Opencode Drive

OpenCode Drive

Use opencode-drive to launch an isolated OpenCode instance and control its TUI and simulated LLM.

Default to a one-shot Effect program. Use defineScript for named, visible, restartable, or manual-launch workflows; it is also Effect-only. Use live commands only for interactive development against a persistent or visible instance.

Catalog State IDs

Browse and copy OpenCode terminal state IDs from:

https://dev.opencode.ai/lab/catalog

Replayable flow states expose canonical <flow-id>/<state-id> addresses, for example:

patch-success-lifecycle/permission-prompt

Reproduce one from an opencode-drive source checkout:

bun run catalog:reproduce -- patch-success-lifecycle/permission-prompt \
  --opencode /path/to/opencode \
  --output /tmp/permission-prompt.frame.json

The command executes the registered recipe only through that checkpoint and writes an opencode-terminal-frame-v1 artifact. Only flows in apps/catalog/scenarios/index.ts are replayable. Browse-only flows and screen cards copy standalone capture IDs instead; do not invent a flow prefix. Use a protocol-compatible OpenCode checkout, ideally the source revision shown by the selected capture set.

To compare a committed local OpenCode branch against current v2 across every catalog state:

bun run catalog:capture -- \
  --opencode /path/to/opencode \
  --revision origin/v2 \
  --revision HEAD

Add repeated --theme flags to capture the same commit pair under multiple themes. Capture resolves detached immutable worktrees, retains earlier sets, and sorts sets by commit time. Uncommitted changes are excluded by design.

Effect Programs

Write drive.ts as a default-exported, fully provided Effect, then run it directly:

import { Effect } from "effect"
import { Llm, OpenCodeDriver } from "opencode-drive"

export default OpenCodeDriver.use(
  {
    project: {
      git: true,
      files: {
        "src/value.ts": "export const value = 1\n",
      },
    },
  },
  ({ ui, llm }) =>
    Effect.gen(function* () {
      yield* llm.queue(Llm.text("The value is 1."))
      yield* ui.submit("Read src/value.ts")
      yield* ui.waitFor("The value is 1.")
      yield* ui.screenshot("result")
    }),
)

opencode-drive run ./drive.ts

run type-checks the module before importing it and requires its default export to be an Effect<unknown, unknown, never>. It accepts exactly one module path; it does not accept --command.* flags or application arguments after --.

OpenCodeDriver.use is the normal lifecycle boundary. It creates an isolated project, starts the server and primary TUI, races the program against backend failure, settles queued LLM work, closes all TUIs, exports recordings, and removes the artifact directory unless keepArtifacts: true is set. Settlement failures fail the program.

Use OpenCodeDriver.make only when explicit settlement is necessary. It requires a scope, and the program must call driver.settle() before leaving that scope.

Deterministic Project DSL

Declare the project, semantic OpenCode configuration, and TUI configuration in OpenCodeDriver.use options:

export default OpenCodeDriver.use(
  {
    project: {
      git: true,
      files: { "README.md": "# Fixture\n" },
    },
    config: {
      autoupdate: false,
      username: "Drive",
    },
    tuiConfig: {
      theme: { name: "opencode", mode: "dark" },
      scroll: { speed: 1 },
    },
    setup: ({ fs, config, tuiConfig }) =>
      Effect.gen(function* () {
        yield* fs.writeFile("src/setup.ts", "export const ready = true\n")
        config.username = "Setup wins"
        tuiConfig.scroll = { speed: 2 }
      }),
  },
  ({ ui }) => ui.screenshot("home"),
)

The DSL is applied in this order:

1. project.files is written into the isolated project.
2. config and tuiConfig are deeply merged over .opencode/opencode.jsonc and .opencode/cli.json fixture values. Objects merge recursively; arrays and scalar values replace existing values.
3. setup runs and may write project files or mutate the merged config and tuiConfig objects. Its mutations take final precedence.
4. Drive writes both configs as stable, formatted JSON. With project.git: true, it creates a repository and commits the complete pre-launch state with fixed Git identity and timestamps.

fs.writeFile is rooted inside the simulated project and creates parent directories. project.git: true refuses to replace existing Git metadata; omit it when prepared fixtures already include a repository.

tuiConfig contains current V2 CLI settings despite retaining its public option name. Drive pins OPENCODE_CONFIG_DIR to the isolated project's .opencode, so cli.json is the fixture's global terminal configuration. Do not write legacy tui.jsonc, or write cli.json manually in setup instead of mutating tuiConfig: the normalized object is written after setup finishes.

