Every file here is an OrcaReplay capture — one harness, seven models, captured across 2026-09-02 and 09-03.
| model | chars | tools | prompt |
|---|---|---|---|
big-pickle |
9,622 | 11 | prompt |
ling-3.0-flash-fin-free |
9,648 | 11 | prompt |
mimo-v2.5-free |
9,630 | 11 | prompt |
nemotron-3-ultra-free |
9,644 | 11 | prompt |
nemotron-3.5-lightning-free |
9,656 | 11 | prompt |
muse-spark-1.2-contributor-free |
10,250 | 11 | prompt · tools |
gpt-5.6-sol |
10,334 | 9 | prompt · tools |
Seven models, three prompts. The first five differ from each other in exactly one line — You
are powered by the model named … — and send a byte-identical tool set, which is why they share one
opencode-tools.json. Muse Spark gets a different opening and the responses
dialect. GPT-5.6-Sol gets a third template entirely, with apply_patch in place of edit and
write. The prompt and the wire format are both chosen per model, and you can diff the files here
to see it.
nemotron-3.5-lightning-free also appears under Hermes — same model, same free
endpoint, a different harness. That pair is the cleanest comparison in the archive.
node capture/capture.mjs opencode --model <id>Details in docs/CAPTURES.md.