drydock-cli 2.7.4__tar.gz → 2.7.5__tar.gz

drydock_cli-2.7.5/.auto_release.lock

@@ -0,0 +1 @@
+3714091

drydock_cli-2.7.5/.claude/scheduled_tasks.lock

@@ -0,0 +1 @@
+{"sessionId":"d6003854-3a6a-46c9-8637-1639ff283487","pid":3603597,"procStart":"355816330","acquiredAt":1777119887143}

{drydock_cli-2.7.4 → drydock_cli-2.7.5}/.gitignore

@@ -202,3 +202,10 @@ tests/playground/*
 logs/
 test_bank_results/
 .pause_auto_release
+.pause_vllm_failover
+
+# Research loop: per-experiment outputs are high-churn and regenerate
+# from results.tsv + config_best.toml. Stop sentinel is intentional state.
+research/staged/
+research/traces/
+research/STOP

{drydock_cli-2.7.4 → drydock_cli-2.7.5}/Admiral.md

@@ -651,3 +651,273 @@ past 24h. **No new code without observed evidence it's needed.**
 
 ***
 
+## 11. Deep Noir integration pilot (2026-04-23, IN FLIGHT)
+
+Goal stated by user: "I want Deep Noir framework to work with Admiral
+and meta-harness to improve drydock and Gemma 4." Not a paper — a
+shippable tool.
+
+### 11.1 Intended architecture
+
+DN plugs in as a **new intervention class** for Admiral, alongside the
+existing prompt/tool-config interventions:
+
+  Admiral detector → classifies failure mode →
+      (a) prompt-level fix (current)
+      (b) activation-space steering hook (new, via DN)
+
+  Meta-harness (`research/experimenter.py`) already mutates scalar knobs
+  (e.g., `wrap_up_warn_at`, `REPEAT_FORCE_STOP_THRESHOLD`); DN extends
+  the search space to `(layer, head_subset, magnitude)` tuples per
+  failure mode. The experimenter scores configs the same way — by
+  medium-hard PRD success rate.
+
+The asset DN needs is a **labeled probe set**: real decision-point
+contexts from drydock sessions, each tagged with the "correct"
+next-token class. Discovery then finds the (L, K, M) that maximizes
+labeled accuracy.
+
+### 11.2 What was built this session
+
+* **Probe extractor** (`/data3/Deep_Noir_1/drydock_probes/build_probes.py`
+  + focused variants): mines `~/.drydock/logs/session/**/messages.jsonl`
+  for every assistant decision point, renders the chat-history context,
+  and labels by whether the turn emitted a tool_call or text. Output:
+  - `drydock_act_ask.jsonl` — 5979 probes, 296 sessions
+  - `drydock_wrap_vs_act.jsonl` — 360 balanced probes (act vs text)
+  - `drydock_1b_compat.jsonl` — 240 probes portable to non-Gemma-4
+    tokenizers
+
+* **Discovery driver** (`discover_steering.py`) — model-agnostic wrapper
+  around DN's `SteeringDiscoveryEngine`. Loads any HF Gemma, runs the
+  5-phase pipeline, emits a `SteeringConfig` JSON.
+
+* **Sanity check on Gemma-3-1B-IT**: Phase-1 layer ranking reproduced
+  the paper's "Gemma uses deep layers" finding on drydock-domain data.
+  Top layers 18–24 match the sentiment results in Figure 1. Pipeline
+  works end-to-end.
+
+* **vLLM logprob eval harness** (`evaluate_wrapup_detector.py`) — was
+  built to ship a logprob-based pre-generation detector without
+  activation hooks. Abandoned; see §11.4.
+
+* **Dependencies** added to `miniconda3` base env (harmless):
+  `transformers @ git+main`, `compressed-tensors`, `peft`,
+  `bitsandbytes`, `accelerate`.
+
+* **Pause sentinel added to `scripts/vllm_failover.sh`**
+  (`/data3/drydock/.pause_vllm_failover`) — the cron was auto-respawning
+  the container and kept clobbering the GPU. Sentinel is user-visible
+  and reversible; remove the file to resume normal failover.
+
+### 11.3 Three blockers hit (honest)
+
+1. **`Gemma-4-26B-A4B-it-AWQ-4bit` checkpoint cannot be loaded in
+   transformers.** The directory is mis-named — the format is
+   `compressed-tensors` with MoE `_packed`/`_scale` expert weights.
+   Transformers (5.6.2 AND git-main 5.7.0.dev0) does not decompose
