fix(interface): offer codex none and pi ultracode, which both CLIs accept

[drewstone]

Problem

reasoningEffortsFor is what the sandbox picker offers. Measured against the pinned CLI binaries, two levels were reachable all along and the picker hid them:

• codex could not reach none — the API enumerates none first in its own invalid-value error, so turning thinking off was expressible the entire time it was unselectable.
• pi could not reach ultracode — pi --thinking accepts max, which canonical ultracode maps to, so pi's top rung was unreachable from the UI.

The docstring above the table was itself a stale hand-written claim: it asserted codex starts at minimal and pi stops at xhigh. Both are wrong against the binaries.

Measured

Against the versions agent-dev-container pins (infra/nix/agent-clis.nix), not a developer PATH — those disagree by a minor version on every harness:

harness	version	accepted
codex	0.145.0	none, minimal, low, medium, high, xhigh, max, ultra
claude-code	2.1.218	low, medium, high, xhigh, max
pi	0.81.1	off, minimal, low, medium, high, xhigh, max
openclaw	2026.7.1-2	off, minimal, low, medium, high, xhigh, adaptive, max
kimi-code (kimi-cli)	1.49.0	boolean --thinking / --no-thinking

Two things worth carrying in the docstring, both recorded:

• codex accepts ultra end-to-end (session header echoes reasoning effort: ultra) even though its own enumeration stops at max. ultra2 is rejected, so this is a real value, not lax parsing.
• claude-code warns and substitutes rather than rejecting. An unsupported --effort prints "ignoring it and using the default effort" and runs anyway — so an overstated set there is a silent downgrade, not an error. Its set must stay conservative.

The upstream probe and its drift gate live in agent-dev-container (scripts/probe-reasoning-capabilities.mjs, docs/reference/reasoning-capabilities.json), where a test now asserts the shipped mapper never emits a value the CLI was not measured to accept.

Verification

• agent-interface: 332/332 tests, tsc --noEmit clean, run in a fresh worktree off main with a clean install.
• Expectations updated in harness-capabilities.test.ts for both sets and for the maxEffort: "high" narrowing case, which now includes none.
• Patch changeset included.

[tangletools]

✅ Auto-approved drewstone PR — e96ef4a1

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-08-04T05:54:53Z}