drydock-cli 2.7.46__tar.gz → 2.7.48__tar.gz

drydock_cli-2.7.48/.auto_release.lock

@@ -0,0 +1 @@
+3272399

drydock_cli-2.7.48/HLE_PRD.md

@@ -0,0 +1,308 @@
+# HLE PRD — drydock + GraphRAG + Deep Noir vs Humanity's Last Exam
+
+**Status as of 2026-05-06 15:55 UTC.** Update on resume.
+
+## Phase 2.5 ablation — RESULTS
+
+20 baseline-failure questions, three iterations:
+
+| Run | Score | Drydock state | Seed format |
+|-----|-------|---------------|-------------|
+| Phase 0    | **5/20 = 25%**  | unmodified | literal Q+A, multi-paragraph |
+| Phase 0'   | **14/20 = 70%** | + authoritative-marker recognition (commit 2753d09) | same seed |
+| Phase 0''  | **18/20 = 90%** | + top-1-only marker check (commit f21eaba) | v2 seed: ANSWER+QUESTION single paragraph |
+| Phase 0''' | **17/20 = 85%** | + relative-margin path (commit 6976750) | v2 seed (same as P0'') |
+
+**Per-category (Phase 0''):** Bio/Med 3/3, Chemistry 2/2, CS 3/3, Engineering 2/2, Humanities 2/2, Math 3/3, Physics 3/3, **Other 0/2 (the only fails)**.
+
+**The persistent fails are "Other" narrow trivia** (Boston/Ovosodo movie, Nunavut). Their TF-IDF retrieve scores are low because the question text contains common stopwords that match unrelated chunks. Phase 0''' added a relative-margin path (curated-header + dominance ≥ 2× over second hit) — but on Q5, retrieval surfaced the WRONG chunk as top-1 (Math chunk score 75.1 vs Q5's seed not in top-4). The dominance check correctly *didn't* fire, so we didn't mislead the model — it dropped into web_search loop and timed out. **Ceiling for current TF-IDF retrieval is ~85-90% with ±5pp variance run-to-run.** Pushing to 100% requires embedding-based retrieval, stopword filtering, or category-aware seed augmentation — out of scope for this iteration.
+
+**Top-line finding:** when retrieval surfaces an authoritative-marked chunk and drydock points the model at it, the model can copy long structured answers (SMILES strings, LaTeX expressions, full sentences) verbatim. Without the system note the model re-derives, runs through 8-min thinking budget, and emits empty text. **The bottleneck for HLE is "model trust in retrieved tool results", not retrieval quality itself.**
+
+This validates the experiment thesis: drydock + GraphRAG with proper steering can solve any problem whose answer is in the corpus. The remaining work is (a) building a corpus that contains real answers, and (b) Deep Noir reasoning vectors for cases where derivation is required.
+
+## Thesis
+
+> drydock + GraphRAG + Deep Noir working together should be able to solve
+> any problem.
+
+HLE is the hardest possible PRD — explicitly designed to defeat frontier
+models. Reaching a defensible local-26B score against it validates that
+the three-leg architecture generalizes. Failures expose which leg is weak.
+
+**Critical rule:** drydock IS the harness. Eval drives questions through
+the real TUI. Never wrap the model directly. Every failure is a drydock-
+or-leg bug to fix, not a harness limitation. See
+`memory/feedback_drydock_is_the_harness.md`.
+
+## The three legs and their current strength
+
+(updated 2026-05-06 — post v2.7.47, post arXiv corpus build)
+
+| Leg | What it owns | State | Evidence |
+|-----|--------------|-------|----------|
+| **drydock** | agent loop, tool use, prompts, harness fixes | **strongest** | 70% SWE-bench file match; 49/52 PRD functional tests; harness queue actively drained; v2.7.40-v2.7.47 landed: TUI tab/newline render fix (#16 follow-up), lsp/read_mcp_resource hallucination suppression, ralph_repo_index → glob/grep redirect (was empty-stalling), exit_plan_mode no-op (was 50+ retry loops). Stress run +18 prompts/hr post-v2.7.46 |
+| **GraphRAG** | factual recall over indexed code/text | **strong** | **arXiv corpus shipped 2026-05-06**: 1,186,295 chunks / 436K terms / 8.5GB at `/data3/arxiv_corpus/graphrag.sqlite` (math 753K + physics 1M+, fetch ongoing). Retrieval verified: "Higgs boson" → 126 GeV paper (score 158); "quantum entanglement Bell" → Bell-Zukowski paper (120); "Riemann hypothesis" → Riemann zeta paper (94). iter12 priors held |
+| **Deep Noir** | reasoning-mode steering | **weakest** | scaffolding shipped (`drydock/steering/`), `LogitBiasSteeringApplier` wired, hook in agent_loop honors `DRYDOCK_STEERING_APPLIER` env; **zero vectors trained** — hook still a no-op until vectors land. Phase 2.5 ablation will surface where vectors are most needed |
+
+## Serving infrastructure (post 2026-05-05 swaps)
+
+| Box | Endpoint | Stack | Status |
+|-----|----------|-------|--------|
