fix(ollama): route models by advertised capability (#11087) — port of #11088 to the release line

[yourspraveen]

⚠️ base-red inherited: #9985

What

Ports #11088 (merged) onto the active release line. Closes the gap you flagged on #11087:

it landed on main, not on the release/v3.8.50 branch, so it will NOT be in the v3.8.50 package about to publish; it rides the following release line (v3.8.51+)

Same precedent as #11165, which ported #11075 from main onto release/v3.8.50.

Why this is worth porting rather than waiting

release/v3.8.51 does not exist yet, and every PR currently open — 30 of them — is cut from release/v3.8.50. Until the next cycle branch appears, the fix is reachable only from main, which no active branch is based on. Meanwhile the defect is live on the release line: every Ollama Local model is flattened to chat, so /v1/embeddings and /v1/images/generations reject models the daemon advertises as capable.

If you would rather this wait for the v3.8.51 branch, say so and I will close it — the branch will keep.

⚠️ No longer a byte-identical port — #11088 carries two defects

This PR started as a byte-identical port and I described it that way below. CI proved that wrong, so the description is corrected here rather than quietly left standing. Two failures showed up that are not in the base-red set, reproduce on main (which already carries #11088), and do not reproduce on the pristine release tip — i.e. they are defects in the merged change, not in the port:

1. OpenAI import leaks image/video models into chat selections — real behavioural regression.

✖ OpenAI import excludes image and video generation models from chat selections
  actual:   [ 'gpt-5.6-sol', 'gpt-image-2', 'sora-2-pro', 'vendor-image-model' ]
  expected: [ 'gpt-5.6-sol' ]

#11088 dropped filterChatSelectableModels from importManagedModels for every provider, but the compensating read-time filter (resolveLocalSyncedEndpointRoute) is gated on isSelfHostedChatProvider. Non-local providers therefore lost the filter with nothing replacing it, and the models are persisted, not merely returned. Fixed in c018e3a45 by scoping the drop to self-hosted providers — the local behaviour #11087 asked for is untouched and the Ollama suite stays 3/3.

2. An unregistered credential site — hard-lease credential, executor, and connection-query inventory has no unclassified site (src/lib/embeddings/service.ts 2 → 3). The new site resolves through getProviderCredentials with the connection allowlist from resolveLocalSyncedEndpointRoute and handles allRateLimited, so it is fenced like its two siblings — the ledger had simply not been told. Registered with a comment saying why.

main still has both. This PR fixes them only on the release line; main needs the same two-line change. Flagging rather than opening a second PR unasked — say the word and it is up in minutes.

Fidelity (as originally written — see the correction above)

The change set is byte-identical to the merged #11088 — I diffed all 11 files against refs/pull/11088/head after applying, and every one matches:

open-sse/services/combo/autoStrategy.ts                     |   6 +-
src/app/api/providers/[id]/models/discovery/helpers.ts      |  38 ++++
src/app/api/providers/[id]/models/route.ts                  |   3 +
src/app/api/v1/images/generations/route.ts                  |  19 +-
src/lib/embeddings/service.ts                               |  59 +++++-
src/lib/providerModels/managedModelImport.ts                |   6 +-
src/lib/providerModels/modelDiscovery.ts                    |   9 +-
src/lib/providerModels/ollamaCapabilities.ts                |  98 ++++++++++
src/lib/providerModels/syncedEndpointRouting.ts             |  31 ++++
tests/unit/ollama-local-capabilities-routing.test.ts        | 205 +++++++++++++

The ported commit re-derived nothing: applied with git apply --3way at the release tip a7e09eda5, clean, no conflicts. The two-file fix above sits in a separate commit on top (c018e3a45), so the port and the correction stay reviewable independently.

Deliberately not carried over: docs/superpowers/plans/2026-08-23-qdrant-configuration-guidance.md and its specs/ sibling. Those are on main only because they rode in with #11088's squash — they came from #11213 and 855243ab1 already removed them from the release branch. They do not belong here (or on main — see the separate cleanup PR).

The changelog fragment is renumbered to this PR in a follow-up commit, matching how #11165 handled its own fragment.

Verification at the release tip

Check	Result
tests/unit/ollama-local-capabilities-routing.test.ts	3/3 pass
npm run typecheck:core	clean
npx eslint on all 10 changed source/test files	clean¹

¹ One no-restricted-imports hit in src/app/api/providers/[id]/models/route.ts — the pre-existing @/lib/localDb import at line 14, already frozen in config/quality/eslint-suppressions.json with count: 1. Untouched by this port; npm run lint (which applies the suppressions, and is what CI runs) is clean.

Fixes #11087 on the release line.

[yourspraveen]

CI triage — 17 inherited, 1 flake, and 2 that were genuinely mine

The run came back red on all four shards. Breaking it down honestly, because two of these were not base-red:

Inherited (17 assertions). Identical to the set reproduced on the pristine tip and reported at #9985 — CLI_TOOLS ×5, Codex 872000 !== 272000 ×6, chatCore combo skip, search-route, i18n pt-BR ×2, settings-i18n-keys, check-deps. Also hits the maintainer's own #11262 unchanged.

Mine (2 assertions). managed-model-import and hard-session-lease-bypass-inventory. I verified these three ways before concluding anything:

	pristine tip a7e09eda5	main (carries #11088)	this branch
managed-model-import	9/9	8/9	8/9 → now 9/9
hard-session-lease-bypass-inventory	3/3	2/3	2/3 → now 3/3

They fail on main too, which means the port carried them faithfully — they are defects in the merged #11088, not port damage. Details and the main-side reproduction are in #11276; the corrected description is in the PR body above, and the fix is c018e3a45, a separate commit so the port and the correction review independently.

I have to correct something I wrote in the original body: "byte-identical to the merged #11088 … nothing was re-derived" was true of the port commit and is no longer true of the PR. Byte-identical turned out to be the wrong goal once the source change had a bug in it.

One flake, deliberately excluded: ttft() measures first-forwarded-chunk latency failed on one #11146 run and nowhere else — 8/8 on the pristine tip, 3× 8/8 on a branch. Not counted against either PR.

Where this leaves the PR

The two real defects are fixed here and the Ollama capability suite is still 3/3, so #11087's goal is intact. The remaining red is the base, which #9985 tracks and which no PR against this branch can currently escape.

main still carries both defects — #11271 cannot reach it from this base. Say the word and the main-side PR is up in minutes.