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

Four 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".
• codex_models.curated.json — hand-curated Codex entries (facts + judgments), source: "curated".

It also updates the accepted subscription model ids and recommended provider defaults in capabilities.py, plus the canonical API-key provider list in sdk/utils/assets.py.

Refresh workflow

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>.

Two copies of pi-ai are installed, and only one of them is the harness. The runner declares @earendil-works/pi-ai directly at an older version, while @earendil-works/pi-coding-agent — the package that actually runs the agent — carries its own newer copy. Resolve the models file THROUGH pi-coding-agent, so the catalog describes the models the harness really accepts:

# From repo root. `realpath` must do the resolving: the path crosses a pnpm symlink, and a `..`
# left in it would be normalized lexically by node's loader back onto the runner's older direct
# copy — silently regenerating against the wrong pi-ai.
MODELS=$(realpath services/runner/node_modules/@earendil-works/pi-coding-agent/../pi-ai/dist/models.generated.js)
node .agents/skills/sync-model-catalog/generate_pi_models.mjs "$MODELS" \
  sdks/python/agenta/sdk/agents/data/pi_models.generated.json

The generator prints the version it read, and records it in the output's _generator.source. Check that line: a regeneration that reports an OLDER version than the file already carried has read the wrong copy — discard it rather than committing a downgrade.

Detect the bump from a lockfile diff on @earendil-works+pi-coding-agent@<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-5-1[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-5-1: an API-key session offers it bare, and the runner resolves it to claude-fable-5-1[1m] on a subscription session, where that is the only offered variant. Drop an id the pinned build no longer offers at all (Claude Code 2.1.280 dropped claude-fable-5); keeping it publishes a picker option that fails at run time. 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).

4. Refresh Codex and OpenAI

Use OpenAI's official model pages for the current general-purpose lineup, ids, pricing, context, and recommended starting models. Add a new model to supported_llm_models["openai"] and PROVIDER_DEFAULT_MODELS["openai"] only when the official page identifies it as a current general-purpose model.

For Codex subscription models, confirm the model is accepted by the pinned Codex app-server or a live authenticated session. Add it to CODEX_MODELS and codex_models.curated.json. For Pi's ChatGPT subscription path, update the Pi package first; regenerate the catalog, then make PI_SUBSCRIPTION_MODELS["openai-codex"] exactly match the pinned package's catalog.

5. Refresh Anthropic API models

Use GET https://api.anthropic.com/v1/models with the user's key when available. The response is newest-first and supplies id, display_name, capabilities, and token limits. Cross-check the official models overview and deprecation page before changing recommendations or removing an id. Keep the current four product tiers in PROVIDER_DEFAULT_MODELS["anthropic"]; do not remove a deprecated model from the canonical list until Anthropic marks it retired.

6. Refresh OpenRouter by weekly usage

Run:

python .agents/skills/sync-model-catalog/audit_provider_models.py

The script calls OpenRouter's public GET /api/v1/models with sort=most-popular, supported_parameters=tools, and output_modalities=text. Use the first 10 results for both supported_llm_models["openrouter"] and PROVIDER_DEFAULT_MODELS["openrouter"], preserving response order. The endpoint's name field is the human-readable label.

The ranking changes with usage. Review removals before applying them because a saved connection may still use a model that fell out of the top 10.

7. Preserve provider display names

Provider probes keep ids in DiscoveryResult.models for compatibility and return names in DiscoveryResult.model_names. OpenRouter supplies name; Anthropic supplies display_name; Gemini supplies displayName. Save selected names in each model's existing extras.name field so the settings list and model picker still show the friendly name after reload. OpenAI's Models API supplies ids only, so its friendly names come from the curated catalog.

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).

Harness bumps change the Daytona snapshot

A bump of a pinned harness version (Pi, Pi ACP, Codex ACP, Claude ACP) changes a Daytona snapshot build input. In the same commit, bump services/runner/config/sandbox-recipe.json version, regenerate services/runner/images/sandbox/daytona/sandbox-recipe-fingerprint.json, and run the fingerprint test (steps in that folder's README.md, "Recipe version"). The new version is a new snapshot name, which must be built in every Daytona account before the runner that asks for it deploys.

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).
• Jobs 4 and 5 when OpenAI or Anthropic announce a model or deprecation.
• Job 6 weekly or before a release; it is safe to run without a credential.

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