+   those into `gate_up_proj` / `down_proj` at load time. Result:
+   `lm_head.weight` and all MoE expert projections are meta-initialized
+   (random). vLLM loads this fine because it has its own
+   compressed-tensors path; HF transformers does not. This is an
+   upstream gap, not something we can patch quickly.
+
+2. **Driver/torch mismatch in the drydock env.** `miniforge3/envs/drydock`
+   has `torch 2.11.0+cu130`; driver is 575 (CUDA 12.9). Incompatible.
+   `miniconda3` base has `torch 2.7.1+cu126` which works. All DN work
+   now uses `/home/bobef/miniconda3/bin/python3`, not the drydock env.
+
+3. **The wrap-up-hijack detector I tried to ship is based on
+   mis-labeled data.** "Previous turn ended; awaiting your next
+   instruction." — the phrase I treated as a Gemma 4 pathology for a
+   day — is actually drydock's OWN filler injected at
+   `drydock/core/agent_loop.py:3006` via `_ensure_assistant_after_tools`.
+   Every match in every session has `prev_role: tool` and
+   `next_role: user`. An Admiral detector that strips this would be
+   stripping drydock's own checkpoint marker. Killed that path.
+   The 5344 tool-call probes (label=1) are still real Gemma 4 behavior;
+   the 69 "I have …" / "The …" text probes are real completion
+   summaries; only the 182 "Previous turn ended" probes were mis-
+   attributed.
+
+### 11.4 Decision point — user's call
+
+Unquantized Gemma 4 bf16 **download complete** at
+`/data3/Deep_Noir_1/models/gemma-4-26b-a4b-it/` (51.6 GB, both
+safetensors shards present, cache cleaned). Option 1 is immediately
+actionable. Three forward paths:
+
+* **Option 1 — Wait for bf16, run DN on it.** transformers loads
+  vanilla bf16 without the compressed-tensors MoE gap. 26 B bf16 is
+  too big for a single 16 GB GPU, so load with `bitsandbytes` 4-bit +
+  `device_map="auto"` across both 4060 Ti's. Hooks then need per-
+  device direction/mask tensors — 1 small DN patch, not days of work.
+  Probe contrast `<|tool_call>` vs `" Previous"` is still useful:
+  steers toward "emit tool_call" vs "end turn silently so drydock has
+  to fill." Highest-value path that actually gets us a tool for the
+  deployment model. ~30–60 min after download finishes.
+
+* **Option 2 — Stop and catalog.** Probes + drivers + GPU plumbing are
+  reusable. Park the pilot until transformers grows MoE+compressed-
+  tensors support for the existing AWQ-4bit checkpoint. Zero further
+  GPU burn; clean resume later.
+
+* **Option 3 — Re-label probes, try a different contrast.** Manually
+  split the 453 non-filler text responses into "legitimate task-
+  complete" vs "premature-stop". Needs 1–2 hr of labeling before any
+  model work. Only justifies itself if Option 1 fails.
+
+**Recommendation**: 1, with 2 as the fallback if bf16 + bnb 4-bit + DN
+still hits some third bug we can't solve today.
+
+### 11.5 Staged infrastructure (what survives a session loss)
+
+* Pause sentinels active: `/data3/drydock/research/STOP`,
+  `/data3/drydock/.pause_auto_release`,
+  `/data3/drydock/.pause_vllm_failover`. **Remove all three** to
+  restore normal automation.
+* vLLM was restarted with `--gpu-memory-utilization 0.90 --max-model-len 32768`
+  (down from 0.95 / 131072) because TP1 was OOMing during CUDA graph
+  capture. Container name unchanged (`gemma4`); to restore the
+  original config, `docker rm gemma4 && /data3/Models/start_gemma4.sh`.
+* Research experimenter is stopped (STOP sentinel). Best-metric
+  snapshot at pause: 4.725. Post-ban-list trend was negative (random
+  fallback was underperforming the banned `wrap_up_warn_at=100` LLM
+  proposal, which suggests the LLM proposer was earning its keep even
+  when it looked narrow).
