Files
openclaw/docs/concepts/agent-runtimes.md
pashpashpash 8f4eaa9c00 Stop heartbeat tool turns from asking for HEARTBEAT_OK (#76338)
* fix heartbeat tool prompt sentinel

* fix: remove agent runtime fallback config
2026-05-03 13:46:26 +09:00

13 KiB

summary, title, read_when
summary title read_when
How OpenClaw separates model providers, models, channels, and agent runtimes Agent runtimes
You are choosing between PI, Codex, ACP, or another native agent runtime
You are confused by provider/model/runtime labels in status or config
You are documenting support parity for a native harness

An agent runtime is the component that owns one prepared model loop: it receives the prompt, drives model output, handles native tool calls, and returns the finished turn to OpenClaw.

Runtimes are easy to confuse with providers because both show up near model configuration. They are different layers:

Layer Examples What it means
Provider openai, anthropic, openai-codex How OpenClaw authenticates, discovers models, and names model refs.
Model gpt-5.5, claude-opus-4-6 The model selected for the agent turn.
Agent runtime pi, codex, claude-cli The low level loop or backend that executes the prepared turn.
Channel Telegram, Discord, Slack, WhatsApp Where messages enter and leave OpenClaw.

You will also see the word harness in code. A harness is the implementation that provides an agent runtime. For example, the bundled Codex harness implements the codex runtime. Public config uses agentRuntime.id; openclaw doctor --fix rewrites older runtime-policy keys to that shape.

There are two runtime families:

  • Embedded harnesses run inside OpenClaw's prepared agent loop. Today this is the built-in pi runtime plus registered plugin harnesses such as codex.
  • CLI backends run a local CLI process while keeping the model ref canonical. For example, anthropic/claude-opus-4-7 with agentRuntime.id: "claude-cli" means "select the Anthropic model, execute through Claude CLI." claude-cli is not an embedded harness id and must not be passed to AgentHarness selection.

Codex surfaces

Most confusion comes from several different surfaces sharing the Codex name:

Surface OpenClaw name/config What it does
Native Codex app-server runtime openai/* plus agentRuntime.id: "codex" Runs the embedded agent turn through Codex app-server. This is the usual ChatGPT/Codex subscription setup.
Codex OAuth provider route openai-codex/* model refs Uses ChatGPT/Codex subscription OAuth through the normal OpenClaw PI runner.
Codex ACP adapter runtime: "acp", agentId: "codex" Runs Codex through the external ACP/acpx control plane. Use only when ACP/acpx is explicitly asked.
Native Codex chat-control command set /codex ... Binds, resumes, steers, stops, and inspects Codex app-server threads from chat.
OpenAI Platform API route for GPT/Codex-style models openai/* model refs Uses OpenAI API-key auth unless a runtime override, such as agentRuntime.id: "codex", runs the turn.

Those surfaces are intentionally independent. Enabling the codex plugin makes the native app-server features available; it does not rewrite openai-codex/* into openai/*, does not change existing sessions, and does not make ACP the Codex default. Selecting openai-codex/* means "use the Codex OAuth provider route" unless you separately force a runtime.

The common ChatGPT/Codex subscription setup uses Codex OAuth for auth, but keeps the model ref as openai/* and selects the codex runtime:

{
  agents: {
    defaults: {
      model: "openai/gpt-5.5",
      agentRuntime: {
        id: "codex",
      },
    },
  },
}

That means OpenClaw selects an OpenAI model ref, then asks the Codex app-server runtime to run the embedded agent turn. It does not mean "use API billing," and it does not mean the channel, model provider catalog, or OpenClaw session store becomes Codex.

When the bundled codex plugin is enabled, natural-language Codex control should use the native /codex command surface (/codex bind, /codex threads, /codex resume, /codex steer, /codex stop) instead of ACP. Use ACP for Codex only when the user explicitly asks for ACP/acpx or is testing the ACP adapter path. Claude Code, Gemini CLI, OpenCode, Cursor, and similar external harnesses still use ACP.

This is the agent-facing decision tree:

  1. If the user asks for Codex bind/control/thread/resume/steer/stop, use the native /codex command surface when the bundled codex plugin is enabled.
  2. If the user asks for Codex as the embedded runtime or wants the normal subscription-backed Codex agent experience, use openai/<model> with agentRuntime.id: "codex".
  3. If the user asks for Codex OAuth/subscription auth on the normal OpenClaw runner, use openai-codex/<model> and leave the runtime as PI.
  4. If the user explicitly says ACP, acpx, or Codex ACP adapter, use ACP with runtime: "acp" and agentId: "codex".
  5. If the request is for Claude Code, Gemini CLI, OpenCode, Cursor, Droid, or another external harness, use ACP/acpx, not the native sub-agent runtime.
You mean... Use...
Codex app-server chat/thread control /codex ... from the bundled codex plugin
Codex app-server embedded agent runtime agentRuntime.id: "codex"
OpenAI Codex OAuth on the PI runner openai-codex/* model refs
Claude Code or other external harness ACP/acpx

For the OpenAI-family prefix split, see OpenAI and Model providers. For the Codex runtime support contract, see Codex harness.

