Skip to content

v0.48.0

Choose a tag to compare

@github-actions github-actions released this 25 Sep 11:20
· 286 commits to main since this release
35f1d4a

Headline: OpenCode as a third agent harness

Threads can now run on OpenCode alongside Claude Code and OpenAI Codex. OpenCode supports a wide range of model providers, so a single thread can use Amazon Bedrock, the Anthropic API, OpenAI, OpenRouter, or a local model. You keep the Agent Threads chat, permission cards, host tools, and harness switching.

Features

  • OpenCode harness (#615). Select OpenCode in Settings → Agent → Agent harness, pick it in the Agents List or Agent Board kickoff selector, or switch an idle thread to it from the footer Harness menu.
    • Each OpenCode thread runs its own local opencode serve (loopback only). Install opencode first. It is auto-detected on $PATH, Homebrew, /usr/local/bin, and ~/.opencode/bin, and the new OpenCode binary path setting covers other locations.
    • You set up providers, models, and credentials on the OpenCode side: opencode auth login, environment variables through Extra environment variables or Secrets (for example ANTHROPIC_API_KEY, OPENAI_API_KEY, AWS_PROFILE + AWS_REGION, or AWS_BEARER_TOKEN_BEDROCK for Bedrock), or ~/.config/opencode/opencode.json.
    • Model IDs use provider/model form (for example openai/gpt-5). The model menu lists what OpenCode reports.
    • Permission modes are applied as the session runs. Read-only tools always run. Accept edits auto-approves edits. Bypass/Auto approve everything. Don't ask denies anything that would prompt, so scheduled runs never stall. Plan uses OpenCode's read-only plan agent.
    • Agent Threads host tools reach OpenCode through a per-session loopback MCP endpoint and follow the same approval rules as Codex.
    • Resume uses the OpenCode session ID. If that session is gone, the thread history is replayed once.
    • Desktop only.

Improvements

  • Provider-neutral harness session contract (#614). Permission modes, MCP server config, and context usage now use harness-neutral types instead of Claude SDK shapes. This is what let a third adapter plug in cleanly. It also makes the Claude mapper tolerate partial SDK context-usage responses.

Notes

  • Billing: OpenCode uses your provider's API key or cloud credentials, and usage is billed by that provider. It does not use a Claude subscription. If you have Claude Pro or Max, keep using the Claude harness.
  • Known gaps in OpenCode threads: no plan-approval card (the plan agent answers in prose), no MCP elicitation, OpenCode task subagents are followed for permissions and questions but not shown as agent runs, no account quota data, and the effort / reasoning-effort setting is ignored.
  • Docs: OpenCode Harness · ADR: docs/adr/0013-opencode-harness.md