ProviderModelMigrator

ProviderModelMigrator moves the final v1 Redux provider/model state into the v2 user_provider and user_model tables. It also migrates provider logos and the legacy Dexie pinned:models value.

Data Sources

Data	Source
Providers and models	Redux state.llm.providers[]
Provider/model settings	Redux state.llm.settings
Pinned models	Dexie pinned:models
Provider logos	Dexie image://provider-<providerId>

Managed CherryAI rows are not copied from Redux. The v2 seeder owns the canonical CherryAI provider/default model, and legacy CherryAI pins are rewritten to that seeded model.

Preset Ownership Projection

Preset-backed rows are deltas, not snapshots. The migration must distinguish released v1 defaults from user changes without consulting the current v2 registry, because the current registry may already have changed.

mappings/v1-provider-model-baseline.json is the pinned final-v1 baseline. It was extracted from v1 revision d316ec5345680f1de511fd6df3a7fbdb3edad151 (v1.9.12) and contains only fields used by provider/model projection. sourceRevision records that provenance. Its provider API flags represent the post-migration-127/129 Redux shape: the three compatibility flags remain on system-provider top-level rows after apiOptions is cleared. Do not replace this comparison with current providers.json, models.json, or provider-models.json values.

Projection rules:

• A provider with presetProviderId = null is custom and keeps its mapped provider configuration. Its models may still match global models.json metadata independently; provider-model overrides remain unavailable without an effective preset provider.
• A preset provider is resolved by providerId first, then presetProviderId. The second lookup is required for custom IDs backed by a known provider type, such as Azure OpenAI.
• Provider fields equal to the final-v1 baseline become null/absent. API features are compared per key, so one changed feature does not freeze baseline-equal sibling flags. Different values remain row-owned deltas.
• Preset-linked custom provider IDs use the final-v1 custom-provider API feature baseline rather than the linked system preset baseline. This preserves explicit post-v1-migration feature choices while dropping untouched custom defaults.
• A model first uses the effective provider's provider-model override when one is available, then matches global models.json metadata. Global matching also applies to fully custom providers. If an effective provider's override has no global entry, the migrator synthesizes the same provider-exclusive preset used at runtime.
• Preset model fields equal to the final-v1 model become null. Every non-null sparse column is authoritative at read time; there is no separate ownership marker.
• If a current preset exists but the model is absent from the final-v1 baseline, ordinary legacy fields have no provable user provenance and remain null. Explicit capabilities[].isUserSelected choices are preserved. endpointTypes is also preserved when the current provider-model registry cannot re-derive the legacy routing metadata, including dynamic models from built-in NewAPI providers and custom providers with legacy type='new-api'. CherryIN models without legacy endpoint metadata restore its prefix routing (anthropic/, google/, or OpenAI-compatible fallback) explicitly.
• The v1 editor's synthetic 0/0 pricing value is treated as equivalent to an absent final-v1 price.

When updating the pinned baseline, use the released final-v1 source revision, not a v2 registry snapshot. Preserve only the legacy fields declared by V1ModelBaseline / V1ProviderBaseline, update sourceRevision, and run the migrator tests that cover baseline-equal values, genuine deltas, custom preset IDs, and provider-exclusive models.

Target Tables and Field Mapping

user_provider

v1 source	v2 target	Transformation
id	providerId	Direct; invalid and duplicate IDs are filtered first
id / type	presetProviderId	Known system IDs use the catalog ID; supported custom provider types link to their preset
name	name	Direct user-owned value
apiHost, anthropicApiHost	endpointConfigs	Convert to endpoint-keyed base URLs, then project preset rows against final-v1 defaults
type	defaultChatEndpoint	Convert through the legacy endpoint map, then store only a final-v1 delta
apiKey, auth settings	apiKeys, authConfig	Normalize keys and provider-specific credentials
legacy API option flags	apiFeatures	Convert supported flags, then store only a final-v1 delta
provider settings	providerSettings	Normalize provider-specific user settings
enabled	isEnabled	Defaults to true

user_model

v1 source	v2 target	Transformation
provider ID + model id	id, providerId, modelId	Build deterministic providerId::modelId identity
effective registry match	presetModelId	Global preset (independent of provider provenance) or synthesized provider-exclusive preset; null for an unmatched custom model
name, description, group	same-named nullable columns	Complete values for custom rows; final-v1 deltas for preset rows
capabilities	capabilities	Normalize capability names; preserve explicit isUserSelected choices
endpoint_type, supported_endpoint_types	endpointTypes	Normalize legacy endpoint aliases
supported_text_delta	supportsStreaming	Defaults to true for custom rows; final-v1 delta for preset rows
pricing	pricing	Normalize to runtime pricing; discard the synthetic empty 0/0 echo
source order	orderKey	Assign fractional keys within each provider

Intentionally Dropped or Re-derived Data

• Preset model fields without a non-null sparse delta are resolved from the current registry at read time. No userOverrides ownership array is stored.
• Registry-only provider endpoint fields (modelsApiUrls, adapterFamily) are not copied into preset rows. A migrated custom relay may retain a main-only adapterFamily hint because no catalog can re-derive it.
• Preset inputModalities, outputModalities, token limits, reasoning, and parameter support are not inferred from the current registry during migration; null delegates them to read-time resolution.
• Legacy isNotSupportEnableThinking has no v2 ApiFeatures target and is dropped.
• Legacy Anthropic web OAuth tokens lived outside this source and are not recoverable; those providers return to the API-key auth path.
• Corrupt identifiers, duplicate rows after the first occurrence, retired providers, invalid pin references, and missing optional logos are omitted as described below.

Endpoint Routing Boundary

Renderer/API endpoint writes contain only user-editable baseUrl values. Legacy adapterFamily routing provenance may exist in the main-process stored shape for custom providers, but it is not part of the shared write DTO. Preset providers always take routing families from the current registry.

Data Quality Handling

Issue	Handling
Missing/empty provider ID	Skip and warn
Duplicate provider ID	Keep the first and warn
Missing/empty model ID	Skip and warn
Duplicate model ID within a provider	Keep the first and warn
Retired provider	Skip and warn
Missing final-v1 preset baseline	Preserve mapped provider values and warn
Invalid pinned model reference	Drop it
Missing optional logo	Keep the provider without a logo

Provider rows and their model rows are inserted in one synchronous withWriteTx transaction. Order keys preserve the prepared provider/model sequence, after the seeded CherryAI row.

Implementation Files

• ProviderModelMigrator.ts — preparation, projection, transaction, pins, and validation
• mappings/ProviderModelMappings.ts — legacy-to-v2 field transforms
• mappings/v1-provider-model-baseline.json — pinned final-v1 ownership baseline
• __tests__/ProviderModelMigrator.test.ts — migration and provenance regressions