provider-concurrency.ts limits concurrent task tool calls per model provider inside the current OpenCode process. It is intended to keep delegated subagents from overloading a local model server or a paid remote provider.

This directory is the global OpenCode plugin directory for this machine. OpenCode auto-discovers *.ts and *.js files under ~/.config/opencode/plugins/, so plugins/provider-concurrency.ts does not need an opencode.json plugin entry.

Restart OpenCode after changing this plugin or opencode.concurrency.json. Plugin and config files are loaded at startup.

The plugin reads opencode.concurrency.json from the same config directory:

{
  "enabled": true,
  "debug": false,
  "mode": "reject",
  "providerLimits": {
    "openrouter": null,
    "ollama": 1
  },
  "agentProviderOverrides": {},
  "agentModelOverrides": {},
  "queueTimeoutMs": null,
  "staleLeaseMs": 600000
}

Fields:

• enabled: disables all limits when set to false.
• debug: logs missing or stale lease details with console.warn.
• mode: reject throws immediately when a provider is full; queue waits for a slot.
• providerLimits: provider to max concurrent task calls. Set a provider value to null to make that provider explicitly unlimited, for example { "openrouter": null }.
• agentProviderOverrides: agent name to provider, for example { "worker-md": "ollama" }.
• agentModelOverrides: agent name to model string with provider prefix, for example { "worker-md": "ollama/worker-md-local" }.
• queueTimeoutMs: max queue wait before throwing in queue mode. Set to null for no queue timeout.
• staleLeaseMs: age after which a lease can be cleaned up before another same-provider acquire. Set to 0 to disable stale cleanup.

The plugin only intercepts OpenCode's task tool. Provider resolution order is:

1. args.model or args.subagent_model when it has a provider/model prefix.
2. The task agent name from subagent_type, agent, subagent, or agentName.
3. agentModelOverrides[agent], parsed as provider/model.
4. agentProviderOverrides[agent].

If no provider is found, the provider has no configured limit, or the provider limit is null, the task runs without throttling.

Current configured text worker in this config:

• worker-md: ollama/worker-md-local, so it resolves to ollama.

Equivalent explicit override:

{
  "agentModelOverrides": {
    "worker-md": "ollama/worker-md-local"
  }
}

• Limits are process-local. They do not coordinate across multiple OpenCode processes or terminals.
• This does not intercept direct provider API calls; it only gates OpenCode task tool execution.
• OpenCode currently exposes tool.execute.before and tool.execute.after for this use. If a task starts but no after hook runs, staleLeaseMs is the recovery mechanism.
• This is a concurrency guard, not a token, billing, health-check, or provider rate-limit scheduler. Provider limits are defense-in-depth; they complement orchestrator-level lifecycle policies like one-worker-at-a-time execution but do not replace them.