Oh My Pi 17.2.3: ai& provider, fixes Anthropic OAuth 1M context breakage
Adds the ai& (aiand) provider — an OpenAI-compatible inference API with dynamic model discovery. Fixes a critical bug where Anthropic OAuth (Claude Pro/Max) requests sent the 1M-context beta on every request, breaking all subagents with a hard 429 on subscription credentials. Fixes headless browser launch storms, reasoning blocks leaking into memories, and several coding-agent bugs.
PUBLISHED2026-08-01
OBSERVED2026-08-21
AGE24d
SOURCES1
- ai& provider (
aiand): OpenAI-compatible API with dynamic model discovery (context windows, capabilities, reasoning, USD pricing), auth viaAIAND_API_KEY. - Critical fix: Anthropic OAuth requests no longer advertise the 1M-context beta on subscription credentials, which were being rejected with 429. Subscriptions get the standard 200k window. This was breaking all subagents on Claude Sonnet 4.6 and similar models.
- Headless browser: Chromium processes now attach to one project-shared daemon-owned browser; concurrent launches share one process; duplicate daemon starts prevented.
- Memory fix: reasoning blocks (
thinking…response) stripped from mnemopicleanOutput, so stored facts no longer contain reasoning prose. - Fixes: OpenAI Codex Responses honor disabled cache retention. Bash auto-background timer no longer delays shutdown. Ephemeral turns preserve prompt cache key. Ask dialog crash on non-string
questionfixed. Codex web search preserves backend error codes. Session-file writing batched for performance.
COMMUNITY
No curated reactions recorded for this event. Facts and takes are kept in separate layers — community context is added by hand, never blended into the record above.