UI And LLM

UI operations are Effects:

• ui.submit(text) types and presses Enter.
• ui.state(), ui.capture(), and ui.matches(text) inspect the terminal. ui.capture() returns the raw terminal frame as data.
• ui.snapshot() returns the versioned semantic tree; ui.getNode(query, options?) polls for one exact semantic match.
• ui.waitFor(textOrPredicate, options?) polls until a match.
• ui.getElement(query, options?), ui.focus(...), and ui.click(...) target interactive elements.
• ui.screenshot(name?) asks OpenCode for the raw frame, saves it as a PNG inside Drive, and returns its absolute path. It requires no media-directory configuration.
• ui.resize({ cols, rows }), ui.press(...), and ui.arrow(...) control the TUI.
• ui.mouse({ action: "move", x, y }) sends real mouse movement at absolute, zero-based terminal cell coordinates. down/up accept button (left by default, middle, right); moves with a held button drag. scroll requires direction. Mouse modifiers are { shift, alt, ctrl }. Older endpoints without ui.mouse fail with UiCapabilityError.

Build deterministic simulated responses with the Llm namespace and schedule them through the driver's llm controller:

yield* llm.queue(
  Llm.reasoning("Checking the fixture"),
  Llm.pause(20),
  Llm.text("The value is 1.", { delay: 2, chunkSize: 15 }),
)

llm.queue(...) declares the next response without waiting. llm.send(...) waits for the next request and completes its response. For ongoing responses, the handler passed to llm.serve returns an Effect Stream; registering the handler is an Effect. Available outputs include text, reasoning, pause, toolCall, raw, finish, and disconnect; a normal response gets finish("stop") when no terminal output is supplied.

import { Stream } from "effect"
import { Llm } from "opencode-drive"

yield* llm.serve((_request, index) =>
  Stream.make(Llm.text(`Response ${index + 1}`)),
)

Use the capability names literally: opencode is the generated OpenCode SDK, tui is the primary frontend process, ui is tui.ui, and tuis launches additional frontend processes.

Additional TUIs share the server and LLM controller:

const secondary = yield* tuis.launch({
  viewport: { cols: 120, rows: 40 },
  recording: true,
})
yield* secondary.ui.screenshot("secondary")

Annotated Recordings

Set tui: { recording: true, pointerOverlay: true } for a minimal animated mouse cursor in exported videos. A configuration object accepts leadMs (180), lingerMs (700), motionMs (500), and curve (0.06). Travel uses a critically damped Motion spring, arriving exactly at input time along a bounded arc. Use curve: 0 for straight travel; held drags never add decorative curvature. Nearby inputs stay connected. This needs an OpenCode build advertising ui.recording.pointer; input timestamps and actual click coordinates come from its recording clock, not Drive-side estimates. Keep the *.pointers.jsonl sidecar with the terminal recording when copying it. exportRecording accepts the same pointerOverlay option.

The cursor animation never sends input or delays the script. To exercise hover along a path, send real intermediate ui.mouse moves. Use state-based waits for the hover result. Clips crop the already-composited raw animation (including approaches to later inputs); holds freeze both terminal pixels and the pointer. Visibility windows are bounded by retained recording time, not extra video time.

With tui: { recording: true }, label moments during the run and the exported MP4 gets a burned-in footer (segment label bottom-left, elapsed timecode and "drive" branding bottom-right):

yield* tui.recording.mark("typing a prompt — 600ms wire latency")
// ... drive the UI ...
yield* tui.recording.mark("") // clears the label
const video = yield* tui.recording.finish()

For programmatic exports, exportRecording accepts annotations (labelled raw-timeline instants), clips ({ fromMs, toMs, speed?, holdMs?, label? } segments that trim/re-speed/freeze, concatenated in order), and footer (false suppresses it; { brand } overrides the branding). See test/manual/tui-regressions/optimistic-create-demo.ts for a complete before/after demo recording script.

Pull Request Before/After

For a visible OpenCode change, default to a matched A/B recording rather than an after-only demo:

1. Write one deterministic Drive script and fixture. A temporary script is fine; do not fork separate before and after scripts.
2. Create an immutable worktree for the PR base and run the script against it with --dev, then run the same script against the change worktree. Keep the viewport, project files, config, simulated LLM output, and interactions identical. Parameterize only the annotation label when needed.
3. Use state-based waits to reach the comparison checkpoint. After transient overlays settle, call tui.recording.mark("BEFORE: …") or tui.recording.mark("AFTER: …") and hold each stable state for the same duration.
4. Trim from the marked checkpoints and concatenate equal-length clips with a shared timeline. Default to side-by-side playback so reviewers can compare both states simultaneously. Preserve dimensions and synchronize meaningful interactions; use a sequential hard cut only when side-by-side rendering makes terminal text illegible.
5. Inspect a frame from each segment before upload. The demo is complete only when both labels are visible, the regression is obvious in the before segment, the fix is obvious in the after segment, and the PR caption states which revision each segment runs plus any simulated behavior.