+* DN source-of-truth: `/data3/Deep_Noir_1/` (separate repo, MIT). All
+  pilot code under `drydock_probes/` subdirectory; no changes to the
+  published paper or experiments.
+
+### 11.6 When resuming
+
+1. Verify bf16 download completed: `du -sh /data3/Deep_Noir_1/models/gemma-4-26b-a4b-it`
+   should be ~49 GB, with both `model-00001-of-00002.safetensors` and
+   `model-00002-of-00002.safetensors` present and `.cache/` empty.
+2. Test bf16 load in transformers (single process, 2-GPU device_map,
+   bnb 4-bit). Confirm LOAD REPORT shows **no MISSING keys**.
+3. If clean: `python3 discover_steering.py --model_path …/gemma-4-26b-a4b-it
+   --probes probes/drydock_wrap_vs_act.jsonl --meta probes/…meta.json
+   --probe_limit 100 --max_layers 5 --max_K 4 --model_kwargs '{"load_in_4bit":true,"device_map":"auto"}'`.
+4. If accuracy gain > 0 on labeled probes, that's the direction to
+   wire into Admiral as a new intervention. The hook itself needs a
+   separate serving path since vLLM doesn't expose activation hooks —
+   likely a sidecar HF-transformers process that Admiral queries only
+   when a detector fires. Design TBD after we see the gain number.
+
+### 11.7 Overnight (2026-04-24) autonomous session — results
+
+User directive: "try them all, if needed, keep going, heading to bed
+now." Executed Option 1 first, then pivoted to a data-driven version of
+Option 3 when Option 1 hit hard blockers.
+
+**Option 1 (DN on Gemma 4 26B) — BLOCKED by loader plumbing.**
+Six load-attempt variants all failed on different axes:
+  - AWQ-4bit (compressed-tensors MoE): `lm_head.weight` + expert
+    projections meta-initialized. Upstream transformers gap, confirmed
+    on both 5.6.2 and 5.7.0.dev0.
+  - bf16 + fp16 dtype + device_map=auto: last layer on `meta`, forward
+    crashed. 26.5B × 2 B = 53 GB doesn't fit 2×16 GB.
+  - bf16 + bnb-4bit + transformers-dev: `Params4bit.__new__()`
+    TypeError on `_is_hf_initialized` kwarg. bnb 0.49.2 missing it.
+  - bf16 + bnb-4bit + transformers 5.6.2 stable: same TypeError
+    (transformers 5.6.2 also passes the kwarg, not just dev).
+  - Same with monkey-patched Params4bit to swallow the kwarg:
+    `Tensor.item() cannot be called on meta tensors` in bnb's
+    state_dict save-path when CPU offload is enabled.
+  - Same with explicit device_map (17 layers GPU0, embeds+13 layers
+    GPU1) and CPU offload disabled: CUDA OOM on GPU 0 at 15 GiB during
+    load-time fp16 intermediate (bnb reads fp16 shard, quantizes on
+    GPU, moves on).
+
+Bottom line: transformers + bnb + Gemma 4 26B MoE is not a working
+combination today on 2×16 GB hardware. Would need either a newer bnb
+release that accepts `_is_hf_initialized` (trivial) AND a loader that
+streams-and-quantizes without a full fp16 intermediate per shard
+(non-trivial), or an alternate serving path (vLLM already loads it
+cleanly, but doesn't expose activation hooks; forking vLLM to add hooks
+is the real DN-on-Gemma-4 path and is a ~week-long project).
+
+**Pivot — session log behavioral mining (this IS the tool, no DN needed).**
+Ran `mine_behavior_patterns.py` across 400 real drydock sessions (14,822
+messages) to count actual model pathologies directly from traces:
+
+```
+IDENTICAL_TOOL_REPEAT        481 fires  (1.20 per session)
+EMPTY_ASSISTANT_AFTER_TOOL   254 fires  (0.64 per session; 61% of sessions hit)
+TOOL_ARGS_IGNORE_RESULT       45 fires  (0.11 per session)
+```
+
+The existing `detect_tool_call_loop(window=3)` already catches
+IDENTICAL_TOOL_REPEAT but fires advisory-only; the model ignores the
+nudge (CLAUDE.md learning #2: "Gemma 4 ignores advisory nudges"). The
+other two patterns have no current detector.
+
+**Two new proposed detectors, code written + validated end-to-end:**
