What's new in Aroha

Security

• —Replies to outbound consults and action requests are now verified before use. The request was always sealed to the peer's key, but the reply came back in plaintext and was never checked — so whatever answered the endpoint could return any text and it rendered attributed to the peer's DID. Both the signature and the sender DID are now checked, and a failing reply is discarded without passing any of its text on.
• —Action requests report an unknown outcome rather than a failure when a reply is discarded — a discarded reply says nothing about whether the peer acted.
• —Outbound consults are now recorded in the audit log. Only the answering side kept records, so an operator could see every question another organisation asked their agent and nothing about what their own agent asked others. Redactions are logged as categories, never as the values that were stripped.

Added

• —Local runtime — one binary. Detects the model engine you already run, gives each agent its own identity and keys, and serves an interface on 127.0.0.1.
• —Retrieval over your own documents, indexed on your machine, with the source named in the answer.
• —Skills in the open SKILL.md format, loaded on demand rather than held in the prompt.
• —Per-person memory, partitioned so one person's facts never surface for another. Off by default.
• —The agent's work is shown as it happens: what it searched, which documents it read, which skills it loaded, what the turn cost, and the exact prompt behind any answer.
• —Activity view over the hash-chained audit log, including whether the chain still verifies.
• —Cross-organisation consults with a provenance panel showing what was actually checked.
• —Eight starting points when creating an agent, so the system prompt is a starting position rather than a blank field.

Fixed

• —Token counts never reached the interface. Usage was parsed after an early return that skipped exactly the frame carrying it — the one with an empty choices array.
• —The memory toggle never took effect: a checkbox reaches FormData as the string "on", while the API accepts only a boolean.
• —Retrieval reported "no matching passages" every turn for agents that had no documents at all.
• —The knowledge panel reported "keyword matching only" on installs that had an embedding model, because the flag is derived from stored chunks and the store was empty.

Website

• —New two-zone design — dark for marketing, light for documentation — with a display typeface and a saturated accent that works on a dark ground.
• —New /runtime page. The runtime had shipped as software with no presence on the site at all.
• —Code samples on dark pages use a dark syntax theme; the previous transparent-background approach never applied, because shiki writes its colours inline.
• —Fixed a duplicate React key on /docs where three roadmap cards shared one destination.

Fixed

• —Made tampered-token credential test deterministic — previously had a 1-in-64 chance of being a no-op when the last base64url character was already the hardcoded replacement.
• —Publish pipeline now publishes each npm package individually to survive rate limits, instead of publishing in parallel and silently dropping packages on 429s.
• —Workflow dispatch on publish-npm is now restricted to main/tags only to prevent accidental re-runs from feature branches.

Changed

• —SDK version badges on website updated to v1.2.0.
• —Conformance suite restructured to catch timing-dependent test failures earlier in CI.

Added

• —@aroha-sdk/mcp-bridge is now GA — stable API for bidirectional MCP ↔ Aroha capability translation.
• —@aroha-sdk/hermes-bridge beta — connects Hermes Agent (Nous Research) and ZeroClaw (Rust MCP) via stdio.
• —Google A2A bridge (@aroha-sdk/a2a-bridge) added to registry.
• —BrandIcon + BrandBadge component system with official SVG paths for 15+ ecosystem partners.
• —MCP logo uses official 3-stroke SVG path (Linux Foundation / Anthropic).

Fixed

• —LangChain brand color corrected to #7FC8FF (was dark navy, invisible on white backgrounds).
• —BrandBadge now applies correct bg tint for stroke-based (MCP) vs fill-based icons.

Stable

• —Aroha Protocol specification v1 — message types, signing scheme, DID format, spending mandate schema.
• —@aroha-sdk/run, @aroha-sdk/core, @aroha-sdk/credentials, @aroha-sdk/orchestrator — stable API.
• —Python aroha package — asyncio-native, decorator-based serve(), Python 3.11+.
• —Registry server open-sourced under MIT — self-hostable, Cloudflare Workers deployment included.
• —Conformance test suite published — third-party SDK authors can validate compatibility.

Alpha (not for production)

• —Settlement and EscrowSettlement — on-chain financial features, API may change.
• —Go SDK and Java SDK — on roadmap for Q3 2026.

Added

• —Aroha Protocol v0.9 specification — subject to change before 1.0.
• —TypeScript SDK alpha — @aroha-sdk/core and @aroha-sdk/run.
• —Agent Hub (hub.aroha-labs.com) — beta registry for discovering agents.
• —Playground — interactive API explorer with sandbox echo, weather, and currency agents.
• —Studio (beta) — no-code agent builder.