drydock-cli 2.6.148__tar.gz → 2.7.0__tar.gz

drydock_cli-2.7.0/.auto_release.lock

@@ -0,0 +1 @@
+2797280

{drydock_cli-2.6.148 → drydock_cli-2.7.0}/CHANGELOG.md

@@ -5,6 +5,91 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [2.7.0] - 2026-04-21
+
+### Added
+
+- **Meta-Harness autoresearch loop** (`research/` subsystem). Opus-driven
+  search over a bounded mutation surface. Fixed 5-minute kernel,
+  append-only results log, contrastive proposer with access to source +
+  top/bottom traces. See `research/domain_spec.md` for the design.
+- **Admiral actuator**: `scripts/stress_watcher.py` sends SIGUSR1 to the
+  stress harness on unrecoverable TUI wedges; the harness respawns its
+  drydock child at the next iteration boundary. Rate-limited. First loop
+  where admiral *acts* on its observations.
+- **Env override surface** for experiments:
+  `DRYDOCK_AUTO_CONTINUE_DISABLE`,
+  `DRYDOCK_STRESS_{SESSION_RESET_EVERY,MAX_CONSECUTIVE_SKIPS_BEFORE_RESET,SEND_PROMPT_RETRY_COUNT,SEND_PROMPT_WAIT_PER_RETRY_SEC}`,
+  `DRYDOCK_ADMIRAL_{REPEAT_WARNING_THRESHOLD,REPEAT_FORCE_STOP_THRESHOLD,EMPTY_RESULT_THRESHOLD,SAME_TOOL_NAME_REPEAT_LIMIT_{BASH,READ}}`,
+  `DRYDOCK_PROMPTS_DIR`. All default to the prior hardcoded values;
+  production behavior unchanged.
+- **`DEPLOYMENT.md`**: known-working hardware + vLLM Docker flags +
+  `~/.drydock/config.toml` template + env var table + diagnostic
+  checklist. Answers the "works on your machine, not mine" question.
+- **`scripts/stress_babysitter.sh`**: hourly monitor cron that
+  auto-restarts long stress runs from the last successful step. Silent
+  when healthy; Telegrams only on state change.
+- **Check 0 empty-result loop detector** in `agent_loop.py`. Fires
+  FORCE_STOP at 3 identical tool calls when the result matches an
+  empty pattern (`total_count: 0`, `no todos`, `no tasks`,
+  `no matches`, blank). Catches todo / task / grep-empty loops that
+  Check 1's threshold-of-8 is too lax for.
+- **Proactive write-oscillation prune** in
+  `_sanitize_message_ordering`. Any path with ≥3 `write_file` history
+  entries gets its older copies pruned before the next LLM call, so
+  context stays below Gemma 4's 131K ceiling instead of waiting for
+  a 400.
+- **Raw-markdown leakage detector** in stress harness + admiral.
+  Surfaces TUI rendering regressions as `stress-alert/raw-markdown-
+  leakage` on the admiral dashboard.
+- **`OVERNIGHT_PROGRESS.md`** session-continuity doc template.
+
+### Changed
+
+- **`should_break_loop`** in `agent_loop.py:876` — the condition
+  `tool_turns == 0` was unreachable since the file's current structure
+  landed. Text-only assistant responses now cleanly close the user
+  turn instead of regenerating until `PER_PROMPT_BUDGET_SEC` times
+  out. **This is a behavior fix users will notice immediately** —
+  questions the model asks now wait for input.
+- **Continue-loop**: env-gated via `DRYDOCK_AUTO_CONTINUE_DISABLE=1`
+  so stress runs (and any workflow hitting text-only prompt loops)
+  can opt out.
+- **`SESSION_ROOT` resolution** in `scripts/shakedown_interactive.py`:
+  now reads the path from `~/.drydock/config.toml`'s
+  `[session_logging].save_dir` instead of hardcoding
+  `~/.drydock/logs/session/`. Fixes false-SKIP reporting on any
+  machine whose drydock config has the newer default
+  `~/.vibe/logs/session/`.
+- **Todo tool** (`todo.py`): empty-read response now escalates.
+  First empty: *"No todos yet. Write one or proceed with the task."*
+  Second+: *"Still no todos. Stop calling read. Empty state will
+  not change on its own."* Closes GitHub #10.
+- **Fresh-start wipe** in stress harness: wipes cwd to fixtures
+  (PRD/AGENTS/functional_tests) instead of accumulating prior
+  runs' half-written artifacts.
+- **Raw-markdown detector** patterns tightened (`**bold**`,
+  `##heading`, `[link](url)` only — dropped bullet + numbered-list
+  patterns that matched rendered lists and generated 45 % false
+  positives).
+
+### Infrastructure
+
+- First clean 1658-prompt stress run completed 2026-04-21.
+  ~1622 successful prompts (run + babysitter auto-restart),
+  35 SKIPs, 0 TIMEOUTs, ~60 admiral-initiated TUI recycles.
+  ~46-hour wall-clock. Reference baseline for future experiments.
+- `scripts/auto_release.sh` supports `DRYDOCK_FORCE_VERSION` for
+  MINOR / MAJOR bumps (the ordinary PATCH-bumper would have taken
+  us to 2.6.149, not 2.7.0).
+
+### Known follow-ups
+
+- Meta-Harness median-of-3 replication before promotion (currently
+  single-sample).
+- TUI Markdown widget regression surfaced by admiral but not yet
+  fixed.
+
 ## [2.4.2] - 2026-03-12
 
