terminal-bench-2

This runbook describes the current AlphaDiana path for terminal-bench-2.

Use it for local AlphaDiana evaluation and smoke validation. It is not a claim of strict Harbor-equivalent leaderboard execution.

For the benchmark-plan direct_llm baseline, use the official terminal-bench-2 repository and Harbor's built-in terminus-2 agent. The April 19, 2026 OpenRouter/Qwen evidence for that official path is summarized below.

Current Support

Mode	Checked-in smoke config
direct_llm	configs/examples/terminal_bench2_directllm_minimax.yaml
opencode	configs/examples/terminal_bench2_opencode_minimax.yaml
openclaw	configs/examples/terminal_bench2_openclaw_minimax.yaml
zeroclaw	configs/examples/terminal_bench2_zeroclaw_minimax.yaml

All four paths use the same local AlphaDiana terminal_bench2 benchmark loader and scorer. This checkout does not ship a Terminal-Bench 2 full-run config.

Podman Task-Container Readiness

On May 15, 2026, Phase 7 validated the opt-in terminal_bench2_opencode Podman task-container path with a five-task official TerminalBench2 pilot.

Entry points:

• Config: configs/smokes/podman_terminal_bench2/terminal_bench2_opencode_pilot.yaml
• Runner: scripts/run_podman_terminal_bench2_readiness.sh

Run shape:

export TERMINAL_BENCH2_DIR=<official-terminal-bench-2-task-root>
export OPENAI_BASE_URL=<openai-compatible-base-url>
export OPENAI_API_KEY=<api-key-or-placeholder>
export OPENAI_MODEL_NAME=<model-name>
export TB2_OPENCODE_RUNTIME_IMAGE=localhost/alphadiana/tb2-opencode-controller:latest
export ALPHADIANA_TB2_LOGS_DIR="$PWD/logs/podman-terminal-bench2-readiness/task-logs"
export PODMAN_TB2_RUN_PREFIX=podman_tb2_$(date +%Y%m%d_%H%M%S)

bash scripts/run_podman_terminal_bench2_readiness.sh validate
bash scripts/run_podman_terminal_bench2_readiness.sh preflight
bash scripts/run_podman_terminal_bench2_readiness.sh pilot
bash scripts/run_podman_terminal_bench2_readiness.sh audit

Passing evidence:

• Run prefix: podman_tb2_20260515_phase7_abslogs
• Pilot run: podman_tb2_20260515_phase7_abslogs_terminal_bench2_opencode
• Tasks: db-wal-recovery, fix-git, overfull-hbox, adaptive-rejection-sampler, break-filter-js-from-html
• Result: 5/5 task JSON rows, all valid_scored, score=0.0, metadata.container_engine=podman, verifier ok, and discoverable artifacts/logs.
• Audit: passed with five clean infrastructure rows.

This supports recommending a larger overnight TerminalBench2 OpenCode Podman campaign. It does not prove Direct x TerminalBench2, OpenClaw/ZeroClaw TerminalBench2 Podman readiness, full-sweep readiness, Podman global default promotion, or any SWE-bench/MMMU-Pro status.

OpenRouter/Qwen pilot status on April 19, 2026:

• direct_llm official Harbor baseline: the initial batch pilot_20260419_qwen35_27b_terminal_bench2_directllm_t3 failed (0/3 verifier rewards, one AgentTimeoutError), but the repaired follow-up archive pilot_20260419_qwen35_27b_terminal_bench2_directllm_t3_repair_r1 is now 3/3 normal trajectories with 3/3 reward=1 on the approved trio. This is repaired official-checkout evidence, not a stock upstream invocation.
• opencode native in-container pilot: pilot_20260420_qwen35_27b_terminal_bench2_opencode_t3_r2 completed 3/3 normal task records (score=0,0,1) and was uploaded to pilot_run/pilot_20260420_qwen35_27b_terminal_bench2_opencode_t3_r2/
• openclaw native in-container pilot: after removing the bad /app/TASK.md prompt assumption, the rerun pilot_20260420_qwen35_27b_terminal_bench2_openclaw_t3_r4 completed 3/3 normal task records with 3/3 score=1 and was uploaded to pilot_run/pilot_20260420_qwen35_27b_terminal_bench2_openclaw_t3_r4/
• zeroclaw native in-container pilot: pilot_20260420_qwen35_27b_terminal_bench2_zeroclaw_t3_repair_r5 completed 3/3 normal task records (score=0,1,1) and was uploaded to pilot_run/pilot_20260420_qwen35_27b_terminal_bench2_zeroclaw_t3_repair_r5/

