Repository navigation
feat(roleplay): start the intelligence chain on OpenCode Go - #91
Conversation
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
|
Important
This repository does not receive automatic reviews because it has fewer than 10 stars.
 |
Roleplay now tries DeepSeek V4.1 Flash, GLM-5.3 and MiMo-V2.6-Pro on OpenCode Go before GLM-5.3 on ClinePass; the rest of the chain is unchanged. In 24 test turns on Go (scene continuation, conflict, out-of-character recall and a 20k-token lorebook lookup) all three answered every turn and kept to the card. DeepSeek V4.1 Flash was fastest (median 5.2 s), wrote nearly as well as GLM-5.3 and used about a thirtieth of its allowance per turn, so it leads. GLM-5.3 wrote the strongest scenes (8.3 s); MiMo-V2.6-Pro was slowest (13.5 s) with the shortest replies. A rejected Go request falls through to the next entry.
a0dcb65 to
142fa66
Compare
Roleplay's
roleplay:intelligencechain (androleplay:auto, since production routes it there) now tries three OpenCode Go models before GLM-5.3 on ClinePass:opencode:deepseek-v4.1-flashopencode:glm-5.3opencode:mimo-v2.6-proThe rest of the chain is unchanged; ClinePass still lists
cline-pass/mimo-v2.6-proin itsclinePassrecommended models, so that entry stays. This is awrangler.jsoncvar change, plus the config test anddocs/roleplay.md.Measurements (2026-10-25, OpenCode Go, 24 streamed turns)
Four scenes: continuation, conflict, out-of-character recall with a word cap, and a 20k-token lorebook lookup. Every turn returned 200, kept to the card, and passed recall and lookup. No
<think>leaked.DeepSeek wrote nearly as well as GLM, was fastest and is the cheapest by far, so it leads. GLM-5.3 wrote the strongest scenes. MiMo gave the shortest replies and had some point-of-view slips. Go meters each model against its own monthly allowance.
Checks
node --testpasses for the roleplay config, intelligence, worker, deploy-pipeline, Cloudflare-worker and runtime-flag suites.wrangler devboots with the new chain and tries the Go entries in order, falling through on each rejection. From a local Worker, OpenCode answers with Cloudflare error 1010, the Worker-signature block. Production avoids that by sending OpenCode through the container, so the first production turn after deploy is the real end-to-end check.OPENCODE_GO_API_KEYsecret has been replaced with a working key.If the first production turn still gets 403 from OpenCode, remove the three entries; that is a var-only revert.