Runtime support

Status: current integrations plus advanced optional runtime helpers. Operators and orchestrators retain ownership of homes, worktrees, runtime selection, processes, and session UX. The helper matrix below is not an aweb lifecycle guarantee or a prerequisite for communication.

Runtime support has three separate layers: using the CLI to materialize an agent home, using it to launch an AI tool, and waking an independently owned session when team activity arrives. Do not treat support in one layer as support in all three.

Supported materialization values are:

claude-code
codex
pi
local-shell

The runtime is selected when the agent is materialized:

[email protected]/developer=claude-code

It is not part of the profile and can differ between two agents using the same profile.

Pi package lifecycle

aw team up installs @awebai/pi when missing but does not auto-update already-installed executable extension code. Pi 0.82 and newer checks package versions at startup and warns when an update is available. Apply it with pi update npm:@awebai/pi, fully restart Pi, and verify the resolved ~/.pi/agent/npm/node_modules/@awebai/pi/package.json version. For pre-0.82 Pi, use npm install @awebai/pi@latest --prefix ~/.pi/agent/npm as the fallback. A global npm upgrade does not affect Pi’s package tree.

Portable polling loop

Any runtime that can execute the CLI can participate in a team by polling:

aw mail inbox
aw chat pending
aw work ready

This is functional coordination but not automatic real-time wake-up. The human or runtime wrapper must decide when the AI session is prompted again.

Hosted chat AIs

An MCP-capable chat AI uses the hosted MCP/OAuth path rather than a materialized local runtime home. MCP tool availability and wake-up behavior depend on the host client. Hosted MCP messages are server-readable hosted messaging, not local E2E messaging.