Local follow-up on April 19, 2026:

• rerun_20260419_qwen35_27b_terminal_bench2_openclaw_timeoutcheck_r2 completed unattended on the timeout-check sample and no longer reported the old inner ctx=16000 mismatch
• rerun_20260419_qwen35_27b_terminal_bench2_zeroclaw_fixgit_r1 no longer reproduced the old missing-repo / missing-site workspace mismatch; the preserved failure moved to the loop detector

Local OpenCode/logprob follow-up on April 25, 2026:

• phase12_opencode_tb2_qwen35_64k_logprobs_t3_parallel_20260425 completed the same three-task local OpenCode path and wrote all three task JSONs.
• All three tasks have no task-level error and saved provider-proxy logprobs with matching float/int16 sidecars (1388, 1242, and 4389 records).
• This was not a clean task-quality pass: bn-fit-modify is valid_scored, while adaptive-rejection-sampler and break-filter-js-from-html ended with response_json.returncode=-1 and score_status=verifier_error after a long local run. The raw log also contains one vLLM HTTP 400 where accumulated OpenCode context plus the local output cap exceeded the model context window by one token.
• terminal_bench2_opencode now wraps the container-local OpenCode command in timeout --kill-after, so future solver timeouts should not leave container-local OpenCode processes continuing to hit the provider after the host-side docker exec timeout.

Prerequisites

Run from the repo root:

source scripts/activate.sh
export PYTHONPATH=$PWD

export OPENAI_BASE_URL=https://models.sjtu.edu.cn/api/v1/
export OPENAI_API_KEY=sk-...
export OPENAI_MODEL_NAME=minimax-m2.5

For the April 19, 2026 OpenRouter/Qwen pilot:

export OPENAI_BASE_URL=https://openrouter.ai/api/v1
export OPENAI_MODEL_NAME=qwen/qwen3.5-27b

Model-Pin Caveat

The general "some smoke configs pin the model in YAML" rule lives in benchmarks/index.md. This is the concrete tb2 case.

configs/examples/terminal_bench2_directllm_minimax.yaml hard-pins agent.config.model: "minimax-m2.5", so it ignores OPENAI_MODEL_NAME (it still reads api_base / api_key from OPENAI_BASE_URL / OPENAI_API_KEY). To run a different model on that config, override the agent config explicitly:

python -m alphadiana.cli run configs/examples/terminal_bench2_directllm_minimax.yaml \
  -o agent.config.model=... \
  -o agent.config.api_base=... \
  -o agent.config.api_key=...

The opencode, openclaw, and zeroclaw tb2 example configs read the model from the environment (${OPENAI_MODEL_NAME}), so for those switching the env vars is enough. The zeroclaw config additionally accepts -o agent.config.logs_base_dir=... when redirecting log output.

You also need:

• Docker
• a local terminal-bench task checkout
• pre-pulled task images
• runtime source images for the native in-container agents: tmlrgroup/alphadiana:v1 for openclaw, alphadiana/tb2-opencode-controller:latest for opencode, and zeroclaw-reasoning:0.6.9 for zeroclaw

Official DirectLLM Baseline

This section is for the official direct-LLM baseline only. It is outside the AlphaDiana runtime, but it is the benchmark-plan meaning of DirectLLM x terminal-bench-2.

