vava-nessa/free-coding-models v0.5.54 on GitHub

🛠️ patch-openclaw.js now respects real model metadata (#133) — the script was destructuring only 3 of 5 fields from each nvidiaNim tuple ([modelId, label, tier, sweScore, ctx]), so it dropped the real ctx and fell back to a hardcoded contextWindow per tier (e.g. all S+ models got 128000, even DeepSeek V4 Flash which actually has 1M context). Now it parses the real ctx string ('200k' → 200000, '1M' → 1_000_000) and persists tier, sweScore, and a source field for traceability.
🧭 fcm_router now shows a real label in the TUI (#140) — added a fcm_router entry to PROVIDER_METADATA so every UI surface (Setup wizard, Settings panel, tool launchers, TUI filters) resolves "Smart Router Daemon" instead of the raw key fcm_router (which users were reading as "Unknown"). The local getProviderLabel() in endpoint-installer.js was the only place that knew about fcm_router; it now falls through to PROVIDER_METADATA so all callers stay in sync.

🧹 Closed 3 stale issues that were already fixed in older releases (verified by git log on the relevant code paths):
- #116 — Docker API 404s (fixed in v0.5.28, commit 2c07925)
- #117 — "pending" status confusing (fixed in v0.5.28, same commit)
- #118 — Probe loop silent stall (fixed in v0.5.36, commit f0dd3d8)
🧪 +18 new unit tests for patch-openclaw.js parseCtx/defaultMaxTokens (11) and provider-metadata.js fcm_router lookup (7). All test runs go through pnpm test (no extra setup needed).