(bug) Model picker shows a false "not available" badge, and Bedrock credential failures hang instead of failing fast

[mmabrouk]

Two frontend/UX issues found while probing custom-provider connections: the model picker mislabels some valid vault models as unavailable, and a Bedrock connection with bad credentials hangs for a long time before failing with an unhelpful error.

1. False "model not available" badge

Steps to reproduce

1. Add a vault connection whose model id doesn't encode a recognizable provider family (for example a custom-provider model id that doesn't match a known provider/model shape).
2. Open the model picker for an agent config.

Observed

The model is flagged with a "not available" badge, even though the backend accepts it and a run against it succeeds.

Where this lives

web/packages/agenta-entity-ui/src/DrillInView/SchemaControls/agentTemplate/useModelHarness.tsx (around lines 483-524) — availability is derived from the shape of the model id rather than from what the backend actually supports for the selected harness.

2. Bedrock bad-credentials run hangs instead of failing fast

Steps to reproduce

1. Connect a Bedrock provider with invalid access key/secret.
2. Run against it and time how long it takes to fail.

Observed

The run hangs for roughly 210 seconds before failing, and then surfaces a generic 500 rather than a clear "invalid credentials" error.

Context

Found while probing custom-provider connections end to end against the dev stack (2026-07-07). Surfaced by the new connect-model drawer in #5096. Related: this is one of five connection gaps found in the same investigation — see also the Pi-side, Claude-side, and resolution-ranking issues.

[linear-code]

AGE-3921

[mmabrouk]

Related custom-provider connection issues from the same investigation: #5114 (Pi rejects custom providers pre-run), #5115 (Claude custom gateway key not injected), #5116 (Claude + Vertex AI credentials), #5117 (connection-resolution shadowing/ambiguity).

[mannietech15]

Hi @mmabrouk, I'd like to work on this.

To keep the PR tightly scoped, I'll start with the false "not available" badge issue (#1) — I can see the relevant logic in useModelHarness.tsx around lines 483-524 where availability is being derived from the model id shape instead of the backend's actual support check. I'll fix it to key off the harness/backend response rather than pattern-matching the id.

If it makes sense, happy to also pick up the Bedrock hang (#2) in a follow-up once the first fix is reviewed, rather than bundling both into one PR.

I'll test locally against the repro steps you listed and follow the PR template.

[mmabrouk]

Hey @mannietech15 , please confirm first that you can run the repo in dev mode locally. We had a number of potential contributors failing to run it locally, and as a result unable to test their changes (and record videos).

[mannietech15]

Confirmed — I've got the repo running locally via Docker Compose (with-web + with-traefik profiles), all services healthy, and the app loads correctly at localhost. Ready to reproduce the issue and get started on the fix, and I'll record a video of it working per your note above.

[mannietech15]

Hey @mmabrouk , I have fixed the issue, and this is the PR: #5522. Please review, and I am open to any suggestions.