Skip to content

CHANGELOG.md

Latest commit

 

History

History
238 lines (182 loc) · 30.7 KB

CHANGELOG.md

File metadata and controls

238 lines (182 loc) · 30.7 KB
Raw

Changelog

All notable changes to NadirClaw will be documented in this file.

[Unreleased]

Added

  • nadirclaw claude hook — structural views for file reads Claude Code would truncate. Claude Code serves at most ~25,000 tokens per Read and truncates the rest, so a large module arrives as a partial page and the agent either misses everything below the cut or pages through the file at full cost. The new opt-in PreToolUse hook answers those reads with a declaration-level view instead: every definition, decorator, docstring excerpt and module constant, each tagged with its original source lines. No model call, no network, no archive — the file is untouched on disk, so recovery is a narrowed re-read. This is separate from optimize/compress, which rewrite the messages array inside the proxy; the hook runs in the agent before the file enters the conversation, so it works without routing through NadirClaw. Only whole-file reads of .py/.pyi past the cap are served; narrowed reads, other languages, unparsable syntax, files under the cap and modules above 4 MiB pass through untouched. install preserves any other PreToolUse hooks and is idempotent; uninstall removes only NadirClaw's entry, and nadirclaw claude uninstall now clears it too.

[0.23.1] - 2026-09-02

Fixed

  • The prompt cache ignored every request parameter except the model and message text, so it could answer a request with a response that does not satisfy it (#90). The key was sha256(model + [role, content]), while the request surface forwarded upstream also carries tools, tool_choice, response_format, reasoning_effort, thinking, max_tokens, temperature, top_p, n, and arbitrary provider extras. Two requests with the same messages therefore shared a cache entry: a plain chat answer could be served to a later request asking for response_format: {"type": "json_schema"} or for tool_calls, and a response generated under a high max_tokens could be served to a request that asked for a low one — silently overriding the client's contract, since the cache sits in front of the provider call. The key now includes every response-shaping field (request_cache_params() collects the declared ones plus everything in model_extra except stream), and message normalization keeps tool_call_id / name / tool_calls so tool-result turns with identical text no longer collide. Non-serializable extras fall back to repr rather than raising.
  • The prompt cache keyed multi-modal messages on their text only, so the same question about a different image was a cache hit. _normalize_message reduced a ChatMessage through text_content(), which concatenates the type: "text" parts and drops image_url / attachment parts entirely — "What is in this image?" plus a photo of a cat and the same wording plus a photo of a dog hashed to one key, and the second request was answered with the first image's description without a provider call. Multi-modal content (a content list) is now folded into the key whole; plain-text messages key exactly as before.

[0.23.0] - 2026-08-31

Added

  • NADIRCLAW_PREFER_ENV_KEYS — put provider env vars ahead of stored credentials. The resolution chain runs OpenClaw stored token → NadirClaw stored token → env var, which is right on a laptop and backwards on a server: a stored personal OAuth or subscription token carries that person's rate limit, so every request the router proxies is billed and throttled against one human. On a machine that also runs Claude Code there was no way to override it short of deleting a credentials file the box still needs. Set the flag and the environment moves to the front, falling back to the stored credential when the relevant env var is unset. Default off, so existing installs are unchanged. get_credential() and get_credential_source() now share one _env_credential() implementation, so nadirclaw status cannot report oauth while the server is actually sending the env key.

