fi v1.0.1
Provider reliability, subscription provider updates, and contributor onboarding improvements. Attached: signed and notarized macOS Apple Silicon DMG and ZIP, unsigned Windows x64 installer, Linux x64 AppImage, and update metadata.
Subscription provider updates
- Kiro: thinking mode, metering and token usage, monthly quota reported as HTTP 429 with the real reset time, local
max_tokensenforcement, and error status on failed tool results. - Cursor: parallel tool calls delivered as one step, turn-end usage, token refresh, clear plan and credit errors, one HTTP/2 session per account, and request timeouts reported as timeouts.
- Copilot: current client and session headers, premium quota reported as HTTP 429 with its reset time, and policy-disabled models hidden.
- Provider
Retry-Afterhints now drive fi's retry delay. - Antigravity and subscription web search: blocked prompts and policy stops end as content filtered; Google Search beside function tools is refused before dispatch; the live model catalog replaces the static list; retry hints on 429/503.
- OpenCode: cached and coalesced workspace catalogs, Google models over the guarded Gemini relay, reasoning controls, clear quota errors for insufficient funds and the free-tier limit, and session headers.
- Codex, Grok and Claude: native Codex User-Agent and thread id, Grok conversation headers and normalized xAI errors, and no
x-client-request-idon Claude.
Provider fixes
- Remove the legacy capitalized DeepSeek entry from FI's default provider list. The standard lowercase
deepseekprovider supports model discovery. - Keep subscription-only providers such as Ollama Cloud in the subscription section instead of the third-party API-provider list.
- Preserve Cline's configured output-token budget when the catalog does not supply model capacity. This addresses Muse requests rejected because
max_output_tokenswas below 16. - Retry transient OpenCode catalog connection failures up to three times, using exponential delays within the existing request deadline. Exhausted retries report the destination host and attempt count.
- Reload partial model catalogs when Retry is selected. Successful provider groups remain usable while the failed providers are queried again.
- Clarify OpenCode free-tier HTTP 403 errors. This update does not grant access rejected by the provider.
- For the canonical OpenCode chat endpoint only, try the Zen chat route after persistent recognized free-tier rejection while preserving the selected model and account. Other endpoints and authentication failures do not use this recovery.
Getting started and contributing
- Explain account connection choices in the README.
- Welcome focused external contributions with FI-specific guidance.
- Improve issue templates with app, platform, provider, and model details.
Builds
- macOS Apple Silicon: rebuilt 2026-11-01 from
cf4c4758a1f22bd681de28a11f4677dd202ec604, which includes the subscription provider updates. Developer ID signed; app and DMG notarized and stapled. - Windows x64 and Linux x64: built from
cfa3b51e941e35dfb45b287120bd68831a3fcf72and do not include the subscription provider updates. Packaging and packaged-runtime smoke checks passed on medicux/fi Actions. Fork build commit78fd8adaafabc4588e18ffb4f51d93a12b979d1adiffers from that source commit only in the repository-specific workflow guard. - Installer hashes are in
SHA256SUMS. Update metadata hashes and sizes match all three platforms.
Validation of the original v1.0.1 build
- Local documentation checks: 21 quick checks and 43 full checks passed.
- Local lint, pre-push typecheck, and secret checks passed.
- OpenCode unit tests: 95 passed. OpenCode session snapshots: 4 passed.
- Shared release version verified across 321 publishable family members.
- Model-picker regression checks: 72 tests passed, including the partial-failure retry snapshot.
Installed macOS v1.0.1 (October build) was verified from its DMG: the third-party picker has lowercase deepseek with model discovery and excludes legacy DeepSeek and Ollama Cloud; Cline Free Muse streamed, created two files, and ran four passing tests; a fresh OpenCode catalog fetch returned 37 models. Switching the same coding session to Codex completed a read-only review and reran all four tests. Windows/Linux checks cover packaged runtimes, not interactive desktop sessions.
Known limitation: OpenCode DeepSeek V4.1 Flash Free generation returned HTTP 403 (POLICY_REJECTED) in the installed app despite successful catalog discovery. Catalog recovery does not establish free-tier inference access.