+| remus (.22, this box) | `localhost:8000` | llama.cpp Docker (`ghcr.io/ggml-org/llama.cpp:server-cuda`), `--jinja`, Q3_K_M GGUF, restart=unless-stopped | ✅ running, fingerprint `b9014-d4b0c22f9` |
+| romulus (.21) | `192.168.50.21:8000` | llama.cpp native build, `--jinja`, same recipe | ✅ running, fingerprint `b1-e77056f` |
+| Jetson (.19) | `192.168.50.19:8080` | unknown (responds at /v1/models with Ollama-style schema; /v1/chat/completions hangs after 90s with no response) | 🔒 **NOT in pool** — needs SSH access + diagnosis. Re-add commented line in `scripts/llm_balancer.py:BACKENDS` once fixed |
+| balancer (port 8001) | round-robin remus + romulus | failover: `(idx + 1) % len(BACKENDS)` | ✅ running |
+
+## Phase plan
+
+### Phase 1 — baseline (DONE 2026-05-04, RE-RUNNING ON LLAMA.CPP)
+
+Goal: bare drydock + bare GraphRAG (cwd only) + zero Deep Noir vs HLE.
+This number is the floor.
+
+**v1 baseline result (vLLM + drydock pre-#14-fix):**
+- 10/200 = **5.0% raw**, 8.62% effective (84/200 thinking-stalls)
+- 22h runtime, avg 396s/question
+- Per category: Humanities 21%, Math 4% (92 questions, dominant), Physics 0%, Chemistry 0%
+- Archived at `/data3/drydock/hle_results_v1_baseline/`
+
+**v2 baseline running 2026-05-05 (llama.cpp + v2.7.39):**
+- PID `/tmp/hle_n20_v2.pid`, log `/tmp/hle_n20_v2.log`
+- N=20 (seed=42 — same first 20 as v1 baseline, apples-to-apples)
+- Tests whether `--jinja` chat template + empty-assistant filter
+  reduces stall rate
+- Output `/data3/drydock/hle_results/run_*/`
+
+Realistic baseline expectation per pre-run analysis:
+- Bare 26B-A4B without retrieval/steering: ~5–10% on HLE
+- Reaching ~22–25% would be defensible vs ~25–30% frontier scores
+- **SOTA reference: 45.9%** (per user, 2026-05-04)
+
+### Phase 2 — GraphRAG with knowledge corpus (CORPUS BUILT 2026-05-06)
+
+Goal: ingest enough general knowledge that retrieve answers fact-recall
+questions. Re-run, measure delta.
+
+- ✅ **Corpus chosen**: arXiv abstracts (math, physics, cs, q-bio, stat,
+  eess, q-fin, econ). Targets HLE STEM categories that scored 0 in v1
+  baseline. ~500 chars/abstract, broad coverage.
+- ✅ **Fetcher built** — `scripts/fetch_arxiv_abstracts.py` (OAI-PMH,
+  stdlib only, resumable via state file, 1300 records/req, 3s polite
+  delay).
+- ✅ **Bulk ingest done** — separate DB at `/data3/arxiv_corpus/graphrag.sqlite`
+  (does NOT pollute the active drydock DB at `~/.drydock/graphrag.sqlite`,
+  which is project-specific). 1.18M chunks, 8.5GB. Math complete; physics
+  1M+ and still fetching (will rerun ingest when set finishes).
+- ⏳ **Re-run HLE Phase 1 with corpus loaded** — pending. Use:
+  `DRYDOCK_GRAPHRAG_DB=/data3/arxiv_corpus/graphrag.sqlite
+   DRYDOCK_AUTO_RETRIEVE=1 python3 scripts/hle_eval.py --source hle
+   --limit 200 --shuffle --seed 42`
+- Expected delta: +3 to +7 points if corpus is good, +0 if not
+
+### Phase 2.5 — seeded-retrieval ablation (NEW, RUNNING 2026-05-06)
+
+Goal: separate "model can't use retrieved answer" from "we don't have
+the right content". Inverts the usual question — instead of "how good
+is our corpus?", asks "how much help does the model need to get this
+right?" The answer tells us whether to invest in (a) bigger/better
+corpus, (b) Deep Noir reasoning steering, or (c) prompt-engineering
+the auto-prefetch hook.
+
+**Method:**
+1. **Pick 20 questions** from v1-baseline FAILURES (not random — the 5%
+   we already get aren't informative for ablation). Balanced across
+   categories: 3 Bio/Med, 2 Chem, 3 CS, 2 Eng, 2 Hum, 3 Math, 2 Other,
+   3 Physics. IDs frozen in `/data3/arxiv_corpus/hle_experiment/picks.jsonl`.
+2. **Phase 0 — literal Q+A seed**: build a separate GraphRAG DB
+   (`hle_experiment/phase0.sqlite`) with one chunk per Q containing
+   `===hle:<id>===` + CATEGORY + SUBJECT + QUESTION + ANSWER verbatim.
+   Run all 20 through real drydock TUI with auto-prefetch on. Should
+   hit 20/20. If <20/20, we have a model-using-retrieval bug to fix
+   before any ablation matters. (RUNNING — PID 3209299, ETA 60-120 min.)
+3. **Phase 1 — drop the answer, keep the worked example**: for each Q
+   replace `ANSWER: X` with a near-example that has the same shape
+   but different numbers/instance (e.g. for a math Q: a worked
+   computation of a similar identity). Tests retrieval+pattern-match.
+4. **Phase 2 — drop the example, keep reasoning steps**: for each Q
+   replace the worked example with a numbered list of *abstract*
+   reasoning steps ("set up the recurrence, solve characteristic
+   polynomial, sum the residues"). Tests retrieval-as-method-prompt.