 ### Added

{drydock_cli-2.6.148 → drydock_cli-2.7.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.6.148
+Version: 2.7.0
 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-2.7.0/drydock/core/prompts/__init__.py

@@ -0,0 +1,56 @@
+from __future__ import annotations
+
+import os
+from enum import StrEnum, auto
+from pathlib import Path
+
+from drydock import DRYDOCK_ROOT
+
+_PROMPTS_DIR = DRYDOCK_ROOT / "core" / "prompts"
+
+
+def _resolve_prompt_path(name: str) -> Path:
+    """Return the .md file for a prompt name.
+
+    If DRYDOCK_PROMPTS_DIR is set and contains a matching file, use that
+    override. Falls back to the shipped prompt. Used by the Meta-Harness
+    research kernel (research/kernel.py) to swap in per-experiment
+    prompt variants without mutating the installed package.
+
+    Missing override file = use shipped (not an error) so experiments
+    can mutate only gemma4 without also overriding every other prompt.
+    """
+    override_dir = os.environ.get("DRYDOCK_PROMPTS_DIR", "")
+    if override_dir:
+        candidate = Path(override_dir) / f"{name}.md"
+        if candidate.is_file():
+            return candidate
+    return (_PROMPTS_DIR / name).with_suffix(".md")
+
+
+class Prompt(StrEnum):
+    @property
+    def path(self) -> Path:
+        return _resolve_prompt_path(self.value)
+
+    def read(self) -> str:
+        return self.path.read_text(encoding="utf-8").strip()
+
+
+class SystemPrompt(Prompt):
+    CLI = auto()
+    EXPLORE = auto()
+    TESTS = auto()
+    BUILDER = auto()
+    PLANNER = auto()
+    DIAGNOSTIC = auto()
+    GEMMA4 = auto()
+
+
+class UtilityPrompt(Prompt):
+    COMPACT = auto()
+    DANGEROUS_DIRECTORY = auto()
+    PROJECT_CONTEXT = auto()
+
+
+__all__ = ["SystemPrompt", "UtilityPrompt"]