Network Chaos

Set network: true in defineScript to route every TUI through a chaos TCP proxy (the TUI is pinned with --server, so reconnects always cross the proxy). The Drive control plane and the opencode SDK stay clean; only the TUI's HTTP and SSE traffic degrades.

export default defineScript({
  network: true,
  run: ({ ui, network }) =>
    Effect.gen(function* () {
      yield* network.set({ latencyMs: 400, jitterMs: 200 })
      yield* network.set({ blackhole: true }) // buffer all bytes until clear
      yield* network.set({ refuseNew: true }) // refuse new connections
      yield* network.killConnections() // drop every open connection
      yield* network.clear() // heal; buffered bytes flush
    }),
})

A quiet blackhole does not raise the TUI's reconnect overlay; that needs a dropped connection (killConnections) while traffic is pending.

Script-Writing Rules Of Thumb

• Timeout severity is split: UiWaitTimeoutError (a waitFor/getElement/ getNode deadline passed) is catchable — branch on "did X appear in time?" with Effect.catchTag("UiWaitTimeoutError", ...). UiTimeoutError (an unanswered UI RPC) aborts the whole run even when caught.
• LLM request bodies carry the entire conversation. Route reply markers by which appears last in the serialized body (lastIndexOf), never by the first includes hit.
• The server projection (opencode.message.list) is the ground truth for whether a prompt landed. The screen and the instance's prompt-history.jsonl can both mislead.
• ctrl+c on an empty composer exits the TUI and kills the run (RpcClientDefect: connection closed); use ctrl+u to clear leftover composer text.
• Effect v4: it is Effect.catch, not Effect.catchAll.
• Isolated-instance state lives under $artifacts/home/...; screenshots land under the run's output/.../generation-N/ directory.
• Set OPENCODE_DRIVE_MEDIA_DIR=$PWD/.drive-output (gitignored) when running probes from an agent: the default media root sits under the system tmpdir with a per-run id in the path, so every run triggers a fresh outside-the-project permission prompt.
• Interrupting a busy session takes two escape presses within 5 seconds (the first arms, the second fires session.interrupt). A single ui.press("escape") is a no-op for interruption.
• server.kill() mid-stream abandons the in-flight served reply (a deliberate detach, not a run failure); server.launch() attaches the LLM stub to the replacement service. Pass OPENCODE_DRIVE_DB=... so both generations share a database.

Scripted TUIs always use an explicit connection to the script-owned server (or the chaos proxy). HTTP address and existing credentials stay stable across server generations, so retained TUIs and SDK clients can reconnect without electing a competing service. After server.kill(), call server.launch(); do not wait for the TUI to create a server on the script's behalf. This does not change non-scripted live launches' managed-service behavior.

Reports And Paths

OpenCodeDriver exports a branded AbsolutePath schema and a compact RunReport containing the artifact root, retention, recording paths, and endpoint compatibility. It also exports decodeAbsolutePath and decodeRunReport for validating unknown values.

driver.settle() returns the report with its recording paths. Use OpenCodeDriver.useReport(options, run) when a safe lifecycle program also needs the report alongside its result.

Drive prefers protocol negotiation and reports explicit legacy fallback. Set opencode.compatibility to "required" when protocol skew must fail before the program runs. Additional built-in tool adapters remain follow-ups.

Arbitrary Dynamic Tools

Use runtime tools.attach for tools that do not have a shipped Drive adapter. Attachment replaces the complete dynamic set without affecting configured static adapters. Take invocations by the model call ID, while Drive owns producer IDs, progress sequences, reconnect replay, and exactly-one terminal commitment.

yield* tools.attach({
  tools: [
    {
      name: "lookup",
      description: "Look up a value",
      inputSchema: {
        type: "object",
        properties: { query: { type: "string" } },
        required: ["query"],
      },
      options: { codemode: false },
    },
  ],
})

const lookup = yield* tools.take("call_lookup")
yield* lookup.progress({ phase: "searching" })
yield* lookup.finish({
  structured: { answer: 42 },
  content: [{ type: "text", text: "42" }],
})

Use awaitCancelled() to observe native interruption. Do not synthesize cancellation or expose transport sequence numbers. Dynamic effective names may not collide with configured shell, webfetch, or websearch adapters.

