fix(design): drop gpt-image-2 tool model on gpt-4o orchestrator (#1771)

[Pablosinyores]

Summary

The Responses API rejects pairing a gpt-4o orchestrator with an image_generation tool spec'd as model: "gpt-image-2" with a generic 400 invalid_request_error. gpt-image-2 is only valid under a gpt-5 orchestrator; with gpt-4o the tool must omit the model field (defaults to gpt-image-1).

This regression landed in v1.43.2.0 (commit 66f3a180, 2026-05-21), which added model: "gpt-image-2" to the tool spec while leaving the orchestrator at gpt-4o. The pairing took all five design subcommands offline:

• design generate
• design variants
• design iterate
• design evolve
• /design-shotgun

Closes #1771.

Fix

Revert the five call sites to the pre-regression shape (drop the tool model field; the API defaults to gpt-image-1 which works under gpt-4o):

• design/src/generate.ts:53
• design/src/variants.ts:73
• design/src/iterate.ts:98 (threaded fresh-prompt call)
• design/src/iterate.ts:145 (threaded follow-up call)
• design/src/evolve.ts:67

Regression tripwire

design/test/image-gen-pairing.test.ts is a static-grep tripwire over every design/src/*.ts. It fails CI if any module contains BOTH the literal model: "gpt-4o" and the literal model: "gpt-image-2" — the unsupported pairing.

Same shape as existing tripwires in browse/test/:

• cdp-session-cleanup.test.ts
• terminal-agent-pid-identity.test.ts
• server-embedder-terminal-port.test.ts

To re-enable gpt-image-2 later, the orchestrator must also be bumped off gpt-4o in the SAME diff — the tripwire allows that because both literals are no longer co-present in one module.

Validation

$ bun test design/test/image-gen-pairing.test.ts
 19 pass
 0 fail
 19 expect() calls
Ran 19 tests across 1 file. [18.00ms]

$ bun test design/test/image-gen-pairing.test.ts design/test/variants-retry-after.test.ts design/test/auth.test.ts design/test/daemon.test.ts design/test/serve.test.ts design/test/gallery.test.ts
 87 pass
 0 fail
 267 expect() calls
Ran 87 tests across 6 files. [7.99s]

design/test/feedback-roundtrip*.test.ts were not run locally — they require a Playwright Chromium install which isn't on this machine. They don't touch the orchestrator/tool model literals, so the fix is orthogonal to that path.

AI assistance disclosure

Patch was drafted with AI assistance. Diff was hand-reviewed against the regression commit 66f3a180; root cause cross-referenced with the issue body (paired with the maintainer's own correction note); tripwire test inverts the failing condition described in the issue and reproduces the failure when reverted (verified locally when generate.ts was initially missed).