Skip to content

CHANGELOG.md

Latest commit

 

History

History
858 lines (597 loc) · 71.7 KB

CHANGELOG.md

File metadata and controls

858 lines (597 loc) · 71.7 KB
Raw

fusionAIze Gate Changelog

v2.7.0 - 2026-08-21

Added

  • Per-model max_input_tokens registry — the provider catalog previously advertised the same flat 262144 input-token floor for every provider. The gateway now records the authoritative input ceiling for its 23 binding model IDs (_MODEL_INPUT_CAPS + get_model_max_input_tokens() in faigate/provider_catalog.py), grounded in the LiteLLM and OmniRoute provider registry reports. Routing (faigate/router.py) resolves the request model's real cap from ctx.model_requested, so an oversized input is rejected at the true model boundary instead of at the uniform floor.

Changed

  • Routing fit and ranking now use the model-specific input cap (get_model_max_input_tokens) in both the dimension-fit filter and the dimension-scoring path, overriding the provider-wide floor when the request model is known.
  • _max_input_token_cap docstring corrected: the provider-wide 262144 value is documented as a floor, not a measured per-model truth. The 413 "advertised" threshold behavior is unchanged.

v2.6.1 - 2026-05-04

Added

  • Cohere provider — Command A/R via OpenAI-compat endpoint (${COHERE_API_KEY})
  • Mistral provider — Mistral Large via OpenAI-compat endpoint (${MISTRAL_API_KEY})
  • Catalog entries for Cohere (catalog.v1.json + provider_catalog.py)

Fixed

  • Removed duplicate cohere block in commented section (active block now lives with other API-key providers)
  • last_reviewed date refreshed on Cohere and Mistral catalog entries

v2.6.0 - 2026-05-04

Added

  • Free provider catalog — 7 zero-cost providers: Pollinations (no API key required), Groq (ultra-fast LPU, 14.4K RPD), Cerebras (wafer-scale, 1M TPD), LongCat (Flash-Lite, 50M tokens/day), NVIDIA NIM (129 models, 40 RPM), Kiro (OAuth via AWS Builder ID, Sonnet 4.5 unlimited), Qoder (OAuth via Google, kimi-k2-thinking unlimited). All providers are commented out in config.yaml by default; operators opt in.
  • Circuit breaker system (faigate/breakers.py) — per-provider state machine (CLOSED → OPEN → HALF_OPEN) with configurable failure threshold, cooldown, and jitter. Integrated into the provider dispatch loop in main.py. State is persisted in the metrics SQLite database (circuit_breakers table). Auto-closes on successful HALF_OPEN probe.
  • Cockpit API endpoints (/api/cockpit/*) — 6 endpoints: GET /health (provider health + circuit snapshot), GET /providers (full config view), GET /circuits (breaker state), POST /circuits/{provider}/reset (force-close), GET /stats (totals, 24h counts, top providers), GET /routes/log (recent request log).
  • Terminal cockpit TUI (faigate/cockpit_tui.py) — Textual-based 4-tab dashboard (Dashboard, Providers, Circuits, Routes) with auto-refreshing data from /api/cockpit/* endpoints. Agent mode (faigate cockpit --agent <endpoint>) outputs JSON for script/programmatic consumption.
  • Circuit breaker response headers — X-faigate-Circuit and X-faigate-Circuit-Provider added to all chat completion responses for client-side circuit awareness.

Changed

  • Routing mode updates — coding-free, coding-fast, and eco profiles now include free providers (pollinations, groq, cerebras) as backup chains.
  • Catalog freshness — last_reviewed dates refreshed across provider catalog entries.
  • Groq and Cerebras catalog entries migrated from track: stable to track: free with updated recommended models and notes.

Fixed

  • Duplicate groq and cerebras entries in _CATALOG merged into single entries.
  • catalog-stale alert false positives resolved by bumping review dates to 2026-05-04.
  • Mock Namespace in test_main_cli updated for subparser compatibility.
  • External offerings/packages catalog env var isolation in test fixtures.

v2.5.0 - 2026-04-27

Fixed

  • Router now honours adaptive hard cooldowns (faigate/router.py): _provider_matches_policy and _validate_health previously consulted only the binary provider.health.healthy flag, which flips back to True after a single successful health probe even if user-facing requests keep failing 4xx. The adaptive RoutePressure tracker (faigate/adaptation.py) classified those persistent failures (auth-invalid, quota-exhausted, model-unavailable, endpoint-mismatch, rate-limited) and recorded a hard cooldown window, but the router never read it — so a structurally broken provider sitting at the top of prefer_providers would be re-selected on every request for the entire cooldown window, burning the user's request budget and producing the same 400 in a tight loop. New helper _provider_in_hard_cooldown(name, ctx) is now consulted in both the policy candidate filter and the fallback-chain iteration. Soft-degrade windows (transport-error, timeout) intentionally stay routable and continue to be handled via the additive adaptation_penalty in ranking. Direct/explicit-model routing (model: "openai-codex" etc.) is unaffected by design — explicit caller intent overrides the adaptive demotion. New unit tests in tests/test_router_cooldown.py cover policy exclusion, recovery after cooldown expiry, and primary→fallback transition. Live regression seen on 2026-04-26: the chatgpt.com Codex backend rejected gpt-5-codex for ChatGPT Plus accounts with HTTP 400, and the unguarded prefer-providers path replayed the same broken provider 5+ times across multiple requests until the operator manually intervened.
  • Codex model-alias mapping is now ChatGPT-plan-tier-aware (faigate/providers.py:_codex_effective_model): the previous mapping unconditionally translated gpt-5.4 → gpt-5-codex upstream. The chatgpt.com Codex backend gates gpt-5-codex to Pro accounts only — Plus accounts must request gpt-5-codex-mini, which produced the HTTP 400 noted above. The new helper _detect_codex_chatgpt_plan_tier() parses chatgpt_plan_type from the cached ~/.codex/auth.json id_token JWT once per process lifetime and routes gpt-5.4 to the variant the account can actually use (gpt-5-codex for Pro, gpt-5-codex-mini for Plus, raw pass-through for unknown/free so the upstream rejects explicitly rather than us guessing wrong).
  • DeepSeek V4 multi-turn safeguard for openai-compat clients (faigate/providers.py): DeepSeek V4 enabled thinking mode by default and made reasoning_content round-trip on assistant messages mandatory when thinking is active. Any openai-compat client that does not track reasoning_content (OpenCode, Codenomad, generic SDKs) breaks on every follow-up turn with HTTP 400 "The reasoning_content in the thinking mode must be passed back to the API.". The send-path now inspects assistant messages and, when any assistant turn lacks reasoning_content, sets body["thinking"] = {"type": "disabled"} (V4 expects an Anthropic-style ThinkingOptions struct, not a boolean — enable_thinking: false is silently ignored). Single-turn reasoning still flows through unchanged. Provider/request extra_body overrides win because they merge in after the auto-set.

Changed

  • DeepSeek V3 → V4 provider migration in the bundled config.yaml: the legacy deepseek-chat and deepseek-reasoner provider blocks are replaced by deepseek-v4-flash (workhorse, 1M ctx, $0.028/$0.14/$0.28 cache-hit/miss/output per 1M tokens) and deepseek-v4-pro (premium reasoning, 1M ctx, $0.145/$1.74/$3.48). Lane canonical models become deepseek/v4-flash and deepseek/v4-pro. All inbound references in prefer_providers, model_shortcuts, static_rules, and degrade chains have been swept. The DeepSeek API still exposes only these two model IDs (GET /v1/models confirmed live on 2026-04-26).
  • Opencode/coding profile auto-router denies Codex providers: added deny_providers: [openai-codex, openai-codex-mini, openai-codex-spark, openai-codex-{high,xhigh,low}, openai-codex-5.4-{low,medium,high,xhigh}] to the opencode and coding client profiles. Explicit invocation via model_shortcuts still works — this only stops the auto-router from picking codex as a long-context fallback when the preferred lanes overflow. With the per-account plan-tier fix above, the deny list is belt-and-suspenders, but it stays as defence-in-depth against future ChatGPT-side gating changes.