Minimal Harbor invocation shape from the upstream repo:

cd /path/to/terminal-bench-2
unset ALL_PROXY all_proxy HTTP_PROXY HTTPS_PROXY http_proxy https_proxy
export OPENROUTER_API_KEY=...

uv run harbor run --path . \
  --agent terminus-2 \
  --model openrouter/qwen/qwen3.5-27b \
  --n-concurrent 3

Local April 19 OpenRouter/Qwen pilot specifics:

• approved trio: adaptive-rejection-sampler, bn-fit-modify, break-filter-js-from-html
• runtime settings: temperature=0.6, top_p=0.95, max_tokens=32768, reasoning_effort=high
• output root: jobs/pilot_20260419_qwen35_27b_terminal_bench2_directllm_t3
• accepted repaired archive: pilot_20260419_qwen35_27b_terminal_bench2_directllm_t3_repair_r1

Observed outcome on the initial official baseline:

• Harbor completed 3/3 trials and preserved all trial artifacts
• all verifier rewards were 0
• adaptive-rejection-sampler ended with AgentTimeoutError
• bn-fit-modify and break-filter-js-from-html both reached the verifier and still scored 0

Local repair follow-up in the same official checkout:

• verifier entrypoints for adaptive-rejection-sampler, bn-fit-modify, and break-filter-js-from-html were normalized to a stable python venv + pip + pytest flow
• Harbor's local JSON parser was patched to ignore benign pre-JSON prefix text instead of surfacing a parser warning
• adaptive-rejection-sampler was rerun after tightening the task contract to the required ars(density_fn, domain, n = ...) interface and switching the task image to a lightweight local R image
• accepted repaired task runs:
  • bn-fit-modify from pilot_20260419_qwen35_27b_terminal_bench2_directllm_t3_r2: normal trajectory, reward=1, exception_type=null
  • break-filter-js-from-html from pilot_20260419_qwen35_27b_terminal_bench2_directllm_break_filter_js_from_html_r4: normal trajectory, reward=1, exception_type=null
  • adaptive-rejection-sampler from pilot_20260419_qwen35_27b_terminal_bench2_directllm_adaptive_rejection_sampler_r6: normal trajectory, reward=1, exception_type=null, verifier 9/9 passed
• the accepted repaired bundle was uploaded to T-MARS/alphadiana-benchmark-results under pilot_run/pilot_20260419_qwen35_27b_terminal_bench2_directllm_t3_repair_r1/

Treat that repaired bundle as the current local smoke-valid signal for the official direct_llm path on OpenRouter/Qwen.

Official leaderboard path (Harbor + Terminus-2)

The sections above are the AlphaDiana container-agent path. To run the official/leaderboard configuration instead (standalone Harbor CLI + the upstream terminus-2 agent), use the path below. Harbor owns the system prompt; the AlphaDiana task-container harness at alphadiana/benchmarks/terminal_bench2/harness/docker.py is not on this path.

vLLM endpoint (Qwen3.5-27B; the serve-side flags match the official spec):

python -m vllm.entrypoints.openai.api_server \
  --model Qwen/Qwen3.5-27B --host 0.0.0.0 --port <port> \
  --trust-remote-code --enable-auto-tool-choice --tool-call-parser qwen3_coder \
  --tensor-parallel-size 2 --gpu-memory-utilization 0.9 \
  --max-model-len 262144 --generation-config vllm \
  --override-generation-config '{"presence_penalty": 1.5}' \
  --served-model-name qwen3.5-27b Qwen/Qwen3.5-27B

--enable-auto-tool-choice --tool-call-parser qwen3_coder is required for terminus-2 to issue tool calls; --generation-config vllm ignores the model's shipped sampling defaults; omitting --reasoning-parser keeps thinking tokens in message.content.

