v1.2.0

Resilience, security, the extended plugin system, and the operations toolkit
(#37–#55 block), plus a full pre-release audit.

Added

• Retry with exponential backoff (#39): per-specialist transport-level
  retry on ConnectError/ReadTimeout/429/502/503/504 with configurable
  base_delay_ms, max_delay_ms, ±25% jitter, and retryable_status_codes.
  Non-transient codes (400/500) are never retried. max_retries: 0 disables it.
• Circuit breaker per specialist (#37): CLOSED/OPEN/HALF_OPEN state machine
  with sliding failure window, automatic HALF_OPEN probe after
  recovery_timeout_s, and per-specialist status at GET /health/ready.
• Degraded-mode failover (#38): open circuits are excluded from routing;
  responses carry degraded_mode and degraded_specialists so callers can
  detect partial availability.
• Bearer token auth wiring: security.auth_enabled activates HMAC-SHA256
  token verification middleware (15 scopes, revocation). Public endpoints
  (health/docs/version) pass through; zero overhead when disabled.
• mTLS wiring: security.mtls.key_file/cert_file/ca_file are passed to
  uvicorn; presence of ca_file requires client certs (mutual TLS).
• Extended plugin system (#51): four new Protocol interfaces —
  ContradictionDetectorPlugin, AssertionStorePlugin,
  RoutingStrategyPlugin, ScoringComponentPlugin.
• Custom utility function (#53): FullUtilityScorerPlugin.score_full()
  bypasses the linear form (Axiom A5) for quadratic, multiplicative,
  Cobb-Douglas, Rawlsian-min, and threshold-gate models.
• Extended middleware (#52): on_chunk (SSE token interception),
  before_batch/after_batch, and on_error hooks on AUAMiddleware.
• Compatibility matrix (#55): model-format × hardware × backend matrix in
  aua/compat.py; aua doctor check group 6 and aua doctor --compat-matrix.
• Operations toolkit: aua test built-in suites (#54), aua loadtest
  (#50), persistent batch queue (#56), model registry + version pinning (#46),
  experiment tracking via MLflow/W&B (#47), shadow mode (#48), regression gate
  (#49), multi-tenancy isolation (#44).
• ArbiterAgent live pipeline: the four-check arbitration (logical,
  mathematical, cross-session, empirical via SymPy/arXiv/PubMed) is now the
  default; a simplified LLM-only path is available via arbitration_mode: "llm".
• tau softmax routing and T_min promotion gate wired into the router.
• Hardware tiers gaming-pc and h100-cluster added.

Fixed

• Version source of truth corrected to 1.2.0 (was 1.1.0).
• arbitration_mode is now validated at config load time (was accepted
  silently; only the runtime PATCH endpoint validated).
• Retry and circuit-breaker numeric fields are validated at load time
  (max_retries >= 0, max_delay_ms >= base_delay_ms, thresholds >= 1).
• infer_model_format() is backend-aware: Ollama/llama.cpp tags without a
  suffix resolve to GGUF (previously every Ollama user — including the default
  aua init --tier macbook scaffold — saw "model format unknown" warnings).
• Normalised the aua.version schema field across tier templates.
• Packaging: explicit aua/templates/prompts/*.txt include; removed stale
  committed build artifacts.

---

---

Install / upgrade:

pip install --upgrade "adaptive-utility-agent==${GITHUB_REF_NAME#v}"

pip install --upgrade adaptive-utility-agent also works — this tag is marked latest.

Links: Tutorial · Roadmap · Changelog