Configuration

What this is. Team reads one optional project-local file, .team/config.json. It overrides the bundled model selections used to dispatch body-loaded agents on Codex CLI and Antigravity CLI. Every other setting lives in the host’s own plugin configuration.

Where it lives

.team/config.json sits at the home project root — the project Team was pointed at when the run began. The same root is used for later slice worktrees and companion repositories. Team reads only that root’s file: it does not search ancestor directories and never creates the file for you.

The file is optional. Without it, dispatch uses the bundled selections. The repository ignores the whole .team/ directory, so a local override stays untracked. Pipeline state is separate and lives under docs/plans/<id>/.

What it configures

Model selection for the hosts that load an agent definition’s body instead of a native agent file. Claude Code’s named-agent dispatch reads native model: and effort: frontmatter and ignores this file entirely. OpenCode makes no model claim through it.

Bundled selections map each Claude tier to a host model:

An override replaces individual bundled selections, not the whole map. Tiers you do not name keep their bundled value.

Schema

{
  "codex": {
    "sonnet": { "model": "gpt-5.6-terra", "reasoning_effort": "medium" }
  },
  "antigravity": {
    "sonnet": { "model": "flash" }
  }
}

• Host keys: only codex and antigravity.
• Tier keys: only opus, sonnet, and haiku.
• Selection: requires model. Codex also accepts reasoning_effort; no other fields are allowed.

Codex

model must be a concrete model ID — not a Claude alias (opus, sonnet, haiku), fable, or inherit. The optional reasoning_effort replaces the agent’s own effort; without it the agent’s effort is preserved. Accepted efforts: low, medium, high, xhigh, max, ultra.

Antigravity

model must be one of inherit, flash_lite, flash, or pro. Antigravity selects effort with the invocation tier, so there is no separate effort field. inherit deliberately chooses the parent session’s model.

How overrides are validated

Before the first dispatch — and again if the host’s provider or model configuration changes — Team reads the running host’s current capabilities. Codex exposes its model catalog and supported reasoning efforts; Antigravity exposes the invoke_subagent model enum.

Team rejects the file when it names an unknown host, tier, or field; omits model; uses a Claude alias as a Codex model; or selects a model, tier, or effort the host does not currently offer. A missing or invalid override is an error, never a silent fall back to a default. The resolver validates selections only — it does not query a provider or spawn an agent.

For the dispatch procedure and the full validation contract, see model selection and cross-host portability.

The pre-prompt guard

A UserPromptSubmit hook validates .team/config.json before a prompt reaches the model. An absent file is valid. A present file that cannot be read, is not valid JSON, or fails the schema above blocks the prompt with exit 2 and shows the reason, so a config the resolver would reject is caught before any work starts rather than at dispatch time. Availability against the running host is still checked at dispatch; the guard covers syntax and schema only.

The guard runs on every host that can express it, and each host names its own failure channel:

• Claude Code — .claude-plugin/plugin.json, UserPromptSubmit, exit 2.
• Codex CLI — hooks/hooks.json, UserPromptSubmit, exit 2, reusing the canonical hooks/validate-team-config.mjs unchanged.
• Antigravity CLI — root hooks.json, PreInvocation, inject-only. That host cannot block a prompt, so an invalid config surfaces an injectSteps ephemeral message and the prompt proceeds.
• OpenCode — a named gap: there is no prompt-block hook, and the nearest candidate is unprobed.

The full matrix, with each cell’s verification status, is hooks-portability.md.