Zen: muse-spark-1.2-contributor-free requires /responses not /chat/completions (500) and has text->tool chain stall #44659

Summary

muse-spark-1.2-contributor-free via Zen POST /zen/v1/chat/completions returns 500 Internal Server Error on all clients, but POST /zen/v1/responses returns 200 and works. Indicates model was moved to Responses API without docs/client update.

After using responses, a second model bug: text (only) -> tool -> text 3-step prompt stalls after step 1. Model says "Next I'll run a command" but finish_reason:stop with 0 function_call, loop exits. Same prompt via mimo-v2.5-free (chat) works, and spark with No text before tool works.

Repro

curl -s https://opencode.ai/zen/v1/chat/completions -H "Authorization: Bearer $ZEN_KEY" -H "User-Agent: opencode/1.0.0 (linux; x64)" -H "HTTP-Referer: https://opencode.ai/" -H "X-Title: opencode" -H "x-opencode-client: cli" -d '{"model":"muse-spark-1.2-contributor-free","messages":[{"role":"user","content":"hi"}],"stream":false}'
# 500 {"type":"error","error":{"type":"error","message":"Internal server error"}}

curl -s https://opencode.ai/zen/v1/responses -H ... -d '{"model":"muse-spark-1.2-contributor-free","input":"hi","stream":false}'
# 200 OK with output_text "Hey there! ..."

# Tool chain via opencode (also Hermes codex_responses)
opencode run "first send me a msg (only text) then run a cmd (any, like echo hello) then send another msg (text) to me" --model opencode/muse-spark-1.2-contributor-free
# -> Step 1 text only, then stops, no tool execution (Build 6.3s then idle)
# vs mimo: executes echo and finishes steps 1-3

Evidence

• Direct curl above (same key, same headers) proves routing.
• Hermes opencode_model_api_mode() and ~/.reasonix both assume chat for spark, so all harnesses hit 500 until patched to codex_responses.
• After patch, responses tool chain still stalls for forced only text first turn. Hermes repro: AIAgent(api_mode=codex_responses) 1 call vs 2 for mimo.

Suggested

• Either revert spark to chat_completions or update Zen docs/model catalog to advertise responses and fix tool training to allow text -> tool interleave (or document text+tool same-turn pattern).
• If kept on responses, ensure Hermes clients map muse-spark* to codex_responses (currently hermes_cli/models.py maps only gpt-*).

Env

opencode 1.18.21, https://opencode.ai/zen/v1, muse-spark-1.2-contributor-free