v2.4.0 - 2026-04-26

Added

  • Metadata catalog sync loop: the gateway now runs a lazy startup refresh plus an optional 24h background tick for the public/private metadata catalog. Failed syncs back off through 5m/15m/1h windows and persist sync state alongside the cached catalog.
  • Catalog sync alerts: provider catalog, health, dashboard, and API surfaces now feed sync failures, stale caches, invalid payloads, and auth issues through the existing build_catalog_alerts alert path.
  • Kilo Auto profile examples: the shipped config and OpenCode example now expose kilo-auto/frontier, kilo-auto/balanced, kilo-auto/free, and kilo-auto/small for FAIGate-backed OpenCode/Codenomad use.
  • Fresh bundled catalog snapshot: refreshed the embedded metadata catalog with newly verified Opus 4.7, Haiku 4.5, GPT-5.5, DeepSeek V4, and OpenRouter Auto pricing metadata.

v2.3.0 - 2026-04-19

Added

  • Brand-first quota widget (/dashboard/quotas): replaces the flat package list with one card per brand (Claude, DeepSeek, Kilo, …). Cards are sorted worst-alert first so the thing about to break is at the top. Each card stacks its sub-packages (session vs weekly, pay-as-you-go vs credits) with a pace marker on every window-based bar and an identity line (OAuth · claude-code, API key · ${KEY}) so you can tell which account feeds the meter.
  • Per-brand detail view (/dashboard/quotas/<brand_slug>): a focused page that shows the same quota panel plus a 24h totals strip, clients-by-profile table, routes/lanes breakdown, and an hourly sparkline. Shares CSS variables with the overview so scanning between them needs no retraining.
  • Read-only brand endpoints (/api/quotas/<slug>/clients, /routes, /analytics): the data feed behind the detail view. 404 on unknown brands (distinguishes "typo" from "no traffic yet"); analytics clamps hours=1..168 and days=1..90 to prevent URL-typo DB scans. Catalog surfaces catalog_tagline so the "Available to add" mini-block can show tier/price/quota shape at a glance.
  • Default landing view (dashboard.quotas.default_view in config.yaml): three options — overview (default), brand:<slug>, cockpit. GET /dashboard/quotas honors the setting via 302; ?view=overview is an always-available escape hatch. A Pin as Home / 📌 Home button sits on every brand card and the detail-page header — one-click promotion, no modal. Writes go through ruamel.yaml round-trip so the 220+ operator comments in a real config.yaml survive a pin toggle (yaml.safe_dump would flatten them). GET/POST /api/dashboard/settings drives the widget and is available to external consumers.
  • Gate Bar 0.1 (macOS menubar companion) at apps/gate-bar/: SwiftUI MenuBarExtra app that reads the same /api/quotas feed and shows fAI · 83% in the menubar (tightest window across all brands, colour-coded). Popover renders brand cards with the web widget's visual vocabulary; footer links to Dashboard ↗ (server-side redirect honours default_view) and Cockpit ↗. Preferences: gateway URL, Cockpit URL, refresh cadence (manual / 1 / 2 / 5 / 15 min). 13 Swift-Testing tests green on both Xcode.app and Command Line Tools via scripts/swift-test.sh. Read-only — every write path links out to the Operator Cockpit. Sparkle auto-update + notifications + code signing + Homebrew cask tracked for 0.2+ (apps/gate-bar/README.md).

Changed

  • Quota widget groups by provider_id first and gates on credential availability so brands without a resolvable API key / OAuth token land in the collapsed "Skipped" block instead of showing a perpetually-empty bar.
  • QuotaStatus payload carries brand, brand_slug, pace_delta, and identity on every package (v1.3 catalog schema). Older v1.2 catalogs still decode through the brand-fallback table in quota_tracker.py.

Upgrade notes

  • New runtime dependency: ruamel.yaml>=0.18.6 (already added to requirements.txt + pyproject.toml). Fresh pip install faigate or brew upgrade faigate picks it up automatically.
  • The Gate Bar macOS app is a separate artifact. v0.1 is source-only — build it with cd apps/gate-bar && ./scripts/install-local.sh for local testing. A notarized Homebrew cask ships with Gate Bar 0.2.

v2.2.3 - 2026-04-18

Fixed

  • Kilo balance polling switched to tRPC: the v2.2.0 implementation probed four REST URLs that Kilo never shipped (all 404/308). Replaced the probe-list with a single tRPC batch call to https://app.kilo.ai/api/trpc/user.getCreditBlocks,kiloPass.getState,user.getAutoTopUpPaymentMethod?batch=1 — the same endpoint CodexBar uses. Correctly sums amount_mUsd across credit blocks for total, uses totalBalance_mUsd for remaining, and converts mUSD → USD. Live-verified against a real account: balance and block-level expiry now visible in /dashboard/quotas.

v2.2.2 - 2026-04-18

Fixed

  • Quota poller ignored concrete provider instances (faigate/quota_poller.py): the dispatcher compared provider_id against literal "deepseek" / "kilocode", so catalog entries using the real router-facing IDs (deepseek-chat, deepseek-reasoner, kilo-sonnet, kilo-opus) silently fell through with "no balance fetcher for provider X". Introduced _provider_family() that collapses concrete IDs to balance-polling families, and updated both the fetcher dispatch and the env_map API-key lookup (now recognises KILOCODE_API_KEY for the kilo family).

v2.2.1 - 2026-04-18

Fixed

  • Slow brew upgrade / pip install caused by source-only pydantic-core: pyproject.toml now caps pydantic, fastapi, httpx, and uvicorn to version ranges that ship prebuilt wheels for all supported Python versions (3.10–3.13) and platforms (incl. macOS 15 Tahoe arm64). Without these caps pip could drift onto a bleeding-edge pydantic-core release with no wheel, forcing a 3–5 minute Rust/cargo source build during upgrade. This is a packaging-only change — no runtime behavior difference.

v2.2.0 - 2026-04-18

Added

  • Unified QuotaStatus abstraction (faigate/quota_tracker.py): single view covering three package types — credits, rolling_window, daily — with EWMA burn rate, runway projection, and a five-level alert classifier (ok / watch / topup / use_or_lose / exhausted)
  • Background balance poller (faigate/quota_poller.py): refreshes provider balances on a fast lane (15m) for packages within 14 days of expiry and a default 1h otherwise; covers DeepSeek via /user/balance and probes Kilo across a candidate endpoint list; atomic JSON persistence; disabled by default (opt-in via quota_poll.enabled)
  • Passive header-capture middleware (faigate/quota_headers.py): dialect-aware parser for x-ratelimit-* (OpenAI / DeepSeek / OpenRouter) and anthropic-ratelimit-* families; hooked into ProviderBackend.complete so rolling-window packages get a free, near-realtime quota signal
  • Operator cockpit endpoints: GET /api/quotas returns the full QuotaStatus list plus latest header snapshots; GET /dashboard/quotas is a self-contained HTML page with color-coded progress bars sorted by alert urgency, polling every 60s
  • 7-provider catalog template at docs/examples/fusionaize-metadata-repo/packages/catalog.v1.json covering Kilo, DeepSeek, Blackbox, Anthropic Pro, OpenAI Plus, Qwen, and Gemini (free + Pro + CLI + Antigravity) with documented field semantics
  • quota_poll config block with sane defaults (enabled: false, interval_seconds: 3600, fast_lane_interval_seconds: 900)

Changed

  • Router package scoring now picks up QuotaStatus.alert == "use_or_lose" and applies a +3 boost so expiring packages with real burn pressure decisively win ties over static-runway packages
  • Router now awards rolling_window and daily package types up to +5 points based on remaining ratio (parity with credits, no provider class dominates solely by type)
  • faigate-install refuses to run alongside an existing Homebrew formula / service or a loaded LaunchAgent / systemd unit; --force override available; defensive timestamped backups before any write