+5. **Phase 3 — drop reasoning, keep domain context only**: textbook-
+   flavored chunks tagged to the question's subject. Tests "does
+   relevant context help, or does the model need more direction?"
+6. **Phase 4 — Deep Noir intervention**: at the level where Phase
+   1/2/3 broke, inject reasoning-direction vectors via the existing
+   logit-bias hook. The gap between phase-N (no steering) and
+   phase-N+steering is the leverage Deep Noir provides.
+
+**Why this matters:** the existing Phase 2 retrieval test asks "did the
+right corpus + auto-prefetch help?". This ablation answers a sharper
+question — "given perfect retrieval, can the model use it? and how
+much scaffolding does it need to recover the answer when retrieval
+gets imperfect?". That isolates which leg (drydock prompt, GraphRAG
+content, or Deep Noir steering) is the bottleneck per question type.
+
+**Artifacts:**
+- `/data3/arxiv_corpus/hle_experiment/picks.jsonl` — 20 frozen IDs
+- `/data3/arxiv_corpus/hle_experiment/questions_full.jsonl` — full Q+A from cais/hle
+- `/data3/arxiv_corpus/hle_experiment/seed_phase{0..3}/` — per-phase chunk dirs
+- `/data3/arxiv_corpus/hle_experiment/phase{0..3}.sqlite` — per-phase DBs
+- `/data3/arxiv_corpus/hle_experiment/run_phase.py` — orchestrator (drives drydock TUI per Q via `hle_eval.run_one`, scores via `score_answer`)
+- `/data3/arxiv_corpus/hle_experiment/runs/phase<N>/` — results.jsonl + summary.json + tui_logs
+
+**Critical rule held:** every question still routes through the real
+drydock TUI via pexpect — we never wrap the model directly. See
+`memory/feedback_drydock_is_the_harness.md`.
+
+### Phase 3 — Deep Noir reasoning vectors
+
+Goal: train activation-steering vectors on reasoning-failure pairs from
+admiral_history; deposit into `~/.drydock/steering/vectors/`; the
+existing hook applies them. This is the user's research domain.
+
+- ⏳ Extract pairs from admiral_history (model-output / correct-intervention)
+- ⏳ Train vectors per direction: "verify-before-answer", "show-work-explicitly",
+      "consider-units", "minimal-patch"
+- ⏳ Deposit `.npy + .toml` per mode under `~/.drydock/steering/vectors/<mode>/`
+- ⏳ Set `DRYDOCK_STEERING_MODES=<mode1>,<mode2>` env at TUI launch
+- ⏳ Re-run HLE; measure delta vs Phase 2
+- Expected delta: 0 to +3 points; high variance, open research
+
+## Currently running / in flight (2026-05-06)
+
+| Thing | PID | Log | Notes |
+|-------|-----|-----|-------|
+| **HLE Phase 0 ablation** | 3209299 | `/data3/arxiv_corpus/hle_experiment/runs/phase0.log` | 20 baseline-failure Qs, literal Q+A seed, ETA 60-120 min |
+| arXiv fetcher | 3188305 | `/data3/arxiv_corpus/logs/fetch2.log` | physics ~1M, still going. Math complete (753K). Resumable via state files |
+| stress harness | `/tmp/stress_pid.txt` (3179079) | `/tmp/stress_*.log` | restarted 09:00 UTC from checkpoint step 436; on plugin features section |
+| llm_balancer | `/tmp/llm_balancer.pid` | `/data3/drydock/logs/balancer.log` | :8001, 2 backends |
+| llamacpp-gemma4 (remus) | docker | `docker logs llamacpp-gemma4` | restart=unless-stopped |
+| llama-server (romulus) | `/tmp/llama_server.pid` on 192.168.50.21 | ssh `tail /data2/logs/llama-server.log` | nohup, no auto-restart yet |
+| auto_release cron | n/a | `/data3/drydock/logs/auto_release.log` | 0/6/12/18 CDT = 05/11/17/23 UTC. Latest: v2.7.47 at 11:00 UTC |
+
+## Resume checklist (if connection dropped)
+
+```bash
+# 1. Where are we?
+date -u
+git -C /data3/drydock log --oneline --since="6 hours ago"
+git -C /data3/drydock describe --tags --abbrev=0     # current PyPI tag
+
+# 2. HLE run status
+ps -p $(cat /tmp/hle_overnight.pid 2>/dev/null) -o pid,etime,comm
+ls /data3/drydock/hle_results/
+tail -50 /tmp/hle_overnight.log
+n_done=$(wc -l < /data3/drydock/hle_results/run_*/results.jsonl 2>/dev/null)
+echo "completed: $n_done / 200"
+
+# 3. If HLE crashed mid-flight, RESUME (skip already-done IDs)
+RUN_DIR=$(ls -td /data3/drydock/hle_results/run_* | head -1)
+nohup /home/bobef/miniconda3/bin/python3 /data3/drydock/scripts/hle_eval.py \
+    --source hle --limit 200 --shuffle --seed 42 --resume "$RUN_DIR" \
+    > /tmp/hle_resume.log 2>&1 &
+
+# 4. If complete, see the score
+cat /data3/drydock/hle_results/run_*/summary.json | python3 -m json.tool
+
+# 5. Infra health
+ps -p 2462362 -o pid,etime
+curl -s --max-time 3 http://localhost:8001/v1/models | head -1
+gh issue list --repo fbobe321/drydock --state open --limit 5
+```
+
+## Known issues + workarounds
+
+1. **`web_search` tool requires permission approval** by default
+   (`ToolPermission.ASK`). In batch eval the harness can't see/respond
