Not able to run Skill with OpenAI API

[felipenonato]

Describe the bug
I'm getting the error when trying to run any skill with the OpenAI model.

Couldn't run Enhance — Invalid schema for response_format 'response': In context=(), 'required' is required to be supplied and to be an array including every key in properties. Missing 'reasoning'.

[percymcn]

This is already fixed on main — it just hasn't been released yet, which is why it still reproduces for everyone on the latest build.

Timeline

• v0.1.0 (the current latest release) was tagged 2026-05-19.
• The fix, b5ee4b4a "fix(skills): nullable reasoning for OpenAI strict mode (PRSM-78)", landed 2026-05-26 — a week after that tag.
• No release has been cut since, so every user on v0.1.0 still hits it on any OpenAI model.

Why the error says exactly Missing 'reasoning'

OpenAI strict structured outputs (the default in @ai-sdk/openai) have no concept of an optional field: every key in properties must also appear in required. Zod's .optional() omits the key from required, so the request is rejected before the model ever runs — which is why it fails on every skill rather than intermittently.

The current OUTPUT_SCHEMA in apps/desktop/src/services/skills-runtime/skill-agent.ts already encodes the right fix:

export const OUTPUT_SCHEMA = z.object({
  markdown: z.string(),
  reasoning: z.string().nullable(),   // not .optional()
});

.nullable() keeps the key in required while still letting the model return null — that's the correct way to express "optional" under strict mode, and it preserves the field's meaning instead of forcing the model to invent a value.

So: @felipenonato, building from main should clear it; otherwise it needs a new release. @nchopra — a tag off main would close this one out.

One note for whoever picks it up: this class of bug is easy to reintroduce, because the schema is valid Zod and valid JSON Schema — it's only invalid under OpenAI's strict dialect, so nothing fails until a real API call. Two of the other repos I've seen hit it hit it twice, the second time on a different key of the same schema. Asserting the compiled schema in a test (or gating it in CI) is worth more than the one-line fix. I maintain a small dependency-free checker for exactly that if it's useful — npx github:percymcn/llm-json-schema --to openai --check schema.json exits non-zero when a schema isn't strict-valid — but the rule above is the part that matters.