drydock-cli 2.6.147__tar.gz → 2.6.148__tar.gz

drydock_cli-2.6.148/.auto_release.lock

@@ -0,0 +1 @@
+2706548

{drydock_cli-2.6.147 → drydock_cli-2.6.148}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.6.147
+Version: 2.6.148
 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-2.6.147 → drydock_cli-2.6.148}/drydock/core/agent_loop.py

@@ -125,8 +125,36 @@ class ToolDecision(BaseModel):
 
 MAX_TOOL_TURNS = 200  # Bug fixes rarely need more than 50 turns; 200 is generous ceiling
 MAX_API_ERRORS = 5
-REPEAT_WARNING_THRESHOLD = 4  # Same exact call 4+ times before warning
-REPEAT_FORCE_STOP_THRESHOLD = 8  # Same exact call 8+ times before force-stop
+
+
+def _admiral_env_int(name: str, default: int) -> int:
+    """Read a DRYDOCK_ADMIRAL_<name> env var at module-load; fall back
+    to the hardcoded default on missing / empty / unparseable. This is
+    the knob the meta-harness kernel writes when running a variant
+    from research/experimenter.py. Production installs never set
+    these vars, so behavior is unchanged for normal drydock users."""
+    import os as _os
+    v = _os.environ.get(f"DRYDOCK_ADMIRAL_{name}", "")
+    if not v:
+        return default
+    try:
+        return int(v)
+    except (TypeError, ValueError):
+        return default
+
+
+# Same exact call N+ times before warning. Env: DRYDOCK_ADMIRAL_REPEAT_WARNING_THRESHOLD
+REPEAT_WARNING_THRESHOLD = _admiral_env_int("REPEAT_WARNING_THRESHOLD", 4)
+# Same exact call N+ times before force-stop. Env: DRYDOCK_ADMIRAL_REPEAT_FORCE_STOP_THRESHOLD
+REPEAT_FORCE_STOP_THRESHOLD = _admiral_env_int(
+    "REPEAT_FORCE_STOP_THRESHOLD", 8)
+# Check 0 (empty-result) threshold. Env: DRYDOCK_ADMIRAL_EMPTY_RESULT_THRESHOLD
+EMPTY_RESULT_THRESHOLD = _admiral_env_int("EMPTY_RESULT_THRESHOLD", 3)
+# Per-tool consecutive-call limits. Env: DRYDOCK_ADMIRAL_SAME_TOOL_NAME_REPEAT_LIMIT_{BASH,READ}
+SAME_TOOL_NAME_REPEAT_LIMIT_BASH = _admiral_env_int(
+    "SAME_TOOL_NAME_REPEAT_LIMIT_BASH", 5)
+SAME_TOOL_NAME_REPEAT_LIMIT_READ = _admiral_env_int(
+    "SAME_TOOL_NAME_REPEAT_LIMIT_READ", 7)
 
 logger = logging.getLogger(__name__)
 
@@ -1607,6 +1635,58 @@ class AgentLoop:
         )
         self.messages.reset(kept)
 
+    def _proactive_prune_write_oscillation(self) -> None:
+        """Prune duplicate writes BEFORE they push the session into a
+        vLLM 400 Bad Request (context overflow).
+
+        `_prune_duplicate_writes` above only fires after the hard-block
+        trips on the Nth write. By that point the model has already
+        written the file 4+ times, the context contains 4+ full copies
+        of the file content, and vLLM has started returning 400s on
+        every call.
+
+        This method runs as part of _sanitize_message_ordering, BEFORE
+        every LLM call. Any path with ≥3 `write_file` assistant-tool-
+        call entries in history gets pruned down to the most recent 2.
+        That leaves the latest attempt + one priored to compare against,
+        without carrying an arbitrary number of historical copies.
+
+        GitHub issue from 2026-04-21 user report: session wrote
+        prepare.py 4× before hitting 15+ 400 errors and giving up.
+        The existing `_prune_duplicate_writes` would have caught this,
+        but only AFTER the hard-block fired (8+ identical calls).
+        By 4× on a 600-line file we're already at ~75K tokens of
+        duplicate content.
+        """
+        PROACTIVE_PRUNE_THRESHOLD = 3
+        try:
+            path_counts: dict[str, int] = {}
+            for msg in self.messages:
+                if msg.role != Role.assistant or not msg.tool_calls:
+                    continue
+                for tc in msg.tool_calls:
+                    if not tc.function or tc.function.name != "write_file":
+                        continue
+                    try:
+                        args = json.loads(tc.function.arguments or "{}")
+                    except (json.JSONDecodeError, TypeError):
+                        continue
+                    target = args.get("file_path") or args.get("path") or ""
+                    if not target:
+                        continue
+                    path_counts[target] = path_counts.get(target, 0) + 1
+            for target_path, count in path_counts.items():
+                if count >= PROACTIVE_PRUNE_THRESHOLD:
+                    logger.info(
+                        "Proactive prune: path %s has %d write_file "
+                        "entries (threshold %d). Pruning older writes "
+                        "before next LLM call to avoid context overflow.",
+                        target_path, count, PROACTIVE_PRUNE_THRESHOLD,
+                    )
+                    self._prune_duplicate_writes(target_path)
+        except Exception as e:
+            logger.debug("Proactive write prune failed: %s", e)
+
     def _truncate_old_tool_results(self) -> None:
         """Shrink old verbose tool results before they bloat context.
 