Fixed

  • faigate-install SIGPIPE bug: install now uses explicit if/elif file-existence checks instead of ls | head -1, which under set -euo pipefail would silently kill the script when head closed the pipe

v2.1.6 - 2026-04-08

Fixed

  • Codex function tool calling: Codex-backed requests now forward OpenAI-style tools and tool_choice to the ChatGPT Codex responses endpoint and translate returned function-call events back into OpenAI-compatible tool_calls, so tool-using clients like Codenomad can execute MCP-style tool flows instead of only seeing text or pseudo-JSON

v2.1.5 - 2026-04-08

Fixed

  • Codex tool transcript handling: assistant tool-call turns and role: "tool" follow-up messages are now normalized before they are sent to the ChatGPT Codex responses endpoint, so Codex-backed clients such as Codenomad no longer fall back just because a tool result appears in the message history

v2.1.4 - 2026-04-08

Fixed

  • Codex OAuth provider startup gating: OAuth-backed Codex providers no longer get skipped during startup just because they do not use a static api_key
  • Codex helper resolution in Brew/libexec installs: faigate-auth is now resolved robustly from the packaged runtime so Codex OAuth refresh works in Homebrew-managed installs
  • Codex responses endpoint handling: explicit empty chat_path values are preserved, so Codex requests stay on chatgpt.com/backend-api/codex/responses instead of drifting back to /chat/completions
  • Codex streaming compatibility: stream=true now uses the same responses adapter as non-stream requests and returns OpenAI-compatible SSE chunks without falling back to another provider

v2.1.3 - 2026-04-08

Added

  • Claude Code OAuth: reads Claude Code credentials from the macOS Keychain, refreshes expired tokens, and injects the required Anthropic OAuth headers on requests
  • Antigravity + Gemini CLI OAuth: activates Google OAuth-backed antigravity and gemini-cli providers via cloudcode-pa.googleapis.com/v1internal

Fixed

  • Qwen Portal inference: corrected the portal base URL, request headers, and required system-message behavior for OAuth-backed inference
  • Gate-namespaced Codex models: model IDs such as faigate/openai-codex-5.4-medium now normalize correctly before provider routing
  • Local artifact hygiene: .codenomad/ and catalog_output.json are no longer tracked release artifacts

v2.1.2 - 2026-04-07

Fixed

  • Codex chat_path bug: openai-compat backend appended /chat/completions to the Codex endpoint URL. Codex base_url IS the full endpoint; per-provider transport bindings now set chat_path: "" for all openai-codex* providers
  • faigate-auth JSON output: OAuthBackend._run_helper() expects JSON on stdout; CLI was printing human-readable text → json.JSONDecodeError on every Codex request. Fixed: token data now printed as JSON to stdout; status messages go to stderr
  • max_tokens limit: OpenCode model limit was 32768, Codex endpoint maximum is 8192 → invalid_request_error. All Codex models now capped at 8192
  • Fallback to paid providers: Removed Sonnet/GPT-4o from Codex degrade_to chains
  • oauth backend validation: Added oauth to _SUPPORTED_BACKENDS in config.py

Added

  • GPT-5.4 reasoning effort control: default_extra_body mechanism in providers.py injects provider-configured fields into every request. New providers: openai-codex-5.4-xhigh (extra_high), openai-codex-5.4-high (high), openai-codex-5.4-medium (medium), openai-codex-5.4-low (low)
  • Codex 5.3 reasoning variants: openai-codex-xhigh → gpt-5.3-codex-xhigh, openai-codex-low → gpt-5.3-codex-low
  • All 20 Codex model variants in _CANONICAL_MODEL_LANES (gpt-5.4, gpt-5.3-codex family, gpt-5.2, gpt-5.1 series, gpt-4.1, gpt-4o, o4-mini)
  • OAuth transport profile: Default transport binding for oauth backend with chat_path: ""
  • OpenCode model menu: 10 Codex model entries covering all reasoning levels

v2.1.1 - 2026-04-06

Added

  • OpenAI Codex OAuth: full ChatGPT OAuth implementation — reads ~/.codex/auth.json, Auth Code + PKCE login flow on port 1455, single-use refresh token handling, JWT exp claim for expiry detection. Inference via chatgpt.com/backend-api/codex/responses. Run: faigate-auth openai-codex

Fixed

  • registry.py: corrected openai-codex base URL to chatgpt.com/backend-api/codex/responses (was incorrectly set to api.openai.com/v1)
  • config.yaml: corrected OAuth endpoints to auth.openai.com/oauth/token and client ID to app_EMoamEEZ73f0CkXaXp7hrann

v2.1.0 - 2026-04-05

Added

  • Shell parity & complete provider coverage: CLI deep-links, --suggest and --link flags, full LLMAIRouter and KiloCode model-level lane coverage (v2.0.0 baseline)
  • Qwen OAuth: production-ready integration via qwen-code CLI credentials; device-code flow, token refresh, coder-model + qwen-code alias
  • Antigravity OAuth: Authorization Code + PKCE flow; reads ~/.gemini/oauth_creds.json; calls Google Generative Language API directly (not through local gRPC proxy)
  • Google Antigravity provider: full registry, catalog, and lane-registry integration for ag/ model family (Claude Opus/Sonnet 4.6, Gemini 3.x variants)
  • CI Gate: single required status check consolidating test, lint, package, and forbid-artifacts; admin bypass disabled
  • Auto-merge bot: PRs merge automatically once CI Gate passes (gh pr merge --auto --squash)
  • pre-commit enforcement: validate-config step in CI; pre-commit install required in onboarding

Changed

  • Antigravity base_url resolved to generativelanguage.googleapis.com/v1beta/openai (was incorrect endpoint)
  • forbid-artifacts check consolidated into ci.yml (removed separate repo-safety.yml)
  • Release title convention enforced in notify-tap workflow (fusionAIze Gate vX.Y.Z)

v2.0.1 - 2026-04-04

Added

  • OAuth wrapper for managed providers: token store, generic OAuth backend, device-code flows for Google, Qwen, and Antigravity; claude_code_oauth() reads token from local claude CLI settings
  • Antigravity provider: full registry, catalog, and lane-registry integration for ag/ model family (Claude Opus/Sonnet 4.6, Gemini 3.x variants via Google Antigravity gateway)
  • Local worker GPU metrics: probe GPU/VRAM usage from Ollama (/api/ps) and vLLM (/metrics); GpuInfo surfaced in discovery output and provider config
  • Dynamic model enumeration: dynamic_models field on DiscoveredWorker; discovered models preferred over static defaults in generate_provider_config
  • Grid worker discovery: reads ~/.faigrid/config.json (JSON format) with fallback to legacy key=value state file
  • Per-client budget limits: cost_limit_usd_day and cost_limit_usd_month fields in client profile config; HTTP 429 returned before routing when threshold is reached
  • Anomaly detection: MetricsStore.get_anomalies() compares recent window to rolling baseline for error rate, latency, cost, and traffic spikes
  • Alerts API: GET /api/alerts with configurable lookback_hours and baseline_hours parameters

Changed

  • google-vertex renamed to google-gemini-cli in registry and catalog (alias preserved for backward compatibility)

v2.0.0 - 2026-04-03

Added

  • Shell parity and intelligence: CLI commands now integrate deeply with dashboard
    • --suggest argument analyzes metrics to recommend relevant CLI commands
    • --link generates dashboard deep‑link URLs with filters preserved
    • All CLI commands (overview, recent, daily, trends) show dashboard links
    • Filter arguments (--provider, --modality, --client‑profile, etc.) work across commands
    • Dashboard links include matching filters for seamless CLI→dashboard navigation
  • Safe config workflows: New faigate-config CLI for config management
    • preview: Preview config changes before applying
    • diff: Show detailed config differences
    • apply: Apply config changes with backup and confirmation
    • validate: Validate config syntax and structure
  • Clipboard integration: --copy flag copies dashboard URLs to clipboard (macOS/Linux/Windows)
  • Scope suggestions: CLI suggests relevant commands based on metrics analysis (failure rates, provider concentration, costs, recent activity)
  • Local worker auto‑discovery: faigate-config discover automatically detects local AI workers (Ollama, vLLM, LM Studio, LiteLLM) and suggests configuration snippets
  • Complete provider coverage: All LLM AI Router custom endpoints now represented in the provider catalog
    • Added missing providers: xAI, Z.AI, Mistral, Groq, HuggingFace, MoonshotAI, MiniMax, Volcano Engine, BytePlus, Qwen, OpenAI Codex, OpenCode Zen, Cerebras, GitHub Copilot, Synthetic, Kimi Coding, Vercel AI Gateway
    • Generic provider support (OpenAI, Anthropic, Google) with config examples
    • KiloCode model‑level access: individual catalog entries for kilo‑auto/frontier, kilo‑auto/balanced, kilo‑auto/free
    • Consistent recommended_model values across all providers
  • Local worker examples: Commented configuration templates for Ollama, vLLM, LM Studio, LiteLLM in config.yaml
  • Enhanced provider catalog: 41 curated provider entries (up from 17) with official source URLs, signup links, and volatility ratings