Simulated Shell Execution

Declare the built-in tools Drive should intercept, then control each invocation from the running program. Unregistered tools remain real. Calls may be accepted in arrival order or by the stable ID supplied in Llm.toolCall, so parallel invocations can progress and settle independently.

import { Effect } from "effect"
import { Llm, OpenCodeDriver } from "opencode-drive"

export default OpenCodeDriver.use({ tools: ["shell"] }, ({ tools, llm, ui }) =>
  Effect.gen(function* () {
    const shells = yield* tools.control("shell")
    yield* llm.queue(
      Llm.toolCall({
        index: 0,
        id: "call_shell",
        name: "shell",
        input: { command: "compile" },
      }),
      Llm.finish("tool-calls"),
    )
    yield* ui.submit("Compile the project")
    const shell = yield* shells.take("call_shell")
    yield* shell.progress(`Running ${shell.input.command}\n`)
    yield* shell.succeed({ output: "Controlled success\n", exit: 0 })
  }),
)

The same declaration and runtime tools capability are available in defineScript. Supported adapters are shell, webfetch, and websearch. Each progress value replaces the visible tool output, so send accumulated text when earlier lines should remain visible. Calls settle exactly once; awaitInterrupted() observes session interruption or transport disconnection.

The callback form remains available for fixed behavior that does not need runtime orchestration:

import { Effect } from "effect"
import { Tool } from "opencode-drive"

const tools = (registry: Tool.Registry) => {
  registry.handle("shell", ({ input, progress }) =>
    Effect.gen(function* () {
      yield* progress(`Running ${input.command}\n`)
      return { output: "Controlled success\n", exit: 0 }
    }),
  )
}

Foreground callback handler Effects are interrupted when OpenCode interrupts the session, the transport disconnects, or Drive shuts down. Detached background shell handlers continue after their launch response and are interrupted when Drive shuts down.

Effect Scripts

Use defineScript with start --script when the workflow must have a stable instance name, be visible, rerun on restart, or explicitly launch and kill its server and TUIs. setup and run return Effects. Operations on fs, ui, llm, tools, server, and tuis also return Effects; there is no Promise API or compatibility shim.

import { Effect } from "effect"
import { defineScript, Llm } from "opencode-drive"

export default defineScript({
  config: { autoupdate: false },
  tuiConfig: { theme: { name: "opencode", mode: "dark" } },
  project: {
    git: true,
    files: { "src/value.ts": "export const value = 1\n" },
  },
  run: ({ ui, llm }) =>
    Effect.gen(function* () {
      yield* llm.queue(Llm.text("The value is 1."))
      yield* ui.submit("Read src/value.ts")
      yield* ui.waitFor("The value is 1.")
    }),
})

Always type-check a script before starting it:

opencode-drive check ./drive.ts
opencode-drive start --name demo --script ./drive.ts

For a new script, run opencode-drive script init ./drive.ts once. It creates a canonical Effect-native starter and refuses to overwrite an existing file. check adds focused migration guidance when it finds Promise-style script callbacks.

The script DSL applies project, config, tuiConfig, and setup with the same deterministic ordering described above. Automatic scripts run again after opencode-drive restart --name demo.

Use launch: "manual" only when the workflow must control server and TUI restarts itself. In manual mode tui and ui are null; run server.launch() before tuis.launch(name). Only one server may run at a time, server.kill() permits relaunch, and a closed TUI name may be reused.

export default defineScript({
  launch: "manual",
  run: ({ server, tuis }) =>
    Effect.gen(function* () {
      yield* server.launch()
      const alice = yield* tuis.launch("alice", { recording: true })
      yield* alice.ui.screenshot("alice")
      yield* alice.close()
    }),
})

Cancellation uses Effect interruption. Interrupting the script or an operation's fiber interrupts in-flight work and runs scoped finalizers; do not introduce AbortSignal or Promise cancellation wrappers.

Prepare An Instance

Use init only when files must be copied into an isolated home or project before a named live or scripted instance starts:

artifacts=$(opencode-drive init --name demo)
cp -R ./fixtures/home/. "$artifacts/"
cp -R ./fixtures/project/. "$artifacts/files/"
opencode-drive start --name demo --dev ~/projects/opencode

The simulated project is under $artifacts/files. A later start --name demo reuses the prepared artifacts; otherwise start initializes them automatically.