@@ -1697,6 +1777,7 @@ class AgentLoop:
         # Proactive context shrinkage runs first so the LLM call sees the
         # smaller payload.
         self._truncate_old_tool_results()
+        self._proactive_prune_write_oscillation()
 
         if not self.messages:
             return
@@ -2315,14 +2396,15 @@ class AgentLoop:
                 if p in s:
                     return True
             return False
-        if len(sigs) >= 3 and all(s == sigs[-1] for s in sigs[-3:]):
+        if (len(sigs) >= EMPTY_RESULT_THRESHOLD
+                and all(s == sigs[-1] for s in sigs[-EMPTY_RESULT_THRESHOLD:])):
             recent_results: list[str] = []
             for msg in reversed(self.messages):
                 if msg.role == Role.tool:
                     recent_results.append(str(msg.content or ""))
-                    if len(recent_results) >= 3:
+                    if len(recent_results) >= EMPTY_RESULT_THRESHOLD:
                         break
-            if (len(recent_results) >= 3
+            if (len(recent_results) >= EMPTY_RESULT_THRESHOLD
                     and all(_looks_empty(r) for r in recent_results)):
                 return "FORCE_STOP"
 
@@ -2430,12 +2512,14 @@ class AgentLoop:
         # Check 2: Same tool called N+ times consecutively with different args
         # Lower thresholds catch loops where the model uses the same tool
         # with slightly different args (e.g., grep with different paths).
+        # Limits are env-overridable via DRYDOCK_ADMIRAL_SAME_TOOL_NAME_REPEAT_LIMIT_*.
         if last_tool in ("bash", "run_command"):
-            same_tool_limit = 5
+            same_tool_limit = SAME_TOOL_NAME_REPEAT_LIMIT_BASH
         elif last_tool in ("grep", "read_file"):
-            same_tool_limit = 7  # investigation tools need some room
+            # investigation tools need some room
+            same_tool_limit = SAME_TOOL_NAME_REPEAT_LIMIT_READ
         else:
-            same_tool_limit = 5
+            same_tool_limit = SAME_TOOL_NAME_REPEAT_LIMIT_BASH
         if (
             len(tool_names) >= same_tool_limit
             and all(n == tool_names[-1] for n in tool_names[-same_tool_limit:])

{drydock_cli-2.6.147 → drydock_cli-2.6.148}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.6.147"
+version = "2.6.148"
 description = "Local-first CLI coding agent — tested with Gemma 4 26B via vLLM"
 readme = "README.md"
 requires-python = ">=3.12"

{drydock_cli-2.6.147 → drydock_cli-2.6.148}/research/kernel.py

@@ -153,6 +153,23 @@ def run_kernel(variant_path: Path, results_tsv: Path,
     for k, v in (cfg.get("env_flags") or {}).items():
         env[str(k)] = str(v)
 
+    # Translate variant TOML sections into DRYDOCK_* env overrides the
+    # harness + agent_loop read at startup. Admiral's KNOB_BOUNDS path
+    # still writes to admiral_tuning.json (see validate_and_collect_knobs
+    # above) because those tunables flow through a different apply hook;
+    # these env overrides cover the constants that live in scripts/ and
+    # drydock/core/ sources, not in admiral's tuning store.
+    for h in (cfg.get("harness_threshold") or []):
+        name = h.get("name")
+        val = h.get("value", h.get("default"))
+        if name and val is not None:
+            env[f"DRYDOCK_STRESS_{name}"] = str(val)
+    for d in (cfg.get("admiral_detector") or []):
+        name = d.get("name")
+        val = d.get("value", d.get("default"))
+        if name and val is not None:
+            env[f"DRYDOCK_ADMIRAL_{name}"] = str(val)
+
     log_path = tmpdir / "tui.log"
     start = time.time()
 

{drydock_cli-2.6.147 → drydock_cli-2.6.148}/scripts/stress_shakedown.py

@@ -223,18 +223,53 @@ def _restore_checkpoint_to_step(session_dir: Path, state_data: dict,
 _API_ERROR_BANNER_RE = re.compile(r"Stopping:\s*\d+\+\s*API errors",
                                   re.IGNORECASE)
 