Changed

  • CLI help text updated with new arguments and examples
  • Dashboard deep links use proper URL encoding and parameter validation
  • Existing CLI commands remain fully backward compatible

v1.21.0 - 2026-04-02

Added

  • Lane family decision factors table in dashboard Routes view with selection path breakdown per family
  • Selection path categorization (same‑lane fallback, downgrade, primary) with colored pills
  • Routes KPIs extended with same‑lane fallback and downgrade traffic percentages
  • Enhanced route explainability for operator trust (v1.21.x roadmap)

Changed

  • Updated dashboard HTML and JavaScript to include new panel and aggregation logic
  • Bumped version to 1.21.0

Fixed

  • No breaking changes; existing routing behavior remains compatible

v1.20.0 - 2026-04-02

Added

  • External metadata integration: promotion alerts, source badges, provider-mix analytics endpoint (Issue #186)
  • Route explainability dashboard and metrics: route_summary column, decision history panel, why_selected and alternatives display (Issue #188)
  • Added /api/analytics/provider-mix endpoint for comparing providers based on external catalog pricing
  • Enhanced dashboard Routes view with "Route decision history" table showing routing explanations
  • Added route_summary field to metrics DB and /api/traces endpoint

Changed

  • Updated config.yaml metrics db_path default to ./faigate.db and enabled metrics by default
  • Extended _build_route_summary function to populate why_selected and alternatives
  • Updated all six log_request calls to include route_summary

Fixed

  • No breaking changes; existing routing behavior remains compatible

v1.17.0 - 2026-04-01

Added

  • Added router integration with offerings catalog for price-aware routing decisions (Phase 2b)
  • Added package scoring based on remaining credits and expiry dates for intelligent routing
  • Added detailed package overview to dashboard with credits, expiry, and provider mapping
  • Added _metadata_packages_detail() function for detailed package insights
  • Enhanced dashboard with package details section and cost projection improvements

Changed

  • Updated router cost estimation to prefer offering-specific pricing over provider defaults
  • Improved provider dimension scoring with package score integration
  • Extended dashboard metadata catalogs summary with package details
  • No breaking changes; existing routing behavior remains compatible

v1.18.0 - 2026-04-01

Added

  • Added cost projection wizard CLI (faigate-stats --project) for estimating costs across providers based on token usage patterns (Phase 2c)
  • Added package management UI with enhanced dashboard views for packages, expiry alerts, and credit tracking
  • Added analytics integration with --trends flag showing daily cost trends and projected monthly spend
  • Added faigate.cost module with provider cost estimation and cross-provider comparison utilities

Changed

  • Extended dashboard alerts view with packages block showing credits, expiry, and expiring-soon warnings
  • Enhanced CLI with --trends flag for historical cost/performance trend analysis
  • No breaking changes; existing routing behavior remains compatible

v1.16.0 - 2026-04-01

Added

  • Added external metadata schemas and catalogs for models, offerings, and packages (Phase 2a)
  • Added environment variables FAIGATE_OFFERINGS_METADATA_FILE and FAIGATE_PACKAGES_METADATA_FILE
  • Added get_offerings_catalog() and get_packages_catalog() public API functions
  • Updated sync-metadata.sh to touch offerings and packages catalog files

Changed

  • Extended provider catalog loading to support offerings and packages catalogs alongside providers
  • No breaking changes; all existing functionality remains unchanged

v1.14.2 - 2026-03-31

Added

  • Added an optional shared provider metadata overlay path so Gate can load repo-backed provider catalog snapshots and merge them into the runtime catalog.
  • Added faigate-provider-metadata-sync plus restart/update integration so provider metadata snapshots can be materialized automatically before Gate restarts.

Changed

  • Documented the fusionAIze-only shared metadata repo shape and shipped example catalog/overlay JSON for Gate, including initial tracked metadata coverage for anthropic-haiku, anthropic-sonnet, and gemini-pro.

v1.14.1 - 2026-03-31

Added

  • (New release baseline)

v1.14.0 - 2026-03-31

Added

  • (New release baseline)

v1.15.0 - 2026-04-01

Added

  • Enhanced provider catalog system with cost truth reporting, priority clusters, and actionable recommendations similar to llmairouter.com
  • Integrated external fusionaize-metadata repository for centralized provider metadata management
  • Added numeric pricing rates lookup from external catalog, overlay, and built-in registry
  • Added dashboard metric cards for cost truth and priority cluster visualizations
  • Added provider discovery recommendations with actionable improvement suggestions

Changed

  • Increased line length to 120 characters (modern standard) across codebase
  • Reframed Claude-native bridge model aliases around routing intent instead of direct frontier spend: built-in Claude Code model ids now resolve to auto, premium, or eco so Gate can still choose the cheapest capable route for the request
  • Tightened the shipped config and integration examples around coding auto modes, so claude, opencode, openclaw, and related coding clients can share clearer coding-auto, coding-fast, and coding-premium entry points instead of muddled provider-first defaults
  • Updated the roadmap and implementation plan to prioritize cost-aware coding auto modes first, then stronger product surfaces and licensing-aware stack boundaries for Gate as a standalone product
  • Added a dedicated dashboard IA document so the next web and shell dashboard work is grouped around operator jobs such as overview, providers, clients, routes, analytics, request log, integrations, and troubleshooting instead of one long admin surface

v1.15.1 - 2026-04-01

Fixed

  • Removed hardcoded user path (/Users/andrelange/Documents/repositories/github/fusionaize-metadata) from external metadata integration
  • Changed default metadata root to non-existent path, preventing unauthorized directory access requests
  • Updated sync-metadata.sh to use ~/.faigate/metadata or FAIGATE_METADATA_DIR environment variable
  • External metadata remains opt-in via FAIGATE_PROVIDER_METADATA_DIR environment variable

Unreleased

v1.13.0 - 2026-03-30

Added

  • Added an optional Anthropic-compatible bridge inside Gate with POST /v1/messages and POST /v1/messages/count_tokens, so Claude-native clients can enter through a dedicated surface without splitting routing, policy, health, or fallback behavior into a second gateway
  • Added an internal canonical request and response layer for bridge traffic, which keeps Anthropic-shaped ingress mapping separate from the existing routing and completion core instead of adding one-off protocol logic directly in the router
  • Added a community claude-code-router hook that can prefer coding-strong, tool-capable, and larger-context routes for Claude Code traffic without making the bridge itself depend on any one routing policy
  • Added bridge-specific validation and release-readiness helpers, including a client-near validation script and an explicit bridge release checklist for opt-in production rollouts
  • Expanded the provider source catalog scope beyond blackbox, kilo, and openai so Gate can also track mirrored official source data for anthropic, deepseek, and google
  • Added local models-endpoint overlays per configured route, which lets Gate compare what a specific key can really see against the mirrored global provider catalog

Changed

  • Hardened Anthropic bridge compatibility for real operator workflows: basic tool_use / tool_result flows now stay on the same execution path, Anthropic version and beta headers survive the bridge, and bridge responses expose the key route-resolution headers needed for debugging
  • Improved quota-aware fallback behavior for Anthropic-shaped traffic by introducing shared quota metadata on routes, which lets Gate avoid blindly retrying another path that is still backed by the same exhausted Anthropic or BYOK quota domain
  • Clarified the Bridge release position across docs: v1.13.0 ships the Anthropic surface as opt-in and production-usable for early adopters, but does not claim full Anthropic, Claude Code, or Claude Desktop parity yet
  • Aligned the doctor and bridge validation tooling with non-default live configs so release validation runs against the same configured DB, env file, and runtime instance instead of silently falling back to repo-local defaults
  • Provider source alerts now distinguish more clearly between global catalog drift and key-specific route/model visibility drift
  • Catalog summaries now include local route counts, local visible model counts, and route-vs-catalog mismatch hints instead of only source freshness and change counts

v1.12.0 - 2026-03-29

Added

  • Added a provider source catalog line with startup refresh, dashboard/API summaries, and operator-facing alert actions across faigate-doctor, faigate-provider-probe, and Quick Setup so model/pricing drift is visible earlier instead of hiding behind stale curated assumptions
  • Added explicit Kilo paid workhorse lanes for Sonnet and Opus plus Kilo-specific routing fit scoring, which lets Gate model premium, balanced, and free Kilo traffic without relying on opaque kilo-auto/* header behavior
  • Added route-preview coverage for Kilo frontier lane selection so operators can see when Gate chose kilo-opus, kilo-sonnet, or kilocode and why
  • Added stronger release automation dry-run coverage so the release helper itself is exercised in CI before a tagged release is cut