Install Harbor + Terminus-2:

export DIRECTLLM_TB2_ROOT=/path/to/terminal-bench-2
curl -LsSf https://astral.sh/uv/install.sh | sh
uv tool install harbor          # installs the `harbor` binary

No official Terminal-Bench 2 campaign manifest is checked into this checkout. Use Harbor directly for that leaderboard path, or add and review a campaign manifest before invoking alphadiana.benchmark_rollout_cli.

The intended campaign command runs harbor run --dataset [email protected] --agent terminus-2 --model openai/qwen3.5-27b ... at temperature=0.0, reasoning_effort=high, top_p=0.95, max_tokens=131072, --n-concurrent 10 (pass@1). Outputs land in $DIRECTLLM_TB2_ROOT/jobs/<run_id>/. Renderer and dispatch live in alphadiana/utils/rollout_campaign.py (_render_official_tb2_command, render_run_command); Harbor upstream is github.com/laude-institute/harbor.

Prepare Tasks

Clone a local task checkout:

git clone --depth=1 https://github.com/laude-institute/terminal-bench.git /tmp/terminal-bench

Set the full-run task root:

export TERMINAL_BENCH2_DIR=/tmp/terminal-bench/tasks

Prepare a deterministic smoke staging directory with one task:

rm -rf /tmp/terminal-bench-smoke-dbwal
mkdir -p /tmp/terminal-bench-smoke-dbwal
cp -a /tmp/terminal-bench/tasks/db-wal-recovery /tmp/terminal-bench-smoke-dbwal/

export TERMINAL_BENCH2_SMOKE_DIR=/tmp/terminal-bench-smoke-dbwal

For the April 19 OpenRouter pilot, the approved staged trio was:

• db-wal-recovery
• fix-git
• break-filter-js-from-html

Example staging flow with environment variables instead of a hardcoded local path:

export TERMINAL_BENCH2_SOURCE_ROOT=/path/to/terminal-bench-2/tasks
export TERMINAL_BENCH2_PILOT_ROOT=/path/to/staged-terminal-bench-2-qwen-t3

rm -rf "$TERMINAL_BENCH2_PILOT_ROOT"
mkdir -p "$TERMINAL_BENCH2_PILOT_ROOT"
cp -a "$TERMINAL_BENCH2_SOURCE_ROOT"/db-wal-recovery "$TERMINAL_BENCH2_PILOT_ROOT"/
cp -a "$TERMINAL_BENCH2_SOURCE_ROOT"/fix-git "$TERMINAL_BENCH2_PILOT_ROOT"/
cp -a "$TERMINAL_BENCH2_SOURCE_ROOT"/break-filter-js-from-html "$TERMINAL_BENCH2_PILOT_ROOT"/

The smoke configs assume TERMINAL_BENCH2_SMOKE_DIR points at a directory whose immediate children are task directories. The full-run configs assume TERMINAL_BENCH2_DIR points at the full task root.

Current loader note on April 22, 2026: pointing TERMINAL_BENCH2_DIR at a normal checkout root is valid. The loader skips non-task directories such as .git and jobs. On the current local April 22 checkout used for the OpenRouter full-run follow-up, that meant 91 immediate directories on disk but 89 loaded task roots.

Pre-pull Task Images

Before any smoke or full run:

python - <<'PY' | sort -u | xargs -r -n1 docker pull
import os, tomllib
from pathlib import Path

for task_toml in Path(os.environ["TERMINAL_BENCH2_DIR"]).glob("*/task.toml"):
    with task_toml.open("rb") as f:
        data = tomllib.load(f)
    image = data.get("environment", {}).get("docker_image")
    if image:
        print(image)
PY

For the default smoke task specifically:

docker pull alexgshaw/db-wal-recovery:20251031

Prepare Runtime Source Images

direct_llm still uses the helper-workspace controller path. The native agents opencode, openclaw, and zeroclaw now run inside a derived task image, so they need runtime source images instead of controller containers.