+
+def _env_int(name: str, default: int) -> int:
+    """Parse an int from the named env var; fall back to `default` on
+    missing / empty / unparseable. Used for DRYDOCK_STRESS_* overrides
+    written by the meta-harness kernel — see research/config_base.toml."""
+    v = os.environ.get(name, "")
+    if not v:
+        return default
+    try:
+        return int(v)
+    except (TypeError, ValueError):
+        return default
+
+
+def _env_float(name: str, default: float) -> float:
+    v = os.environ.get(name, "")
+    if not v:
+        return default
+    try:
+        return float(v)
+    except (TypeError, ValueError):
+        return default
+
 # Matches raw markdown syntax that would only be visible to the user
 # if the TUI's markdown widget failed to render it. User's 2026-04-20
 # feedback: "text is still a big blob on the screen, not formatted
 # nicely." If the screen shows literal `**bold**` or `##heading`
 # instead of the rendered equivalents, these patterns fire.
+#
+# KEEP patterns whose syntax disappears after a working render:
+#   - `**bold**`  rendered as ANSI-bold without the asterisks
+#   - `##heading` rendered with styling, NO literal `##`
+#   - `[x](y)`    rendered as underlined text; the brackets + paren gone
+#
+# EXCLUDE patterns whose syntax legitimately remains visible:
+#   - `- bullet`  renderers often keep `-` (or use `•`, which our regex
+#                 wouldn't match either way)
+#   - `1. foo`    numerals + period remain as display text — literally
+#                 the correct rendering — so matching them generated
+#                 a 45% false-positive rate on the overnight run
+#                 (rec-checks fired raw_md=33 on fully-rendered
+#                 "1. Setup: Created a dummy test...").
 _ANSI_ESCAPE_RE = re.compile(r"\x1b\[[0-9;?]*[a-zA-Z]")
 _RAW_MARKDOWN_PATTERNS = [
-    re.compile(r"\*\*\w[^*]{1,80}\*\*"),   # **bold words**
-    re.compile(r"(?m)^#{1,6}\s+\w"),        # ##heading
-    re.compile(r"(?m)^\s*[-*]\s+\w"),       # - bullet (first char on line)
-    re.compile(r"(?m)^\s*\d+\.\s+\w"),      # 1. numbered
-    re.compile(r"\[[^\]]{1,40}\]\([^)]{1,80}\)"),  # [link](url)
+    re.compile(r"\*\*\w[^*]{1,80}\*\*"),            # **bold words**
+    re.compile(r"(?m)^#{1,6}\s+\w"),                 # ##heading
+    re.compile(r"\[[^\]]{1,40}\]\([^)]{1,80}\)"),    # [link](url)
 ]
 
 
@@ -653,15 +688,17 @@ def run(cwd: Path, pkg: str, prompts_file: Path, max_per_prompt: float,
     timed_out = 0
 
     cur_section: str | None = None
-    SESSION_RESET_EVERY = 15  # /clear every N prompts to bound context
-    # Number of consecutive SKIPs that triggers an ESC+/clear force
-    # reset. This catches stuck states the banner detector misses:
-    # "admiral directive + no-tool-call" loops, silent model hangs on
-    # huge context, spinner-for-20-minutes etc. — any case where the
-    # harness can't get prompts through for multiple iterations in a
-    # row means drydock is wedged regardless of whether it's showing
-    # the API-errors banner or not.
-    MAX_CONSECUTIVE_SKIPS_BEFORE_RESET = 2
+    # Thresholds may be overridden via env for the meta-harness
+    # experimenter (see research/config_base.toml [[harness_threshold]]).
+    # Source defaults are unchanged — we fall back if the env var is
+    # unset, empty, or unparseable.
+    SESSION_RESET_EVERY = _env_int("DRYDOCK_STRESS_SESSION_RESET_EVERY", 15)
+    MAX_CONSECUTIVE_SKIPS_BEFORE_RESET = _env_int(
+        "DRYDOCK_STRESS_MAX_CONSECUTIVE_SKIPS_BEFORE_RESET", 2)
+    SEND_PROMPT_RETRY_COUNT = _env_int(
+        "DRYDOCK_STRESS_SEND_PROMPT_RETRY_COUNT", 3)
+    SEND_PROMPT_WAIT_PER_RETRY_SEC = _env_float(
+        "DRYDOCK_STRESS_SEND_PROMPT_WAIT_PER_RETRY_SEC", 120.0)
     consecutive_skips = 0
     # When resuming, step counter starts at the resumed step + 1 so the
     # printed log lines stay aligned with the original run.
@@ -722,10 +759,14 @@ def run(cwd: Path, pkg: str, prompts_file: Path, max_per_prompt: float,
         prev_writes = watcher.count_writes()
         print(f"\n[{i:>3}/{total}] {prompt[:80]}{'...' if len(prompt) > 80 else ''}")
 
-        ok = send_prompt_and_confirm(child, prompt, watcher,
-                                     max_retries=3, wait_per_retry=120.0)
+        ok = send_prompt_and_confirm(
+            child, prompt, watcher,
+            max_retries=SEND_PROMPT_RETRY_COUNT,
+            wait_per_retry=SEND_PROMPT_WAIT_PER_RETRY_SEC,
+        )
         if not ok:
-            print(f"          SKIP: TUI did not accept after 3 retries")
+            print(f"          SKIP: TUI did not accept after "
+                  f"{SEND_PROMPT_RETRY_COUNT} retries")
             skipped += 1
             consecutive_skips += 1
             results.append({"i": i, "prompt": prompt[:60], "status": "skipped"})

drydock_cli-2.6.147/.auto_release.lock

@@ -1 +0,0 @@
-2652274