* feat(nvidia): add NVIDIA provider with onboarding flow
Add the NVIDIA build.nvidia.com API as a bundled provider. Default model
is nvidia/nvidia/nemotron-3-super-120b-a12b: first segment is the provider
id, remaining "nvidia/nemotron-3-super-120b-a12b" is the literal upstream
model id (which happens to start with "nvidia/" because NVIDIA is also the
model maker).
Supporting core change: introduce a provider capability flag
nativeIdsIncludeProviderPrefix so providers whose native catalog ids
intentionally include their provider prefix (OpenRouter) opt into self-prefix
dedupe in modelKey, without hardcoding provider names in core. Providers
whose ids merely happen to start with their own name (NVIDIA) leave the flag
unset and get the full <provider>/<model-id> concatenation.
- extensions/nvidia/*: new plugin, catalog, onboarding, tests, docs
- extensions/openrouter/index.ts: declare nativeIdsIncludeProviderPrefix
- src/plugins/types.ts: add field to ProviderPlugin
- src/plugins/registry.ts: populate self-prefix set on registration
- src/agents/provider-self-prefix.ts: sync accessor used by modelKey
- src/agents/model-ref-shared.ts: modelKey consults the flag
- test updates for affected surfaces
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor(model-picker): simplify literal-prefix display to label-only
* fix(model-picker): pass workspaceDir/env to allowlist literal-prefix resolution
* chore: untrack generated baseline JSON artifacts (gitignored)
* fix(nvidia): show literal model ref in picker and onboarding notes
* fix(nvidia): show hint whenever display label differs from stored config
* fix(nvidia): drop redundant hint from Keep current label
* fix(nvidia): restore literal double-prefix display labels
* fix(picker): handle literal-prefix fast path
* fix(picker): show literal keep label
* fix(docs): update nvidia provider docs
* fix(nvidia): update test helper imports
* fix(changelog): add nvidia provider entry
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
DeepSeek models had no provider-policy-api.ts, so materializeRuntimeConfig
filled contextWindow with DEFAULT_CONTEXT_TOKENS (200k) and cost with zeros
for all DeepSeek models. This caused premature session compaction at ~125k
instead of using the full 1M window, and zero-cost display for v4 models.
Add a normalizeConfig surface that hydrates missing contextWindow, maxTokens,
and cost from the bundled DeepSeek model catalog for matching model ids.
Explicit user overrides are preserved.
Fixes#74245
* fix(memory): add LIKE fallback when FTS5 MATCH throws and log silent search errors
When searchKeyword FTS5 MATCH fails (e.g. unicode61 tokenizer rejects
certain query patterns), the search now falls back to a LIKE-based query
instead of silently returning zero results. The four .catch(() => [])
sites in the search orchestrator now log warnings so failures are
visible in diagnostics.
Fixes#74036
* fix(memory): split LIKE fallback into per-token clauses and log MATCH errors