feat: upstream model autodetection + one long-context rule for every family (v4.8.57)

[askalf]

What

Operator direction (2026-06-10): model availability should be autodetected, and all models should use the same method for context windows.

1. Upstream model autodetection

New src/model-catalog.ts: GET /v1/models now asks api.anthropic.com/v1/models what actually exists (authenticated the same way the request path is — OAuth bearer, or x-api-key in ANTHROPIC_UPSTREAM_API_KEY mode), normalizes the listing (claude-only; legacy claude-3-x dropped; CC-style short ids preferred over dated duplicates; deterministic family-rank/version-desc order; unknown future families kept — the next fable-like launch appears on the next refresh with no dario release), and serves it from a stale-while-revalidate TTL cache (1h, DARIO_MODEL_CATALOG_TTL_MS; failed fetches back off 5 min). Every failure path falls back to the baked list — cold start offline, broken auth, upstream 4xx/5xx, garbage listing — so the route always 200s and behaves exactly as before when upstream is unreachable. Prewarmed at startup so the first client call is answered from cache.

2. One long-context rule

[1m] handling was hand-sprinkled: the listing carried claude-fable-5[1m] but no opus/sonnet variants, and each <family>1m alias was a hand-picked pin — already drifted (#389 bumped opus to 4-8; opus1m silently stayed on 4-7).

Now:

• longContextEligible() is the single rule: every claude family takes a [1m] variant except haiku (real CC never offers 1M haiku).
• withLongContextVariants() generates the advertised variants from that rule — the listing now carries claude-opus-4-8[1m], claude-sonnet-4-6[1m], etc., exactly like fable's.
• <family>1m aliases derive from resolve(<family>) + '[1m]' — the pair can never disagree again.

Wire mechanics unchanged (and already uniform): [1m] stays a client-side label — base id + context-1m-2025-08-07 on the wire, existing billing-rejection auto-retry untouched.

⚠️ Behavior change

opus1m now resolves to claude-opus-4-8[1m] (was the stale claude-opus-4-7[1m]). Want the old target? Send claude-opus-4-7[1m] explicitly.

Tests

• New test/model-catalog.mjs — 59 assertions: the eligibility rule, variant generation/adjacency, upstream normalization (dated-id dedupe, legacy-generation filter, unknown-family retention), family resolution + alias derivation, cache success/fallback/TTL/backoff, x-api-key auth mode, payload shape.
• test/provider-prefix.mjs updated for the derived opus1m.
• Full suite 85/85; tsc clean.

Risk

Low. The hot request path only gains a synchronous cache read in resolveClaudeAlias. /v1/models is the only route that awaits the catalog, and it cannot throw. Offline/air-gapped deployments serve the identical baked list as today.

[github-actions]

Compat test: ✅ PASSED

Ran node test/compat.mjs against dario proxy --passthrough on the self-hosted runner for commit 0182c02270c8cfa2b80ce68b493dcf7beec11e29.

Output

============================================================
  dario Compatibility Validation (--passthrough via dario)
  2026-06-10T21:29:33.439Z
============================================================

--- Anthropic Messages API (Hermes) ---
✅ #1 Anthropic non-stream: "COMPAT OK" | thinking_injected=false | service_tier=-
✅ #2 Anthropic stream: 8 events | order=correct | "**STREAM COMPAT**"
✅ #3 SSE framing: event:/data: pairs correctly paired

--- Passthrough Verification ---
✅ #4 No thinking injection: No thinking block in response (passthrough clean)
✅ #5 Client betas preserved: Client-requested thinking honored (status=200)

--- Tool Use (OpenClaw) ---
✅ #6 Tool use: tool=get_weather | input={"location":"Tokyo"}
✅ #7 Tool use stream: 10 events | tool_use=true | input_json_delta=true

--- OpenAI Compat ---
✅ #8 OpenAI non-stream: "OPENAI COMPAT"
✅ #9 OpenAI stream: 3 chunks | [DONE]=true | "**STREAM OPENAI**"

--- Header Visibility ---
✅ #10 Header visibility: request-id=true | ratelimit=true (12 headers)

============================================================
  RESULTS: 10 passed, 0 failed, 0 warnings
============================================================

Full workflow run