feat(inference): AWS Bedrock direct/SigV4 support + non-bridge profile #1940

Activity

1. github-actions

   added

   state:triage-neededOpened without agent diagnostics and needs triageOpened without agent diagnostics and needs triage

   on Jun 17, 2026
2. st-gr commented on Jul 30, 2026

   

   st-gr

   on Jul 30, 2026

   ContributorAuthor

   More actions

   #1801 and #1802 were closed as duplicates into this issue. Consolidating is the right call — SigV4 signing is item 1 here and {region} substitution is item 4, so this issue is the superset. Two implementation details from #1801 aren't captured above though, and I'd rather they not have to be rediscovered. Recording them here.

   1. ResolvedRoute has nowhere to carry AWS credentials.

   Item 1 says the router should "consume the four AWS credential map entries," but the route struct the router actually forwards with holds exactly one secret. On main, crates/openshell-router/src/config.rs:45:

   pub struct ResolvedRoute {   // config.rs:40
       ...
       pub api_key: String,     // config.rs:45 — single secret, no map
       pub auth: AuthHeader,
       ...
   }

   So SigV4 needs a carrying mechanism before signing can be wired up. Two options:

   • add secret_credentials: HashMap<String, String> to ResolvedRoute, or
   • model the AWS fields explicitly (access key / secret / session token / region).

   The map is more general and keeps future SigV4-authed AWS providers from needing another struct change; explicit fields are more type-safe and make a missing credential a compile error rather than a runtime None. Either works — worth deciding before the signing code, since it determines whether AuthHeader::SigV4 needs to carry service alone or more.

   Note RouteConfig (config.rs:20) also has an api_key: Option<String> — that's the serde config type, distinct from the resolved route, and would likely need the same treatment for file-configured routes.

   2. Test plan for the signer.

   • Sign against the SigV4 test vectors in the AWS docs — fixed canonical-request + date pairs give an exact expected Authorization value, which pins the whole chain (canonical request → string-to-sign → signing key → header) rather than just asserting the header is non-empty.
   • Assert x-amz-security-token is present iff a session token is present. Emitting it unconditionally with an empty value is the easy mistake here, and AWS rejects that.

   One convenient hook: the AuthHeader::None doc comment (crates/openshell-core/src/inference.rs:21-29) already documents that aws-bedrock uses None because "SigV4 signing is deferred to a follow-up PR." That comment should be updated when this lands, otherwise it will contradict the new variant. The injection switch to extend is crates/openshell-router/src/backend.rs:250-258 — currently three arms (Bearer / Custom / None), and signing has to happen before it since SigV4 covers the whole request, not just a header slot.
3. johntmyers commented on Sep 17, 2026

   

   johntmyers

   on Sep 17, 2026

   Collaborator

   More actions

   📋 triage-agent

   Closing this router-specific proposal as superseded. #3195 removed the managed inference router, route protocols, and inference.local; generic proxy-side SigV4 capability landed separately in #1638. If an importable direct-Bedrock profile or AWS event-stream handling is still required, it should be tracked as provider-profile and generic L7 work against provider-native endpoints, without restoring the deleted router design.
4. johntmyers

   closed this as not plannedon Sep 17, 2026
5. purp

   removed

   state:triage-neededOpened without agent diagnostics and needs triageOpened without agent diagnostics and needs triage

   on Oct 1, 2026