Changed

  • Hardened release automation end-to-end: local release scripts now validate versions more strictly, verify package metadata coherency, and point to the dedicated fusionAIze/homebrew-tap repository instead of assuming a local formula in this repo
  • Release artifact publishing now validates tag/version alignment before publishing and reuses prebuilt Python artifacts for PyPI instead of rebuilding a second time in the publish job
  • Reframed the legacy blackbox-free route as a low-cost burst path rather than a guaranteed free path, because the currently working curated model is not reliably the :free SKU for every key
  • Updated Kilo defaults and examples to use the current gateway base URL without the stale /v1 suffix

v1.11.2 - 2026-03-29

Changed

  • Restored Python 3.10 compatibility for the release update helpers by falling back cleanly when datetime.UTC is unavailable, which fixes the release CI collection failure that blocked the v1.11.x line
  • Realigned the runtime version surfaces so package metadata, faigate --version, and the FastAPI app version all move together again instead of reporting mixed 1.10.1 / 1.11.x values

v1.11.1 - 2026-03-29

Changed

  • Established the current release-driven changelog baseline so subsequent v1.11.x+ entries stay grouped by real user-visible behavior instead of placeholder release notes

v1.11.0 - 2026-03-29

Added

  • Gemini 3 / 3.1 Modernization — full rollout of Google Gemini 3 Flash and Gemini 3.1 Pro (High/Low) models across all surfaces:
    • google/gemini-flash now resolves to gemini-3-flash
    • google/gemini-flash-lite now resolves to gemini-3-flash-lite
    • google/gemini-pro-high and google/gemini-pro-low added for high-reasoning and balanced Gemini 3.1 Pro lanes
  • Flexible Model Labels — centralized model version mapping in lane_registry.py:
    • New _ACTIVE_MODEL_VERSIONS and _MODEL_VERSION_LABELS lookup tables
    • Decouples canonical lanes (e.g. openai/gpt-4o) from concrete provider strings (e.g. gpt-4o)
    • Ensures consistent labels and IDs across config, catalog, wizard, and dashboard
  • Updated Catalog & Wizard — provider_catalog.py and wizard.py refactored to use dynamic model helpers:
    • Removes hardcoded Gemini 2.5 strings from recommendation logic
    • Synchronizes curated aliases and notes with the new versioned metadata
  • Release Automation — new scripts/faigate-release Python tool to automate version bumping, changelog updates, and Homebrew formula syncing.

Changed

  • lane_registry.py now exports get_active_model_id(canonical_id) and get_active_model_label(canonical_id) for shared use.
  • Anthropic default model in wizard updated to claude-3-5-sonnet-20241022 (Sonnet 4.6).

v1.10.1 - 2026-03-25

Added

  • hooks/adapters/grok_api_adapter.py — OpenAI-compat adapter that bridges faigate's virtual grok-xai provider to Grok's web interface (no XAI API key required):
    • Exposes GET /v1/models and POST /v1/chat/completions on port 8091
    • Translates OpenAI messages array → single Grok prompt (system prompt + history + last user message)
    • Maps model names: grok-3 → grok-3-auto, grok-3-fast, grok-4, grok-4-mini-*
    • Full streaming via SSE (uses Grok-Api's stream_response token list)
    • Runs from fusionAIze/grok-api-hook fork for stability
    • GET /health endpoint for liveness checking
  • Virtual provider registration — community hooks can now register providers programmatically without config.yaml entries:
    • register_virtual_provider(name, config) in hooks.py — validates base_url + model, sets safe defaults
    • get_virtual_providers() — returns all registered virtual providers
    • load_community_hooks() now passes register_virtual_provider as optional second arg when hook's register() accepts it
    • main.py startup merges virtual providers into _providers after config-defined providers (config wins on name collision)
  • grok-wrapper updated — now registers grok-xai as a virtual provider (tier: mid, cost: free, latency: slow) pointing to http://127.0.0.1:8091/v1 with full lane + capabilities metadata

Changed

  • hooks/grok-wrapper.py now uses two-arg register(register_fn, register_provider_fn) signature
  • All references updated to use fusionAIze/grok-api-hook fork

v1.10.0 - 2026-03-24

Added

  • Cache Intelligence layer — all providers now carry precise cache metadata in config.yaml and the routing engine uses it for cost estimation and scoring:
    • cache.min_prefix_tokens — minimum stable prefix required for cache activation (provider-specific: 64 for DeepSeek, 1024 for Google, 1024 for Anthropic)
    • cache.ttl_seconds — provider cache lifetime (0 = unknown; Google: 3600 s, Anthropic: 300 s)
    • cache.max_cached_tokens — maximum prefix retained in cache (64k for DeepSeek, 1M for Google, 32k for Anthropic)
    • cache.cache_read_discount — actual cost ratio of cached vs fresh input tokens (e.g. DeepSeek: 0.07, Anthropic: 0.10, Google Flash: 0.04)
    • cache.cache_write_surcharge — write cost multiplier for explicit-mode caches (Anthropic: 1.25, others: 1.0)
    • Cache threshold in router.py is now provider-aware (max(64, min_prefix_tokens) per provider)
    • Cache scoring bonus in router.py rewards providers where cache is likely to activate given the current request shape
    • /route endpoint response now includes a cache_intelligence block with activation forecast, estimated savings, TTL, and write surcharge
  • Community / plugin hook system — faigate now supports dropping custom Python hooks into a directory without editing core code:
    • load_community_hooks(plugin_dir) in hooks.py scans *.py files in the configured dir and calls their register(register_fn) entry point
    • community_hooks_dir config key added to request_hooks section; community hooks are loaded before hook-name validation so plugin names pass cleanly
    • Community hooks are logged at startup and exposed in /health response under community_hooks
    • get_community_hooks_loaded() utility function for observability
  • hooks/grok-wrapper.py — first community hook example, ships with faigate:
    • Automatically routes requests to grok-xai when the model name is a known Grok variant (grok-3, grok-3-mini, grok-3-fast, grok-2, …) or starts with grok-
    • Also triggered by X-Faigate-Grok: 1 / true / yes request header
    • Includes full install instructions and self-documenting docstring
  • grok-xai provider template in config.yaml section B — full lane + cache + capabilities metadata for xAI Grok 3 (commented out; activate with XAI_API_KEY)
  • Updated request_hooks config comment block to document community_hooks_dir, example usage, and grok-wrapper installation steps

v1.9.2 - 2026-03-24

Fixed

  • Fixed routing_mode from request hooks being silently discarded — _sanitize_routing_hints now accepts and preserves routing_mode as a first-class hint field; _merge_routing_hints propagates it through the hook pipeline so X-faigate-Mode header overrides correctly reach the scoring layer
  • Fixed X-faigate-Mode: eco and X-faigate-Mode: premium being ignored on short prompts — _evaluate_heuristic_match now bypasses short-message and general-default rules when an explicit routing_mode hint is present, allowing mode-specific scoring to run regardless of token count

