Loop

Keep an opencode session working automatically — on a fixed interval or at a pace the model sets for itself — instead of re-typing the same prompt every time the session goes idle. /loop re-injects your prompt into your own session each time a turn finishes, carrying the full conversation context forward, until the task is done, you stop it, or a safety cap trips.

Why use loop

• Keep an agent grinding on a task. Refactor a module, chase a flaky test, or work through a checklist across many turns without re-typing the prompt each time the session goes idle.
• Run recurring checks on a cadence. Re-run the test suite, a build, or a monitoring script every few minutes and get a fresh report each time.
• Let the model pace itself. Skip the fixed interval for open-ended or exploratory work and let the model call schedule_wakeup only when it actually has something worth checking back on.
• Walk away without babysitting. The loop pauses itself on a permission prompt or an error, and hard caps guarantee it can never run forever.

Installation

The plugin is a self-contained ESM module whose entry file is loop.js. Register it in your opencode.json under the singular "plugin" key.

For a local clone or copied checkout, register loop.js by an absolute path or by a path relative to the opencode.json file that declares it:

{
  "plugin": [
    "/path/to/opencode-loop/loop.js"
  ]
}

To change the safety caps, use the plugin tuple form:

{
  "plugin": [
    [
      "./plugins/loop/loop.js",
      {
        "maxIterations": 50,
        "maxWallClockMinutes": 1440
      }
    ]
  ]
}

maxIterations must be a positive integer from 1 to 50 and defaults to 50. maxWallClockMinutes must be a positive integer from 1 to 1440 and defaults to 60. Invalid fields independently fall back to their safe default and produce a redacted warning diagnostic; a valid field still applies. Restart opencode after changing these options. Each loop snapshots the effective caps when it starts, so later caller-side mutation cannot change an active loop.

For the npm package, install it where opencode resolves plugins and register the package name:

{
  "plugin": [
    "@mcrescenzo/opencode-loop"
  ]
}

The "plugin" value is an array, so add the entry alongside any other plugins you already register. After registering or changing the plugin, restart opencode — running sessions keep already-loaded plugin code.

Register this plugin once. Do not load both a path copy and the published package in the same opencode config. If another plugin or user command already owns /loop, this plugin leaves that command definition and execution alone.

This package requires Node 20.11 or newer for its local test/package scripts. The runtime plugin dependency is @opencode-ai/plugin@^1.17.7; the current release candidate was smoke-tested against opencode 1.17.13. opencode does not install dependencies for path-spec plugins, so run bun install or npm install in this directory before using a path registration during development.

Quick start

/loop command syntax

Form	Behavior
/loop <interval> <prompt>	Fixed-interval. <interval> is <n>s|m|h (e.g. 5m), clamped to 5s–24h and rejected when it cannot fit within the effective wall-clock limit. Idle-gated: the next iteration fires at least <interval> after the previous one finishes, so an overrun delays rather than stacks.
/loop <prompt>	Dynamic / self-pacing. Each iteration may call the schedule_wakeup tool with delaySeconds (clamped to 60–3600) to continue. Finishing a turn without calling schedule_wakeup ends the loop.
/loop status	Show the active loop: mode, iteration count, and elapsed time. Alias: info.
/loop stop	Stop the active loop. Aliases: off, cancel, clear, end.

Examples:

/loop 5m run the test suite and report any failures
/loop keep refactoring this module until it is clean, then stop
/loop status
/loop stop

Looped content is re-injected as plain prompt text. Re-running slash commands (e.g. /loop 5m /some-command) isn't supported yet.

What you'll see

Running:

/loop 5m run the test suite and report any failures

starts the loop and immediately runs iteration 1 with that prompt — the command's own turn is the first iteration, so the session just gets to work. A toast confirms Loop started (fixed, 300s).. From then on, every time the session goes idle, the plugin waits until at least 5 minutes have passed since the last run and re-injects the same prompt, carrying conversation context forward, until you stop it or a safety cap trips.