Prepare the checked-in sources once:

docker pull tmlrgroup/alphadiana:v1
docker image inspect alphadiana/tb2-opencode-controller:latest >/dev/null
docker pull zeroclaw-reasoning:0.6.9

The first native-agent smoke/full run automatically builds a derived alphadiana-tb2-runtime:<agent>-<fingerprint> image from the task image plus the selected runtime source image.

Early OpenRouter full-run evidence on April 22, 2026 uses nvidia/nemotron-3-nano-30b-a3b:free:

• full_20260422_openrouter_nemotron_3_nano_30b_a3b_terminal_bench2_directllm_r1 and ..._openclaw_r1 already wrote tb2_adaptive-rejection-sampler.json as valid_scored with reward=0, metadata.verifier_status=ok, and metadata.verifier_reward_observed=true
• ..._zeroclaw_r1 also advanced with normal valid_scored reward-0 task JSONs
• ..._opencode_r1 is less quiet so far: its first task wrote score_status=verifier_error with metadata.verifier_status=missing_reward
• the first task still takes roughly minutes before its JSON appears because of container bring-up plus verifier startup; do not classify the run as stalled during that gap alone

For ZeroClaw, prefer putting large temporary files on a data disk before running:

export TMPDIR=/data/$USER/tmp/alphadiana-tb2
mkdir -p "$TMPDIR"

Runtime Model

AlphaDiana now uses two different TB2 execution contracts:

• direct_llm: helper-workspace controller mode. The model sees tb2-exec, tb2-copy-from, tb2-copy-to, and tb2-test.
• opencode, openclaw, zeroclaw: native in-container mode. AlphaDiana derives a runtime image from the task image, starts that task container directly, and runs the agent CLI inside it.

For the native agents:

• the model sees the live task filesystem directly
• tb2-exec / tb2-copy-* are not exposed to the model
• /tests/test.sh and reward.txt stay unchanged
• the outer harness still runs verification once at the end

Smoke Runs

Validate the smoke configs first:

python -m alphadiana.cli validate configs/examples/terminal_bench2_directllm_minimax.yaml
python -m alphadiana.cli validate configs/examples/terminal_bench2_opencode_minimax.yaml
python -m alphadiana.cli validate configs/examples/terminal_bench2_openclaw_minimax.yaml
python -m alphadiana.cli validate configs/examples/terminal_bench2_zeroclaw_minimax.yaml

Run the three smoke configs:

python -m alphadiana.cli run configs/examples/terminal_bench2_directllm_minimax.yaml --redo-all
python -m alphadiana.cli run configs/examples/terminal_bench2_opencode_minimax.yaml --redo-all
python -m alphadiana.cli run configs/examples/terminal_bench2_openclaw_minimax.yaml --redo-all
python -m alphadiana.cli run configs/examples/terminal_bench2_zeroclaw_minimax.yaml --redo-all

April 19 OpenRouter/Qwen 3-task pilot commands:

python -m alphadiana.cli validate configs/examples/terminal_bench2_openclaw_minimax.yaml \
  -o run_id=pilot_20260419_qwen35_27b_terminal_bench2_openclaw_t3 \
  -o output_dir=./results \
  -o benchmark.config.tasks_dir="$TERMINAL_BENCH2_PILOT_ROOT" \
  -o benchmark.config.max_tasks=3 \
  -o agent.config.model_name=qwen/qwen3.5-27b

python -m alphadiana.cli validate configs/examples/terminal_bench2_opencode_minimax.yaml \
  -o run_id=pilot_20260419_qwen35_27b_terminal_bench2_opencode_t3 \
  -o output_dir=./results \
  -o benchmark.config.tasks_dir="$TERMINAL_BENCH2_PILOT_ROOT" \
  -o benchmark.config.max_tasks=3 \
  -o agent.config.model_name=qwen/qwen3.5-27b \
  -o agent.config.model=custom/qwen/qwen3.5-27b \
  -o agent.config.streaming=true \
  -o max_concurrent=2

