@thoth-agents/pi-openai-fast

Pi extension that adds a priority-tier -fast sibling next to every OpenAI model Pi can reach through the Responses APIs.

For each physical model whose api is openai-responses or openai-codex-responses, the extension lists <provider>/<id>-fast inside the same provider. This covers the built-in openai and openai-codex providers and providers added by other extensions (for example openai-codex-2), without naming any of them. Selecting a fast variant sends the request to the physical <provider>/<id> with that provider's own credentials and service_tier: "priority".

Install

Install manually; the Thoth installer does not manage this extension.

pi install npm:@thoth-agents/pi-openai-fast

How it works

• On every session_start (startup, reload, new, resume, fork) the extension reads the model registry and registers one Pi virtual model per eligible base model. The sync is idempotent and unregisters only variants it registered whose base model has disappeared.
• A variant mirrors its base model's name (suffixed (fast)), thinking levels, context window, max tokens and input types.
• Pi calls the variant's route() for each agent-loop request. It resolves the base model from the registry at request time and returns it with the requested thinking level (clamped to what the base supports). A missing base model raises a clear error.
• Routes for reasons user, continuation and retry arm a one-shot token. The next before_provider_request adds service_tier: "priority" to a shallow copy of the payload only when payload.model equals the routed base id. Any payload consumes the token, so a stale token never grants priority to a later request.
• When priority is applied, a one-shot pricing marker corrects the next matching Codex assistant message's displayed cost to priority pricing (×2, or ×2.5 for exactly gpt-5.5). Already-priced usage is left unchanged. Any new route or assistant message clears the marker; the original message is never mutated.
• OpenAI API pricing follows the service tier the server reports; the extension does not adjust openai-responses usage costs.
• The extension never calls registerProvider; existing providers, catalogs, oauth and model refresh callbacks are untouched.

Limits

• Models are discovered dynamically; there is no allowlist or configuration.
• direct requests (compaction summaries, extension calls) and requests for other models chosen by other packages (for example a compaction-model package) keep the standard tier, even when they use the same base model.
• Fast variants and Codex cost correction also work in pi-subagents children when this package is enabled for lifecycle passthrough.
• Priority entitlement and latency are decided by the provider. An unsupported tier surfaces as a normal provider error.
• Models whose id already ends in -fast, and bases whose -fast id would collide with an existing physical model, get no variant.

Development

From the workspace root:

pnpm install --frozen-lockfile
pnpm --filter @thoth-agents/pi-openai-fast run typecheck
pnpm --filter @thoth-agents/pi-openai-fast run test

Credits

The idea of a priority-tier fast model for Codex is inspired by 2h2d-co/pi-openai-codex-fast (MIT, commit b2258b0bc9e2c03fd40d51143530882917aabb00). This package is a new implementation, not a fork, and uses Pi virtual models instead of a separate provider.

License

MIT