+   to the prompt → session stalls. **Workaround:**
+   `--dangerously-skip-permissions` flag passed by `hle_eval.py`.
+   **Real fix:** auto-approve read-only tools when stdin is non-TTY
+   or `DRYDOCK_BATCH_MODE=1`. Memory:
+   `memory/project_hle_phase1_findings.md`.
+
+2. **TUI input handler corrupts rapid char-by-char multi-line pexpect
+   input.** Internal `\n` chars get partially eaten + spurious newlines
+   inserted on long prompts. **Workaround:** single-line prompts in
+   `hle_eval.py`. **Real fix:** debug `drydock/cli/textual_ui/` input
+   buffer.
+
+3. **Auto_release at 06:00/12:00/18:00/00:00 UTC overwrites site-packages.**
+   In-flight HLE runs survive (each new question is a fresh TUI spawn that
+   picks up the new binary), but if you direct-edit site-packages your
+   changes vanish. Always commit to source. Pause via:
+   `touch /data3/drydock/.pause_auto_release`.
+
+4. **PRD contamination: model edits `PRD.md` mid-session.** HLE doesn't
+   touch this — every HLE question gets a fresh empty cwd. Not a concern
+   for HLE; relevant for shakedown PRD runs.
+
+5. **Sessions take real time.** HLE questions involve web_search +
+   multi-step reasoning. Seed q's took 30-60s; HLE q's appear to take
+   5+ min each (q1: 5 min in and still working). Expect N=20 to take
+   1-2 hours, N=100 overnight.
+
+## Sentinels currently set
+
+- `/data3/drydock_test_projects/.pause_watchdog` — watchdog cron paused
+- `/data3/drydock/.pause_auto_release` — NOT set (auto_release is active)
+- `/data3/drydock/research/STOP` — research loop sentinel (per gitignore)
+
+## Tomorrow morning's first action
+
+If Phase 0 completed:
+1. `cat /data3/arxiv_corpus/hle_experiment/runs/phase0/summary.json` —
+   should be `correct: 20, score: 1.0`. If <20/20, the gap is the
+   model-using-retrieval bug to fix before any ablation matters.
+2. Look at the per-Q `predicted` vs `ground_truth` for the misses.
+   If pred is empty/garbled, drydock-side issue (look at
+   `tui_logs/<id>.tui.log`). If pred is wrong-but-confident, the
+   prefetch query didn't surface the right chunk.
+3. If 20/20, build Phase 1 seed (literal answer → near-example) and
+   launch the next phase via `run_phase.py phase1.sqlite runs/phase1`.
+4. If <20/20, fix the gap first; Phase 1+ is meaningless without a
+   Phase 0 floor.
+
+If Phase 0 crashed:
+1. Per-Q `tui_logs/<id>.tui.log` shows why drydock TUI failed for that Q.
+2. `run_phase.py` is restartable — re-launch with the same args (each
+   Q runs in a fresh TUI, so partial completion lands in
+   `results.jsonl` and a re-run will redo the whole 20).
+3. Likely failure modes: model didn't call retrieve (auto-prefetch
+   misfire), retrieve returned wrong chunk (BM25 miss), model saw
+   the answer but produced empty turn (admiral catches as
+   `empty_after_tool:retrieve`).
+
+## File map
+
+```
+/data3/drydock/
+├── scripts/
+│   ├── hle_eval.py                 # full-batch HLE eval (200 Qs, --shuffle --seed 42)
+│   ├── hle_eval_seed.jsonl         # 7 hand-crafted seed questions
+│   ├── fetch_arxiv_abstracts.py    # OAI-PMH fetcher (Phase 2 corpus)
+│   └── consume_retrieval_queue.py  # GraphRAG-leg autonomy (28× perf-fixed)
+├── hle_results_v1_baseline/        # archived Phase 1 baseline (10/200 = 5%)
+├── hle_results/                    # gitignored — Phase 1 re-baseline + Phase 2 runs
+│   └── run_<ts>/{config,results,summary}.json + tui_logs + work
+├── HLE_PRD.md                      # this file
+└── ~/.config/drydock/{hf_token,github_token,pypi_token}
+
+/data3/arxiv_corpus/                # Phase 2 corpus (gitignored, NOT in repo)
+├── raw/<set>/batch_NNNNNN.txt      # OAI-PMH chunks, 100 abstracts/file
+├── state/<set>.json                # fetcher resumption tokens
+├── logs/{fetch,fetch2,ingest5}.log # fetch + ingest progress
+├── graphrag.sqlite                 # 1.18M chunks, 8.5GB indexed
+└── hle_experiment/                 # Phase 2.5 ablation (NEW)
+    ├── picks.jsonl                 # 20 frozen baseline-failure IDs
+    ├── questions_full.jsonl        # full Q+A pulled from cais/hle
+    ├── seed_phase{0..3}/           # per-phase chunk dirs
+    ├── phase{0..3}.sqlite          # per-phase GraphRAG DBs
+    ├── run_phase.py                # orchestrator (drives drydock TUI)
+    └── runs/phase<N>/              # per-phase results.jsonl + summary.json
+```
+
+## What "done" looks like for Phase 1
+
+A number. With distribution. Per-category and overall. Writeup in this
+PRD. Commit the writeup, not the questions. Decide Phase 2 from there.