python -m alphadiana.cli run configs/examples/terminal_bench2_openclaw_minimax.yaml \
  -o run_id=pilot_20260419_qwen35_27b_terminal_bench2_openclaw_t3 \
  -o output_dir=./results \
  -o benchmark.config.tasks_dir="$TERMINAL_BENCH2_PILOT_ROOT" \
  -o benchmark.config.max_tasks=3 \
  -o agent.config.model_name=qwen/qwen3.5-27b \
  2>&1 | tee logs/pilot_20260419_qwen35_27b_terminal_bench2_openclaw_t3.log

python -m alphadiana.cli run configs/examples/terminal_bench2_opencode_minimax.yaml \
  -o run_id=pilot_20260419_qwen35_27b_terminal_bench2_opencode_t3 \
  -o output_dir=./results \
  -o benchmark.config.tasks_dir="$TERMINAL_BENCH2_PILOT_ROOT" \
  -o benchmark.config.max_tasks=3 \
  -o agent.config.model_name=qwen/qwen3.5-27b \
  -o agent.config.model=custom/qwen/qwen3.5-27b \
  -o agent.config.streaming=true \
  -o max_concurrent=2 \
  2>&1 | tee logs/pilot_20260419_qwen35_27b_terminal_bench2_opencode_t3.log

Observed results for that pilot:

• pilot_20260419_qwen35_27b_terminal_bench2_opencode_t3: 3/3 task records, all score=1
• pilot_20260419_qwen35_27b_terminal_bench2_openclaw_t3: 3/3 task records, tb2_fix-git -> score=1, tb2_db-wal-recovery -> score=0, tb2_break-filter-js-from-html -> score=0
• The OpenClaw pilot captured low context window: ... ctx=16000 on all three tasks and needed manual watchdog interruption on the first two tasks. Treat that path as experimental on OpenRouter/Qwen.
• the initial strict ZeroClaw smoke-plan alignment attempt smoke_20260420_qwen35_27b_terminal_bench2_zeroclaw_align_r1 was abnormal because AlphaDiana still let ZeroClaw auto-enable its own internal Docker sandbox inside the TB2 controller image, which hid the mounted control workspace and broke the ./tb2-exec contract
• after forcing security_sandbox_enabled=false on terminal_bench2_zeroclaw, the repaired smoke_20260420_qwen35_27b_terminal_bench2_zeroclaw_align_r2 completed normally as a reward-0 task record, and the replacement rerun pilot_20260420_qwen35_27b_terminal_bench2_zeroclaw_t3_repair_r3 completed 3/3 with normal task JSONs (2 reward-0 failures, 1 reward-1 pass)
• the April 20 in-container migration smokes supersede the native-agent controller-specific caveats: smoke_20260420_qwen35_27b_tb2_openclaw_incontainer_r2 completed 1/1 on break-filter-js-from-html with score=1, smoke_20260420_qwen35_27b_tb2_opencode_incontainer_r2 completed 1/1 on break-filter-js-from-html with a normal reward-0 trajectory, and smoke_20260420_qwen35_27b_tb2_zeroclaw_incontainer_r2 completed 1/1 on db-wal-recovery with a normal reward-0 trajectory after sanitizing runtime logs out of the top-level assistant text
• the full April 20 native TB2 reruns are now the accepted pilot evidence: pilot_20260420_qwen35_27b_terminal_bench2_opencode_t3_r2 wrote 3/3 normal task JSONs and was uploaded, pilot_20260420_qwen35_27b_terminal_bench2_openclaw_t3_r4 wrote 3/3 normal task JSONs after fixing the in-container prompt contract to stop assuming /app/TASK.md, and pilot_20260420_qwen35_27b_terminal_bench2_zeroclaw_t3_repair_r5 wrote 3/3 normal task JSONs and supersedes the older controller-mode repair_r3 archive

Smoke success means:

• the task loads
• the selected agent path runs
• /tests/test.sh runs
• a scored JSONL result is written

It does not mean the agent is competitive across the full benchmark.

