Changelog

What's shipping

What shipped, version by version. v0.3.12 is the current release: the router, the safety model and the one-command install are live. Expect this list to grow.

  1. v0.3.12 current September 2026
    • Gemini tool-call history now preserves thought signatures, preventing malformed follow-up requests.
    • Transient Mistral capacity failures stay model-scoped instead of unnecessarily disabling the whole provider.
    • Completion, media context and concurrent cooldown safeguards are more accurate and harder to bypass.
    • Cloudflare Workers AI joins the strict-zero provider catalog, and coding-model qualification is refreshed.
  2. v0.3.11 September 2026
    • Anthropic-compatible clients can use the new Messages API through the same strict-zero routing core.
    • Loopback clients no longer need a generated API token, while exposed listeners remain authenticated.
    • The local Admin Control Center now presents routing activity and operational state more clearly.
  3. v0.3.10 September 2026
    • Bounded production-quality feedback now contributes evidence to model scoring without weakening strict-zero eligibility.
    • The local Admin Control Center has a denser, responsive operational design.
    • Pro installation verifies a signed release attestation before pip executes the wheel.
    • Authentication, streaming termination, cooldown classification and stale-socket recovery are more robust.
    • Source installation now passes editable options to pip in a compatible order.
  4. v0.2.0 August 2026
    • Windows joins macOS and Linux as a service host: a per-user scheduled task that registers and runs without elevation.
    • /health answers from the cached catalog and reports its own freshness, so polling it no longer refreshes every provider on the request thread.
    • Credential stores name the reason the OS refused a key instead of silently writing it to a plaintext file.
  5. v0.1.8 August 2026
    • ficelle demo failover kills a provider mid-call and shows the reroute, so the behaviour everything else rests on is visible in one command.
    • Request budgets are metered per provider, and the daemon learns each account’s own cap instead of assuming a tier.
    • Discovery stops probing a provider for a modality or structured output its own catalog already refused.
  6. v0.1.6 August 2026
    • Provider-aware pacing protects free-tier quotas, and vanished catalog models are pruned safely with an undo trail.
    • Bad client requests no longer cool healthy models, while a saturated model no longer pauses its whole provider.
    • The Requests view streams live, charts real elapsed time, and exports both the OpenAI base URL and full chat URL.
  7. v0.1.5 August 2026
    • Streaming upstream failures now stay inside the fallback loop, and oversized requests use their dedicated cooldown.
    • The Control Center exposes the local endpoint immediately and warns before routing edits are lost.
  8. v0.1.4 August 2026
    • Control Center restyled to match the site: canonical brand marks, shared colour tokens at WCAG-AA, pill controls.
    • Verified in-app updates: ficelle update --check and --install, with SHA-256 verification, backup and rollback.
  9. v0.1.3 July 2026
    • Strict-zero routing by default: free pricing required, tools required, no paid fallback.
    • OpenRouter and Nous strict-zero provider pools.
    • Credential-gated free-tier providers (NVIDIA NIM, Mistral, Groq, Gemini, Cerebras, SiliconFlow), dormant without a key.
    • Cooldowns, quarantine, and the anti-false-free guard.
    • Benchmarks and canaries for tool-call / JSON / vision capability.
    • Local admin UI and machine-readable /admin/status.json.
    • macOS LaunchAgent and Linux systemd --user lifecycle via the ficelle CLI.
    • Standalone generic install, with Hermes as an optional first-class integration.
    • Ficelle Pro licensing and private-wheel upgrade path.