ldgr-programbench

ldgr-programbench prepares and runs a real, bounded ProgramBench attempt with the harness selected in the user's LDGR configuration. It creates a usable workspace, invokes the harness through agentctl, packages the candidate, evaluates it with ProgramBench, and retains the attempt as LDGR evidence.

It is an on-host, validator-visible demonstration. It is not an official ProgramBench submission, a clean-room run, an independent evaluation, or a score suitable for general model ranking.

Quick start

Install the signed adapter, enter the directory that should own the run, and use the defaults:

ldgr adapter install programbench
mkdir -p ~/repos/programbench
cd ~/repos/programbench

ldgr programbench setup
ldgr programbench reproduce
ldgr programbench verify
ldgr programbench report

reproduce defaults to:

• the current directory as --benchmark-root;
• sharkdp__hyperfine.327d5f4 as one bounded demonstration task;
• default_harness from ~/.ldgr/config.toml or ~/.ldgr/config.json;
• ./programbench-runs/<timestamp> for output;
• ProgramBench evaluation after the harness returns.

The singular --benchmark-root spelling is canonical. The plural --benchmarks-root spelling is an accepted compatibility alias:

ldgr programbench reproduce --benchmark-root ~/repos/programbench
ldgr programbench reproduce --benchmarks-root ~/repos/programbench

Use --harness codex (or another harness already selected in the user config) to override the configured default. Use --dry-run to write and display the complete run plan without downloading task images or invoking the harness.

Command responsibilities

• setup checks Docker, uvx, agentctl, and the selected harness; installs ProgramBench through uvx; pulls the task image; and writes local setup files.
• reproduce performs a new attempt. It does not require a previous submission, evaluation artifact, or historical archive.
• verify validates and summarizes the newest local evaluation result, or the run selected with --results.
• report renders the same evidence with the interpretation boundary attached.
• custody is the explicitly separate legacy command for checking retained historical artifacts.

See docs/programbench-run.md for the installed run guide. Historical classifications remain available under docs/historical/, but they no longer define or block the normal reproduction workflow.

Numerical sequence protocol

When LDGR Core numerical telemetry is explicitly enabled, the reproduction command queues only the ProgramBench reproduction state machine /sequences/programbench-reproduction/v1 through Core. Adapter states represent inputs prepared, reproduction prepared, attempt recorded, and evidence finalized. Benchmark names, paths, commands, logs, prompts, and result contents are not encoded or transmitted.