[Epic] Models/embeddings subsystem

[kriszyp]

Tracking epic for the embeddings/models pipeline: license-enforcement dev-mode carve-out, dimension-mismatch detection, local-embedding dev docs, lazy model pulling, startup backend logging, Ollama auto-detect, multimodal input support, and a derived-caching table for @embed.

---

Filed by Claude (Opus 4.8), issue-backlog triage & grouping pass.

[kriszyp]

Note: #750 is thematically part of this cluster but already lives as a native sub-issue of #510 (scope.models) — left there rather than re-parented, since it already has an established, more specific home.

Filed by Claude (Opus 4.8), issue-backlog triage & grouping pass.

[heskew]

Triage pass over open models-titled issues that had no parent on either epic. Three now linked here:

• models: generation backends drop provider options (Ollama num_ctx) → silent prompt truncation #1775 — generation backends drop provider options (Ollama num_ctx) → silent misbehavior. Backend passthrough gap; sits with the other Ollama items.
• ollama backend: generate() returns empty content for thinking models (response tokens land in 'thinking', never read) #1805 — ollama backend generate() returns empty content for thinking models.
• Models backends: no retry/backoff on transient failures, and @embed surfaces an unhelpful sanitized error #1594 — no retry/backoff on transient backend failures, and @embed surfaces an unhelpful sanitized error. PR fix(models): retry transient backend failures with backoff; actionable @embed errors (#1594) #1770 is in flight for it (draft).

Placed here rather than under #510 because none is a #510 acceptance criterion — they're backend/subsystem quality, which is what this epic groups. Trivially reversible if you'd rather they sit elsewhere.

Two other issues from the same sweep went differently:

• [Models] /v1 gateway: input-validation hardening — 400-vs-500 envelope consistency, json_schema nesting, 'developer' role #1857 (/v1 gateway input-validation hardening) — closed. Every item was already fixed on main by the feat(models): OpenAI-compatible /v1/* gateway as built-in Resources (#631, Phase 4 of #510) #1616 review round; verified against the code, evidence on the issue.
• [Models] toolMode 'auto' does not resolve tools against scope.resources (only caller-supplied toolHandlers) — #510 criterion, untracked #1740 (toolMode: 'auto' resolving against scope.resources) — linked under Add unified model-access API (scope.models) #510, since it is one of its acceptance criteria.

#1235 (5.1 GA readiness: models/vector subsystem audit findings) already has #1658 as its parent, so I left it alone — flagging in case it belongs here instead.

---

🤖 Posted by Claude on Nathan's behalf