+
+  `drydock/admiral/detectors_proposed.py` (NEW, NOT wired into
+  `run_all` — observation mode from §10 still binding).
+
+  1. `detect_empty_after_tool` — fires when the last assistant turn
+     had no content and no tool_calls, or when drydock's filler
+     `Previous turn ended; awaiting your next instruction.` sits where
+     the assistant response should be (retrospective equivalent).
+     Replay validation: **254 fires in 244 of 400 sessions (61%)**.
+     Intervention: directive telling the model its next turn MUST
+     produce a tool call or a text summary — no silent hand-back.
+
+  2. `detect_retry_after_error` — fires when an identical tool call
+     follows a tool result that looks like an error. Fires at
+     window=2 (earlier than the existing detector's window=3).
+     Replay validation: **32 fires in 23 of 400 sessions (5.8%)**.
+     Intervention: directive embedding the error snippet and naming
+     the repeated tool.
+
+**Data + artifacts written:**
+  - `/data3/Deep_Noir_1/drydock_probes/mine_behavior_patterns.py`
+  - `/data3/Deep_Noir_1/drydock_probes/validate_proposed_detectors.py`
+  - `/data3/Deep_Noir_1/drydock_probes/results/session_behavior_stats.json`
+  - `/data3/Deep_Noir_1/drydock_probes/results/session_behavior_report.md`
+  - `/data3/Deep_Noir_1/drydock_probes/results/proposed_detector_validation.json`
+  - `drydock/admiral/detectors_proposed.py` — in drydock tree, not
+    wired; audit before enabling.
+  - `tests/test_admiral_proposed.py` — **9/9 tests pass**.
+    Cover empty-assistant + drydock-filler cases, healthy-sequence
+    negative, retry-with-different-args negative, retry-after-success
+    negative, both-detectors-with-run_proposed positive/negative.
+
+### 11.8 What to do when you wake up
+
+1. Read `session_behavior_report.md` — the one-page summary of what
+   the 400-session mine found.
+2. ~~Review `drydock/admiral/detectors_proposed.py`. If you like the two
+   detectors, wire them in by adding both functions to the tuple in
+   `drydock/admiral/detectors.py::run_all`.~~ **Wired in 2026-04-24.**
+   Both functions are now called from `run_all`; advisory-only; existing
+   `interventions.py` handles the Findings. Watch for `empty_after_tool:*`
+   and `retry_after_error:*` codes in the intervention log.
+3. Decide whether to escalate `EMPTY_ASSISTANT_AFTER_TOOL` from
+   advisory to state-mutating. Given the 61%-of-sessions fire rate,
+   dropping the empty+filler pair from context and re-rolling the
+   turn (with the directive injected) is the aggressive move. But
+   this breaks the "never stop, only guide" rule — your call.
+4. Pause sentinels still active:
+   - `/data3/drydock/research/STOP` — stops `research_babysitter.sh`
+     respawn of the experimenter. Remove to resume the research loop.
+   - `/data3/drydock/.pause_auto_release` — stops the 6-hour PyPI
+     auto-release cron. Remove when you want to publish again.
+   - `/data3/drydock/.pause_vllm_failover` — stops the 5-min
+     respawn-if-down cron. **I added this sentinel tonight; the guard
+     itself is a 3-line patch in `scripts/vllm_failover.sh`. Keep
+     both or revert both together.** Remove sentinel to resume.
+5. vLLM container is back up on the original start config (0.95
+   util, 131072 ctx). Drydock TUI usable immediately.
+6. For the DN path itself: park it until either (a) transformers
+   lands proper MoE+compressed-tensors decompression for the
+   existing AWQ-4bit checkpoint, or (b) we commit to forking vLLM
+   to expose activation hooks. Neither is a this-week task; the
+   probe dataset + drivers are cold-storage-ready under
+   `/data3/Deep_Noir_1/drydock_probes/`.
+
+***
+