Fixed

  • /v1/messages forwarded parameters the routed model cannot accept, so cheap tiers 400 on every request — clients pick request parameters from the model id they can see, which behind the router is an alias (nadir-auto), so they assume the newest capabilities. When routing then selected an older, cheaper model, the request carried parameters that model rejects and every request failed (#83). The passthrough now reconciles the request against the 400 it gets back and retries within the same request, bounded to 4 attempts: thinking: {"type":"adaptive"} is rewritten to {"type":"enabled","budget_tokens":N}, an unsupported effort hint is dropped, and system/developer turns are folded into the top-level system field (the same fix already applied to the OAuth completion path in 0.21.1). Each fix is remembered per model, so the discovery costs one round trip per model per process instead of one per request. Applied fixes are recorded in requests.jsonl as modifiers_applied entries and signalled by an X-NadirClaw-Params-Reconciled response header. Verified against a live Anthropic subscription: claude-haiku-4-5 went from 3 failed upstream calls per Claude Code request to a single successful one.
  • /v1/chat/completions had the same 400 on its Anthropic OAuth path — the reconcile above only covered /v1/messages, so OpenClaw and Codex users routed to an older model still hit adaptive thinking is not supported on this model (#83). The direct Anthropic call behind the OpenAI-compatible endpoint now runs the same bounded reconcile loop and shares the per-model cache, and the fixes it applies are recorded as modifiers_applied on the request log. Two smaller corrections to the same mechanism: a reconciler that matched a 400 but declined to rewrite anything is no longer cached against the model, and the final attempt of the loop no longer applies a fix it has no round trip left to send.
  • Clamping max_tokens could strand thinking.budget_tokens above it, turning one recoverable 400 into an unrecoverable one — Anthropic requires 1024 <= budget_tokens < max_tokens, but the reconcile loop lowered max_tokens to the routed model's ceiling (#73) without revisiting a thinking budget the client had sized for the ceiling it thought it had. A request with max_tokens: 100000 and budget_tokens: 60000 routed to a model capped lower was clamped into a body that 400s on budget_tokens, an error no reconciler matches, so the loop gave up and surfaced the failure. The clamp now lowers the budget to fit under the new ceiling, or drops thinking outright when no value clears Anthropic's 1024 floor. Reachable both from a client-supplied budget and from the adaptive downgrade above, whose budget is sized before a later attempt clamps max_tokens.
  • serve could not start without a tty. On first run it asked "No configuration found. Run setup wizard?" unconditionally. With no tty click.confirm raises Abort and the process exits 1, so the shipped Dockerfile (CMD ["nadirclaw", "serve", "--host", "0.0.0.0"]) could never boot a fresh container, and the same failure hit systemd units and CI. The prompt is now gated on sys.stdin.isatty(); headless starts fall through to the existing "Starting with defaults" path.
  • Python 3.10 installs were broken by litellm 1.98.0. That release imports NotRequired from typing in llms/anthropic/experimental_pass_through/context_management/editors/compact.py, but typing.NotRequired only exists on 3.11+, so import nadirclaw raised ImportError on every fresh 3.10 install — while litellm still declares Requires-Python: >=3.10. The dependency is now capped below 1.98.0 for Python < 3.11 only; 3.11 and 3.12 continue to track the latest litellm. The cap comes off once upstream fixes the import.

Internal

  • CI ran no server tests. .github/workflows/ci.yml had excluded tests/test_server.py since the workflow was first added, so every test covering /v1/messages, /v1/chat/completions, routing headers, request logging, and the parameter reconcile above was skipped on every push and pull request. The exclusion is removed; the file passes on 3.10–3.12.

[0.22.0] - 2026-07-20

Changed

  • License changed from MIT to the PolyForm Noncommercial License 1.0.0. NadirClaw is now free for any noncommercial use (personal, research, education, evaluation, and noncommercial organizations). Commercial use requires a separate commercial license, available via getnadir.com. This applies to new versions; releases previously published under MIT remain available under MIT. The bundled wide_deep_asym_v3 classifier weights and the cascade-verifier-v1 snapshot are now released under the same noncommercial terms.

Added

  • v3+complex-gate is the default wide-deep routing. The bundled wide-deep classifier now defaults to the v3 checkpoint under a Neyman-Pearson complex gate (complex_gate_v1.pt, τ=0.12) with a head medium/simple split, matching the routing that runs in Nadir Pro. Unlike asym, v3 keeps a real P(simple), and the head split under-routes less than the legacy companion LR. The gate is on by default for v3 and off for asym/symmetric, so legacy behaviour is intact. Tunable via NADIR_COMPLEX_GATE=0, NADIR_GATE_THRESHOLD, and NADIR_MS_SPLIT=companion. Measured on RouterArena (n=2,479): 8.2% miss-complex, ~41% cost reduction vs always-premium. NadirClaw's global default classifier stays binary; see MODEL_CARD.md.