{drydock_cli-2.6.148 → drydock_cli-2.7.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.6.148"
+version = "2.7.0"
 description = "Local-first CLI coding agent — tested with Gemma 4 26B via vLLM"
 readme = "README.md"
 requires-python = ">=3.12"

drydock_cli-2.7.0/research/config_best.toml

@@ -0,0 +1,145 @@
+[target]
+model = "mistral-vibe-cli-latest"
+task = "unknown"
+
+[[knob]]
+name = "per_prompt_budget_sec"
+default = 300
+min = 300
+max = 3600
+mutable = true
+why = "Hard ceiling on a single user turn. Too low wastes work; too high lets loops run wild."
+value = 600
+
+[[knob]]
+name = "hard_stop_tool_calls"
+default = 100
+min = 30
+max = 250
+mutable = true
+why = "HARD_STOP_CALLS in agent_loop. Cap on tool-call depth per user turn."
+
+[[knob]]
+name = "wrap_up_warn_at"
+default = 40
+min = 10
+max = 100
+mutable = true
+why = "First progressive nudge. Earlier = tighter turns, may lose quality."
+
+[[knob]]
+name = "stop_now_warn_at"
+default = 60
+min = 20
+max = 150
+mutable = true
+why = "Second progressive nudge. Gemma often ignores the wrap-up; this is the harder prod."
+
+[[knob]]
+name = "temperature"
+default = 0.3
+min = 0.0
+max = 0.9
+mutable = true
+why = "Sampling temperature. Lower = more deterministic; higher = more exploration."
+
+[[knob]]
+name = "loop_detector_window"
+default = 3
+min = 2
+max = 6
+mutable = true
+why = "How many consecutive identical tool calls trigger the loop advisor."
+
+[[knob]]
+name = "struggle_threshold"
+default = 20
+min = 10
+max = 40
+mutable = true
+why = "Reads-without-a-write before admiral fires the struggle directive."
+
+[env_flags]
+DRYDOCK_AUTO_CONTINUE_DISABLE = "1"
+
+[prompts.gemma4]
+source_path = ""
+mutable = true
+why = "Single highest-leverage mutable target. Governs how Gemma 4 interprets every turn."
+
+[prompts.cli]
+source_path = ""
+mutable = true
+why = "Alternate non-Gemma system prompt. Rarely the right target but available for cross-model experiments."
+
+[[harness_threshold]]
+name = "MAX_CONSECUTIVE_SKIPS_BEFORE_RESET"
+default = 2
+min = 1
+max = 5
+mutable = true
+why = "Force ESC+/clear after N consecutive SKIPs. Lower = aggressive recycling; higher = tolerant of transient wedges."
+
+[[harness_threshold]]
+name = "SESSION_RESET_EVERY"
+default = 15
+min = 5
+max = 50
+mutable = true
+why = "Periodic /clear cadence in prompts. Tight cadence bounds context but costs reset overhead every N prompts."
+
+[[harness_threshold]]
+name = "SEND_PROMPT_RETRY_COUNT"
+default = 3
+min = 1
+max = 6
+mutable = true
+why = "How many times send_prompt_and_confirm re-types before marking SKIP. Each retry is ~120s."
+
+[[harness_threshold]]
+name = "SEND_PROMPT_WAIT_PER_RETRY_SEC"
+default = 120
+min = 30
+max = 300
+mutable = true
+why = "Seconds to wait for user-msg confirmation before retyping. Tighter = faster SKIP, may miss slow prompts."
+
+[[admiral_detector]]
+name = "REPEAT_WARNING_THRESHOLD"
+default = 4
+min = 2
+max = 8
+mutable = true
+why = "Consecutive identical tool calls before a WARNING. Lower = early nudge; higher = tolerant of legit repetition."
+
+[[admiral_detector]]
+name = "REPEAT_FORCE_STOP_THRESHOLD"
+default = 8
+min = 4
+max = 20
+mutable = true
+why = "Consecutive identical tool calls before FORCE_STOP. Must be >= REPEAT_WARNING_THRESHOLD."
+
+[[admiral_detector]]
+name = "EMPTY_RESULT_THRESHOLD"
+default = 3
+min = 2
+max = 6
+mutable = true
+why = "Check 0 (empty-result loop). Fire FORCE_STOP after N identical-call + empty-result in a row."
+
+[[admiral_detector]]
+name = "SAME_TOOL_NAME_REPEAT_LIMIT_BASH"
+default = 5
+min = 3
+max = 10
+mutable = true
+why = "Same-tool repetition limit for bash. Different from read/grep which need more headroom."
+
+[[admiral_detector]]
+name = "SAME_TOOL_NAME_REPEAT_LIMIT_READ"
+default = 7
+min = 4
+max = 12
+mutable = true
+why = "Same-tool repetition limit for read_file + grep. Investigation tools legitimately repeat more."