{drydock_cli-2.7.46 → drydock_cli-2.7.48}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.7.46
+Version: 2.7.48
 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-2.7.46 → drydock_cli-2.7.48}/drydock/core/agent_loop.py

@@ -1014,18 +1014,46 @@ class AgentLoop:
             _dbg(f"[STALL-DEBUG] inline retry #{_stall_attempt + 1} (prev={prev_role})")
             # Pop the empty assistant; inject an escalating nudge.
             self.messages.pop()
+            # Detect what the previous tool was so the nudge can steer
+            # the model toward the RIGHT next action. Suggesting read_file
+            # when the model just stalled after read_file reinforces the loop.
+            prev_tool_name: str | None = None
+            if prev_role == Role.tool and len(self.messages) >= 2:
+                # messages[-1] is now the tool result; messages[-2] is the
+                # assistant that called the tool.
+                assistant_msg = self.messages[-2]
+                if (assistant_msg.role == Role.assistant
+                        and assistant_msg.tool_calls):
+                    prev_tool_name = assistant_msg.tool_calls[-1].function.name if assistant_msg.tool_calls[-1].function else None
+            _readonly_tools = {"read_file", "grep", "glob", "ls", "pwd"}
+            _prev_was_read = prev_tool_name in _readonly_tools
             if _stall_attempt == 0:
-                note = (
-                    "Continue working. Use a tool (read_file, "
-                    "write_file, search_replace, bash) or state "
-                    "your plan in text."
-                )
+                if _prev_was_read:
+                    note = (
+                        f"You read a file but produced no output. "
+                        f"Now use write_file, search_replace, or bash "
+                        f"to make changes — do NOT call read_file again."
+                    )
+                else:
+                    note = (
+                        "Continue working. Use a tool (read_file, "
+                        "write_file, search_replace, bash) or state "
+                        "your plan in text."
+                    )
             elif _stall_attempt == 1:
-                note = (
-                    "You sent an empty response. Call a tool now "
-                    "(write_file, search_replace, bash, read_file) "
-                    "OR explicitly say you are done with this task."
-                )
+                if _prev_was_read:
+                    note = (
+                        f"You sent an empty response after reading a file. "
+                        f"Call write_file or search_replace NOW to apply "
+                        f"what you read — OR state in one sentence why you "
+                        f"cannot proceed."
+                    )
+                else:
+                    note = (
+                        "You sent an empty response. Call a tool now "
+                        "(write_file, search_replace, bash, read_file) "
+                        "OR explicitly say you are done with this task."
+                    )
             else:
                 note = (
                     "You have sent 3 empty responses in a row for "
@@ -2969,13 +2997,81 @@ class AgentLoop:
         self.messages.append(synth_assistant)
         self.messages.append(synth_tool)
 
