Files
Assistant/architecture/adrs/ADR-0007-openai-advisory-binding.md
m b61612f543
CI / test (push) Successful in 12s
CI / test (pull_request) Successful in 11s
arch: accept OpenAI advisory binding
2026-08-16 20:37:38 +03:00

4.2 KiB

ADR-0007 — OpenAI Creator advisory binding

Status: Accepted
Date: 2026-08-16
Accepted under: ARCH-000010

Decision

  1. The first real provider class for creator-advisory-v1 is OpenAI at the fixed HTTPS origin https://api.openai.com, using POST /v1/responses.
  2. Binding revision 1 targets the immutable model snapshot gpt-5.4-mini-2026-03-17 with reasoning effort low. An alias or another snapshot is a different binding revision and receives no inherited promotion.
  3. Requests set store: false, are synchronous and single-turn, and contain no tools, functions, conversations, background mode, streaming, files, images, web/file search, MCP, hosted execution, previous_response_id or provider memory.
  4. Mode A permits only the Creator message explicitly entered for that request plus a fixed data-free system instruction. Automatic excerpts and PRIVATE_ALLOWED context are prohibited until a separately evidenced ZDR/MAM binding revision is accepted and promoted.
  5. Before each Mode A request the Creator UI must disclose that the message leaves the private network and provider abuse-monitoring retention may be up to 30 days. Submission acknowledges that request only and is not durable blanket consent.
  6. SECRET, UNKNOWN, credentials, authorization/routing identifiers and ambient context never enter the provider payload. Trusted Actor, Principal, Scope, session, workspace, task and binding identifiers remain local authorization/routing data and are not model context.
  7. The API key belongs to a dedicated OpenAI project and may be mounted read-only only into cognitive-worker as a secret file. It is absent from Git, databases, images, environment variables, logs, browser state and other containers.
  8. Adapter implementation must use the standard library, fixed origin, normal TLS verification, disabled redirects and ambient proxies, strict request/response types, a 30-second deadline, no automatic retry and no request/response body logging.
  9. Provider output is advisory text only. Tool/executable/unknown output items, invalid JSON, origin changes, oversize output or malformed text fail closed and are never partially displayed.
  10. Initial canary limits are one in-flight request, five accepted requests per minute and fifty per day for the single Creator. A provider project spend limit and alert are required before any production key is mounted.
  11. Production activation requires adversarial fake-server tests, a synthetic/public evaluation record, shadow evidence and explicit promotion of the exact binding revision. This ADR alone does not authorize a real key, network egress or production call.

Rationale

The existing provider-neutral boundary makes cognition replaceable and non-canonical. A dated snapshot preserves reproducibility and prevents silent model drift. Mode A enables a narrow explicit advisory use while making the provider's default retention boundary visible and forbidding automatic private context. Isolation and fail-closed mapping preserve the Assistant when cognition is absent or malformed.

Consequences

  • A provider adapter may be implemented under adapters/models and exercised only against adversarial local fake servers before promotion.
  • The cognitive worker still has no canonical owner, database role, Commit/Executor credential or external mutation path.
  • Existing creator-advisory-v1 contracts do not change.
  • The deterministic fake remains the only production binding until a separate evaluation/credential/promotion task succeeds.
  • A ChatGPT or Codex subscription does not substitute for a dedicated API project, API billing controls or API key.

Rollback

Disable or retire the OpenAI binding and resolve all requests to deterministic-fake, returning PROVIDER_UNAVAILABLE. No canonical repair or migration is needed because advisory request/response state is not persisted.

Revisit triggers

  • OpenAI publishes or withdraws a model snapshot.
  • ZDR/MAM is approved for the exact API project.
  • Context excerpts, conversation persistence, tools, streaming or multimodal input are requested.
  • Evaluation shows the snapshot, effort, limits, latency or cost are unsuitable.
  • Provider policy, retention, region, endpoint or origin changes.