OpenCode hooks

If stagelint is installed in your project or through a tool manager, invoke it the way that tool does, such as npx stagelint, uv run stagelint or mise exec -- stagelint.

OpenCode takes a plugin rather than a config entry. Put this in .opencode/plugins/stagelint.js:

.opencode/plugins/stagelint.js

const maxAttempts = 5
const attempts = new Map()

export const Stagelint = async ({ $, client }) => ({
  event: async ({ event }) => {
    if (event.type === 'session.deleted') {
      attempts.delete(event.properties.info.id)
      return
    }

    if (event.type !== 'session.status' || event.properties.status.type !== 'idle') {
      return
    }

    const id = event.properties.sessionID
    const { data } = await client.session.get({ path: { id } })
    if (data?.parentID) {
      return
    }

    const res = await $`stagelint --unstaged --quiet`.nothrow()
    if (res.exitCode === 0) {
      attempts.delete(id)
      return
    }

    const n = (attempts.get(id) ?? 0) + 1
    if (n > maxAttempts) {
      return
    }
    attempts.set(id, n)

    await client.session.prompt({
      path: { id },
      body: { parts: [{ type: 'text', text: res.stderr.toString().trim() }] },
    })
  },
})

What each part does

The plugin fires on session.status when the status turns idle, which is when the agent loop ends. Subagents run in their own sessions that go idle too, so the plugin looks the session up and skips any that has a parentID. A subagent has already handed its result back to the parent, and the parent session goes idle afterwards anyway.

--unstaged runs your commands against the working tree rather than the index, because the agent's edits are not staged yet. Nothing is stashed and nothing is staged. --quiet prints only the output of failed commands, which is what gets sent back as the next message.

$ is Bun's shell, which throws on a non-zero exit. That is exactly what a failing lint run produces, hence .nothrow(). Without it the plugin throws before it can report anything.

OpenCode has no loop limit of its own. Each prompt starts a new turn, and every turn ends idle, so a failure the agent cannot fix would prompt it forever. The plugin counts prompts per session and stops after maxAttempts until the checks pass again.