[diegosouzapw]

What

Generic OpenAI-compatible / custom-node discovery had no metadata-mapping layer for
CLIProxyAPI-style reasoning-level metadata (nothing was ever captured — not "lost").

Owner-scoped this to two layers (not the report's original five):

1. Normalization (src/lib/providerModels/modelDiscovery.ts): detectSupportedThinkingEfforts
   now also parses supported_reasoning_levels (entries as {effort} objects or plain
   strings) and thinking.levels, alongside the existing reasoning.supported_efforts
   handling (feat(providers): capture upstream reasoning.supported_efforts at sync so synced openai-compatible models become selectable with effort levels #7694). A shared parseEffortList helper normalizes/dedupes/canonicalizes
   (max→xhigh, unknown native tiers like ultra pass through), and drops malformed
   entries per-record via Zod safeParse — never throws, never fails a whole sync
   (Hard Rule chore(env): sync .env.example with current .env structure #7). Precedence is preserved and tested: flat supportedThinkingEfforts
   (verbatim) > reasoning.supported_efforts > supported_reasoning_levels >
   thinking.levels.
2. Catalog negotiation (src/app/api/providers/[id]/models/discoveryClientVersion.ts,
   wired into route.ts): a default-OFF, per-connection opt-in
   (providerSpecificData.discoveryClientVersionEnabled +
   discoveryClientVersion) appends client_version to the model-list URL only.
   Unrelated generic OpenAI-compatible upstreams see byte-identical requests unless they
   explicitly opt in; the param never touches any inference URL.

Out of scope per the owner's narrowing (documented follow-ups, not this PR):
persistence-schema, catalog-exposure, combo work (appendSyncedEffortVariants /
mergeComboCapabilities already consume supportedThinkingEfforts generically), and the
token-budget range fields (thinking.min/max/zero_allowed/dynamic_allowed).

How tested

TDD (Hard Rule #18): tests/unit/model-discovery-reasoning-levels.test.ts written first,
confirmed failing (missing discoveryClientVersion module), then made to pass — 13/13
green:

• {effort} object entries, plain-string entries, thinking.levels
• dedupe + max→xhigh canonicalization + unknown native tier retained
• malformed entry inside an otherwise-valid array dropped, no throw
• fully malformed shape degrades to undefined, no throw
• 3 precedence-order assertions (regression guard for feat(providers): capture upstream reasoning.supported_efforts at sync so synced openai-compatible models become selectable with effort levels #7694)
• client_version absent by default / present only when opted in / never when disabled

Also re-ran the pre-existing #7694/#8072 suites
(sync-reasoning-supported-efforts-7694.test.ts,
combo-builder-effort-variants-8072.test.ts) — 24/24 still green, confirming no
precedence regression.

Gates run in-branch:

• node --import tsx/esm --test tests/unit/model-discovery-reasoning-levels.test.ts — green
• npm run typecheck:core — clean
• npm run typecheck:noimplicit:core — pre-existing errors only, none in touched files
• npx eslint on touched files — clean
• npm run check:cycles — clean
• node scripts/check/check-test-discovery.mjs / check-file-size.mjs — OK
• check:complexity-ratchets — unchanged base-red (2167/956, matches documented baseline drift, not introduced by this PR)
• Grep-confirmed client_version only reaches the model-list uniqueEndpoints array, never any inference/executor path

Closes #8347