feat(plugins): support reviewed OAuth AI providers

[LIghtJUNction]

Summary

Reviewed plugin bundles can now declare named OpenAI-compatible AI providers and public OAuth clients through extensions.net.codewhale.providers. The existing provider route, model catalog, Chat Completions client and streaming path consume these declarations; no companion proxy process or provider-specific host patch is required.

The host owns PKCE authorization, ephemeral loopback callbacks, state/issuer validation, secure credential storage and serialized token rotation. Plugin code never receives token material. Credentials bind the exact provider, endpoint and complete OAuth descriptor. Each actual request rechecks the review receipt, binds the effective endpoint/OAuth/public headers to that reviewed declaration, resolves the current bearer and refuses redirects. Candidate routes and in-memory overrides cannot borrow another declaration’s authorization. Disabled/revoked plugins invalidate already-built clients. Runtime API key writes cannot replace plugin OAuth. Client construction and generic config-key reads never expose the bearer. Read-only readiness and diagnostic probes never refresh credentials; reviewed public headers do not inherit global connection headers. Model-only chooser preferences persist through the existing selection owner without persisting plugin authority.

Includes auth plugin-login --provider <id> / plugin-logout, provider/model declarations, an example bundle and authoring documentation. This is a declarative standard public-client PKCE/OpenAI Chat extension; arbitrary executable auth/transport callbacks, device grants, confidential clients and remote revocation are explicitly outside its scope. The provider capability changes the activation policy to 5/6, intentionally requiring fresh review of older plugin receipts.

No-Issue: standalone provider-neutral plugin integration requested by an external integrator.

Testing

421 distinct tests passed on the actual TUI unit-test binary (filters overlap; counts are not additive):

Filter	Passed	Failed	Ignored
plugin_oauth	10	0	0
plugins::	243	0	1
config_persistence::	35	0	0
oauth::	106	0	0
provider_readiness::	26	0	0
runtime_api::secrets::	8	0	0
from_candidate	2	0	0

The host fixture loads an installed, trusted, enabled manifest, stores an expired bound credential, then exercises real HTTP token refresh, model discovery, Chat Completions and SSE with the actual client. It also checks readonly diagnostics, isolation from ambient headers, valid-token endpoint tampering, revocation during a delayed refresh and cached-client/recovery-probe revocation. Plugin tests cover manifest validation, trust/staging, declaration binding and save/reload of model selection.

For this <16 GiB host, the documented package-local low-memory approach was used. For the first four filters:

CARGO_BUILD_JOBS=1 scripts/dev-cargo.sh test -p codewhale-tui --lib --locked \
  --config 'profile.test.package.codewhale-tui.incremental=false' \
  --config 'profile.test.package.codewhale-tui.codegen-units=4' \
  --config 'profile.test.package.codewhale-tui.debug=0' \
  <filter> -- --test-threads=1

The remaining filters reused that same compiled TUI test binary directly with <filter> --test-threads=1. The cold target took 30m06s; the real host OAuth test executed successfully, rather than being inferred from compilation. Runtime tests preceded a final Clippy-only equivalent let-chain cleanup; final static checks cover the cleaned source.

Final static checks passed:

cargo fmt --all -- --check
CARGO_BUILD_JOBS=1 scripts/dev-cargo.sh clippy -p codewhale-tui --lib --tests --locked -- \
  -D warnings -A clippy::uninlined_format_args \
  -A clippy::too_many_arguments -A clippy::unnecessary_map_or
CARGO_BUILD_JOBS=1 scripts/dev-cargo.sh clippy -p codewhale-cli -p codewhale-app-server --locked -- \
  -D warnings -A clippy::uninlined_format_args \
  -A clippy::too_many_arguments -A clippy::unnecessary_map_or
python3 scripts/check-blocking-calls-budget.py
python3 -m json.tool docs/examples/plugins/oauth-provider/plugin.json
git diff --check

Both Clippy runs use the current CI's three allowances (rather than the older five documented in CONTRIBUTING). The blocking-call budget remains 755 sites across 209 files. The compiler invocations also used process-local jemalloc retention settings for this memory-constrained host; no product or system configuration changed.

Full workspace/all-features gates and interactive browser/production authorization were not run locally. These results demonstrate hermetic host integration, not an authorization grant against a live third-party issuer.

Checklist

• plugins/providers.rs adapts the existing manifest/activation owner to the existing named route/model catalog; OAuth extends the current secure-store owner
• Updated docs and a runnable declaration example
• Added tests for protocol, credential rotation and revocation boundaries
• Manual interactive TUI / production OAuth acceptance
• Full workspace Clippy and full workspace suite (upstream CI)

[chatgpt-codex-connector]

You have reached your Codex usage limits for security reviews. Please try again later.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Hmbown]

this is awesome thank you so much for this!