{drydock_cli-2.7.4 → drydock_cli-2.7.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.7.4
+Version: 2.7.5
 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-2.7.4 → drydock_cli-2.7.5}/PRD.md

@@ -1,7 +1,7 @@
 # DryDock — Local CLI Coding Agent
 
 **Repository:** https://github.com/fbobe321/drydock
-**PyPI:** https://pypi.org/project/drydock-cli/ (v2.6.107)
+**PyPI:** https://pypi.org/project/drydock-cli/ (v2.7.4)
 **License:** Apache 2.0 (fork of [mistralai/mistral-vibe](https://github.com/mistralai/mistral-vibe))
 
 ## Vision
@@ -10,11 +10,19 @@ Best-in-class local coding agent. Build, debug, and ship software using local
 LLMs on consumer hardware — and **prove it actually works** with a test
 harness that drives the real TUI like a real user, not a tool-call counter.
 
+Drydock is a **living harness that adapts to tasks and models**. Admiral
+(in-process supervisor) watches each session and applies detector-driven
+interventions; the Meta-Harness experimenter drives an overnight auto-tuning
+loop against the stress test and proposes configuration or prompt mutations
+from trace data. Both run with a local-LLM-first, air-gapped posture —
+proposers and analyzers default to the operator's own vLLM instance, and
+cloud escalation is gated behind an explicit env flag.
+
 ## Current Status
 
 - **Active model:** Gemma 4 26B-A4B-it-AWQ-4bit (MoE, 4B active params per
   token, ~70 tok/s) via vLLM Docker on 2x RTX 4060 Ti 16GB
-- **Version:** 2.6.107 on PyPI. The v2.6.79–v2.6.107 wave shipped 28+
+- **Version:** 2.7.4 on PyPI. The v2.6.79–v2.6.107 wave shipped 28+
   loop-prevention iterations: Read-before-Write contract, read dedup,
   terse results, token sanitization, system-reminder framing, JSON
   sanitizer, field-aware path cleaner, fake-tool-call text detection,
@@ -25,6 +33,11 @@ harness that drives the real TUI like a real user, not a tool-call counter.
   80K compact threshold, and finally the **session-reset-every-10-
   prompts pattern** (v2.6.107) borrowed from
   [Adversarial Code Review](https://asdlc.io/patterns/adversarial-code-review/).
+  The v2.6.108–v2.7.4 wave added **Admiral** (in-process supervisor,
+  Phase 1→3b + empty-after-tool and retry-after-error detectors), the
+  **Meta-Harness experimenter** (Karpathy-style overnight auto-tuning
+  with airgap-locked proposer), and a batch of rendering + tool-call
+  serialization fixes from user-reported GitHub issues #8–#13.
 - **Users can:** build projects from PRDs, fix bugs, review code, refactor
   — through the TUI only (headless mode is gone)
 
@@ -64,9 +77,15 @@ Rules of engagement:
 **Best run so far:** v2.6.102 reached 96/201 prompts before tool calls
 started returning `<user_cancellation>` infinite loop.
 
-**Current run:** v2.6.107 with session-reset-every-10 pattern. Cron
-job sends hourly status to Telegram (`scripts/stress_telegram_status.py`).
-If a run completes 201 prompts, the goal extends to 2000 prompts.
+**Current focus:** The stress test is the historical benchmark, but
+active development has shifted to Admiral (in-process supervisor that
+actuates on detectors without touching the harness) and the Meta-Harness
+experimenter (overnight config-mutation loop that optimises drydock's
+own knobs against the suite). Both feed the same goal — drydock that
+holds up under unattended multi-hour workloads — but they exercise
+different axes. Hourly Telegram status cron
+(`scripts/stress_telegram_status.py`) is still wired; bring it online
+with a fresh run when a 2.7.x release should be pressure-tested.
 
 The pattern of failure modes hit (and shipped fixes for):
 - Tight identical-call loops (v2.6.79–v2.6.95)
@@ -232,6 +251,87 @@ is codified in memory as
 - `auto_release.sh` and `watchdog.sh` cron jobs respect pause flags
   at `.pause_auto_release` and `.pause_watchdog` for manual debugging
 
