v0.0.5

0.0.5 — 2026-04-26

Client reuse, gateway observability, hot-reload, identity hardening, and concurrency safety across the SDK.

Gateway

• Per-request URL metadata — /m/{mode}/{metadata?}/{endpoint}/{path} route now accepts an optional metadata segment: comma-separated key=value pairs (e.g. /m/prod/agent=cortex,task=42/openai/v1/chat/completions). Metadata is URL-decoded and threaded through to usage records, surfacing in /usage responses and uploads. Route matching tries the 4-segment form first, falling back to the legacy 3-segment form.
• GATEWAY_DUMP_DIR full-call auditing — set the GATEWAY_DUMP_DIR environment variable to write a JSON dump of every proxied request+response to that directory (one {timestamp}.json per call). Covers streaming (chunks accumulated alongside real-time forwarding), non-streaming, and translate-path streaming. Dump failures silently degrade — they never break the proxy.
• Upstream header forwarding — response headers from upstream providers are now forwarded wholesale (minus RFC 7230 hop-by-hop headers and x-gateway-* namespace), replacing the previous hardcoded allowlist of x-request-id / openai-organization / anthropic-ratelimit-requests-remaining.
• Health endpoint auth — /health now respects gateway auth config, returning 401 unless the path is listed in public_paths.
• Translate-path stream usage recording — when the gateway translates streaming responses (Anthropic↔OpenAI protocol bridge), usage is now extracted from accumulated SSE chunks and recorded via the usage tracker. Previously, translate-path streaming usage was silently lost.
• Config hot-reload — GatewayServer.reload_config() swaps the full gateway configuration in place without dropping the HTTP server, preserving bound host/port and in-memory health/usage trackers. _config_watcher_loop() polls the config file's mtime every 1s and triggers reload on change. The watcher is wired into gateway.start() by default; opt out with watch_config=False.
• DeepSeek thinking block injection — when Anthropic extended thinking is enabled and the upstream is DeepSeek, the gateway now injects empty thinking blocks into assistant messages that lack them (_ensure_thinking_blocks()). DeepSeek requires reasoning_content on every assistant turn in multi-turn conversations; clients that strip empty thinking="" blocks would otherwise get 400 errors.

Provider Adapters — Client Reuse

All provider adapters now cache SDK client instances, only rebuilding when the API key or base URL changes. This eliminates per-request client construction overhead and reuses HTTP connection pools.

• Anthropic (anthropic_.py) — caches anthropic.Anthropic / AsyncAnthropic keyed on API key. Multiple system-role messages are now joined with "\n\n" instead of only keeping the last one.
• OpenAI (openai_.py) — caches openai.OpenAI / AsyncOpenAI with a composite key of (api_key, base_url, default_headers_tuple).
• Google (google_.py) — caches genai.Client keyed on API key. http_options.timeout is now set from the SDK-level timeout parameter (previously ignored).
• Compatible providers (compatible_.py) — new _CachedCompatibleMixin class that caches clients for DeepSeek, Mistral, xAI, Groq, Together, MoonshotAI, and Qwen. All share a single _resolve_api_key() helper.

API Client

• StatusAPI — now caches httpx.Client / httpx.AsyncClient instances instead of creating one per request (async with httpx.AsyncClient(...) → instance reuse). Adds close() and aclose() methods with atexit cleanup. Client instances are lazily created on first use.

Usage & Upload

• Shared thread pool executor — UsageUploader replaces per-upload threading.Thread(daemon=True).start() with a class-level ThreadPoolExecutor(max_workers=2, thread_name_prefix="aistatus-upload"), initialized once via double-checked locking. Submissions use executor.submit(); the executor is registered for atexit shutdown with wait=False.
• Identity field truncation — name (200 chars), organization (200 chars), and email (254 chars) are now truncated in upload payloads to prevent oversized requests.
• Metadata in usage records — UsageTracker.record_usage() accepts an optional metadata: dict[str, str] parameter. Non-reserved keys are merged into the usage record and fanned out through both local storage and the uploader.
• Locked JSONL appends — UsageStorage.append() now acquires fcntl.LOCK_EX before writing and releases with LOCK_UN afterward, making concurrent writes from multiple processes safe.
• Local timezone for "today" — period_since("today") now computes midnight in the local timezone instead of UTC, so daily usage summaries align with the user's calendar day.

Pricing

• Atomic cache writes — _write_file_cache() now writes to a temp file via tempfile.mkstemp() then calls os.replace() for atomic rename, preventing corruption when multiple processes write concurrently.

Router

• Non-streaming usage recording — route_stream() now records usage for both code paths: the direct streaming path (accumulating usage from chunks) and the non-streaming fallback (when the adapter lacks acall_stream). Previously, only the streaming path recorded.
• System message dedup — when both the system parameter and the messages array contain a system-role message, duplicates are now merged (prepended with "\n\n") instead of inserted twice.
• 429 retry hardening — retry path now respects allow_fallback: when retry fails and fallback is disabled, raises ProviderCallFailed instead of silently continuing. Added log.warning() on retry failure.

Protocol Translation

• Non-text content warnings — anthropic_request_to_openai() now logs a warning when non-text content blocks are dropped during Anthropic→OpenAI translation, and emits a "[Unsupported non-text content omitted …]" placeholder in the output message.
• Input tokens in terminal SSE — message_delta terminal events now include input_tokens alongside output_tokens in the usage field.
• SSE finished guard — a finished flag prevents emitting duplicate terminal events when both [DONE] and finish_reason arrive in the same stream.

Health Tracker

• Smarter cooldown clearing — record_success() now only clears the cooldown when there are no recent errors within the sliding window, preventing a single success from prematurely reactivating a flapping backend.

Config

• Thread-safe singleton — get_config() now uses double-checked locking (threading.Lock()) for safe concurrent access.