Added

  • Three new providers using existing API keys — no new credentials required:
    • anthropic-sonnet: Claude Sonnet 4.6 (claude-sonnet-4-6) — quality-workhorse lane, cost_tier: standard (~$3/MTok input), high reasoning and tool strength, degrade-to: haiku → deepseek-chat
    • anthropic-haiku: Claude Haiku 3.5 (claude-haiku-3-5) — fast-workhorse lane, cost_tier: cheap (~$0.80/MTok input), degrade-to: deepseek-chat → gemini-flash
    • gemini-pro: Gemini 2.5 Pro (gemini-2.5-pro) — quality-workhorse lane, cost_tier: premium, high reasoning and context strength (1M token context window), degrade-to: gemini-flash → deepseek-reasoner
  • Added proper lane: metadata block to gemini-flash-lite (was missing, preventing correct cluster-based scoring)
  • Updated global fallback_chain to reflect full provider depth: deepseek-chat → anthropic-haiku → gemini-flash → deepseek-reasoner → anthropic-sonnet → gemini-pro → openai-gpt4o → openrouter-fallback → anthropic-claude → kilocode → blackbox-free
  • Updated routing_modes.eco to explicitly prefer anthropic-haiku, gemini-flash-lite, gemini-flash, deepseek-chat
  • Updated routing_modes.premium to prefer anthropic-sonnet, openai-gpt4o, gemini-pro, anthropic-claude, deepseek-reasoner
  • Updated opencode client profile to include high quality tier so Sonnet/Gemini Pro are reachable in auto mode
  • Added anthropic-sonnet, anthropic-haiku, gemini-pro to model_shortcuts
  • Enabled routing_modes (was enabled: false)

v1.9.1 - 2026-03-24

Fixed

  • Corrected version strings in __init__.py and main.py which incorrectly reported 1.8.0 after the v1.9.0 release
  • Fixed load_dotenv() in config.py so the service correctly resolves faigate.env from the config directory (/opt/homebrew/etc/faigate/) instead of searching upward from the package installation path, which caused all providers to start in unresolved-key state when run via Homebrew launchd

Added

  • Added mode-override-header request hook that reads the new X-faigate-Mode header and maps it to a routing posture (auto, eco, premium, free, plus common aliases quality, save, cheap, balanced, standard); unknown values are silently ignored
  • Updated _routing_posture() to check hook_hints.routing_mode before profile_hints.routing_mode so the header override takes effect end-to-end
  • Expanded opencode signal groups from 5 to 10 by adding devops (kubernetes, terraform, helm, ci/cd, …), testing (unit tests, integration tests, pytest, jest, tdd, coverage, …), security (jwt, oauth, xss, sql injection, csrf, rbac, …), and database (schema, query optimization, index, replication, sharding, …) — each group triggers short_complex escalation to reasoning lanes when ≥ 2 groups fire on a brief prompt
  • Added plural keyword forms (unit tests, integration tests) to the testing signal group so word-boundary matching no longer silently drops plural prompts
  • Extended _OPENCODE_COMPLEXITY_HINTS and _OPENCODE_COMPLEXITY_RULE_KEYWORDS with event sourcing, cqrs, kubernetes, terraform, infrastructure as code, unit test, integration test, jwt, sql injection, vulnerability, schema, replication, and query optimization terms
  • Added short_complex and prefer_providers bypass guards in _evaluate_heuristic_match so brief cross-domain prompts and explicit provider-preference requests skip the short-message and general-default fallthrough rules and reach the profile scoring layer

v1.9.0 - 2026-03-23

Changed

  • Strengthened short-but-risky opencode prompt detection so brief architecture, queue/backpressure, and rollout-planning requests escalate out of cheap lanes earlier instead of being flattened by generic short-query heuristics
  • Expanded route-preview explainability with structured complexity reasons plus explicit why_not_selected guidance for cheaper alternatives, so operators can now see why a lower-cost lane lost on reasoning depth, benchmark fit, freshness, or runtime pressure

v1.8.0 - 2026-03-22

Changed

  • Started the adaptive-orchestration runtime line with canonical model-lane and provider-route metadata in config, wizard, runtime inventory, and provider-catalog surfaces
  • Added the first lane-aware router scoring slice so quality, balanced, eco, and free postures now influence candidate ranking through lane cluster, benchmark cluster, route type, and runtime pressure instead of only provider tier
  • Added Same-Lane-Route fallback preference before weaker cluster downgrades when a compatible alternate route exists for the same canonical model
  • Added an in-memory adaptation state for rate-limit, quota, timeout, and latency pressure so hot routes can be demoted conservatively at runtime
  • Persisted routing explainability fields such as canonical_model, route_type, lane_cluster, selection_path, and decision_details into metrics and route traces
  • Expanded candidate cards, client scenarios, and provider dashboard drilldowns so operators can now see route mirrors, degrade chains, canonical lanes, and runtime penalties directly in Gate

v1.7.1 - 2026-03-22

Changed

  • Tightened the terminal header rendering again so interactive screens no longer insert apparent blank spacer lines between the three wordmark rows
  • Fixed the client-scenario apply flow so choosing Write config now returns cleanly to the calling menu after the confirmation step instead of dropping operators straight back into the same scenario list

v1.7.0 - 2026-03-22

Changed

  • Added internal Gate drilldowns for client quickstarts, provider discovery, and dashboard details so operators no longer need to leave the menu just to open one parameterized view
  • Expanded client scenarios into lane-based explanations with explicit quality, reasoning, workhorse, budget, and fallback roles so templates like opencode / balanced now explain why kilocode, blackbox-free, or openrouter-fallback are in or out
  • Added family-coverage hints for scenario output so operators can see when a provider family currently has only one quality or balanced slot and would need separate provider entries for richer Opus / Sonnet / Haiku-style splits
  • Refined the shell header color segmentation again to match the tighter blue / yellow / blue / green brand grouping across all three wordmark rows

v1.6.3 - 2026-03-22

Changed

  • Hardened config-wizard merge writes so existing configs with null sections such as client_profiles.rules, routing_policies.rules, or request_hooks.hooks now merge into real runtime config safely instead of failing mid-write
  • Closed the remaining Homebrew helper-parity gap so user-facing commands such as faigate-config-wizard, faigate-status, faigate-restart, faigate-logs, faigate-start, faigate-stop, faigate-update, and faigate-auto-update ship through the Brew formula too
  • Refined the terminal wordmark again with the new three-color brand palette and kept the inline version sourced dynamically from the current package version

v1.6.2 - 2026-03-22

Changed

  • Fixed the config-wizard write path so guided Write config flows persist the actual runtime config instead of accidentally writing the purpose/client/suggestions summary payload back into config.yaml
  • Added an explicit doctor warning when config.yaml appears to contain wizard summary keys, which makes accidental miswrites easier to catch before restart and rollout work
  • Restored executable bits for packaged helper scripts such as faigate-config-wizard, faigate-config-overview, faigate-provider-discovery, and the onboarding/client helper scripts so Brew-installed helper entrypoints no longer fail with Permission denied

v1.6.1 - 2026-03-20

Changed

  • Fixed the packaged faigate-dashboard helper so the shipped script keeps its executable bit and the Brew-installed dashboard no longer fails with Permission denied
  • Polished the interactive terminal wordmark again so the large I aligns with the intended shape and the current version now appears inline at the right edge of the logo in the same subdued tone as the subtitle

v1.6.0 - 2026-03-20

Added

  • Added faigate-provider-setup plus matching Quick Setup / Configure menu entries so operators can add known providers, custom OpenAI-compatible upstreams, and local workers before dropping into the purpose-aware config wizard
  • Added faigate-provider-probe so configured sources can be checked against config, env, and the live /health payload before client rollout begins
  • Added faigate-client-scenarios plus matching menu entries so operators can apply named templates such as opencode / eco, opencode / quality, n8n / reliable, or cli / free instead of thinking only in raw profile-mode edits
  • Added faigate-dashboard plus a new top-level Dashboard menu section so operators now get one shell-native performance view for traffic, latency, spend, token volume, provider/client hotspots, and action-oriented alerts

Changed

  • Tightened the onboarding docs and main README around the new provider-source-first UX so first setup now reads more like Provider Setup -> Provider Probe -> API Keys -> Full Config Wizard -> Client Scenarios -> Validate -> Client Quickstarts

  • Renamed the old FOUNDRYGATE STATS CLI banner to fusionAIze Gate Stats so the terminal metrics surfaces stay on-brand

  • Expanded client scenarios with clearer budget, best when, and tradeoff guidance so operators can pick templates by intent instead of only by routing-mode names

  • Expanded the new dashboard with budget, quota, and routing-pressure hints so it now helps answer whether traffic should shift, a cheaper scenario is worth trying, or a provider likely needs more budget