[0.21.1] - 2026-06-25

Added

  • Opt-in Claude Code identity injection for OAuth tokens (NADIRCLAW_CLAUDE_CODE_IDENTITY=1) — Anthropic gates premium models (Sonnet/Opus) behind subscription/OAuth tokens (sk-ant-oat*) unless the request leads with the official Claude Code identity system block. The real Claude Code client always sends it; raw API/SDK callers omit it and get a bare rate_limit_error on those models while Haiku works (#74). When enabled, /v1/messages and the OAuth completion path prepend "You are Claude Code, Anthropic's official CLI for Claude." as the first system block — only for Bearer/OAuth tokens (no effect on sk-ant-api* keys), only when not already present, preserving any caller-supplied system prompt after it. The decision is recorded as claude_code_identity on the request log. Default off, since it changes the system prompt the model sees.

Fixed

  • OAuth completion path sent system turns as chat messages — the direct Anthropic OAuth call in /v1/chat/completions forwarded role: "system" messages inside the messages array, which Anthropic's /v1/messages API rejects (system must be a top-level field). System/developer turns are now collected into the top-level system field before forwarding (#74).

[0.21.0] - 2026-06-24

Added

  • /v1/messages/count_tokens endpoint — Anthropic-native clients (Claude Code, the official anthropic SDK) call count_tokens to size requests before sending; this previously 404'd, so clients silently fell back to approximate local token estimation. The new route resolves the model through the same router as /v1/messages and forwards to Anthropic's real count_tokens, returning {"input_tokens": N} verbatim. Non-billable, so it is excluded from cost/budget recording (#72).

Fixed

  • /v1/messages traffic was invisible to metrics and budget — record_request short-circuited on type != "completion", dropping every /v1/messages log entry (type="messages"). Since Claude Code and the Anthropic SDKs talk to /v1/messages, all of that traffic was missing from Prometheus counters and the budget tracker. The recorder now accepts messages entries, and the /v1/messages handler computes cost via the budget tracker and stamps status/latency/token counts on both the streaming (usage recovered from the SSE message_start/message_delta events) and non-streaming paths (#71).
  • max_tokens was not reconciled against the routed model's output ceiling — when the router rewrote model to a tier whose max-output is lower than the client-supplied max_tokens, Anthropic returned an intermittent 400 that was proxied straight back. /v1/messages now detects a max_tokens: N > M 400, clamps max_tokens to the reported ceiling M, and retries once (both streaming and non-streaming), tagging the non-streaming response with X-NadirClaw-MaxTokens-Clamped: true (#73).
  • nadirclaw test failed for Claude subscription tokens — the command called litellm.completion() directly, which sends sk-ant-oat* subscription/OAuth tokens as x-api-key instead of Authorization: Bearer + the oauth-2025-04-20 beta header the server uses. It now probes Anthropic models through the same OAuth path as the running server, so nadirclaw test reflects real server behavior (#74).

[0.20.0] - 2026-06-12

Added

  • Context-optimizer compression upgrades (#65):
    • Pluggable backend — NADIRCLAW_OPTIMIZE_BACKEND selects native (default, built-in stdlib pipeline) or headroom (opt-in, delegates to the Apache-2.0 headroom-ai package via pip install nadirclaw[headroom]). Headroom is lazy and fail-open: if it is not installed or raises, the optimizer transparently falls back to native and the request never fails. Per-request override via optimize_backend in the body.
    • Progressive (staged) compression — --optimize progressive / NADIRCLAW_OPTIMIZE=progressive runs an escalation ladder (native_safe → native_aggressive → headroom_structural → headroom_ml) that stops as soon as NADIRCLAW_OPTIMIZE_TARGET_TOKENS is met. With no budget set it stops after native_aggressive (dependency-free, lossless); Headroom stages are skipped silently when headroom-ai is absent; the lossy ML stage runs only when NADIRCLAW_OPTIMIZE_ALLOW_LOSSY is on. Tunable via NADIRCLAW_OPTIMIZE_MAX_STAGE. New library entrypoint nadirclaw.optimize.compress_progressive().
    • Columnar JSON-array packing (json_array_pack, aggressive mode) — rewrites homogeneous arrays of same-keyed objects (DB results, API list responses, large tool outputs) into a header plus one value-array per row, emitting each key once. Information-lossless and deterministically reversible; ~68% vs pretty-printed JSON. Never runs in safe mode.
    • Native CCR (nadirclaw/ccr.py) — deterministic offload + nadir_retrieve fetch-back loop that moves oversized content out of the prompt behind a retrieve handle, fully reversible because the originals are kept server-side. Library-only for now (not yet wired into nadirclaw serve).
    • Apache-2.0 attribution for headroom-ai in THIRD_PARTY_NOTICES.md; docs, benchmarks, and tests (ccr, progressive, json_array_pack, backends, code-safety).

