feat(agent-adapter): advertise approval tiers before session start

[AprilNEA]

Closes CODE-463.

The gap

The pre-session agent catalog landed in 38c6389f with the whole schema → engine → client → UI chain, but only pi ever overrode startCatalog. claude-code, opencode and grok-build inherited the base default, and codex (which gained a catalog later, for the effort work) returned a hardcoded policies: []:

startCatalog(_opts?: AgentStartCatalogOptions): Promise<AgentStartCatalog> {
  return Promise.resolve({ models: [], policies: [] });
}

new-session-surface.tsx gates the approval-policy control on catalog && catalog.policies.length > 0, so approvalPolicy resolved to undefined and the New Task surface rendered no picker for four of the five agents. In-session reflection was never affected — claude-code still advertises availablePolicies once a session starts — which is why this reads as "the picker disappeared on New Task" rather than "the feature is gone".

What this does

Fills in the missing half, so every adapter advertises the tiers a session would actually run under, and honors a pick made before the session exists.

agent	pre-session policies	initial pick
claude-code	the five tiers + the resolved permissions.defaultMode as defaultPolicyId	outranks the settings default; unknown id degrades to it with an error event
codex	the three tiers + acceptEdits	counts as an explicit pick, so it may override a config.toml sandbox from the first turn, exactly like a live switch; unknown id degrades to acceptEdits
opencode	selectable agents (the agent axis) + models, both directory-scoped, read over the shared history server	outranks a resumed session's recorded agent; an unselectable pick degrades to the catalog default with an error
grok-build	the fixed bypassPermissions tier	n/a (visibility only — set-approval-policy still rejects)
pi	already shipped	already shipped

opencode's selectable-agent and reachable-model mapping moved to two module-level pure functions so the live onStart fetches and the pre-session read share one definition instead of drifting.

A failed pre-session read degrades to an empty axis rather than blocking session creation, matching the live best-effort catalog fetches.

Notes

• grok-build is beyond the original scope (it postdates the work this restores), but leaving it out would have shown the picker for four of five agents.
• No wire change — AgentStartCatalogSchema already carries policies/defaultPolicyId, and the engine forwards the catalog verbatim.

Verification

pnpm check:ci and pnpm test both pass (2011 passed / 5 skipped). New coverage: pre-session catalog contents per adapter, the initial-pick path, and the degrade-with-error path; codex additionally asserts that a new-session pick overrides a configured sandbox while an unknown one does not.

Not yet driven in the running app — the adapter contract is covered by tests, and the engine/client legs of the chain were already live for pi.

[linear-code]

CODE-463

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5f436c454a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".