{drydock_cli-2.6.148 → drydock_cli-2.7.0}/research/experimenter.py

@@ -135,10 +135,34 @@ def mutate_opus(cfg: dict) -> tuple[dict, str] | None:
     elif target == "env_flag":
         flags = new.setdefault("env_flags", {})
         flags[name] = str(value)
+    elif target == "prompt":
+        # Materialize the proposed prompt to a candidate-local file and
+        # point the variant's source_path at it. The kernel will copy
+        # this into the isolated HOME's DRYDOCK_PROMPTS_DIR at spawn.
+        # Staging dir mirrors the variant toml: one sibling per exp.
+        # Using a stable-per-variant name (no exp_id here yet — caller
+        # provides it via the outer staged path). We write into the
+        # per-variant staged_prompts subdir and let the kernel resolve.
+        prompts = new.setdefault("prompts", {})
+        prompt_entry = prompts.setdefault(name, {})
+        prompts_dir = STAGED_DIR / "prompts"
+        prompts_dir.mkdir(parents=True, exist_ok=True)
+        # Use a content-hash so concurrent / retried variants don't
+        # stomp each other. Kernel reads the file by source_path.
+        import hashlib as _hashlib
+        h = _hashlib.sha1(str(value).encode()).hexdigest()[:12]
+        target_file = prompts_dir / f"{name}_{h}.md"
+        target_file.write_text(str(value))
+        prompt_entry["source_path"] = str(target_file)
+        prompt_entry["mutable"] = True
     else:
         return None
 
-    desc = f"opus {target}:{name} -> {value} ({reason})"
+    if target == "prompt":
+        size = len(str(value))
+        desc = f"opus prompt:{name} -> [{size} chars] ({reason})"
+    else:
+        desc = f"opus {target}:{name} -> {value} ({reason})"
     return new, desc
 
 

{drydock_cli-2.6.148 → drydock_cli-2.7.0}/research/kernel.py

@@ -93,14 +93,46 @@ def setup_isolated_home(tmpdir: Path, cfg: dict) -> tuple[Path, Path, Path]:
     drydock_dir.mkdir(parents=True)
     cwd.mkdir(parents=True)
 
