Phase 5 — Launch

Goal: validate the job, then run it on explicit confirmation.

1. Dry-run (always first)

uv run ./scripts/run.sh --job <slug> --dry-run

This resolves the manifest, prints it, and exits — no Docker, no dataset mutation, no model calls. It runs the same preflight the real run does: model_connection validation, protocol-bridge check, extra_body/direct incompatibility check, and a harness-mount preflight. Fix any error and re-run until it resolves cleanly. Typical dry-run failures:

• model_connection=direct is incompatible with this job → the model has extra_body or needs a protocol bridge. Set model_connection: local_proxy (Phase 4).
• Model config not found / no top-level 'model:' → job's model: slug doesn't match a configs/models/<slug>.yaml.
• missing local harness mount image → build it: scripts/harness/build-harness-mounts.sh --harness <family>/<version>, or pass AUTO_BUILD_HARNESS_MOUNT=1 on the real run.
• auth/connection errors → an env-var name the model references is missing from .env (back to Phase 3).

2. Confirm, then run for real

Ask the user to confirm before the real run — it is time- and cost-intensive (real model calls, Docker builds, every task). Only after explicit confirmation:

uv run ./scripts/run.sh --job <slug>

--help lists available jobs and every env knob.

Env knobs (from run.sh --help)

• SHARDS=N — parallel Harbor shards (default 1). Also engaged automatically when the job has a task_selection.
• SHARD_CONCURRENCY=N — concurrency within each shard (default: job's n_concurrent_trials, else 2).
• NO_FORCE_BUILD=1 — skip Docker image rebuild (faster re-runs).
• DISABLE_VERIFICATION=1 — agent rollout only; skip task verification + the host-side LLM judge.
• AUTO_BUILD_HARNESS_MOUNT=1 — build a missing local harness mount image.
• PROXY_PORT=N — preferred job-private proxy port (default 3456; auto-bumps if busy).
• SHARED_PROXY=1 — reuse one long-lived shared proxy on :3456 instead of a per-job one. Start it first: scripts/proxy/proxy-shared.sh start.

Both SHARDS and SHARD_CONCURRENCY must be positive integers — run.sh rejects non-numeric/zero values early.

Where results land

<bench.jobs_dir>/<slug>/<timestamp>/ (bench jobs_dir defaults to results). A run dir contains config.json, result.json, job.log, and per-trial dirs (patches, verifier output, trajectories). With record_full_io, per-trial agent/requests.jsonl is populated after the run.

Guardrails

• Never kill proxies by port/pattern — run.sh stops only the job-private proxy it started (verified PID + config). Leave shared proxies alone.
• Do not launch the real run without user confirmation.

When the run finishes, hand off to analysis — see 06-analysis.md.