Fixed

  • Whitespace normalization corrupted unfenced source code — the whitespace_normalize transform collapsed the leading indentation of raw (unfenced) code arriving as file-read tool outputs, flattening nested Python/YAML/diffs into invalid syntax while reporting it as "savings". It now preserves leading indentation and only collapses interior multi-spaces, in both safe and aggressive modes (#65).

[0.19.3] - 2026-06-12

Fixed

  • OpenAI OAuth login failed with authorize_hydra_invalid_request — nadirclaw auth openai login sent redirect_uri=http://127.0.0.1:1455/auth/callback in the authorize request while the callback server bound and printed localhost. OpenAI's Hydra authorization server exact-matches redirect_uri against the client allow-list and rejected the 127.0.0.1 variant before the login screen. Now uses localhost consistently, matching the callback server and the Antigravity/Gemini flows (#67, #69).

[0.18.0] - 2026-05-25

Added

  • Application Default Credentials (ADC) for Gemini — when no GOOGLE_API_KEY is set, the Gemini path now falls back to google.auth.default() so users can authenticate via GOOGLE_CLOUD_PROJECT / GOOGLE_CLOUD_LOCATION (Vertex AI / gcloud-managed creds) instead of pasting a key. Original work by @froody (#57).
  • nadirclaw status displays the mid-tier model when one is configured, alongside simple/complex (#57).

Fixed

  • Gemini streaming was broken — _dispatch_model_stream consumed _stream_gemini (an async generator) with a plain for loop, which would raise TypeError: 'async_generator' object is not iterable on any actual streaming Gemini call. Now uses async for, and chunk / finish_reason parsing is robust to the google-genai SDK returning enum-like objects (#57).
  • savings no longer crashes on None values in the request log for selected_model and tier — these show up for failed / aborted requests and previously broke the report (#57).

[0.17.0] - 2026-05-15

Added

  • Configurable embedding backends for the centroid classifier — NADIRCLAW_EMBEDDING_BACKEND (default sentence-transformers; also ollama via /api/embed), NADIRCLAW_EMBEDDING_MODEL, NADIRCLAW_EMBEDDING_API_BASE, and NADIRCLAW_CENTROID_DIR. Custom centroid directories require a centroid_metadata.json (schema-versioned, with prototypes_hash for traceability) so users never silently mismatch a self-built centroid against a different encoder. nadirclaw build-centroids gains --backend, --model, --api-base, --output-dir flags. Original work by @clawSean (#50).
  • Optional prompt-injection guard — nadirclaw/prompt_guard.py. Heuristic detection of 7 patterns (instruction override, role reassignment, prompt extraction, JSON role confusion, delimiter injection, encoded payloads, DAN/jailbreak). NADIRCLAW_PROMPT_GUARD: log (default) / warn / block. Scans only user/tool messages — system/assistant treated as trusted. Original work by @pradumna-gautam (#55, supersedes #31).
  • Optional PII redactor — nadirclaw/pii_redactor.py. Detects email, US phone, SSN, and Luhn-validated credit-card numbers. NADIRCLAW_PII_REDACTION: none (default) / log_only / redact. Non-streaming responses only. Original work by @pradumna-gautam (#55).

