feat(backend): add LLM registry read cache with pub/sub refresh

[ntindle]

Why

Part 2 of 9 of the LLM registry restack (see #13605 for the schema and full context). This layer gives every backend pod a fast, always-warm view of the registry so later parts (public catalog endpoint, copilot routing, admin API) never touch the DB on hot paths.

What

backend/data/llm_registry/ — re-cut of @Bentlybro's #12359 core layer:

• Two-level cache: L1 module-global dict under an asyncio.Lock; L2 Redis via @cached(ttl=300, shared_cache=True) on the DB fetch. All read functions are sync L1 lookups.
• Read API: get_model, get_all_models, get_enabled_models, get_schema_options, get_default_model_slug, get_all_model_slugs_for_validation, and new get_route(surface, mode, tier) for the routing cells.
• Cross-pod refresh: llm_registry:refresh channel; refresh_llm_registry + long-lived subscribe task wired into the rest_api lifespan (fail-soft startup, clean cancellation on shutdown).

Still dormant: nothing consumes these reads yet.

How (deltas vs the original branch)

• Pub/sub rewritten to sharded SPUBLISH/SSUBSCRIBE via connect_sharded_pubsub_async — prod Redis is cluster mode; classic pub/sub broadcasts to every node. Mirrors event_bus.AsyncRedisEventBus.
• ModelMetadata NamedTuple import from blocks/llm.py replaced with a standalone frozen RegistryModelMetadata — decouples the registry from the module it will eventually replace, and NamedTuples serialize as JSON arrays (wrong for API payloads).
• RegistryModel carries the new schema columns (kind, visibility, minSubscriptionTier, source, catalogRemovedAt, fallbackModelSlug).

Verification

• 30/30 tests pass locally (backend/data/llm_registry/) — record mapping edge cases, refresh/L1 population incl. routes, cache clearing both fetches, sharded pub/sub happy path + handshake filtering + reconnect + close-on-exit
• poetry run format + poetry run lint clean

Checklist

• Data-layer additions have no user-facing routes yet (no user ID checks needed — no user-scoped data involved; registry rows are global config)
• Out-of-scope changes: none

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: bf56286d-46b8-49e8-9b28-0dcb20a45a4c

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feat/llm-registry-v2-core

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 76.01%. Comparing base (30ee04c) to head (2a3a7f0).

Additional details and impacted files

@@                       Coverage Diff                       @@
##           feat/llm-registry-v2-schema   #13606      +/-   ##
===============================================================
- Coverage                        76.02%   76.01%   -0.02%     
===============================================================
  Files                             2688     2688              
  Lines                           204282   204280       -2     
  Branches                         19654    19654              
===============================================================
- Hits                            155305   155278      -27     
- Misses                           44639    44662      +23     
- Partials                          4338     4340       +2     

Flag	Coverage Δ
platform-frontend-e2e	30.98% <ø> (-0.19%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Components	Coverage Δ
Platform Backend	82.83% <ø> (ø)
Platform Frontend	49.97% <ø> (-0.08%)	⬇️
AutoGPT Libs	∅ <ø> (∅)
Classic AutoGPT	28.43% <ø> (ø)

🚀 New features to boost your workflow:

• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[ntindle]

After a post-build design review we pivoted from the DB-registry to catalog-as-code: one canonical, schema-validated catalog file is the source of truth (models + costs + copilot routing), updated by PR (catalog-only diffs may ride hotfix→master; the /review bot covers the lane), propagated by CD, with LaunchDarkly demoted to optional cohort experiments and git history as the audit log. Rationale: the admin machinery mostly bought instant-mutation and non-engineer authoring we don't need, and LD is too unreliable to be the load-bearing control layer.

Replaced by #13622 — the same read interface and L1 cache, loaded from the catalog file; Redis L2 and pub/sub are unnecessary when the data only changes at deploy.