feat(claude-native): enumerate a gateway provider's live models in the launch picker

[btli]

Related issue

Closes #7875

Summary

The claude-native launch picker is built from the claude -p /model probe, then claude_model_catalog() drops every bare claude-* row whenever ANTHROPIC_BASE_URL's host is not anthropic.com. That hostname heuristic protects the namespaced gateways (Bedrock, Databricks) but wrongly empties the picker for a generic Anthropic passthrough (e.g. a LiteLLM Anthropic passthrough) that serves bare canonical ids — the picker reads "Models unavailable" even though turns route fine.

This mirrors the pi pattern from #7007 / #7713 (closed #7019 / #6185):

• A key/gateway/local anthropic provider records a listing_provider (ResolvedModelProvider) descriptor on its ClaudeNativeUcodeConfig — metadata only, no build-time I/O, compare=False, hash=False (like pi's PiProviderConfig.listing_provider).
• claude_model_catalog() enumerates that endpoint's <base>/v1/models through the shared model-catalog fetcher (_fetch_openai_compatible_listing for a gateway, _fetch_anthropic_listing for a real anthropic key), off-thread and fail-open.
• The gateway's own listing is the source of truth: a bare claude-* id — concrete or a claude-* / * wildcard — marks the gateway canonical, so the probe's claude-* rows are kept and any concrete listed ids are surfaced; a namespaced listing (OpenRouter's anthropic/claude-*) leaves them dropped, since Claude Code can't drive those.
• claude_catalog_fingerprint() keys on the listing endpoint (kind/family/base_url — never the credential) so the persisted catalog re-probes when the configured gateway changes.

This replaces the need for any per-host flag or hostname special-casing: a LiteLLM gateway configured as a providers: entry "just works", and namespaced/unsupported gateways are correctly excluded.

Test Plan

• uv run pytest tests/test_claude_native.py → 396 pass (9 new): _gateway_claude_listing for wildcard / concrete / namespaced (OpenRouter) / unreachable; the descriptor is recorded by _provider_config_for_native_claude; the catalog keeps canonical rows when the listing serves them, drops them when namespaced, surfaces+dedupes concrete ids; and the fingerprint keys on the endpoint. The keep/drop pair differ only in the listing result and produce opposite outcomes (mutation check).
• ruff check / ruff format --check / pyrefly check clean.
• Verified live: the target LiteLLM gateway's /v1/models lists claude-* and POST /v1/messages with claude-sonnet-4-5-… returns 200; OpenRouter's /api/v1/models lists only anthropic/claude-* (no bare claude-*), so it is correctly excluded.

Demo

• Visual demo attached below
• Non-visual evidence provided below or in Test Plan
• Not applicable — no behavioral change

Type of change

• Bug fix
• Feature
• UI / frontend change
• Refactor / chore
• Docs
• Test / CI
• Breaking change

Test coverage

• Unit tests added / updated
• Integration tests added / updated
• E2E tests added / updated
• Manual verification completed
• Existing tests cover this change
• Not applicable

Coverage notes

Manually verified the live gateway listings (LiteLLM lists claude-* and serves claude-sonnet-4-5 over /v1/messages; OpenRouter lists only anthropic/claude-*). Unit tests cover the listing classifier, the descriptor recording, all three catalog outcomes (keep / drop / surface-concrete), and the fingerprint key.

Changelog

Claude Code sessions routed through a configured Anthropic-passthrough gateway (e.g. LiteLLM) now list the gateway's models in the launch picker instead of showing "Models unavailable".

🤖 Generated with Claude Code

Issues

Resolves OMNI-9136

[omni-resolve-agent]

Resolve-agent reviewed this contributor PR as the candidate fix. Future actionable maintainer review feedback may be remediated automatically.

[omni-resolve-agent]

Resolve-agent review outcome: fixed.

Summary
PR #7876 (btli) mirrors pi #7007: ClaudeNativeUcodeConfig gains a metadata-only listing_provider descriptor recorded by _provider_config_for_native_claude; claude_model_catalog() enumerates the endpoint's own /v1/models off-thread and fail-open via the shared model-catalog fetchers — a bare claude-* id (concrete or wildcard) marks the gateway canonical so probe rows are kept and concrete listed ids are surfaced/deduped, a namespaced-only listing (OpenRouter) still drops them, an unreachable/empty listing fails open; claude_catalog_fingerprint() keys on the non-secret listing endpoint. Bedrock/Databricks builders record no descriptor, so the hostname heuristic still protects them

Root cause
claude_model_catalog() classifies the endpoint by hostname alone (_serves_canonical_anthropic_ids / _ambient_env_is_non_anthropic_gateway): any non-anthropic.com ANTHROPIC_BASE_URL host is treated as a namespaced gateway, so every probe row resolving to a bare claude-* id is dropped and the configured default is refused, persisting an empty catalog that the New Chat Claude picker renders as 'Models unavailable' — even for a generic Anthropic-passthrough gateway whose /v1/models lists those exact ids and whose /v1/messages routes them

CI
PR head (fork draft): security/hygiene/DCO/Trivy lanes pass; heavy lanes (Pytest, E2E UI, Pre-commit, Docker) skipping on the draft fork and the Maintainer Approval gate is pending — expected, not diff-caused. Locally on the PR head: repro e2e passes and tests/test_claude_native.py is 399-pass green (ruff/pyrefly unreachable from this sandbox; GitHub CI owns them after publication)

Automated review
no real Polly review has run on this PR (no polly-review-bot comment; fork PRs' automatic check is a credentials-gated no-op). Dispatching one is a GitHub write owned by the workflow publisher — this implementation-only job performed no GitHub writes, so approval is withheld and the review verdict ships as findings for the publisher to post