Runtime ownership

Different runtimes own different amounts of the loop.

Surface OpenClaw PI embedded Codex app-server
Model loop owner OpenClaw through the PI embedded runner Codex app-server
Canonical thread state OpenClaw transcript Codex thread, plus OpenClaw transcript mirror
OpenClaw dynamic tools Native OpenClaw tool loop Bridged through the Codex adapter
Native shell and file tools PI/OpenClaw path Codex-native tools, bridged through native hooks where supported
Context engine Native OpenClaw context assembly OpenClaw projects assembled context into the Codex turn
Compaction OpenClaw or selected context engine Codex-native compaction, with OpenClaw notifications and mirror maintenance
Channel delivery OpenClaw OpenClaw

This ownership split is the main design rule:

  • If OpenClaw owns the surface, OpenClaw can provide normal plugin hook behavior.
  • If the native runtime owns the surface, OpenClaw needs runtime events or native hooks.
  • If the native runtime owns canonical thread state, OpenClaw should mirror and project context, not rewrite unsupported internals.

Runtime selection

OpenClaw chooses an embedded runtime after provider and model resolution:

  1. A session's recorded runtime wins. Config changes do not hot-switch an existing transcript to a different native thread system.
  2. OPENCLAW_AGENT_RUNTIME=<id> forces that runtime for new or reset sessions.
  3. agents.defaults.agentRuntime.id or agents.list[].agentRuntime.id can set auto, pi, a registered embedded harness id such as codex, or a supported CLI backend alias such as claude-cli.
  4. In auto mode, registered plugin runtimes can claim supported provider/model pairs.
  5. If no runtime claims a turn in auto mode, OpenClaw uses PI as the compatibility runtime. Use an explicit runtime id when the run must be strict.

Explicit plugin runtimes fail closed. For example, agentRuntime.id: "codex" means Codex or a clear selection/runtime error; it is never silently routed back to PI.

CLI backend aliases are different from embedded harness ids. The preferred Claude CLI form is:

{
  agents: {
    defaults: {
      model: "anthropic/claude-opus-4-7",
      agentRuntime: { id: "claude-cli" },
    },
  },
}

Legacy refs such as claude-cli/claude-opus-4-7 remain supported for compatibility, but new config should keep the provider/model canonical and put the execution backend in agentRuntime.id.

auto mode is intentionally conservative. Plugin runtimes can claim provider/model pairs they understand, but the Codex plugin does not claim the openai-codex provider in auto mode. That keeps openai-codex/* as the explicit PI Codex OAuth route and avoids silently moving subscription-auth configs onto the native app-server harness.

If openclaw doctor warns that the codex plugin is enabled while openai-codex/* still routes through PI, treat that as a diagnosis, not a migration. Keep the config unchanged when PI Codex OAuth is what you want. Switch to openai/<model> plus agentRuntime.id: "codex" only when you want native Codex app-server execution.

Compatibility contract

When a runtime is not PI, it should document what OpenClaw surfaces it supports. Use this shape for runtime docs:

Question Why it matters
Who owns the model loop? Determines where retries, tool continuation, and final answer decisions happen.
Who owns canonical thread history? Determines whether OpenClaw can edit history or only mirror it.
Do OpenClaw dynamic tools work? Messaging, sessions, cron, and OpenClaw-owned tools rely on this.
Do dynamic tool hooks work? Plugins expect before_tool_call, after_tool_call, and middleware around OpenClaw-owned tools.
Do native tool hooks work? Shell, patch, and runtime-owned tools need native hook support for policy and observation.
Does the context engine lifecycle run? Memory and context plugins depend on assemble, ingest, after-turn, and compaction lifecycle.
What compaction data is exposed? Some plugins only need notifications, while others need kept/dropped metadata.
What is intentionally unsupported? Users should not assume PI equivalence where the native runtime owns more state.

The Codex runtime support contract is documented in Codex harness.

Status labels

Status output may show both Execution and Runtime labels. Read them as diagnostics, not as provider names.

  • A model ref such as openai/gpt-5.5 tells you the selected provider/model.
  • A runtime id such as codex tells you which loop is executing the turn.
  • A channel label such as Telegram or Discord tells you where the conversation is happening.

If a session still shows PI after changing runtime config, start a new session with /new or clear the current one with /reset. Existing sessions keep their recorded runtime so a transcript is not replayed through two incompatible native session systems.