50 lines
3.0 KiB
Markdown
50 lines
3.0 KiB
Markdown
|
# AIUI provider setup follow-up
|
||
|
|
|
||
|
|
Status: implementation in progress; not deployed or accepted.
|
||
|
|
|
||
|
|
The app's browser key vault did not configure the node's authoritative Claude
|
||
|
|
ledger. Embedded chat delegates to the node's tool loop, whose provider selection
|
||
|
|
also ignored the frontend's Claude/OpenRouter picker. The backend retired the
|
||
|
|
OpenRouter relay while that picker still offered it. Generic502/503 errors were
|
||
|
|
classified as missing keys and sent users back to settings.
|
||
|
|
|
||
|
|
Implementation scope: offer setup in trusted dashboard chrome before first use;
|
||
|
|
keep credentials out of the iframe's chat, prompts, history and browser storage;
|
||
|
|
use private atomic node credential writes; persist an explicit provider choice;
|
||
|
|
retain the existing tool permissions and outbound privacy screen. Explicit
|
||
|
|
provider selection must not silently send a failed request to a different cloud
|
||
|
|
provider. Routstr funding and allowance remain separate, deliberate actions.
|
||
|
|
|
||
|
|
OpenAI support uses its standard API key, not a presumed Codex subscription key.
|
||
|
|
For the existing node tool loop, the Chat Completions API retains the same
|
||
|
|
message/tool-result representation and existing egress checks. This is an
|
||
|
|
intentional integration choice, not a claim that it is the newer Responses API.
|
||
|
|
The HTTP adapter has a fixed HTTPS destination, no redirects or retries, a bounded
|
||
|
|
completion and response, store:false, and errors that do not echo upstream bodies.
|
||
|
|
The operator supplies the model ID; no inference is issued merely by saving a key.
|
||
|
|
|
||
|
|
Official documentation checked2026-10-06:
|
||
|
|
- https://developers.openai.com/api/reference/overview (server-side bearer credentials)
|
||
|
|
- https://developers.openai.com/api/reference/resources/chat/subresources/completions/methods/create
|
||
|
|
(max_completion_tokens, tool calls, store)
|
||
|
|
- https://developers.openai.com/api/docs/guides/streaming-responses
|
||
|
|
(Responses recommendation and distinction from Chat Completions)
|
||
|
|
|
||
|
|
Qualification so far: all 1,244 dashboard tests and 363 AIUI tests pass before
|
||
|
|
final toolbar placement/funding refinements; latest focused checks pass 15
|
||
|
|
dashboard and 24 AIUI tests. Both production bundles build. Chromium390/1440px
|
||
|
|
verifies first-use setup, private fixture key/model save, no key in localStorage,
|
||
|
|
retained unsent draft and no page errors. Visual review found an overlapping
|
||
|
|
mobile setup button; moved setup into the existing model menu. That final layout
|
||
|
|
still needs rebuilt-browser qualification. No fixture key reached a real provider.
|
||
|
|
|
||
|
|
Explicit local selection now stays local; Claude/OpenAI selection cannot silently
|
||
|
|
fall through. Routstr selection persists and retains the existing budget checks.
|
||
|
|
Payment-required responses request the funding view, while temporary502/503 and
|
||
|
|
rate limits do not claim credentials are missing. Reopening ecash funding reloads
|
||
|
|
the address, including repeated opens on the same tab.
|
||
|
|
|
||
|
|
Remaining: isolated backend results, final browser/funding/key-error checks,
|
||
|
|
actual provider compatibility and node deployment. No paid inference tests or new
|
||
|
|
wallet spending have been performed or authorized by this implementation work.
|