+### Admiral — in-process supervisor
+Admiral is a Textual asyncio task attached to the running `AgentLoop`.
+It polls message history every 5 s, runs a battery of detectors over
+the recent window, and when a finding qualifies it injects a
+`<system-reminder>` via the existing `_inject_system_note` channel.
+Never stops the session — interventions are advisory, matching the
+project rule in `feedback_no_tool_errors_for_loop_detection.md`.
+
+Shipped (v2.6.137 → v2.7.4):
+
+- **Phase 1 — heuristic detectors.** `detect_tool_call_loop` (3+
+  identical calls in a row) and `detect_struggle` (20+ non-write calls
+  since the last write). Append-only audit log at
+  `~/.drydock/logs/admiral_history.log`; 60 s dedup window so the same
+  directive doesn't fire repeatedly.
+- **Phase 2 — local-LLM meta-analysis.** When a heuristic finding has
+  no canned directive, `llm_analyzer.analyze` asks the local model
+  (via the session's own backend, temp=0.1) for a DIRECTIVE or a
+  STUMPED response given the last 12 turns plus the detector code.
+  If STUMPED, `opus_escalator` escalates to Claude Opus (Anthropic SDK
+  if key present, `claude -p` subprocess as fallback) — **only when
+  `ADMIRAL_ALLOW_CLOUD_ESCALATION=1`**; default posture is local-only
+  to preserve the air-gap pitch.
+- **Phase 3a — hyperparameter adaptation.** `task_classifier` labels
+  each session (build / bugfix / explore / refactor); `metrics.py`
+  writes one JSONL line per finished session; `tuning.py` maintains
+  per-(model, task) knob overrides at `~/.drydock/admiral_tuning.json`.
+  `apply_to_agent_loop` writes private `_admiral_*` attributes the
+  main loop reads — wall-clock budget, hard-stop, warn thresholds,
+  temperature, detector windows. Safe no-op until a tuning file exists.
+- **Phase 3b — gated code proposals.** Admiral can propose a code-level
+  change (not just a directive); proposals are written to
+  `~/.drydock/admiral_proposals/` and surface in `admiral_probe`. No
+  auto-merge — user reviews and applies.
+- **Phase 4+ detectors (v2.7.4 / 2026-04-24).** `detect_empty_after_tool`
+  fires when an assistant message after a tool result is empty or pure
+  filler; `detect_retry_after_error` fires when the model retries the
+  exact failing call after the previous attempt errored. Both wired
+  into `run_all`; retroactive validation on 50 sessions from
+  `~/.vibe/logs/session/` shows empty 68% / retry 2%, matching the
+  mine baseline (61% / 5.8%).
+- **Directives rubric** (`Admiral.md`) — the contract Admiral holds
+  drydock to. Used in Phase 2 analyzer prompts so directive quality
+  stays stable across runs.
+- **Observability — `admiral_probe.py`.** Read-only HTTP probe on
+  `0.0.0.0:8878` that exposes telemetry as JSON (`history_tail`, PID
+  list, drydock version). Pure log-tailer — no drydock imports, so
+  restarts are unnecessary on detector code changes.
+
+### Meta-Harness — overnight auto-tuning
+Karpathy-style autoresearch scaffold at `research/`. The experimenter
+runs a shakedown batch, scores it, and evolves drydock's configuration
+between rounds. Proposals come from either a blind random mutator or
+an LLM-backed `proposer.py` that reads source + recent traces +
+`results.tsv` and suggests ONE contrastive mutation per round.
+
+- **Mutation surface.** `research/config_base.toml` exposes ~16 knobs
+  spanning Admiral detector thresholds, harness timings, session-reset
+  cadence, and Gemma-4-specific prompts. Target rotation is mandatory
+  when one knob class dominates ≥70% of the recent 15 rounds.
+- **Airgap-locked proposer.** Default transport is local vLLM via
+  `DRYDOCK_RESEARCH_PROPOSER_URL` (the same balancer the TUI uses) and
+  `DRYDOCK_RESEARCH_PROPOSER_MODEL=gemma4`. Opus is gated behind
+  `--proposer opus` AND an explicit `ADMIRAL_ALLOW_CLOUD_ESCALATION=1`,
+  so a self-tuning loop can't silently phone home in an air-gapped
+  deployment. Rule recorded in memory as
+  `feedback_local_proposer_only.md`.
+- **Experimenter safeguards.** Duplicate `(target, name, value)`
+  proposals are rejected up front; `mutate_random` reads ban-lists from
+  `config_base` (retired knobs don't resurface); no-op integer samples
+  are rejected; each promising round gets median-of-3 replication to
+  separate signal from variance.
+- **Self-heal.** `research_babysitter.sh` hourly cron restarts a stuck
+  experimenter. `DRYDOCK_FORCE_VERSION` env var lets the experimenter
+  request minor/major bumps without waiting for the normal auto-release
+  cadence.
+- **Kernel overlay.** Per-round configs are applied as a TOML
+  parse+rewrite on the operator's real `~/.drydock/config.toml` (not an
+  append), so the test config always sits on top of the user's actual
+  providers/credentials and can be rolled back cleanly.
+
 ### Read-only bash auto-accept
 The bash tool's default allowlist now covers most read-only commands
 without requiring per-user configuration:
@@ -298,6 +398,120 @@ order:
 Auto-created if none exist. AGENTS.md is essential for devstral
 (it loops without one) but Gemma 4 works without it.
 
