pi hardcodes reasoning:true → non-reasoning models (gpt-4o) no-op instantly

[gewenyu99]

Found testing an OpenAI sonnet-class model (openai/gpt-4o) via pi, per the parity suite (#528).

What

Model	Type	pi result
claude-sonnet-4-6	reasoning	✅ full integration (~10 min)
openai/gpt-5	reasoning	✅ works but too slow (>15 min, #789)
openai/gpt-4o	non-reasoning	❌ no-op in 12s — agent "completes" with zero code

The gateway accepts openai/gpt-4o (a plain completion returns gpt-4o-2024-08-06). So it's not the model id and not capability — pi itself choked.

Cause

harness/pi/index.ts registers every gateway model with reasoning: true (line ~247), unconditionally. gpt-4o isn't a reasoning model, so pi's reasoning-configured request produces no usable tool loop — the run races through the screens in 12s and writes nothing (but the loose E2E check still says PASS because deps were added).

Impact

pi can currently only drive reasoning models. That rules out the cheaper/faster sonnet-class OpenAI tier (gpt-4o), which is exactly what you'd want as a cost/speed alternative — flagship gpt-5 is too slow (#789). Fix: set reasoning per-model in the registration (or infer it) so non-reasoning models work.

Not a #701 blocker (pi opt-in/off; default/anthropic unaffected), but it narrows pi's usable model set to reasoning models only.

[gewenyu99]

I think this is a small switch board config update

[gewenyu99]

Fixed (via the switchboard, configurable + defaulted)

Root cause confirmed by proof: pi sent reasoning_effort (from reasoning: true) → gateway litellm.UnsupportedParamsError for the non-reasoning gpt-4o → run no-op'd. Same model with tools + no reasoning param returns tool_calls fine — so it was never a capability issue.

Fix (#701, 1cd5e35): the model's reasoning trait is now resolved through the switchboard, not hardcoded in the harness.

• New switchboard/models.ts — a per-model capability table (MODEL_CAPABILITIES) + modelCapabilities(id).
• Configurable: add/override a model in the table.
• Defaulted: unknown models default by transport — reasoning on for anthropic-messages, off for openai-completions (a reasoning openai model opts back in via the table). So a new non-reasoning model works out of the box.
• The pi harness reads modelCapabilities(modelId).reasoning instead of true.
• Tests added (sonnet/opus/haiku/gpt-5 → reasoning; gpt-4o + unknown-openai → not).

Verified: pi + openai/gpt-4o went from a 12s no-op (0 code) to a real integration — outro "✔ Successfully installed PostHog!", wrote app/instrumentation.ts + Stripe events. Closing once the final clean-PR run lands.

[gewenyu99]

Fixed in #701 — reasoning is now a per-model trait in the switchboard capability matrix (switchboard/models.ts modelCapabilities().reasoning), no longer hardcoded in the pi harness. Non-reasoning openai models default to reasoning off.