Security

  • Production hardening baseline — recommended for anyone exposing nadirclaw serve beyond localhost. Original work by @pradumna-gautam (#30).
    • CORS: explicit allowlist via NADIRCLAW_CORS_ORIGINS; localhost regex default; never wildcard + credentials.
    • Auth: constant-time token comparison via hmac.compare_digest to defeat timing-side-channel guessing.
    • Security headers on every response: X-Content-Type-Options: nosniff, X-Frame-Options: DENY, Referrer-Policy: strict-origin-when-cross-origin, Cache-Control: no-store on /v1/*, opt-in HSTS via NADIRCLAW_HSTS=true.
    • Bounds validation on ChatCompletionRequest: caps on messages (500), max_tokens (100K), temperature (0–2), top_p (0–1), n (1–8) — closes a cost-amplification surface.
    • Sanitized validation errors — Pydantic internals no longer leak to clients; full details still server-side logged.
    • Async logging — SQLite writes moved off the event loop into a ThreadPoolExecutor, with done_callback exception logging and shutdown(wait=True) on SIGTERM so queued entries drain instead of dropping.
    • Prompt truncation — 500-char default in SQLite request logs (configurable via NADIRCLAW_LOG_PROMPT_TRUNCATE); API-key shaped tokens (sk-…, AIza…, ghp_…, gho_…, xox[bpars]-…) redacted from logged system prompts.

[0.16.0] - 2026-05-14

Added

  • Anthropic-compatible /v1/messages endpoint — Anthropic-native clients (Claude Code) now route through NadirClaw. The proxy classifies, rewrites the model field, forwards to api.anthropic.com, and pipes SSE streaming through byte-for-byte (#51).
  • Seamless Claude Code integration — nadirclaw claude onboard / shim / uninstall. Onboarding detects models, maps them into tiers, persists ANTHROPIC_BASE_URL + ANTHROPIC_MODEL into ~/.claude/settings.json, and installs a launchd / systemd auto-start unit (#51).
  • Live model detection — onboarding queries Anthropic's /v1/models using the stored token (Bearer for subscription tokens, x-api-key for API keys) instead of a hardcoded list; --interactive lets you pick a model per tier (#51).
  • Pluggable complexity classifier — NADIRCLAW_COMPLEXITY_ANALYZER=binary (default, ~10ms centroid) or distilbert (3-class fine-tuned DistilBERT predicting simple/mid/complex natively). The DistilBERT artifact downloads from the Hugging Face Hub on first use with a graceful fallback to binary (#51, #52).
  • Pro upsell surfaces — nadirclaw savings / serve / report and the README now surface Nadir Pro at high-intent moments with attribution-tagged URLs; new demo/cost_vs_opus.py zero-API-key demo (#53).
  • Enriched /v1/models — responses now include Anthropic-style type / display_name / description / created_at alongside the OpenAI-style fields.

Fixed

  • ANTHROPIC_BASE_URL is written as the bare host (Claude Code appends /v1/messages itself; a /v1 suffix produced a broken /v1/v1/messages path) (#51).
  • Updated the stale Claude model fallback list from the 4.5/4.1 generation to the 4.6 family (#51).

[0.15.0] - 2026-05-09

Added

  • nadirclaw update-models command — writes refreshable model metadata to ~/.nadirclaw/models.json, optionally merging a published registry JSON via --source-url or NADIRCLAW_MODEL_REGISTRY_URL.
  • Local model metadata overrides — the router now merges ~/.nadirclaw/models.json and user-managed ~/.nadirclaw/models.local.json into the runtime model registry.
  • DeepSeek V4 explicit aliases — added deepseek-v4, deepseek-v4-flash, and deepseek-v4-pro while preserving the existing deepseek alias for deepseek/deepseek-chat.
  • Model pool weighted load balancing — pool tier configuration with weighted round-robin across multiple models in the same tier (#36).
  • Selective context compression module — opt-in compression for tool-heavy contexts (#40).
  • Complex coding detection and enhanced reasoning markers — improved tier classification for coding-heavy prompts and Chinese reasoning markers (#38).
  • Upgrade-only session cache for agent frameworks — caches routing decisions per session to avoid repeated downgrades on multi-turn agent flows (#27).
  • Agent role detection for AI coding assistants — recognizes Claude Code / Cursor-style system prompts and routes accordingly (#37/#45).
  • Fallback reasons logging — failed fallback attempts now record ordered per-model fallback_reasons with compact error types and sanitized messages (#47).
  • Provider health-aware fallback routing — optional NADIRCLAW_PROVIDER_HEALTH=true mode tracks in-process model health and tries healthy fallback candidates before cooling-down ones; debug snapshot via /internal/provider_health (#48).

