require "opencode-ruby"

client  = Opencode::Client.new(base_url: "http://localhost:4096")
session = client.create_session(title: "My session")

reply = client.stream(session[:id], "Explain monads in two sentences.") do |part|
  print part["content"] if part["type"] == "text"
end

puts
puts reply.full_text
puts "(#{reply.tool_parts.size} tool calls, #{reply.parts_json.size} parts total)"

Three lines of setup, four lines of work. Block fires every time a part appears, grows, finalizes, or (for tool calls) advances state. The final return value is a typed Opencode::Reply::Result you can persist or inspect.

Install

# Gemfile
gem "opencode-ruby"

Or:

gem install opencode-ruby

Then require "opencode-ruby".

Configuration

client = Opencode::Client.new(
  base_url: "http://localhost:4096",   # or ENV["OPENCODE_BASE_URL"]
  password: "secret",                   # or ENV["OPENCODE_SERVER_PASSWORD"]
  timeout:  120                         # or ENV["OPENCODE_TIMEOUT"], seconds
)

Multi-tenant apps construct multiple clients with different base_urls — each Opencode::Client holds its own Net::HTTP connection, no shared state.

Core API

Configured and parent-linked sessions

OpenCode can create a session under an existing parent and select its agent, model, metadata, workspace, and permission policy in the same request:

child = client.create_session(
  title: "Destination curator",
  parent_id: parent_session_id,
  agent: "destination-list-curator",
  model: "openai/gpt-5.5",
  metadata: { run_id: "9" },
  workspace_id: workspace_id,
  permissions: permission_rules
)

Model strings use OpenCode's provider/model form; a preformatted model hash with providerID and id keys is also accepted. These configured-session fields first appeared in OpenCode 1.16.1. That is an API-introduction note, not evidence that every later server is compatible; use the certified compatibility evidence for deployment choices.

Streaming (the headline)

reply = client.stream(session_id, "What's 2 + 2?") do |part|
  case part["type"]
  when "text"      then print part["content"]
  when "reasoning" then # ignore, or render in a separate UI
  when "tool"      then puts "  [tool: #{part['tool']} → #{part['status']}]"
  end
end

reply.full_text       # => "2 + 2 = 4."
reply.tool_parts      # => array of terminal tool-call parts
reply.reasoning_text  # => the model's hidden reasoning, if any
reply.parts_json      # => the full ordered parts array, ready for persistence

stream waits for OpenCode's initial server.connected SSE readiness frame before it submits the asynchronous prompt. If the event connection drops afterward, the client reconnects only the subscription; it never reposts the prompt. This prevents both missed fast responses and duplicate turns.

Synchronous send (no streaming)

result = client.send_message(session_id, "Quick yes/no: is Ruby fun?")
# result is the OpenCode response hash; see API docs for fields.

Updating session permissions

client.update_session(session_id, permissions: permission_rules)

OpenCode appends PATCHed permission rules and evaluates the last matching rule. Hosts should send a complete ordered policy and fingerprint it so the same policy is not appended on every turn. This endpoint first appeared in OpenCode 1.16.1; that version floor describes endpoint availability, not compatibility with every later server.

Lower-level event firehose

If you need raw SSE events (every server tick, todo update, prompt asked/replied), use stream_events directly:

client.stream_events(session_id: session_id) do |event|
  puts event[:type] # "message.part.delta", "todo.updated", "session.status", ...
end

Orchestrators that submit their own async prompt must do so through on_subscribed; the callback runs only after the first server.connected frame and at most once across automatic reconnects:

client.stream_events(
  session_id: session_id,
  on_subscribed: -> { client.send_message_async(session_id, prompt) }
) do |event|
  reply.apply(event)
end

If the callback raises, stream_events propagates that error and does not retry it. This is intentional: a timed-out prompt response is ambiguous, so reposting could duplicate the model turn and its cost.

Interactive prompts

When the agent uses the question or permission tools, opencode emits question.asked / permission.asked events. Answer them via:

client.reply_question(request_id: "que_...", answers: [["yes"]])
client.reply_permission(request_id: "per_...", reply: "always")

Error model

Every method that hits the network raises Opencode::Error (or a subclass) on failure. Catch the parent or the specific subclass:

begin
  client.health
rescue Opencode::ConnectionError      # server unreachable
rescue Opencode::TimeoutError         # client-side timeout
rescue Opencode::SessionNotFoundError # 404 on a session
rescue Opencode::StaleSessionError    # no session event arrived after the prompt
rescue Opencode::IdleStreamError      # mid-turn SSE wedge
rescue Opencode::ServerError          # 5xx
rescue Opencode::BadRequestError      # 4xx other than 404
rescue Opencode::Error                # catch-all
end

Instrumentation

Want to see what the gem is doing? Plug in an adapter. Default behaviour is silent no-op — the gem ships zero opinion about your observability stack.

# stdout for debugging:
Opencode::Instrumentation.adapter = ->(name, payload, &block) {
  puts "[#{name}] #{payload.inspect}"
  block.call
}

# ActiveSupport::Notifications in a Rails app:
Opencode::Instrumentation.adapter = ->(name, payload, &block) {
  ActiveSupport::Notifications.instrument(name, payload, &block)
}

Event names emitted today:

Event	Payload
`opencode.request`	`:method`, `:path`

Want this in a Rails app?

See examples/conversation_recipe.rb for a ~60-line plain-ActiveRecord blueprint covering session lifecycle (with_lock, update_columns mid-stream snapshots, CAS-safe finalize). Drop it into your app and adapt.

If enough Rails developers do that and want it as a one-liner, we'll ship opencode-rails with acts_as_opencode_session. File an issue if that's you — your issue is the signal.

Position against `opencode_client`

Want every OpenCode endpoint auto-generated from the OpenAPI spec? Use opencode_client. This gem is the hand-rolled idiomatic alternative — smaller surface, opinionated defaults, block-form streaming. Pick whichever fits how you want to write Ruby.

Compatibility

Ruby ≥ 3.2
Runtime dependency: activesupport (>= 6.1) — not Rails. ActiveSupport is a standalone helpers gem (blank?, present?, presence, truncate, etc.).

OpenCode server compatibility is evidence-based, not an open-ended SemVer promise. OpenCode's HTTP, SSE, and runtime behavior can change independently of a numeric version floor, so this gem does not infer compatibility from a constraint such as >= 1.15.

Use the OpenCode compatibility corpus as the source of current evidence. Its image matrix records the exact gem commit and immutable OpenCode OCI digest exercised by the ruby-rest-sse profile. Its runtime tuples add the exact consumer commit and canary evidence needed for promotion and rollback. A version label or image tag is provenance only, never the execution coordinate.

A combination absent from passing evidence is unverified, not necessarily incompatible. Run the corpus against the exact gem, image digest, and consumer commit before adopting or promoting it. See the certification policy for the full process.

Development

bundle install
bundle exec rake test

The smoke suite covers Client end-to-end against WebMock-stubbed OpenCode endpoints, including subscription-before-prompt ordering and reconnect-without-repost.

The repository contains a tag-triggered release.yml workflow intended for RubyGems trusted publishing, but its RubyGems trusted-publisher registration is not configured as of 0.0.1.alpha7; that release was published manually. A v* tag push therefore does not currently guarantee publication. Before a future release, verify the RubyGems registry result explicitly. Once the workflow is registered as a trusted publisher for the release environment, it can build, attest, and publish without a long-lived RubyGems API key.

License

MIT. See LICENSE.

README.md

opencode-ruby