| name | geobench-run-aiterator |
| description | Run GeoBench with aiterator headless exec mode. Use when the user asks to run aiterator, aiter exec, or /geobench-run-aiterator. Shared reset, isolation, grade, and import live in the geobench-run skill.
|
GeoBench via aiterator
Read geobench-run first. --harness is aiterator.
This harness is the exception to the container default: aiterator is a local Go binary that is not installed in the geobench-agent image, so it runs host-native. Note that in the run notes.
Invoke
Set AITERATOR_BIN to the host path of the executable. Rebuild it from the aiterator checkout with go build -o "$AITERATOR_BIN" ./cmd. --cd sets the task workspace;
the prompt is @ plus the instruction path — do not paste instruction.md
onto the command line.
"${AITERATOR_BIN:?Set AITERATOR_BIN}" exec --cd "$workspace_win" --model "<model-id>" \
--effort low --full-access --json --timeout 3600 \
--history "$logdir_win/<task-id>.jsonl" "@$instruction_win"
From Git Bash, --cd, --history, and the @ instruction path all want
Windows paths (cygpath -w).
--full-access is the unattended mode: approvals off, no workspace
confinement. Exit code is non-zero on provider error or timeout; continue the
remaining tasks.
--history writes the per-task session JSONL into the run's
results/agent-logs/... directory for provenance. --json events on stdout
can be captured to the same directory.
--effort is low or high. Record it in --notes (effort high). Default
low if the user did not name an effort.
Models
--model must be a canonical id or provider slug routed in
provider_routing.toml (repo or ~/.aiterator/provider_routing.toml).
Currently routed via openrouter on this host:
--model | Notes |
|---|
qwen/qwen3.6-plus | default model in the aiterator repo config (openrouter) |
kimi-k2.6 | routes to moonshotai/kimi-k2.6 (openrouter) |
gpt-5.5 | routes to the codex provider using the stored Codex OAuth credential; works even though provider_route pins openrouter (the pin is a preference, not a filter) |
Grade --model as the same id passed to --model.
Pitfalls
OPENAI_API_KEY must be set in the shell (exported in ~/.bashrc for Git
Bash; PowerShell sessions need it set explicitly) — without it exec fails
fast with a credential error naming the missing variable (exit 1).- Model ids must be routed in
provider_routing.toml; unrouted ids fail fast
in exec unless plain OPENAI_API_KEY/OPENAI_BASE_URL env semantics
apply.
- Pasting instruction text as argv → quoting breakage. Use
@$instruction.
--full-access has no workspace confinement and runs outside the container:
the isolation caveats in geobench-run apply in full.