Drive uses an in-memory OpenCode database by default. For a script that restarts the OpenCode service and must recover the same sessions, set OPENCODE_DRIVE_DB to a file-backed path. Relative paths resolve inside the isolated run's OpenCode data directory:

OPENCODE_DRIVE_DB=restart.sqlite \
  opencode-drive start --name restart-demo --script ./restart.ts

Live Interaction

Use live commands to inspect or iterate on a persistent instance. Headless start requires a unique --name; visible instances may omit it. Headless start detaches after the instance is ready, so do not add &. Always stop the instance when finished.

opencode-drive start --name demo

opencode-drive send --name demo \
  --command.ui.type '{"text":"Explain this project"}' \
  --command.ui.enter

opencode-drive send --name demo --command.ui.state
opencode-drive send --name demo --command.ui.capture
opencode-drive send --name demo --command.ui.screenshot
opencode-drive stop --name demo

Image commands have one simple distinction:

• --command.ui.capture prints the raw terminal frame as JSON.
• --command.ui.screenshot saves that frame as a PNG and prints its absolute path.

Drive chooses the PNG directory. Do not configure a media directory or expect the OpenCode endpoint to provide one. Run --command.ui.screenshot as its own send invocation when the path is needed on stdout; a multi-command batch reports only whether the batch succeeded.

send executes command flags from left to right. JSON-valued commands take one JSON argument. Supported commands are:

• --command.ui.type '{"text":"..."}'
• --command.ui.press '{"key":"p","modifiers":{"ctrl":true}}'
• --command.ui.enter
• --command.ui.arrow '{"direction":"down"}'
• --command.ui.focus '{"target":12}'
• --command.ui.click '{"target":12,"x":4,"y":1}'
• --command.ui.mouse '{"action":"move","x":20,"y":8}'
• --command.ui.resize '{"cols":120,"rows":40}'
• --command.ui.screenshot or --command.ui.screenshot '{"name":"home"}'
• --command.ui.state
• --command.ui.snapshot
• --command.ui.capture
• --command.ui.matches '{"text":"OpenCode"}'
• --command.ui.recording.finish

Use meta for the terminal Alt modifier. For example: --command.ui.press '{"key":"down","modifiers":{"meta":true}}'.

Start with --record to record a headless live instance. stop finishes the recording, exports the MP4, performs owner cleanup, and prints the path.

Add --keypress-overlay when the video should show KeyCastr-style pills for the agent's semantic hotkeys, Enter presses, and arrow navigation. It requires --record; batched typed text is deliberately omitted.

Add --pointer-overlay for real mouse input with an eased cursor in the video. It also requires --record and a pointer-capable OpenCode build.

opencode-drive start --name demo --record --keypress-overlay
opencode-drive stop --name demo

dir prints a live instance's artifact directory, and list lists active instances:

opencode-drive dir --name demo
opencode-drive list

Report Drive Issues

Drive is a work in progress. While using it, file a GitHub issue on anomalyco/opencode-drive for anything that surprises you or that you wish existed; do not just work around it silently. Check gh issue list first and comment on an existing issue instead of opening a duplicate.

Worth an issue:

• Wrong, confusing, or silent failures: an empty error message, a press string that typed literally instead of sending a key, a timeout with no frame, a hang in stop or prune.
• Stale dependencies or protocol drift: a pinned @opencode-ai/* client that cannot decode a current server response.
• Documentation gaps: something this skill or the README should have said.
• Wish-list items: a capability you had to hand-roll in a script (a filesystem operation, a frame-to-text helper, an assertion pattern) that other scripts would reuse.

Keep OpenCode bugs out of the Drive tracker. When Drive merely exposed a server or TUI defect, note the Drive-side workaround in the probe or README and report the defect in anomalyco/opencode (or Kit's todo tracker when a public issue is premature).

gh issue create -R anomalyco/opencode-drive --label bug \
  --title "press: string with '+' is typed literally instead of parsed as a chord" \
  --body "Drive <version>. Script excerpt, expected vs actual, log path."

Include the Drive version, the OpenCode revision under test, the smallest script excerpt that shows it, expected versus actual, and the artifact or log path. Never include credentials or private project data.

Prune

prune removes inactive artifact directories. To remove one instance's artifacts, pass the instance name supplied to init or start, not the generated run-* artifact directory name:

opencode-drive prune --name demo

# Force removal of all artifact directories, including active ones.
opencode-drive prune --force