+## Recent fixes (April 16–24, 2026)
+
+### v2.7.4 + unreleased (April 24) — GitHub issue triage batch
+Four user-reported bugs from live TUI usage, each with regression
+tests that fail without the fix:
+
+- **Issue #13 — JSONDecodeError on `/v1/chat/completions`,
+  `/compact` won't clear it.** A bash tool result containing a raw
+  `\x00` or `\x1b` ridden into the conversation history survived
+  compaction. `ensure_ascii=True` escaped control chars in the OUTER
+  body, but vLLM's tool-call parser re-parses
+  `tool_calls.function.arguments` as JSON and choked on the literal
+  control byte in the unescaped string. Fix: recursive
+  `_strip_control_chars` over the whole payload in `OpenAIAdapter`.
+  Mirrors `ReasoningAdapter._sanitize_content` but covers `content`,
+  `arguments`, and tool descriptions in one pass.
+- **Issue #12 — missing whitespace/newlines after hidden
+  reasoning/tool-call blocks.** `_break_walls_of_text` only fires
+  when called per-chunk with >200 chars, so streaming chunks never
+  trip it. The final accumulated content could still be a long
+  no-newline blob. Fix: override `AssistantMessage.stop_stream` to
+  run wall-rescue once on the full content at stream end and replay
+  the rescued text into the markdown widget.
+- **Issue #11 — `task(task=..., agent=...)` printed as plain text.**
+  Gemma 4 sometimes emits a real tool call as Python-syntax text
+  instead of using the `tool_calls` protocol; nothing ran and the
+  TUI sat idle. The v2.6.91 fake-tool-call nuker handled
+  `call:name{...}` but not `name(arg=...)`. Added a regex branch that
+  matches when the entire content (after peeling optional `thought`
+  prefix) is a function-call shape with no real tool_calls, so the
+  agent loop's recovery nudge fires.
+- **Issue #10 — `todo` loop on `navygpt` (Gemma-derived model).** The
+  auto-disable set for loop-prone tools gated on `"gemma" in name`;
+  navygpt slipped through under its own alias. Broadened the
+  substring hint list to cover known Gemma-derived names and checked
+  both `name` and `alias`. Until ModelConfig grows a
+  `model_family` field, the list stays explicit.
+
+### v2.7.4 (April 24) — Admiral Phase-4 detectors wired
+`detect_empty_after_tool` and `detect_retry_after_error` (shipped as
+proposals earlier) are now invoked by `admiral.detectors.run_all` so
+they actually fire in live sessions. 9 new unit tests +
+retroactive validation on 50 real sessions (empty 68%, retry 2%,
+matches 400-session mine baseline).
+
+### v2.7.x research improvements (April 19–22)
+- **Meta-Harness integration** — Karpathy-style overnight auto-tuning
+  loop shipped in v2.6.146. Experimenter mutates config, runs a
+  shakedown batch, logs trace, scores, iterates.
+- **Airgap-locked proposer** — local vLLM by default, Opus gated
+  behind an explicit env flag. Three layers of protection so a
+  self-tuning loop can't silently phone home. See
+  `feedback_local_proposer_only.md`.
+- **Prompt mutation surface** wired through proposer → experimenter
+  → kernel → drydock, so the experimenter can evolve prompt text in
+  addition to numeric knobs.
+- **Proposer deduplication** — duplicate `(target, name, value)`
+  proposals rejected; `mutate_random` respects ban-lists in
+  `config_base` (retired knobs don't resurface); no-op integer
+  samples rejected; mandatory target rotation when one knob class
+  dominates ≥70% of recent 15 rounds.
+- **Median-of-3 replication** — each promising round replicated
+  twice more to separate signal from variance before it's promoted
+  to `config_best`.
+- **research_babysitter.sh** hourly cron restarts a stuck
+  experimenter; `DRYDOCK_FORCE_VERSION` env var lets the experimenter
+  request out-of-band version bumps.
+
+### v2.6.145+ (April 20) — admiral watcher sees raw-markdown leakage
+`scripts/stress_watcher.py` tails the stress log for
+`[rec-check] ... raw_md=N`, fires a `stress-alert` when ≥3 of the
+last ≥10 rec-checks leak raw markdown. Part of the "others on
+different machines" observability lift.
+
+### v2.6.144 (April 19) — auto-"Continue." wedge fix + stress hardening
+- `DRYDOCK_AUTO_CONTINUE_DISABLE` env gate — the agent-loop
+  auto-"Continue." injection was wedging stress runs on text-only
+  prompts. Env gate lets the stress harness disable it without a code
+  change. See `project_drydock_auto_continue_loop.md`.
+- `break user turn on text-only assistant response` — when the model
+  emits a text-only turn with no tool calls, the user-turn ends cleanly
+  instead of being re-prompted into another tool-call attempt.
+- **stress_watcher** catches orphan/retry/skip patterns and actuates
+  via SIGUSR1 — the harness respawns the TUI on signal so a wedged
+  session doesn't require a babysitter.
+
+### v2.6.141–v2.6.143 (April 17–18) — Admiral Phase 1 through 3b
+Three-week arc (in a week): in-process supervisor → local-LLM
+meta-analysis → hyperparameter adaptation + gated code proposals.
+Details in the Admiral section above. v2.6.142/143 fixed
+stress-harness memory leaks that were masquerading as drydock leaks;
+v2.6.141 shipped the observation-mode audit after the user's
+"let's slow down" directive.
+
+### v2.6.137–v2.6.140 (April 17) — Admiral.md + observability
+- `Admiral.md` Directives rubric — the contract Admiral holds drydock
+  to, referenced by Phase 2 analyzer prompts.
+- `admiral_probe.py` read-only HTTP probe on `0.0.0.0:8878` for the
+  dashboard tab (192.168.50.21 box pulls `/api/admiral`).
+- `.vibe` → `drydock` refactor across the repo. Log paths moved to
+  `~/.drydock/logs/`; `~/.vibe/logs/` still tailed for
+  back-compatibility.
+
+### v2.6.134–v2.6.136 (April 17) — user-reported rendering/stop fixes
+- **Issue #8** — walls of text in TUI assistant responses. First pass:
+  `_break_walls_of_text` promotes inline bold headings and numbered
+  lists to real blocks. Reopen fix landed in v2.6.137 with a more
+  aggressive pass for the long-prose-no-newlines case.
+- **Issue #9** — per-prompt limits too aggressive on hard tasks.
+  Budget raised.
+- **"when to stop vs continue"** — Gemma 4 was treating mid-turn
+  planning as a stop signal. Prompt + nudge adjustments so the model
+  only declares done when a real tool call chain has completed.
+
 ## Recent fixes (April 13–15, 2026)
 