  • Added a dedicated adaptive-orchestration roadmap that sketches the path from lane metadata to scoring, live adaptation, benchmark freshness, and budget-/quota-aware routing through the v1.10.x and v1.11.x lines

v1.5.1 - 2026-03-20

Changed

  • Reworked the interactive config wizard candidate screen so purpose/client selection now shows compact Ready now, More options if you add keys, and Optional specialty add-ons cards instead of a raw provider metadata dump
  • Improved the client quickstart surfaces so the menu and client helper now show a clearer Best next step hint and friendlier Preset matches wording instead of implying that Presets 0 means something is broken
  • Clarified the API-key helper so provider base URL overrides are explicitly labeled as optional upstream overrides, reducing confusion between local Gate client URLs and upstream provider endpoints
  • Nudged the terminal logo spacing closer to the intended fusionAIze Gate wordmark in interactive screens

v1.5.0 - 2026-03-20

Changed

  • Fixed the standalone shell helpers on macOS/Homebrew so service status, logs, and service-manager labels now recognize the Brew-managed homebrew.mxcl.faigate path instead of assuming only the manual LaunchAgent path
  • Fixed faigate-menu model listing so it parses the /v1/models payload correctly instead of trying to read JSON through a broken stdin pipeline
  • Fixed faigate-auto-update on macOS's default Bash 3.2 by removing the mapfile dependency from its payload parsing path
  • Fixed user-facing helper scripts so --help exits safely instead of accidentally triggering live install/update logic in shell environments that only wanted usage text
  • Improved faigate-health, faigate-update-check, and faigate-menu so operators now see compact human-readable summaries before diving into raw payloads
  • Added a service-manager mismatch warning when /health responds but the configured manager reports a stopped or missing service, which helps catch stale old runtimes still bound to the same port
  • Polished the terminal header to align more closely with the intended fusionAIze Gate visual identity in interactive terminals
  • Added a dedicated Quick Setup happy path and summary cards for gateway, config, providers, and clients in the main menu flows
  • Updated the client helper and the Clients menu so operators see compact recommendation cards first and can drill into one client without dumping the full cross-client quickstart wall every time
  • Added first Next step receipts after the key guided actions in the shell flow so wizard, validation, restart, and client-setup paths now end with a short operator-oriented “what to do next” block

v1.4.5 - 2026-03-19

Added

  • Added a first faigate-menu control center with a shared terminal UI, the new fusionAIze Gate header, and consistent q/c navigation across status, configure, explore, validate, control, and update menus
  • Added faigate-api-keys and faigate-server-settings so API keys, host, port, and log-level changes have a Gate-native interactive path instead of living only in external orchestration layers
  • Added faigate-routing-settings so the global default routing mode and client-profile routing defaults can be reviewed and adjusted from the same Gate-native control flow
  • Added faigate-client-integrations plus a Clients section in faigate-menu so OpenClaw, n8n, opencode, and generic CLI quickstarts can be reviewed and driven through client-scoped wizard flows
  • Added faigate-config-overview plus a clearer Current Config / Guided Setup / Direct Settings split inside faigate-menu so configuration flows now map more cleanly to the later Grid-style orchestration model

Changed

  • Aligned helper scripts such as faigate-health, faigate-status, faigate-update-check, faigate-auto-update, and faigate-doctor around shared config/env/port resolution so repo, packaged, and later Grid-driven flows can behave consistently
  • Extended install and Homebrew helper exposure so the new menu/config helpers can ship through the same operator-facing paths as the existing scripts
  • Expanded faigate-status, faigate-logs, and faigate-restart so service control now carries clearer service-manager context, recent-vs-live log flows, and restart verification instead of only raw process-manager commands
  • Polished faigate-menu with compact runtime/config snapshots in the main and control/config submenus plus short inline tips so the shell UX stays self-orienting between steps

v1.4.0 - 2026-03-19

Changed

  • Renamed the product branding from FoundryGate to fusionAIze Gate across the repository, documentation, examples, and operator-facing surfaces
  • Renamed the technical runtime slug from foundrygate to faigate, including the Python package, npm CLI package, helper scripts, example file names, service templates, and Homebrew formula path
  • Moved the repository references from typelicious/FoundryGate to fusionAIze/faigate and aligned env prefixes, headers, and operational examples with the new FAIGATE_ / x-faigate-* naming
  • Completed the first release-prep baseline for the rebrand so future releases, installs, and documentation no longer depend on the old names

v1.3.0 - 2026-03-19

Added

  • Added a first faigate-config-wizard helper that suggests an initial config.yaml from the API keys already present in .env
  • Added first-class routing_modes and model_shortcuts config blocks so virtual model ids such as auto, eco, premium, free, or custom names can participate in routing
  • Added wizard candidate listing and conservative config merging so operators can select multiple provider candidates during first setup or later catalog-driven updates
  • Added config-aware wizard update suggestions so existing installs can see recommended_add, recommended_replace, and recommended_keep groups before applying provider changes
  • Added wizard recommended_mode_changes suggestions so existing client profiles can be nudged toward the current purpose-aware routing defaults without silently rewriting them
  • Added an apply suggestions wizard flow so selected provider and client-mode recommendations can be merged into an existing config without manual copy/paste
  • Added a wizard dry-run change summary so operators can preview added providers, model replacements, fallback changes, and client-mode changes before writing config updates
  • Added optional wizard write-backup snapshots so config updates can keep a local pre-change copy before overwriting config.yaml
  • Added a built-in faigate-config-wizard --help flow so first setup, catalog review, update suggestions, dry-run previews, and backup-aware writes are all discoverable directly from the CLI
  • Added optional provider-catalog discovery metadata and env-backed signup-link overrides so future CLI or control-center surfaces can show disclosed provider links without mixing link configuration into normal config files
  • Added first CLI surfacing of disclosed provider discovery links in onboarding and doctor outputs, always alongside a link-neutral recommendation policy signal
  • Added faigate-provider-discovery for one compact text/JSON discovery view that later browser or control-center work can consume
  • Added discovery-link filters for CLI and API views so operators can narrow provider links by offer_track, link_source, or disclosed_only

Changed

  • client_profiles can now choose a default routing_mode, letting one client keep the global mode while another uses a different or custom mode by default
  • GET /v1/models, route previews, and runtime response headers now expose configured routing modes and resolved shortcut/mode metadata
  • faigate-doctor, onboarding reports, and the provider-catalog API now surface curated model-drift, source-confidence, volatility, and catalog-freshness alerts for configured providers
  • Provider catalog entries now distinguish direct providers from aggregators and wallet routers, track auth modes such as api_key, byok, and wallet_x402, and keep community watchlists explicitly secondary to official sources
  • faigate-config-wizard can now filter candidates by purpose and client, accept multi-select provider input, and merge selected providers back into an existing config instead of forcing a full rewrite
  • Tightened the roadmap and user-facing docs around v1.3.0 so guided setup, catalog-assisted updates, and future recommendation-link work stay transparent and clearly separated from ranking logic
  • Provider discovery metadata now carries an explicit link-neutral recommendation policy so provider-link configuration can never be mistaken for a ranking signal

v1.2.3 - 2026-03-19

Changed

  • Hardened the Homebrew formula so native Python extensions such as pydantic-core and watchfiles are built from source with extra Mach-O header padding on macOS instead of relying on the vendored wheel layout
  • Strengthened the formula test so it validates the wrapped faigate --version entrypoint instead of only importing the package inside libexec
  • Fixed the Python service entrypoint so python -m faigate.main and the Brew-managed wrapper both execute the runtime correctly
  • Clarified in the README, workstation guide, and troubleshooting docs that active Python virtualenvs can shadow the Brew-installed faigate binary

v1.2.1 - 2026-03-19

Changed

  • Switched the Homebrew formula baseline from python@3.13 to python@3.12 to reduce macOS packaging friction around vendored native Python wheels
  • Clarified in the README and workstation docs that brew install faigate resolves cleanly after tapping fusionAIze/faigate, while the fully qualified install path remains the safest first-run example

v1.2.0 - 2026-03-19

Added

  • Added a workstation operations guide for Linux, macOS, and Windows runtime layouts
  • Added a macOS launchd LaunchAgent example for local workstation installs
  • Added Windows PowerShell and Task Scheduler starter examples for local workstation installs
  • Added platform-aware runtime helper scripts so macOS can use the same faigate-install / start / stop / status flow style as Linux
  • Added a project-owned Homebrew formula plus brew services guidance for packaged macOS workstation installs
  • Added explicit FAIGATE_CONFIG_FILE config discovery and faigate --config / --version support so service wrappers and packaged installs can point to config outside the repo
  • Added a helper-level onboarding smoke test for explicit config/env/python wiring