Checking in with:

/loop status

prints something like:

Active /loop: fixed (300s), status running.
Iterations: 1/50. Elapsed: 12s.
Limits: 50 iterations, 60m wall clock.
Prompt: run the test suite and report any failures

Safety

• A loop stops automatically at its effective iteration or wall-clock cap. The defaults are 50 iterations and 60 minutes; plugin options may narrow iterations to 1–50 or set wall time to 1–1440 minutes. The time cap is enforced before scheduling or injecting the next iteration.
• It pauses while a permission prompt is open and resumes when the prompt is answered; a rejected permission ends the loop.
• Closing or deleting a session clears that session's in-memory loop state.
• If a running iteration reports session.error, the loop pauses and stays paused until you run /loop stop (an error-pause is sticky). If the plugin cannot dispatch the next prompt through opencode, it stops that loop instead and emits a diagnostic record.
• The loop drives the current session, carrying conversation context forward.
• A longer wall-clock cap does not guarantee 50 iterations. Model completion, errors, rejected permissions, restart or session closure, and manual stop can all end a loop sooner.

For AI agents

If you're the model executing /loop, this plugin intercepts the command before it reaches you — seeing this command's raw markdown as a normal prompt means the plugin isn't loaded. In dynamic mode, call schedule_wakeup with a delaySeconds (60–3600) to continue the loop past this turn, or just finish the turn to end it. Iteration caps, wall-clock caps, and permission/error pausing are enforced by the plugin, not by you.

Diagnostics environment variables

The plugin writes local structured JSONL diagnostic records (opencode.plugin.diagnostic.v1). Records are generated by this plugin only and are not uploaded by the plugin. The redaction heuristics are adapted from the author's local opencode diagnostics tooling and are kept self-contained in this package. Default storage root:

Linux:   ${XDG_STATE_HOME:-~/.local/state}/opencode/plugin-diagnostics/<project>-<hash>/loop/
macOS:   ~/Library/Application Support/opencode/plugin-diagnostics/<project>-<hash>/loop/
Windows: %LOCALAPPDATA%\opencode\plugin-diagnostics\<project>-<hash>\loop\

• OPENCODE_PLUGIN_DIAGNOSTICS_DIR — override the diagnostics root directory with an absolute path or ~ path (useful for tests or local debugging).
• OPENCODE_PLUGIN_DIAGNOSTICS_DISABLED=1 — disable diagnostics writes entirely.

Diagnostic files are named loop-YYYY-MM-DD-<pid>.jsonl and rotate before they exceed roughly 1 MB.

Hooks

loop.js registers five opencode plugin hooks — config, event, command.execute.before, tool, and dispose. See the full hook-by-hook breakdown, including exactly what each one does, in https://github.com/mcrescenzo/opencode-loop/blob/main/docs/design.md.

Development

Run npm test (node --test tests/*.test.mjs) before opening a pull request; use Node, not bun test, for validation. There's intentionally no committed package-lock.json. The full contributor workflow — dependency license review, npm run smoke:package, and CI details — lives in https://github.com/mcrescenzo/opencode-loop/blob/main/CONTRIBUTING.md.

Support & Status

This is an unofficial opencode plugin, a community project not affiliated with or endorsed by opencode.ai. Public compatibility is limited to the opencode/plugin API versions documented above until newer versions are tested. Report bugs and compatibility issues at https://github.com/mcrescenzo/opencode-loop/issues.

The loop is in-memory only: closing or restarting opencode ends every active loop with no respawn.

To uninstall, remove @mcrescenzo/opencode-loop or the local loop.js path from the plugin array in opencode.json, then restart opencode. Existing loops are in-memory only, so restart also clears any active loop state.

Contributor-facing design rationale and invariants live in the repository: https://github.com/mcrescenzo/opencode-loop/blob/main/AGENTS.md and https://github.com/mcrescenzo/opencode-loop/blob/main/docs/design.md. Security and contribution guidance live at https://github.com/mcrescenzo/opencode-loop/blob/main/SECURITY.md and https://github.com/mcrescenzo/opencode-loop/blob/main/CONTRIBUTING.md.