Full Runs

There is no checked-in Terminal-Bench 2 full-run YAML. The example configs are bounded smoke entry points. For a full evaluation, create and review a dedicated config that points benchmark.config.tasks_dir at the complete task checkout and sets the intended task selection, runtime, timeouts, output directory, and concurrency.

Recommended concurrency:

• direct_llm: max_concurrent: 4
• opencode: max_concurrent: 2
• openclaw: max_concurrent: 1
• zeroclaw: max_concurrent: 1

Adjust only if the local machine has enough Docker and API capacity.

Result Interpretation

The current AlphaDiana terminal_bench2 scorer is binary:

• reward.txt == "1" means pass
• missing or non-1 reward means fail

The JSONL score comes from that reward path.

As of April 22, 2026, current main no longer fabricates metadata.reward="0" for metadata.verifier_status="skipped_duplicate" when no reward file was actually observed. Historical artifacts such as full_20260422_terminal_bench2_opencode_deepseek_chat_r2/tasks/tb2_adaptive-rejection-sampler.json remain useful audit evidence, but the intended current behavior is the post-fix reproducer fixproof_after_20260422_tb2_opencode_deepseek_fast2_t1, which records metadata.reward=null, metadata.verifier_reward_observed=false, and score_status=verifier_error for that bookkeeping path.

Current Config Semantics

Smoke configs:

• live under configs/examples/
• use TERMINAL_BENCH2_SMOKE_DIR
• intentionally run one staged task

A full-run config is not checked in. A project-specific full config should use TERMINAL_BENCH2_DIR and scan the intended task directories under that root.

For the current checked-in smoke setup, the canonical staged task is db-wal-recovery.

The April 19 OpenRouter/Qwen pilot used the approved trio db-wal-recovery, fix-git, and break-filter-js-from-html instead of a single staged task.

ZeroClaw Reproduction Notes

terminal-bench-2 does not use ROCK. The formal ZeroClaw smoke path is now the TB2 native in-container path:

• AlphaDiana starts a derived TB2 runtime image for the selected task
• ZeroClaw runs directly inside that task container
• the task JSON is normal as long as the run writes a scored record, even if the verifier reward is 0

Reproduce The 2026-04-20 In-Container Smoke

Prepare the staged smoke task and runtime source image first:

export OPENAI_BASE_URL=https://openrouter.ai/api/v1/
export OPENAI_API_KEY=sk-...
export OPENAI_MODEL_NAME=Qwen/Qwen3.5-27B
export PYTHONPATH=$PWD
export TERMINAL_BENCH2_SMOKE_DIR=/path/to/terminal-bench-smoke-dbwal
export TMPDIR=/data/$USER/tmp/alphadiana-tb2
mkdir -p "$TMPDIR"

docker pull alexgshaw/db-wal-recovery:20251031
docker pull zeroclaw-reasoning:0.6.9

Run the smoke:

python -m alphadiana.cli run configs/examples/terminal_bench2_zeroclaw_minimax.yaml \
  -o run_id=smoke_20260420_qwen35_27b_tb2_zeroclaw_incontainer_r2 \
  -o output_dir=./results/smoke_20260420_qwen35_27b_tb2_zeroclaw_incontainer_r2

Observed local verification on 2026-04-20:

• run_id: smoke_20260420_qwen35_27b_tb2_zeroclaw_incontainer_r2
• result: 1/1 completed, predicted=0, no error
• execution mode: derived in-container TB2 runtime image
• the assistant trajectory contains normal task reasoning after runtime-log sanitization; it no longer exposes tb2-exec/tb2-copy-*