## Why Starting a turn discarded the discovery cache and forced MCP to rebuild from a fresh inventory even when the executor context was unchanged. An MCP refresh could also discover a different inventory from the one the step had already captured. ## What changed Keep a per-thread cache keyed by environment, executor session, working directory, and sandbox. Explicit invalidation reloads discovery on the next lookup and lets in-flight readers finish with their captured results. Use the step's captured inventory for MCP reconciliation and bind it with the current MCP policy. Apply the binding's plugin policy when selecting skills and hooks. ## Testing Cover cache reuse across turns, lazy invalidation during in-flight discovery, and MCP refresh with a captured inventory after a server has been revoked. GitOrigin-RevId: 8d4087dffdb3d6feef2456918a08ae7724cf18e4
Codex MCP runtime
This crate manages Codex's MCP connections, tool calls, and their captured authority.
Sandbox request metadata
An MCP server advertises the experimental codex/sandbox-state capability to
receive sandbox metadata on tool calls. The deprecated capability name
codex/sandbox-state-meta opts into the same contract during migration.
Codex supplies these entries in the request's _meta object:
| Key | Value |
|---|---|
codex/sandbox-state |
Opaque policy to pass to codex sandbox --sandbox-state-json. |
codex/sandbox-state-meta |
An identical copy of that policy under the deprecated name. |
codex/sandbox-launcher |
When available, an object with an absolute codexExecutable path to a verified full Codex CLI on the MCP server's host. |
Use the launcher path to invoke sandbox --sandbox-state-json with the entire
state value serialized as one argument, followed by -- and the command to run.
The state contains neither launcher discovery data nor deprecation messages.
Preserve unknown policy fields by forwarding the original value rather than
deserializing and reconstructing it. The two state names are aliases, not separate
policies; new consumers should prefer codex/sandbox-state.
Launcher metadata is currently available only for local stdio servers. An HTTP endpoint, including loopback, or a remote executor does not establish a shared filesystem with the harness. Codex omits the launcher when its path is unknown, not representable as UTF-8, or fails the cached capability probe. Absence does not authorize a PATH search or unsandboxed fallback. Servers that require a launcher must fail closed or require an explicit operator opt-out.
The state and launcher describe the issuing call. Retain both when starting a
long-lived subprocess or caching policy-bound results; reject changed authority
before reusing them. Tool arguments supplied by the model are not authority.
See sandbox_metadata.rs for the wire types and
binding.rs for execution-environment selection.