-    # Drydock config — point session_logging at the isolated session dir.
-    # Everything else falls back to drydock defaults.
-    (drydock_dir / "config.toml").write_text(
-        '[session_logging]\n'
-        f'save_dir = "{session_root}"\n'
-        'session_prefix = "session"\n'
-        'enabled = true\n'
-    )
+    # Drydock config — parse the USER'S real ~/.drydock/config.toml
+    # (so providers + models + tool permissions + MCP servers carry
+    # through) and overwrite ONLY [session_logging] with the
+    # per-experiment path. First pass just appended a second
+    # [session_logging] table, which tomllib correctly rejects with
+    # "Cannot declare ('session_logging',) twice". Parse → mutate →
+    # re-serialize is the only safe overlay.
+    real_config = Path.home() / ".drydock" / "config.toml"
+    base: dict = {}
+    if real_config.is_file():
+        try:
+            with real_config.open("rb") as f:
+                base = tomllib.load(f)
+        except (tomllib.TOMLDecodeError, OSError):
+            base = {}
+    base["session_logging"] = {
+        "save_dir": str(session_root),
+        "session_prefix": "session",
+        "enabled": True,
+    }
+    try:
+        import tomli_w
+        with (drydock_dir / "config.toml").open("wb") as f:
+            tomli_w.dump(base, f)
+    except ImportError:
+        # Fallback: if tomli_w isn't available, write just the minimal
+        # config. Drydock will start but without the user's providers
+        # — experiments will fail, but at least with a clear error.
+        (drydock_dir / "config.toml").write_text(
+            '[session_logging]\n'
+            f'save_dir = "{session_root}"\n'
+            'session_prefix = "session"\n'
+            'enabled = true\n'
+        )
+    # Trust the isolated cwd so drydock doesn't pop the modal — this
+    # is what shakedown_interactive's harness handles inline, but from
+    # a fresh HOME there's nothing to answer yet.
+    trusted = drydock_dir / "trusted_folders.toml"
+    if not trusted.exists():
+        trusted.write_text(f'trusted_folders = ["{cwd}"]\n')
 
     # Admiral tuning — write variant knobs keyed on (model, task).
     knobs = validate_and_collect_knobs(cfg)
@@ -170,6 +202,31 @@ def run_kernel(variant_path: Path, results_tsv: Path,
         if name and val is not None:
             env[f"DRYDOCK_ADMIRAL_{name}"] = str(val)
 
+    # Variant prompt overrides. For each [prompts.<name>] with a
+    # non-empty source_path, copy into the isolated HOME's prompts
+    # dir. DRYDOCK_PROMPTS_DIR tells drydock (see core/prompts/__init__.py)
+    # to check there before the shipped prompt. Unset prompts fall
+    # through to the installed package defaults.
+    prompts_override_dir = home / ".drydock" / "prompts"
+    prompts_override_dir.mkdir(parents=True, exist_ok=True)
+    any_prompt_override = False
+    for prompt_name, entry in (cfg.get("prompts") or {}).items():
+        if not isinstance(entry, dict):
+            continue
+        source_path = entry.get("source_path", "")
+        if not source_path:
+            continue
+        try:
+            src = Path(source_path)
+            if src.is_file():
+                dest = prompts_override_dir / f"{prompt_name}.md"
+                dest.write_bytes(src.read_bytes())
+                any_prompt_override = True
+        except OSError:
+            continue
+    if any_prompt_override:
+        env["DRYDOCK_PROMPTS_DIR"] = str(prompts_override_dir)
+
     log_path = tmpdir / "tui.log"
     start = time.time()
 

{drydock_cli-2.6.148 → drydock_cli-2.7.0}/research/proposer.py

@@ -19,12 +19,20 @@ Scope (v1 — overnight build):
 
 Contract: returns either
   {
-    "target": "knob" | "harness_threshold" | "admiral_detector" | "env_flag",
+    "target": "knob" | "harness_threshold" | "admiral_detector"
+              | "env_flag" | "prompt",
     "name": "<field-name>",
-    "value": <numeric or string>,
+    "value": <numeric or string; for target=prompt, the FULL new
+             prompt text as a string>,
     "reason": "<≤200 char human-readable>",
   }
 or None on any failure — experimenter then falls through to random.
+
+For target="prompt", `name` must match a [prompts.<name>] entry with
+mutable=true in config_base.toml (currently "gemma4" or "cli"). The
+value is the complete replacement prompt. The kernel writes it to
+the candidate's isolated prompts dir and the TUI picks it up via
+DRYDOCK_PROMPTS_DIR.
 """
 from __future__ import annotations
 
