opencode-action

Run an OpenCode agent from GitHub issue and pull request comments.

Quick start

1. Add a provider secret

In Settings → Secrets and variables → Actions, add the API key for your model provider. The example below uses OPENCODE_API_KEY.

2. Add the workflow

Create .github/workflows/opencode.yml:

---
name: OpenCodeon:
issue_comment:
types: [created]pull_request_review_comment:
types: [created]permissions:
contents: writeissues: writepull-requests: writeid-token: writejobs:
opencode:
if: contains(github.event.comment.body, '/oc') || contains(github.event.comment.body, '/opencode')runs-on: ubuntu-lateststeps:
- name: Checkout repositoryuses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1with:
persist-credentials: false
- name: Run OpenCodeuses: dceoy/opencode-action@e5a3599d8b64f9adda6a03a18c8b0b9d077ffebb # v0.7.0env:
OPENCODE_API_KEY: ${{ secrets.OPENCODE_API_KEY }}GITHUB_TOKEN: ${{ github.token }}with:
model: opencode-go/kimi-k3

3. Comment on an issue or pull request

/opencode explain this issue

The shorter /oc trigger also works:

/oc fix this

The default setup exchanges the workflow OIDC token for an OpenCode GitHub App token, which requires id-token: write.

Reusable workflows

For smaller caller workflows, this repository provides reusable workflows for the mention bot and pull request reviews:

Workflow	Purpose
opencode-bot.yml	Run OpenCode from trusted issue or pull request comments, or from a fixed prompt.
opencode-review.yml	Run the bundled /review-pr flow for pull_request events.

See Reusable workflows for caller examples, inputs, secrets, and permission requirements.

Models and secrets

Set model to a provider/model value and pass the corresponding API key:

Provider	Example model	Secret
OpenCode	opencode-go/kimi-k3	OPENCODE_API_KEY
OpenRouter	openrouter/openrouter/free	OPENROUTER_API_KEY
Anthropic	anthropic/claude-opus-5	ANTHROPIC_API_KEY
OpenAI	openai/gpt-5.6-sol	OPENAI_API_KEY
Custom provider	myprovider/my-model	Provider-specific

The provider account must have sufficient credits or quota. For providers not built into OpenCode, see Custom providers.

Sakura AI Engine model synchronization

The optional Sakura model synchronization workflow discovers chat-capable Sakura AI Engine models and opens or updates a pull request when the catalog changes. Configure the SAKURA_AI_ENGINE_API_KEY repository secret before enabling it.

The workflow uses the repository-provided GITHUB_TOKEN with contents: write and pull-requests: write; no additional GitHub token secret is required. GitHub does not create new workflow runs for events triggered by GITHUB_TOKEN, so pull requests created by this workflow do not automatically trigger pull_request workflows. See GitHub's GITHUB_TOKEN documentation.

Inputs

Input	Default	Description
model	Required	Model in provider/model format.
agent	build	Primary agent. A slash command can override it.
prompt	Event comment	Fixed prompt to use instead of the triggering comment.
mentions	/opencode,/oc	Comma-separated trigger phrases.
variant	-	Provider-specific reasoning effort; leave empty unless supported. See Custom providers.
share	false	Share the OpenCode session.
use-github-token	false	Use the workflow token instead of the default App-token flow.
opencode-version	latest	OpenCode version to install. /review-pr requires 1.2.14+; the bundled Sakura provider's chunkTimeout needs 1.2.25+ (older pins fall back to the request timeout).
use-bundled-toolkit	true	Use the bundled agents, commands, skills, and configuration.
timeout-minutes	60	Stop OpenCode after this many minutes.
oidc-base-url	https://api.opencode.ai	OIDC exchange URL for a custom GitHub App installation.

When use-github-token: true, keep GITHUB_TOKEN in env and grant only the permissions needed for the task.

Outputs are opencode-version and cache-hit. cache-hit is empty on review-only runs (prompt: /review-pr), which always skip the cache and install fresh.

Pull request reviews

Set prompt: /review-pr to run the bundled read-only review through a dedicated permission-constrained primary agent. The command loads the internal pr-review skill, which builds a change/risk map, dispatches fresh read-only child sessions, independently validates candidate findings, and posts confirmed findings inline when they can be anchored to changed lines. Use /review-pr rather than loading pr-review directly when the enforced read-only boundary is required.

An unscoped review creates a small set of dynamic, risk-driven discovery tasks instead of routing to fixed specialist agents. Explicit aspects such as security, tests, docs, performance, or simplify constrain the selected review lenses. Discovery and validation run in separate fresh child sessions; the current OpenCode v1-compatible implementation uses one hidden review-worker definition for those sessions.

See Pull request reviews for setup, supported review aspects, submission behavior, and security guarantees.