT (not start).
+        So we can't match on working_directory during the session.
+
+        Strategy: Pick the newest session dir created after our start_time.
+        If messages.jsonl exists in it (= drydock is actively writing), use it.
+        """
         if self.session_dir is not None:
             return self.session_dir
         if not SESSION_ROOT.exists():
             return None
-        for entry in sorted(SESSION_ROOT.iterdir(), reverse=True):
+
+        # Find the newest dir that was created (birth time) at-or-after since
+        candidates = []
+        for entry in SESSION_ROOT.iterdir():
             try:
-                if not entry.is_dir() or entry.stat().st_mtime < self.since - 5:
+                if not entry.is_dir():
                     continue
-                meta = json.loads((entry / "meta.json").read_text())
-                wd = meta.get("environment", {}).get("working_directory", "")
-                if str(self.cwd) == wd:
-                    self.session_dir = entry
-                    return entry
+                # Use directory creation time (birth) via stat
+                st = entry.stat()
+                # Directory's ctime usually = creation. Use mtime as fallback.
+                dir_ctime = st.st_ctime
+                if dir_ctime < self.since - 2:
+                    continue
+                candidates.append((dir_ctime, entry))
             except Exception:
                 continue
-        return None
+
+        if not candidates:
+            return None
+
+        # Pick the newest. Take it even without messages.jsonl — drydock
+        # may not flush messages until exit. Since only one drydock process
+        # should be running, the newest dir created after our start is ours.
+        candidates.sort(reverse=True)
+        newest_ctime, newest_entry = candidates[0]
+        self.session_dir = newest_entry
+        return newest_entry
 
     def refresh(self) -> int:
         sd = self.find_session()
@@ -329,27 +352,23 @@ def ralph(cwd: Path, pkg: str, max_iterations: int = 5,
         # Type the prompt — drydock creates a session when it receives input
         type_prompt(child, initial_prompt)
 
-        # Now wait for the session to appear — 3 min is generous. Drydock can
-        # be slow to spin up when the cwd has many files or GPU is busy.
+        # Now wait for the session to appear — 5 min. Drydock can take 3-4
+        # minutes to create the session dir when GPU is busy from a prior run.
         print(f"  Waiting for session to appear...", end="", flush=True)
         session_found = False
-        for i in range(180):  # up to 180s (3 min)
+        for i in range(300):  # up to 300s (5 min)
             drain_pty(child, 1.0)
             if watcher.find_session():
                 session_found = True
                 break
-            if i == 60:
-                print(f" (still waiting at 60s, retyping prompt)", end="", flush=True)
-                time.sleep(2)
-                type_prompt(child, initial_prompt)
-            elif i == 120:
-                print(f" (still waiting at 120s, retyping again)", end="", flush=True)
+            if i > 0 and i % 60 == 0:
+                print(f" (still waiting at {i}s, retyping)", end="", flush=True)
                 time.sleep(2)
                 type_prompt(child, initial_prompt)
         if session_found:
             print(f" found: {watcher.session_dir.name}")
         else:
-            print(f" NOT FOUND after 180s")
+            print(f" NOT FOUND after 300s")
             return 1
 
         # Wait for user message to actually appear in session log

drydock_cli-2.6.38/.auto_release.lock

@@ -1 +0,0 @@
-837529