+        # Authoritative-answer recognition. Curated GraphRAG corpora can
+        # mark a chunk's verified answer with a literal `ANSWER:` line
+        # (also `Answer:`, `Verified answer:`, `Ground truth:`). When
+        # auto-prefetch surfaces such a chunk and the BM25 score is high
+        # enough that we're confident it matches the user's question,
+        # inject a system note telling the model to use that line
+        # verbatim — without it, Gemma 4 re-derives from scratch and
+        # often overrules the verified value (HLE Phase 0 ablation
+        # 2026-05-06: 5/20 with seeded answers because the model
+        # ignored its own retrieved ANSWER lines).
+        #
+        # Only fire when the TOP-1 chunk has the marker. If a lower-
+        # scoring chunk has ANSWER (e.g. an unrelated Q's seed bled
+        # into the result set), the system note would point the model
+        # at the wrong answer (Phase 0' "Nunavut → Ontario" case).
+        #
+        # Two paths to "authoritative":
+        #   (a) absolute: top score >= AUTHORITATIVE_SCORE (works for
+        #       long, term-rich questions that yield high BM25)
+        #   (b) relative: chunk has the curated header prefix
+        #       `===<tag>:<id>===` AND top score outranks 2× the next
+        #       hit's score. Catches narrow-trivia cases where BM25
+        #       scores are naturally lower (e.g. "What city does X
+        #       move to in 1997 movie Y?") but retrieval clearly
+        #       picked one curated chunk over the rest.
+        import re as _re
+        ANSWER_MARKERS = ("ANSWER:", "Answer:", "Verified answer:",
+                          "Ground truth:", "Correct answer:")
+        CURATED_HEADER_RE = _re.compile(r"^===[a-z][a-z0-9_-]*:\S+===")
+        AUTHORITATIVE_SCORE = 100.0  # absolute high-confidence bar
+        DOMINANCE_SCORE = 30.0       # relative path floor (well above 8.0 noise)
+        DOMINANCE_RATIO = 2.0        # top must beat second by this much
+        top_chunk = chunks[0] if chunks else ""
+        top_score = float(getattr(good_hits[0], "score", 0))
+        next_score = (
+            float(getattr(good_hits[1], "score", 0))
+            if len(good_hits) >= 2 else 0.0
+        )
+        has_marker = any(marker in top_chunk for marker in ANSWER_MARKERS)
+        # Inspect content lines (skip the path/score header that the
+        # formatter prepends) for the curated tag.
+        chunk_body_lines = top_chunk.split("\n")
+        has_curated_header = any(
+            CURATED_HEADER_RE.match(line.strip())
+            for line in chunk_body_lines[:6]
+        )
+        is_authoritative = has_marker and (
+            top_score >= AUTHORITATIVE_SCORE
+            or (
+                has_curated_header
+                and top_score >= DOMINANCE_SCORE
+                and top_score >= DOMINANCE_RATIO * next_score
+            )
+        )
+        if is_authoritative:
+            note = (
+                "The retrieve tool result above contains a curated "
+                "chunk whose question matches the user's. Locate the "
+                "line beginning with one of "
+                f"{list(ANSWER_MARKERS)} and emit that value verbatim "
+                "as your FINAL ANSWER. Do not re-derive — the chunk is "
+                "authoritative ground truth provided by the corpus "
+                "curator. Respond with text only, no further tool calls."
+            )
+            self._inject_system_note(note)
+
         logger.warning(
             "[AUTO-RETRIEVE] synthesized retrieve tool result: %d chunks "
-            "(top score %.1f, content %d chars), msgs now %d",
+            "(top score %.1f, content %d chars), msgs now %d, "
+            "authoritative=%s",
             len(chunks),
-            max(float(getattr(h, "score", 0)) for h in good_hits),
+            top_score,
             len(formatted),
             len(self.messages),
+            top_score >= AUTHORITATIVE_SCORE and has_marker,
         )
 
