feat(trace): audit model calls across routes

[KB (KB-syntheticsciences)]

Summary

• capture bounded provider response IDs for managed, BYOK, and ChatGPT/Codex subscription calls
• retain each provider request ID on its exact step and consented model.response trace
• add openscience debug audit for content-free local call identities across a session and its workers
• regenerate the public SDK and document the command and privacy behavior

Verification

• bun run check:fast
• 44 focused provider, step-link, audit, and usage-logging tests
• bun run test:sdk (31 passed)
• docs validation and tests (7 passed)

Companion server audit: https://github.com/synthetic-sciences/atlas/pull/383

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
openscience	Ready	Preview	Oct 10, 2026 10:27am UTC

[KB (KB-syntheticsciences)]

Local live Codex validation completed with the existing ChatGPT OAuth login and openai-codex/gpt-6.1-sol. The test exposed and fixed two gaps: OpenAI error responses use x-oai-request-id, and internal message-summary calls were present in telemetry but absent from the local audit. The local content-free call ledger now retains internal/background calls and de-duplicates visible step records. A final live session produced two local audit calls whose HTTP call IDs and provider resp IDs exactly matched the two model.response telemetry records.