Compute: disable GLM thinking for skill outputs

[Franlinozz]

Summary

• Fix production skill output still rendering [object Object] by normalizing the backend compute response path, not the already-correct modal renderer.
• Add chat_template_kwargs: { enable_thinking: false } to non-streaming 0G Compute chat requests so GLM returns final message.content instead of reasoning-only payloads.
• Record the production response-shape decision in docs/DECISIONS.md.

Production evidence

Authenticated production invoke for Novaris Agent #27 returned:

{
  "skillId": "text.summarize",
  "output": { "summary": "[object Object]" },
  "compute": {
    "chatId": "7d7c7f61-6644-4684-a5bc-19a1389f4784",
    "model": "zai-org/GLM-5-FP8",
    "provider": "0xd9966e13a6026Fcca4b13E7ff95c94DE268C471C"
  },
  "receipt": {
    "txHash": "0x47b63c7735ef98aa1e56f0d9490d58af67e69795f5f0d32642307b213bc30347",
    "status": "minted"
  }
}

Raw provider capture showed GLM thinking was enabled by default:

• message.content: null
• message.reasoning: "..."
• finish_reason: "length"

So Edge shaped String(response) into [object Object] before the frontend saw it. The modal was reading the right field path: result.output.summary.

Verification

• pnpm -F @apogee/compute-client typecheck
• pnpm -F @apogee/compute-client build
• pnpm -F @apogee/compute-client test
• pnpm -F @apogee/edge typecheck
• pnpm -F @apogee/edge build
• pnpm -F @apogee/web test
• Production-like local ComputeClient call using Railway env and patched package returned real content: Apogee is a 0G autonomous-agent runtime for deploying agents, running skills, and minting on-chain proofs of useful work.

Scope

No changes to receipt minting, ownership checks, policy enforcement, status gates, auto-promote logic, Pilot endpoint, Aurora/Vesper/Helix, or heartbeat runtime.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
apogee	Ready	Preview, Comment	May 19, 2026 10:32am