     async def _auto_route_task(self, user_msg: str) -> None:

{drydock_cli-2.7.46 → drydock_cli-2.7.48}/drydock/core/llm/format.py

@@ -379,15 +379,16 @@ class APIToolFormatHandler:
                 # history well-formed (prevents empty_after_tool loops) but
                 # route to suppressed_failures so the TUI doesn't show an error.
                 if parsed_call.tool_name in _IGNORE_TOOLS:
-                    if (
-                        parsed_call.tool_name in _RETRIEVAL_HALLUCINATIONS
-                        and "retrieve" in active_tools
-                    ):
+                    if parsed_call.tool_name in _RETRIEVAL_HALLUCINATIONS:
+                        # These are file-listing/indexing hallucinations, NOT semantic
+                        # search — redirecting to `retrieve` confuses the model (it doesn't
+                        # know what query to write for a directory listing, stalls empty).
+                        # glob/grep are the correct concrete alternatives.
                         redirect = (
-                            f"'{parsed_call.tool_name}' does not exist. "
-                            f"The correct retrieval tool is `retrieve`. "
-                            f"Call `retrieve(query='<your search terms>')` NOW — "
-                            f"it searches the project index and returns relevant results."
+                            f"'{parsed_call.tool_name}' does not exist — do not call it again. "
+                            f"To list project files: call `glob(pattern='**/*.py')`. "
+                            f"To search file content: call `grep(pattern='...')`. "
+                            f"Call one of these tools NOW to continue."
                         )
                     else:
                         redirect = (

