geobench-run-codex

name	geobench-run-codex
description	Run GeoBench with Codex CLI. Use when the user asks to run Codex, gpt-5.6-luna, gpt-5.6-sol, gpt-5.6-terra, or /geobench-run-codex. Shared reset, isolation, grade, and import live in the geobench-run skill.

GeoBench via Codex CLI

Read geobench-run first. --harness is codex.

Invoke

Runs in the geobench-agent container. Prompt on stdin: - after exec is the stdin slot, and docker run -i forwards the pipe into the container.

export MSYS_NO_PATHCONV=1
prompt="$(<"$root/tasks/$task/instruction.md")"
printf '%s' "$prompt" | docker run --rm -i -v geobench-agent-home:/home/agent \
  --security-opt seccomp=unconfined \
  -v "$(cygpath -w "$root/tasks/$task/workspace"):/work" \
  -v "$(cygpath -w "$root/tasks/$task/instruction.md"):/instruction.md:ro" \
  -w /work geobench-agent \
  codex exec - --cd /work --model "<model-id>" \
    -c 'model_reasoning_effort="high"' \
    --approve-for-me --skip-git-repo-check --ephemeral --json

--approve-for-me is the unattended workspace-write mode. Do not add --sandbox; the container is the sandbox.

On the Windows Docker Desktop host, seccomp=unconfined is required because Codex's bundled bwrap sandbox otherwise cannot create its user namespace. This is a runtime option only; the container remains the unprivileged agent user and sees only the mounted task workspace, instruction, and geobench-agent-home. Use --rm and one task container at a time so completed containers are pruned while the login volume remains available.

If the outer session, token budget, or provider session stops, keep the run directory and its timing/log files as provenance, remove only the exact stale task container, regenerate the tasks, and rerun this same model×harness combo. Do not use its partial workspaces for another model; a valid queue checkpoint is the matching finished verify artifact. The shared launcher detects provider rate_limited responses and aborts the suite checkpoint without spending retries on every remaining task.

Models

Ids that have completed a suite on this host:

Model	Effort	--notes
gpt-5.6-luna	max or high	effort max / effort high
gpt-5.6-sol	high	effort high
gpt-5.6-terra	high	effort high

Pass the effort the user named. Do not inherit a leftover max when testing a smaller model.

Pitfalls

• Missing - after exec → Codex does not take the piped prompt.
• Missing -i on docker run → the pipe never reaches Codex, empty run.
• Combining --approve-for-me with --sandbox → rejected by this Codex version.