 ### v2.6.107 (April 15) — session-reset every 10 prompts in stress harness
@@ -535,13 +749,25 @@ Second claude-code pass produced two more wins:
 
 ## Roadmap
 
-### Currently in flight (April 15, 2026)
-- **200-prompt stress test** running against v2.6.107 (PID
-  1614232 as of writing). Hourly Telegram status via cron. If the
-  current run completes 201 prompts cleanly, goal extends to 2000.
+### Currently in flight (April 24, 2026)
+- **GitHub issue triage** — 4 user-reported bugs (#10–#13) fixed this
+  week, all with regression tests. Remaining open issues should be
+  picked up on sight; `feedback_proactive_github_issues.md` codifies
+  the rule.
+- **Admiral observation mode** — after the "let's slow down" directive
+  (v2.6.141), Admiral is wired but the policy cadence is deliberate.
+  New detectors (`detect_empty_after_tool`, `detect_retry_after_error`)
+  landed in v2.7.4 but won't fire in operator-running TUIs until the
+  `.pause_auto_release` flag is lifted or they `pip install -e` the
+  source tree.
+- **Meta-Harness experimenter** — overnight auto-tuning loop runs
+  against the shakedown suite. Active work: proposer rigor
+  (dedup, rotation, no-op rejection, median-of-3 replication),
+  airgap lock verification. The `research/STOP` sentinel holds the
+  loop during manual debugging.
 - **Telegram notifications**: `scripts/stress_telegram_status.py`
-  fires every hour via crontab; reports prompt progress, session
-  health, dup-ratio, max-consecutive, error count.
+  still wired; currently idle. Reattach when a long stress run
+  should be pressure-tested against a 2.7.x release.
 
 ### Likely to retire after v2.6.107 proves out
 With session-reset every 10 prompts, several earlier safety nets

{drydock_cli-2.7.4 → drydock_cli-2.7.5}/drydock/admiral/detectors.py

@@ -96,8 +96,17 @@ def detect_struggle(messages: Sequence[LLMMessage], threshold: int = 20) -> Find
 
 def run_all(messages: Sequence[LLMMessage]) -> list[Finding]:
     """Run every detector and return all findings that fire."""
+    from drydock.admiral.detectors_proposed import (
+        detect_empty_after_tool,
+        detect_retry_after_error,
+    )
     results: list[Finding] = []
-    for fn in (detect_tool_call_loop, detect_struggle):
+    for fn in (
+        detect_tool_call_loop,
+        detect_struggle,
+        detect_empty_after_tool,
+        detect_retry_after_error,
+    ):
         f = fn(messages)
         if f is not None:
             results.append(f)