Changed

  • Updated the README quickstart so Linux, macOS, Windows, and Homebrew paths are visible earlier
  • Replaced the weak PyPI workflow badge with clearer workstation and Homebrew badges

v1.1.0 - 2026-03-16

Added

  • Added richer client usage reporting in GET /api/stats and the dashboard, including per-client tokens, failures, success rate, and aggregate client totals
  • Added a second wave of AI-native starter templates for Agno, Semantic Kernel, Haystack, Mastra, and Google ADK
  • Added client highlight summaries to GET /api/stats and the built-in dashboard for top request, token, cost, failure, and latency signals
  • Added a third wave of AI-native starter templates for AutoGen, LlamaIndex, CrewAI, PydanticAI, and CAMEL

Changed

  • Tightened static and heuristic match semantics so combined fields now behave as cumulative constraints unless any: is used explicitly
  • Tightened policy match semantics so client_profile acts as an additive constraint inside one rule instead of bypassing sibling static or heuristic fields

v1.0.0 - 2026-03-15

Added

  • Added dashboard CSP hashes plus stricter response-security defaults for the no-build operator UI
  • Added stronger provider base URL validation so non-local upstreams must use https
  • Added reduced leakage of upstream provider failure details in client-facing error payloads
  • Added a separate npm CLI package under packages/faigate-cli for basic health, model, update, and route-preview checks
  • Added a documented v1.0.0 security review with mitigations and residual-risk notes
  • Added functional API coverage for upstream error sanitization on top of the earlier dashboard and request-boundary hardening tests
  • Streamlined the root README into a shorter landing page and moved deeper API, configuration, and operations detail into dedicated docs pages

v0.9.0 - 2026-03-15

Added

  • Added conservative response-security headers plus a dashboard CSP for the no-build operator UI
  • Added explicit security config controls for JSON body size, upload size, and bounded routing-header values
  • Added functional API coverage for dashboard headers, JSON request limits, upload limits, and sanitized routing-header behavior

v0.8.0 - 2026-03-15

Added

  • Added faigate-onboarding-report plus a testable onboarding report module for many-provider and many-client readiness checks
  • Added faigate-onboarding-validate so onboarding blockers can fail fast in local setup and CI-style validation flows
  • Added built-in OpenClaw, n8n, and CLI quickstart examples to the onboarding report and integration docs so client onboarding can stay copy/paste friendly
  • Added staged provider-rollout reporting and fallback/image readiness warnings so many-provider onboarding is easier to phase safely
  • Added a client matrix to the onboarding report so profile match rules and routing intent are visible before traffic goes live
  • Added starter example files for OpenClaw, n8n, and CLI clients under docs/examples/ so onboarding can begin from copy/pasteable templates
  • Added starter provider snippets for cloud, local-worker, and image-provider setups under docs/examples/
  • Added matching provider .env starter files for cloud, local-worker, and image-provider onboarding flows
  • Added provider env placeholder checks to faigate-doctor so missing .env values are surfaced before rollout
  • Added --markdown output to faigate-onboarding-report so onboarding state can be pasted into issues, PRs, or hand-off notes
  • Added delegated OpenClaw request and generic AI-native app profile starters to round out the v0.8.x onboarding path

v0.7.0 - 2026-03-12

Added

  • Added stronger update-alert metadata to GET /api/update, including update type, alert level, and recommended action for operators and dashboard consumers
  • Added an opt-in auto_update policy block plus faigate-auto-update so controlled deployments can gate helper-driven updates without enabling silent self-updates
  • Added GET /api/operator-events plus operator-event metrics for update checks and helper-driven auto-update attempts
  • Added dashboard cards and tables for operator-side update checks and apply attempts
  • Added provider-health rollout guardrails so helper-driven auto-updates can block when gateway health is already degraded
  • Added update_check.release_channel and auto_update.rollout_ring so operators can distinguish stable vs preview checks and tighter rollout rings
  • Added auto_update.min_release_age_hours so helper-driven auto-updates can wait for a release to age before becoming eligible
  • Added auto_update.maintenance_window so helper-driven auto-updates can stay inside explicit local maintenance hours
  • Added auto_update.provider_scope so rollout-health guardrails can evaluate only a selected provider subset
  • Added auto_update.verification so helper-driven auto-updates can run a post-update check and emit a rollback hint on failure

v0.6.0 - 2026-03-12

Added

  • Added modality-aware metrics and filters so stats, traces, recent requests, and the dashboard can distinguish chat, image_generation, and image_editing
  • Added POST /api/route/image for dry-run preview of image-generation and image-editing routing decisions
  • Added optional image provider metadata (max_outputs, max_side_px, supported_sizes) so image-capable providers can be ranked against n and size
  • Added top-level capability coverage to GET /health plus GET /api/providers for filtered provider inventory and dashboard coverage views
  • Added shared request validation for image-generation, image-editing, and image-route preview payloads so invalid size, n, and scalar fields fail fast before provider calls
  • Added optional image.policy_tags plus request-side image-policy hints so image routing can prefer providers tagged for quality, cost, balanced, batch, or editing

v0.5.0 - 2026-03-12

Added

  • Added contract: image-provider plus OpenAI-compatible POST /v1/images/generations and POST /v1/images/edits paths for image-capable providers
  • Added a shipped Dockerfile and tag-driven release-artifacts workflow for Python distributions, GHCR images, and optional PyPI publishing
  • Added public community-health and security baseline files: Code of Conduct, Security Policy, issue templates, PR template, Dependabot, and CodeQL
  • Added generic onboarding helpers (faigate-bootstrap, faigate-doctor) and a publish-dry-run workflow for GHCR and Python package validation
  • Added cached release update checks via GET /api/update, the dashboard, and faigate-update-check

v0.4.0 - 2026-03-12

Changed

  • Added optional request_hooks with a small built-in hook registry for per-request provider preferences, locality hints, and profile overrides
  • Added a dedicated routing layer for hook-provided hints before client-profile defaults
  • Added dry-run route output for applied hooks, effective request metadata, and candidate ranking details
  • Added provider route-fit metadata for context_window, token limits, and cache behavior
  • Added filtered stats, recent-request, and trace queries for provider, client, layer, and success views
  • Hardened the built-in dashboard with provider health, client breakdowns, route traces, URL-persisted filters, summary cards, and escaped rendering
  • Deepened provider scoring so routing now considers health, latency, recent failures, cache alignment, and request headroom instead of only first-fit dimension checks
  • Hardened request hooks with sanitized body updates and routing hints plus optional fail-closed behavior via request_hooks.on_error

v0.3.0 - 2026-03-12

Changed

  • Rebranded the public documentation around the fusionAIze Gate product name
  • Completed the technical rename from earlier runtime identifiers to faigate
  • Added validated provider capability metadata with normalized local/cloud and streaming defaults
  • Added an optional policy layer for capability-aware provider selection on auto requests
  • Added an explicit local-worker provider contract for network-local OpenAI-compatible runtimes
  • Added optional client profiles for caller-aware routing defaults based on request headers
  • Added a dry-run route introspection endpoint at POST /api/route
  • Added enriched route traces and client/profile breakdowns in metrics, stats, and CLI output
  • Added startup and /health probing for contract: local-worker providers via GET /models
  • Added built-in client_profiles presets for openclaw, n8n, and cli
  • Added a repository AGENTS.md and a documented Git workflow for main, feature/*, review/*, and hotfix/*
  • Aligned release guidance around semantic-style x.y.z versioning with v0.3.0 as the first fusionAIze Gate-branded release

Docs

  • Reworked the README into a more generic, portable open-source landing page
  • Added clearer API, configuration, deployment, and helper script documentation
  • Added release process documentation, roadmap updates, and a lightweight release checklist template
  • Added architecture, integrations, onboarding, and troubleshooting docs for external users