Sync Model Catalog

Sync model catalog

Keeps the agent model catalog honest. The catalog is the curated decoration over each harness's accepted model set: clean labels, one-sentence descriptions, real pricing, and 1-5 ratings, keyed by the id the harness accepts. It is published additively next to the ids-only models map on the harness capability record (capabilities.py).

Design and rationale: docs/design/agent-workflows/projects/model-catalog-schema/{design.md,plan.md}.

What it owns

Three JSON data files under sdks/python/agenta/sdk/agents/data/, loaded by sdks/python/agenta/sdk/agents/model_catalog.py:

• pi_models.generated.json — machine-generated from pi-ai. Objective facts only (name / pricing / context_window / modalities), source: "pi_generated". Never hand-edit.
• pi_models.curated.json — human overlay for the generated file (id -> {label?, description?, ratings?}), merged onto the generated facts at load. Survives regeneration. Its additions list holds whole entries for models the pinned pi-ai release predates (see below).
• claude_models.curated.json — hand-curated Claude alias entries (facts + judgments), source: "curated".

It never edits capabilities.py logic — only these data files.

The three jobs

1. Regenerate the Pi file (on a pi-ai version bump)

The generator reads the pinned pi-ai models.generated for the providers Agenta reaches (the vault-mapped providers plus openai-codex) and emits one entry per model. pi-ai provider names are mapped to Agenta's vocabulary (google->gemini, together->together_ai); ids are <agenta-provider>/<pi-model-id>.

# From repo root. Point at the pinned pi-ai in the runner's node_modules (the .pnpm path includes
# the version — resolve it with the glob).
MODELS=$(ls services/runner/node_modules/.pnpm/@earendil-works+pi-ai@*/node_modules/@earendil-works/pi-ai/dist/models.generated.js | head -1)
node .agents/skills/sync-model-catalog/generate_pi_models.mjs "$MODELS" \
  sdks/python/agenta/sdk/agents/data/pi_models.generated.json

Detect the bump from a lockfile diff on @earendil-works+pi-ai@<version>. The _generator field in the output records the exact pi-ai version. The curated overlay is untouched — only the .generated.json is rewritten, so the merge on load re-applies the human judgments.

Adding a model pi-ai does not carry yet

A model released after the pinned pi-ai snapshot has no generated entry, so it has no catalog entry at all: the picker can only show its bare id, and PROVIDER_DEFAULT_MODELS drops it (a curated default must exist in the catalog or the accepted set). The overlay cannot fix this, because an overlay key only decorates an id the generated file already has.

Put the whole entry in the additions list of pi_models.curated.json instead, with source: "curated" and every fact sourced from the vendor's own pages. Do not hand-edit the generated file. A generated entry of the same id always wins at load, so the addition retires itself the moment a regeneration carries the model. Prune superseded additions after job 1.

2. Sync Claude to the live accepted set (needs a running runner)

claude_models.curated.json must cover the stable request values the Claude picker can send. Probe live sessions by reading the model config options (the same getConfigOptions call allowedModels uses in services/runner/src/engines/sandbox_agent/model.ts), but do not copy one session's set blindly. Account entitlements and promotions can add or remove context-hinted variants such as claude-fable-5[1m] while keeping the same model family.

Use the stable bare canonical id when the runner can safely widen it to the session's hinted option. For Fable, publish claude-fable-5: it matches a bare live option exactly and the runner resolves it to claude-fable-5[1m] when that is the only offered variant. Do not publish the friendly forms fable or fable[1m]; the harness does not recognize that model family under those ids. Requires an authenticated Claude session, so this is a manual/periodic step, not a CI gate.

3. Refresh curated metadata from current public sources (before a release / on demand)

Labels, descriptions, and ratings state a model's current standing, which a language model's training data gets wrong (the Anthropic frontier is Fable 5, above Opus, as of mid-2026). Look up the current lineup, pricing, and relative standing from the vendor's pages and announcements (WebSearch + WebFetch), then propose updated descriptions and ratings for a human to confirm. Never write a rating from memory. Validate the 1-5 range and flag any entry whose facts you could not verify. Ratings: higher is better on every axis; cost is cost-efficiency (5 = cheapest).

Validate

The pydantic loader enforces the schema (including the 1-5 rating range) on load, and the unit test locks coverage and the overlay merge:

cd sdks/python && uv run --no-sync python -m pytest \
  oss/tests/pytest/unit/agents/connections/test_model_catalog.py -q

A malformed data file fails loud there (and at import in capabilities.py, which then publishes an empty catalog rather than crashing /inspect).

When to run

• Job 1 on a pi-ai bump (automatable from a lockfile diff).
• Jobs 2 and 3 before a release or on demand (need a live session and a web lookup).

The skill writes files and a proposal; a human reviews the curated changes and commits.