Model Registry

The Model Registry is the source of truth for per-model metadata in Hivekeep: context window, input modalities (image / PDF), reasoning support, tool-calling, pricing, and a human-readable label. Providers handle transport (streaming, auth, the live model list); the registry owns the catalogue metadata.

Open it from the Models icon in the left activity bar (admin only) or from Settings → Models → Open the model registry. It is a full-width table of every model across every configured provider.

Where the metadata comes from

Metadata is auto-filled from models.dev, an MIT-licensed, community-maintained database of LLM model metadata. On reconcile, each (provider, model id) is matched against models.dev and the matched entry’s fields are baked into the registry row.

Match confidence is one of:

Confidence	Meaning
exact	The id is in models.dev verbatim.
normalized	Matched after stripping release markers (dates like -2025-08-07, -preview, -latest).
family	Only a base-family match (low confidence), flagged for review.
none	Not in models.dev, flagged for review.

The priority for each field is admin override (pinned) → the provider’s own API hint → models.dev → default, with one exception: when models.dev carries an explicit reasoning effort list for a model, that list wins over the provider’s hint (provider hints are often name-pattern heuristics; the curated per-model list is more accurate). The effective value is computed at reconcile time and stored on the row, so what you see in the table is exactly what an Agent gets at runtime.

Labels

Models are shown by a human-readable label (the models.dev name, e.g. Claude Haiku 4.5) instead of the raw id (claude-haiku-4-5-20251001), everywhere a model name appears: the model picker, the conversation header, etc. The id is always still visible as a secondary line. Set a custom label in a model’s edit dialog; leave it blank to fall back to the models.dev label (and then the id).

Curation: enable / disable

Toggle any model on or off, directly in the table or in its edit dialog. Disabled models are hidden from every model picker, so you can switch off the ones you don’t want and keep only those you care about.

• Disabling is safe: the chat path never blocks, so an Agent already configured on a disabled model keeps working: the model is only hidden from pickers.
• A newly-discovered model with an uncertain match lands disabled and flagged to review. Confirming it (the ✓ on the row, or saving its dialog) clears the flag and enables it.

Use the status filter (enabled / disabled / to-review / unmapped) with the bulk actions (Enable all, Disable all, Confirm reviews) to curate at scale: e.g. filter To review then confirm them all, or filter a provider then disable the lot.

Reviewing an uncertain match

When a match is low-confidence (family) or absent (none), the row is flagged review and disabled. To resolve it:

• Confirm (✓) if the auto-match looks right: clears the flag and enables it.
• Remap (in the edit dialog) to point the row at the correct models.dev entry, searchable across the whole catalogue. Useful for subscription/CLI providers (Claude Pro/Max, Codex) whose ids map onto the base provider’s entries.

Models genuinely absent from models.dev (niche or brand-new) stay unmapped; set their metadata manually, or refresh the snapshot once models.dev has them.

Overriding metadata

Open a model to edit any field: label, context window, max output, pricing, image / PDF / tool-call support, reasoning. Each field you change is pinned: it survives future resyncs (everything else keeps tracking models.dev). Saving an unchanged dialog pins nothing.

• Manual mode freezes the whole row: nothing auto-syncs.
• Reset to auto drops every pin/override and re-derives the row from models.dev.

Keeping it fresh

The bundled models.dev snapshot is baked into the build. Two buttons keep things current without a release:

• Resync: re-match every provider against the current snapshot (picks up new model ids the provider now lists, clears stale rows).
• Update models.dev: download the latest models.dev catalogue (persisted to the data dir, so it survives restarts), then resync. Use this to pick up models that models.dev only just added.

Reconciliation also runs automatically: when a provider is created or (re)validated, and on a periodic background cron.

Reasoning-aware effort selectors

Every place you pick a reasoning effort (the chat composer, Agent settings, cron forms, task dialogs) only offers the levels the selected model actually supports, from the registry’s reasoning metadata. The full ladder is minimal, low, medium, high, xhigh, max:

• A model with an explicit effort list (e.g. low → xhigh on recent OpenAI models) shows exactly those levels.
• A reasoning model with no granularity (e.g. Kimi K2.5) shows a single Enabled toggle.
• A model with no reasoning support shows Off with a hint.
• No model context (project defaults) or an unknown model falls back to the generic ladder; the provider clamps the request to the closest supported level at run time either way.

When you switch a model and its stored effort is out of range, the selector clamps it to the nearest supported level.

Capability-driven upload gating

Because the registry knows each model’s image / PDF support, the chat composer won’t let you attach an image or PDF to a model that explicitly can’t read it: it skips the file and tells you. Files of a type the model accepts (and plain text files, always) pass through. Capability unknown fails open (no block).

API

Method	Endpoint	Description
GET	/api/models	List every registry row (joined with its provider).
GET	/api/models/:id/candidates	models.dev keys for the remap picker.
PATCH	/api/models/:id	Edit a model (pins changed fields; clears review).
POST	/api/models/:id/mode	Switch auto ↔ manual.
POST	/api/models/:id/remap	Re-point at a models.dev entry (or clear).
POST	/api/models/:id/reset	Drop all overrides → fully auto.
POST	/api/models/bulk	Bulk enable / disable / confirm.
POST	/api/models/resync	Reconcile every provider against the snapshot.
POST	/api/models/refresh-snapshot	Download the latest models.dev catalogue, then resync.

See also: Supported Providers, Model Selection, and Token Usage & Cost (the registry’s pricing feeds the cost estimates there).