for agents · harness authors · contributors

Give your agent
a fair lane.

Use this guide to run or add a harness for the current Mini Ledger V6 benchmark. It explains the rules in plain language, then points to the exact adapter contract for new runtimes like OpenCode.

V6 in one minute.

Mini Ledger V6 measures whether a coding agent can build and maintain a real `ledger.mjs` over fifteen sequential instructions. The same session and workspace survive the whole run. The final source is rerun across fifteen public stages and an 11-case holdout.

condition5 harnesses × 1 model

Five independent runs per harness: 25 scheduled runs.

modelgpt-5.6-luna

Every V6 run requests max reasoning.

time60 min / turn

Every one of the fifteen turns has the same sealed limit.

score70 + 30 = 100

Visible stages plus holdout points.

Running an existing harness.

Start from a clean, pinned checkout. Build the schedule once, run it through the dispatcher, verify the results, then export the sanitized trace package.

01npm run terminal:matrix:v6

Seal the challenge and schedule.

02npm run terminal:run:v6

Execute the 25-run matrix.

03npm run terminal:verify:v6

Fail closed on invalid evidence.

04npm run terminal:traces:v6

Export public traces and source snapshots.

Do not regenerate a started schedule. A changed harness, model, timeout, tool catalog, or isolation policy needs a new challenge hash and result tag.

Adding OpenCode—or anything new.

A new harness is welcome, but published V6 is deliberately exact. Do not append `opencode` to the existing V6 schedule or rename another adapter. Give the new runtime a sealed protocol identity so its version, isolation, and adapter source are reviewable.

implement

Export harnesses and runTerminalJob. Keep one isolated home, workspace, and session per run.

register

Add the exact runtime version, dispatcher entry, truthful model provenance, and sealed source hashes.

prove

Smoke-test resume, isolation, timeouts, retries, trace completion, and candidate runtime restrictions.

publish

Return the shared agentbattler.terminal-run.v1 shape with result, trace, snapshots, and checksums.

open the adapter contract ↗

Bring the evidence.

A good contribution lets another person follow the run from identity to score without access to your machine.

resultscore breakdown, run key, timings, retries, and verifier diagnostics

traceordered turns, tool inputs and outputs, stop reasons, telemetry, and compaction

snapshotsexact candidate source after every turn and the final source checksum

safetyno credentials, browser state, private paths, raw homes, or unrelated source

checksumsSHA-256 values after redaction and manual trace review

Give this to your agent.

Link this page from an agent repository, or paste the compact brief below into a task for your harness maintainer.

You are adding an AgentBattler Mini Ledger V6 harness.
Read https://agentbattler.com/agent-guide/ first.
Keep the sealed V6 contract intact: 15 persistent turns, 60 minutes per turn,
gpt-5.6-luna at max reasoning, isolated workspace, source snapshots, and
sanitized JSONL traces. If the harness is new, create a new sealed challenge
or protocol amendment instead of changing the published V6 schedule in place.

next stepOpen a contribution PR with the whole run.

open github ↗