[0.14.0] - 2026-04-03

Added

  • Thinking/reasoning token passthrough — transparently forwards thinking parameters and extracts reasoning content from all provider paths:
    • Request forwarding: reasoning_effort (OpenAI o-series), thinking (Anthropic extended thinking), thinking_config (Gemini), and response_format are now passed through to LiteLLM, Anthropic OAuth, and Gemini native paths.
    • Response extraction: reasoning_content (DeepSeek), thinking blocks (Anthropic), and thought parts (Gemini) are captured from LLM responses and included in choices[].message.
    • Usage reporting: completion_tokens_details.reasoning_tokens surfaced when providers report thinking token counts.
    • Works in both streaming (real SSE and fake/cached SSE) and non-streaming response formats.
  • 15 new tests covering thinking parameter forwarding, response extraction, JSON serialization safety, and streaming passthrough.

[0.13.0] - 2026-03-20

Added

  • Context Optimize — new preprocessing stage that compacts bloated context before LLM dispatch, reducing input token cost by 30-70%. Two modes:
    • safe — five deterministic, lossless transforms: JSON minification, whitespace normalization, system prompt dedup, tool schema dedup, chat history trimming.
    • aggressive — all safe transforms + diff-preserving semantic deduplication. Uses sentence embeddings (all-MiniLM-L6-v2) to detect near-duplicate messages (cosine similarity >= 0.85), then extracts only the unique diff phrases using difflib.SequenceMatcher. Refinements survive dedup — "return values, not indices" is preserved even when 90% similar to an earlier message.
  • Accurate token counting with tiktoken — uses cl100k_base BPE tokenizer instead of len//4 heuristic. Falls back gracefully if tiktoken is not installed.
  • Shared sentence encoder — lazy-loaded SentenceTransformer singleton in nadirclaw/encoder.py for aggressive mode. No import cost when using safe mode or off.
  • nadirclaw optimize command — dry-run CLI tool to test context compaction on files or stdin. Supports --mode safe|aggressive and --format text|json.
  • --optimize flag on nadirclaw serve — set optimization mode at startup (off, safe, aggressive).
  • Per-request optimize override — pass "optimize": "safe" in the request body to override the server default for individual requests.
  • Optimization metrics — tokens_saved, original_tokens, optimized_tokens, and optimizations_applied logged per request in JSONL, SQLite, and Prometheus. Web dashboard shows aggregate savings.
  • New env vars: NADIRCLAW_OPTIMIZE (default: off), NADIRCLAW_OPTIMIZE_MAX_TURNS (default: 40).
  • 60 automated tests covering safe transforms, aggressive semantic dedup, accuracy preservation, edge cases, and roundtrip integrity.

Changed

  • SQLite schema: added columns optimization_mode, original_tokens, optimized_tokens, tokens_saved, optimizations_applied (auto-migrated on startup).

[0.7.0] - 2026-03-02

Added

  • nadirclaw test command — probes each configured model tier with a short live request and reports latency, response, and pass/fail. Exits with code 1 on failure so it works in CI. Supports --simple-model, --complex-model, and --timeout overrides.
  • classify --format json — new --format text|json flag on nadirclaw classify. JSON output includes tier, is_complex, confidence, score, model, and prompt. Composable with jq.
  • Multi-word prompt support for classify — nadirclaw classify What is 2+2? now works without quoting. Previously only the first word was captured.

Changed

  • nadirclaw savings now prefers SQLite — mirrors nadirclaw report: reads from requests.db when available, falls back to requests.jsonl. Previously only JSONL was read, giving empty or stale results for users without a JSONL file.
  • nadirclaw dashboard now prefers SQLite — same fix as savings; dashboard no longer shows empty data when only requests.db exists.
  • SessionCache LRU eviction is now O(1) — replaced List[str] + list.remove() (O(n) per cache hit) with collections.OrderedDict + move_to_end() / popitem(last=False), both O(1). Affects routing.py.
  • ModelRateLimiter.get_status is now thread-safe — all reads of _limits, _hits, and _default_rpm are now taken inside the lock, eliminating a potential data race under concurrent requests.