@@ -38,7 +46,7 @@ import tomllib
 from pathlib import Path
 
 RESEARCH_DIR = Path(__file__).resolve().parent
-DRYDOCK_ROOT = RESEARCH_DIR.parent
+DRYDOCK_ROOT = RESEARCH_DIR.parent  # /data3/drydock — repo root
 OPUS_MODEL = "claude-opus-4-7"
 OPUS_TIMEOUT_SEC = 120
 MAX_TRACES_EACH_SIDE = 3          # top-N + bottom-N contrastive
@@ -206,6 +214,27 @@ def _build_prompt(config_base: dict, config_best: dict,
         for r in bottom
     )
 
+    # Include the current gemma4.md baseline when the proposer might
+    # want to mutate it. Opus needs to see what it's replacing.
+    def _current_prompt(name: str) -> str:
+        prompt_file = DRYDOCK_ROOT / "drydock" / "core" / "prompts" / f"{name}.md"
+        try:
+            return prompt_file.read_text()
+        except OSError:
+            return f"<{name}.md unreadable>"
+
+    prompts_cfg = config_base.get("prompts", {})
+    mutable_prompts = {n: e for n, e in prompts_cfg.items()
+                       if isinstance(e, dict) and e.get("mutable")}
+    prompt_bodies_section = ""
+    if mutable_prompts:
+        prompt_bodies_section = "\n# MUTABLE PROMPT BODIES (current)\n\n"
+        for name in mutable_prompts:
+            body = _current_prompt(name)
+            prompt_bodies_section += (
+                f"## prompts.{name}\n\n```\n{body}\n```\n\n"
+            )
+
     return f"""You are Meta-Harness Proposer for drydock. Your job is to \
 propose ONE mutation to the stress-harness config that will improve the \
 `effective_rate` metric (done_per_minute, cliff at >50% failure).
@@ -221,7 +250,8 @@ within the bounded surface declared in config_base.toml.
     "harness_thresholds": [h["name"] for h in config_base.get("harness_threshold", [])],
     "admiral_detectors": [d["name"] for d in config_base.get("admiral_detector", [])],
     "env_flags": list(config_base.get("env_flags", {}).keys()),
-}, indent=2)}
+    "prompts": list(mutable_prompts.keys()),
+}, indent=2)}{prompt_bodies_section}
 
 # CURRENT-BEST CONFIG
 
@@ -348,6 +378,14 @@ def _validate_mutation(m: dict, config_base: dict) -> bool:
     if target == "env_flag":
         flags = config_base.get("env_flags", {})
         return name in flags and isinstance(value, str)
+    if target == "prompt":
+        prompts = config_base.get("prompts", {})
+        entry = prompts.get(name)
+        if not isinstance(entry, dict) or not entry.get("mutable"):
+            return False
+        # Prompt value must be a non-empty string; set a sanity cap so
+        # a hallucinated 200KB prompt can't blow up context.
+        return isinstance(value, str) and 0 < len(value) <= 50_000
     return False
 
 

drydock_cli-2.7.0/research/results.tsv

@@ -0,0 +1,2 @@
+ts	exp_id	git_commit	config_sha	metric	done	skip	timeout	recycle	elapsed_s	note
+1776806910	exp_1776806574_0001	74075e4	81d102021a	0.000	0	6	0	0	329.5	opus knob:per_prompt_budget_sec -> 600 (No trace history yet. Seed search from a tighter per-turn budget (600s vs defaul

