agent-nuvira 1.58.1 → 1.58.2

package/README.md

@@ -2024,6 +2024,7 @@ npx tsc --noEmit
 | **v1.53.0** | Aug 2026 | Per-action "learned from real usage" telemetry everywhere (chat/execute/plan/edit/skill/learn/ci/doctor all write the registry; dashboard panel + `models status --verbose` show who killed/verified what); daily timeline chart in the dashboard; recovery loop (a later real success un-parks + re-verifies a recovered provider); hermetic E2E failover test (`tests/e2e/`) proving "registry learns the block, next pick skips it"; 2,996 tests |
 | **v1.54.0** | Aug 2026 | VS Code extension telemetry attribution — the extension tags every IDE-driven LLM call with `BUFF_TELEMETRY_ACTION` (`ide-chat` / `ide-inline` / `ide-<command>`) so IDE usage gets its own rows in the per-action "learned from real usage" registry log + dashboard panel; 3,002 tests |
 | **v1.55.0** | Aug 2026 | `buff models unblock <provider>` escape hatch — manually release a registry-blocked provider (demote unavailable→unverified, clear quota parks + ledger cooldown, live re-probe with honest `stillBlocked`); also fixed the pre-existing `models status --json` / `refresh --json` flag-shadowing bug; 3,011 tests |
+| **v1.58.2** | Aug 2026 | Models dashboard "Failed to fetch" repeat fix — the panel fetch is now resilient: per-fetch timeouts, independent health/registry fetches (one failing never hides the other), auto-retry with backoff on transient network failures, and fast self-healing re-poll after a failure (no more stuck error banner); 3,161 tests + 48 dashboard component tests |
 | **v1.58.1** | Aug 2026 | Dashboard mirrors the CLI `model explain` guarantees — the Auto Router panel renders the M2.x chips (🎯 fit / 📏 measured·📐 estimated / ⏳ ctx) on every provider row, served by `/api/routing` + `/api/all`; docs sync (`MODELS_EXPLAIN_DEMO.md` dashboard section); 3,161 tests + 45 dashboard component tests |
 | **v1.58.0** | Aug 2026 | Nuvira-Router P2: capability-aware scoring (`routing.capabilityFit` gate), wire-token measured-cost inputs (real `usage` tokens beat the 2,000/500 estimate; `models explain` shows 📏 measured vs 📐 estimated), multi-account key rotation (multiple `apiKeys` per provider, dead accounts parked + skipped, `tests/e2e/key-rotation`), governance constraints (`routing.governance` — provider/model allow-allow & deny lists, admin per-call max-cost cap, PII-domain block; hard policy violations refuse to serve with `PIIPolicyError`/`GovernancePolicyError`), context-length preflight (`routing.contextFit` — nominal window vs estimated payload, `⏳ ctx N%` chip in `models explain`); also ships the dashboard `--force` + `--port` fixes previously staged as v1.57.0 (never published); 3,159 tests + 42 dashboard component tests |
 | **v1.57.0** | Aug 2026 | `buff dashboard --force` — detect a STALE dashboard on the port (API/SSE mismatch: /api/model-registry answers SPA HTML instead of JSON) and offer to restart it: probe classifies port state (current dashboards and non-dashboard processes are never touched), confirms, finds the PID, kills it, waits for the port to free, and re-binds a fresh server; also fixed the pre-existing `--port` bug — the server bound the import-time 3030 default, now it resolves the bind at call time from explicit overrides; 3,032 tests + 42 dashboard component tests |

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-nuvira",
-  "version": "1.58.1",
+  "version": "1.58.2",
   "description": "Agent-Nuvira: Multi-agent AI coding CLI — plan, write, review, test, and publish code with local models (Ollama) or cloud APIs (Groq, NVIDIA NIM, Google Gemini, OpenRouter)",
   "type": "module",
   "main": "dist/index.js",