[KooshaPari]

Summary

getActiveProvidersWithSyncedModel in src/lib/db/models.ts only reads from key_value(namespace='syncedAvailableModels') and ignores key_value(namespace='customModels'), so models added through the provider dashboard's "model picker" are written to one store but read from another. Chat dispatch then returns Model '…' is not available in the active live catalog for provider '…' for perfectly valid models that the user just successfully added.

Motivation

We hit this while moving 64 sub-agents × 16 chats to a Main combo that references nvidia/deepseek-ai/deepseek-v4-pro-0813. The model is fully in the live catalog (verified via GET /api/providers/nvidia/models) and is added to combos via the dashboard model picker, yet the dispatch path returns 400 invalid_request_error and excludes the provider with terminalReason: Model 'deepseek-ai/deepseek-v4-pro-0813' is not available in the active live catalog for provider 'nvidia'. The only workaround is to re-run POST /api/providers/{id}/sync-models, which rewrites syncedAvailableModels from the upstream catalog — but that is destructive (it dropped 78 non-free models the first time we ran it) and is not what the user did.

Steps to reproduce:

1. Add any model through the provider dashboard "model picker" (it gets stored in key_value(namespace='customModels'))
2. Add that model to a combo (or reference it via provider/model)
3. Send a chat request through the combo
4. Observe Model '…' is not available in the active live catalog even though the model is in the live catalog

Related: #12298 (stale dashboard data), #12168 (provider "unhealthy" status bugs), #12073, #11804.

Proposed Fix

In src/lib/db/models.ts:528-552 (getActiveProvidersWithSyncedModel and its callers), union the customModels namespace with syncedAvailableModels when building the per-provider live model set. Specifically:

// pseudocode for the fix
const syncedRows = await keyValueGetAll("syncedAvailableModels");        // current behaviour
const customRows = await keyValueGetAll("customModels");                 // NEW
const merged = new Map<string, Set<string>>();
for (const row of [...syncedRows, ...customRows]) {
  const set = merged.get(row.provider) ?? new Set<string>();
  for (const m of parseModels(row.value)) set.add(m);
  merged.set(row.provider, set);
}
return merged;

The dashboard write path (useProviderModels.ts, managedModelImport.ts, POST /api/provider-models) already keeps customModels accurate, so no write-side changes are needed.

Optional follow-up: surface customModels separately in the dashboard so users can see why a model is in scope (e.g. "pinned" vs "discovered via sync"), and add a "resync" button that does not drop pinned models.

Testing

$ sqlite3 storage.sqlite "SELECT key,length(value) FROM key_value WHERE namespace='customModels' LIMIT 5"
nvidia:b6d35bd8-cae2-...   842
nvidia:key-2               842
...
$ sqlite3 storage.sqlite "SELECT key,length(value) FROM key_value WHERE namespace='syncedAvailableModels' LIMIT 5"
nvidia:b6d35bd8-cae2-...   12344  # 82 models

$ curl -X POST http://100.96.135.160:20128/v1/chat/completions \
       -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
       -d '{"model":"nvidia/deepseek-ai/deepseek-v4-pro-0813","messages":[{"role":"user","content":"hi"}]}'
# BEFORE: {"error":{"message":"Model '...' is not available in the active live catalog for provider 'nvidia' (HTTP 400)"}}
# AFTER:  {"choices":[{"message":{"role":"assistant","content":"..."}}]}

Checklist

• bun test src/lib/db/models.test.ts passes (will need new test covering the merge)
• npm run lint passes
• Conventional Commits: fix(db): union customModels with syncedAvailableModels in dispatch path
• Docs updated if behavior changed
• No new debug code

Diff characteristics

• 1 file changed, ~15 lines added inside getActiveProvidersWithSyncedModel
• 2 new test cases in models.test.ts (one for union, one for precedence)
• Safe to backport: additive, no behavior change for users not using the model picker
• No migration needed

[Copilot]

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

[KooshaPari]

Closing per H-C plan (retracting DB dispatch fix that depends on schema migrations not yet in upstream). Will re-cherry-pick after Ravi thread resolves or when upstream merges the migration PRs this depends on. PR pairs with #12614 (i18n), #12612 (CVE), #12592 (deps), #12586 (braces), #12470 (backfill), #12667 (docs) which remain open.

Refs: WAITING-handoff 2026-09-03.