drydock_cli-2.7.0/research/staged/exp_1776806574_0001.toml

@@ -0,0 +1,145 @@
+[target]
+model = "mistral-vibe-cli-latest"
+task = "unknown"
+
+[[knob]]
+name = "per_prompt_budget_sec"
+default = 300
+min = 300
+max = 3600
+mutable = true
+why = "Hard ceiling on a single user turn. Too low wastes work; too high lets loops run wild."
+value = 600
+
+[[knob]]
+name = "hard_stop_tool_calls"
+default = 100
+min = 30
+max = 250
+mutable = true
+why = "HARD_STOP_CALLS in agent_loop. Cap on tool-call depth per user turn."
+
+[[knob]]
+name = "wrap_up_warn_at"
+default = 40
+min = 10
+max = 100
+mutable = true
+why = "First progressive nudge. Earlier = tighter turns, may lose quality."
+
+[[knob]]
+name = "stop_now_warn_at"
+default = 60
+min = 20
+max = 150
+mutable = true
+why = "Second progressive nudge. Gemma often ignores the wrap-up; this is the harder prod."
+
+[[knob]]
+name = "temperature"
+default = 0.3
+min = 0.0
+max = 0.9
+mutable = true
+why = "Sampling temperature. Lower = more deterministic; higher = more exploration."
+
+[[knob]]
+name = "loop_detector_window"
+default = 3
+min = 2
+max = 6
+mutable = true
+why = "How many consecutive identical tool calls trigger the loop advisor."
+
+[[knob]]
+name = "struggle_threshold"
+default = 20
+min = 10
+max = 40
+mutable = true
+why = "Reads-without-a-write before admiral fires the struggle directive."
+
+[env_flags]
+DRYDOCK_AUTO_CONTINUE_DISABLE = "1"
+
+[prompts.gemma4]
+source_path = ""
+mutable = true
+why = "Single highest-leverage mutable target. Governs how Gemma 4 interprets every turn."
+
+[prompts.cli]
+source_path = ""
+mutable = true
+why = "Alternate non-Gemma system prompt. Rarely the right target but available for cross-model experiments."
+
+[[harness_threshold]]
+name = "MAX_CONSECUTIVE_SKIPS_BEFORE_RESET"
+default = 2
+min = 1
+max = 5
+mutable = true
+why = "Force ESC+/clear after N consecutive SKIPs. Lower = aggressive recycling; higher = tolerant of transient wedges."
+
+[[harness_threshold]]
+name = "SESSION_RESET_EVERY"
+default = 15
+min = 5
+max = 50
+mutable = true
+why = "Periodic /clear cadence in prompts. Tight cadence bounds context but costs reset overhead every N prompts."
+
+[[harness_threshold]]
+name = "SEND_PROMPT_RETRY_COUNT"
+default = 3
+min = 1
+max = 6
+mutable = true
+why = "How many times send_prompt_and_confirm re-types before marking SKIP. Each retry is ~120s."
+
+[[harness_threshold]]
+name = "SEND_PROMPT_WAIT_PER_RETRY_SEC"
+default = 120
+min = 30
+max = 300
+mutable = true
+why = "Seconds to wait for user-msg confirmation before retyping. Tighter = faster SKIP, may miss slow prompts."
+
+[[admiral_detector]]
+name = "REPEAT_WARNING_THRESHOLD"
+default = 4
+min = 2
+max = 8
+mutable = true
+why = "Consecutive identical tool calls before a WARNING. Lower = early nudge; higher = tolerant of legit repetition."
+
+[[admiral_detector]]
+name = "REPEAT_FORCE_STOP_THRESHOLD"
+default = 8
+min = 4
+max = 20
+mutable = true
+why = "Consecutive identical tool calls before FORCE_STOP. Must be >= REPEAT_WARNING_THRESHOLD."
+
+[[admiral_detector]]
+name = "EMPTY_RESULT_THRESHOLD"
+default = 3
+min = 2
+max = 6
+mutable = true
+why = "Check 0 (empty-result loop). Fire FORCE_STOP after N identical-call + empty-result in a row."
+
+[[admiral_detector]]
+name = "SAME_TOOL_NAME_REPEAT_LIMIT_BASH"
+default = 5
+min = 3
+max = 10
+mutable = true
+why = "Same-tool repetition limit for bash. Different from read/grep which need more headroom."
+
+[[admiral_detector]]
+name = "SAME_TOOL_NAME_REPEAT_LIMIT_READ"
+default = 7
+min = 4
+max = 12
+mutable = true
+why = "Same-tool repetition limit for read_file + grep. Investigation tools legitimately repeat more."