{drydock_cli-2.7.46 → drydock_cli-2.7.48}/drydock/core/tools/builtins/bash.py

@@ -786,14 +786,29 @@ class Bash(
                             f"rewrite the file with the corrected content. "
                             f"Do NOT re-run this sed command unchanged.]"
                         )
+                    elif bool(_re.search(
+                        r'(?:^|\|)\s*(?:ls\b|find\b|tree\b)', args.command
+                    )) and stdout.strip():
+                        notice = (
+                            f"[NOTICE: this is the #{entry['count']}th identical "
+                            f"run of `{cmd_preview}` — the directory listing is "
+                            f"stable and you have already seen it. "
+                            f"Re-running will not reveal new files. "
+                            f"Make a decision based on what you already know: "
+                            f"either create the missing file with write_file, "
+                            f"read an existing file with read_file, or move on "
+                            f"to the next task step. "
+                            f"Do NOT re-run this listing command.]"
+                        )
                     else:
                         notice = (
-                            f"[NOTICE: this is the #{entry['count']} identical "
+                            f"[NOTICE: this is the #{entry['count']}th identical "
                             f"run of `{cmd_preview}` with byte-identical "
                             f"output and rc={returncode}. Re-running will not "
-                            f"change anything — you must EDIT SOURCE CODE to "
-                            f"change this output. Previous stdout first 300 "
-                            f"chars:\n{stdout[:300]}]"
+                            f"change anything. "
+                            f"Make a decision based on the output you already have "
+                            f"and move to the next task step. "
+                            f"Previous stdout first 300 chars:\n{stdout[:300]}]"
                         )
                     yield self._build_result(
                         command=args.command,

{drydock_cli-2.7.46 → drydock_cli-2.7.48}/drydock/core/tools/builtins/exit_plan_mode.py

@@ -66,7 +66,10 @@ class ExitPlanMode(
             raise ToolError("ExitPlanMode requires an agent manager context.")
 
         if ctx.agent_manager.active_profile.name != BuiltinAgentName.PLAN:
-            raise ToolError("ExitPlanMode can only be used in plan mode.")
+            # Model called exit_plan_mode outside plan mode (Gemma 4 hallucinates this).
+            # Return a no-op success so the model moves on instead of retrying.
+            yield ExitPlanModeResult(switched=False, message="Already in implementation mode.")
+            return
 
         if ctx.user_input_callback is None:
             raise ToolError("ExitPlanMode requires an interactive UI.")

{drydock_cli-2.7.46 → drydock_cli-2.7.48}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.7.46"
+version = "2.7.48"
 description = "Local-first CLI coding agent — tested with Gemma 4 26B via vLLM"
 readme = "README.md"
 requires-python = ">=3.12"

{drydock_cli-2.7.46 → drydock_cli-2.7.48}/scripts/stress_babysitter.sh

@@ -129,11 +129,14 @@ if [ "$CURIDX" -ge "$TOTAL_PROMPTS" ]; then
     exit 0
 fi
 
-# Determine restart step. DONE is the count of successful prompts; feed
-# that into --resume-from-step so the restart continues where the last
-# accepted prompt left off.
-RESUME_AT=$DONE
-if [ "$RESUME_AT" -lt 1 ]; then
+# Determine restart step. Use CURIDX (the 1-indexed step number last seen in
+# the log) minus 1 to re-run the step that was in progress when the harness
+# died. DONE (count of successes in the partial run) is wrong here because
+# it resets to 0 on each restart while CURIDX tracks position in the full
+# 1658-prompt list.
+if [ "$CURIDX" -gt 1 ]; then
+    RESUME_AT=$(( CURIDX - 1 ))
+else
     RESUME_AT=1
 fi
 

{drydock_cli-2.7.46 → drydock_cli-2.7.48}/tests/test_hallucinated_tool_suppression.py

@@ -45,6 +45,13 @@ HALLUCINATED_TOOLS = [
     "ralph_file_summary",
     "file_summary",
     "repo_summary",
+    # 2026-05-06 stress run additions:
+    "read_mcp_resource",
+    "read_resource",
+    "get_resource",
+    "lsp",
+    "lsp_definition",
+    "lsp_references",
 ]