Fixed

  • auth status indentation — the "no credentials" help block was over-indented (12 spaces) and the provider hint strings were misaligned. Fixed to consistent 4-space indentation.
  • Removed redundant load_dotenv() in serve — settings.py already loads ~/.nadirclaw/.env at import time; the extra bare load_dotenv() call in the serve command was a no-op that could cause confusion when debugging env resolution.

[0.6.1] - 2026-02-28

Fixed

  • OpenClaw onboard: register nadirclaw provider without overriding the agent's primary model

[0.6.0] - 2026-02-26

Added

  • Configurable fallback chains — when a model fails (429, 5xx, timeout), cascade through a configurable list of fallback models. Set NADIRCLAW_FALLBACK_CHAIN to customize the order.
  • Real-time spend tracking and budget alerts — every request's cost is tracked by model, daily, and monthly. Set NADIRCLAW_DAILY_BUDGET and NADIRCLAW_MONTHLY_BUDGET for alerts at configurable thresholds. New nadirclaw budget CLI command and /v1/budget API endpoint.
  • Prompt caching — LRU cache for identical prompts. Configurable TTL (NADIRCLAW_CACHE_TTL, default 5min) and max size (NADIRCLAW_CACHE_MAX_SIZE, default 1000). New nadirclaw cache CLI command and /v1/cache API endpoint. Toggle with NADIRCLAW_CACHE_ENABLED.
  • Web dashboard — browser-based dashboard at /dashboard with auto-refresh. Shows routing distribution, per-model stats, cost tracking, budget status, and recent requests. Dark theme, zero dependencies.
  • Docker support — official Dockerfile and docker-compose.yml. docker compose up gives you NadirClaw + Ollama for a fully local zero-cost setup.

Changed

  • Fallback logic upgraded from simple tier-swap to full chain cascade
  • Request logs now include per-request cost and daily spend
  • Budget state persists across restarts via budget_state.json

[0.3.0] - 2025-02-14

Added

  • OAuth login for all major providers: OpenAI, Anthropic, Google Gemini, Google Antigravity
  • Interactive Anthropic login — choose between setup token or API key
  • Gemini OAuth PKCE flow with browser-based authorization
  • Antigravity OAuth with hardcoded public client credentials (matching OpenClaw)
  • Provider-specific token refresh (OpenAI, Anthropic, Gemini, Antigravity)
  • Atomic credential file writes to prevent corruption
  • Port-in-use error handling for OAuth callback server
  • Test suite with pytest (credentials, OAuth, classifier, server)
  • CONTRIBUTING.md and CHANGELOG.md

Changed

  • Version is now single source of truth in nadirclaw/__init__.py
  • Credential file writes use atomic temp-file-and-rename pattern
  • Token refresh failures return None instead of silently returning stale tokens
  • OAuth callback server binds to localhost (was 127.0.0.1)

Fixed

  • Version mismatch between __init__.py, cli.py, server.py, and pyproject.toml
  • README references to nadirclaw auth gemini-cli (now nadirclaw auth gemini)
  • OAuth callback server getting stuck (now uses serve_forever())

[0.2.0] - 2025-01-20

Added

  • OpenAI OAuth login via Codex CLI
  • Credential storage in ~/.nadirclaw/credentials.json
  • Environment variable fallback for API keys
  • nadirclaw auth command group

[0.1.0] - 2025-01-10

Added

  • Initial release
  • Binary complexity classifier with sentence embeddings
  • Smart routing between simple and complex models
  • OpenAI-compatible API (/v1/chat/completions)
  • SSE streaming support
  • Rate limit fallback between tiers
  • Gemini native SDK integration
  • LiteLLM support for 100+ providers
  • CLI: serve, classify, status, build-centroids
  • OpenClaw and Codex onboarding commands