chenyu/opencode/README.md

metadata

language:
  - en
tags:
  - robotics
  - mujoco
  - simulation
  - agent-generated

OpenCode MuJoCo generation outputs (evidential export)

Finished and cancelled generation sessions from Chenyu's OpenCode share of the MuJoCo campaign in BX126/dev_simreadybench. Fully separate from the Codex export under chenyu/codex/ and the Windows-Codex part, and from any handoffs.

Sessions ran on opencode-go/deepseek-v4.1-flash and opencode-go/glm-5.3-flash (OpenCode Go subscription); the preserved early sessions ran on the free-tier opencode/deepseek-v4.1-flash-free and opencode/glm-5.3-flash-free. Every row carries its own full model string in run.json and in manifest.jsonl; export_summary.json groups sessions by that exact real model name and never counts free sessions as Go sessions.

Each track/engine/item/run_id directory reconstructs the original run folder:

• output.tar.gz — the unchanged Agent-generated output directory (unpack → output/);
• session.tar.gz — everything else of the run folder: staged input/ (and tools/), the CLI event log generation.jsonl, the exact rendered prompt submitted_prompt.txt, token_usage.json when present, evaluator_protocol.json, and the untrimmed original run.json (contract name, sha256, protocol hash). Unpacking both archives side by side reproduces the run folder verbatim;
• run.json, files.json (output file hashes and symlinks), session_files.json (session-archive file hashes and symlinks).

manifest.jsonl records per session: run id, full model name, budget, execution status, both archives' byte sizes and sha256, and whether the row is a normal finished session or a cancelled attempt (kept and labelled). export_summary.json groups them by model. The campaign's state and configuration are snapshotted per publish batch under campaign/snapshots/<utc>/ (also as the live campaign/{state,config}.json); each snapshot includes batch_runs.jsonl, the exact sessions that batch covers. Historical snapshots are kept.

A budget_expired or cli_failed status exposes the enforced end or cancellation of a generation session — none of these is benchmark success and none is an independent quality PASS. Independent evaluation is not included in this export. Failed self-checks, refinement attempts and checkpoints inside generated output are preserved as they were made.

Nothing from outside the run folders is uploaded: no container credentials or temporary auth caches, no model keys, no benchmark per-item ground truth, no export preparation or scratch files. The original local run folders remain unchanged.