v2026.04.25 (v1.14.24)

What's New

• Per-agent maxOutputTokens override: Agent frontmatter can now cap model output tokens per agent (e.g. maxOutputTokens: 32768 on a reviewer agent), with safe fallback to the model's limit. Fixes silent swallow when set via agent frontmatter (config schema unknown-key fallback).

Fixes

• Provider: resolve useMaxCompletionTokens per-request, not per-SDK — the transform was keyed on the model captured when the SDK was first instantiated, but a single SDK instance is shared across all models under the same provider config. Mixed gateways (e.g. Anthropic + OpenAI reasoning models behind one endpoint) now correctly apply per-model overrides. Anthropic models can opt into max_tokens (which they require) via options.useMaxCompletionTokens: false to avoid a silent ~2K truncation.
• Provider/tool: resolve model alias for useMaxCompletionTokens and isGptPatch — when a config key differs from the SDK-facing api.id (aliased models), both the fetch wrapper and the tool registry now resolve via direct lookup → unique api.id match → provider-level fallback. Multiple aliases sharing the same api.id no longer apply an arbitrary override silently.
• Reviewer: resolve narration rule contradiction in .opencode/agent/reviewer.md — phase 1/2 emit guidance was self-contradictory across sections.
• Upstream rebase to v1.14.24.

Install

Try without installing:

npx opencode-ai-nofc

Install globally:

npm install -g opencode-ai-nofc
opencode

Or download binaries below.

Based on upstream opencode v1.14.24.