release aidev SIG 4/5

Hermes Agent v0.11.0: Ink TUI rewrite, pluggable transports, native Bedrock, GPT-5.5 via Codex OAuth

The interactive CLI is rewritten in React/Ink over a Python JSON-RPC backend, and format conversion plus HTTP transport are extracted from run_agent.py into a pluggable agent/transports/ layer that native AWS Bedrock (Converse API) is built on. GPT-5.5 becomes reachable through ChatGPT Codex OAuth and five new inference paths land; note that auto routing for side tasks now defaults to your main model, where aggregator users were previously routed to a cheap provider-side default.

PUBLISHED2026-04-23
OBSERVED2026-08-11
AGE3mo
SOURCES1
  • hermes --tui: React/Ink rewrite with Python JSON-RPC backend (tui_gateway) — sticky composer, live streaming with OSC-52 clipboard, stable picker keys, status bar with per-turn stopwatch and git branch, /clear confirm, light theme, subagent spawn observability overlay (~310 commits)
  • Transport ABC: AnthropicTransport, ChatCompletionsTransport, ResponsesApiTransport, BedrockTransport, each owning its own format conversion and API shape; native AWS Bedrock via the Converse API on top
  • New inference paths: NVIDIA NIM, Arcee AI, Step Plan, Google Gemini CLI OAuth, Vercel ai-gateway (pricing + dynamic discovery); Gemini routed through the native AI Studio API
  • GPT-5.5 over ChatGPT Codex OAuth, with live model discovery feeding the model picker
  • QQBot as 17th platform (QQ Official API v2): QR scan setup wizard, streaming cursor, emoji reactions, DM/group policy gating
  • Plugin surface: register_command, dispatch_tool, pre_tool_call veto, transform_tool_result, transform_terminal_output, image_gen backends, custom dashboard tabs; bundled disk-cleanup plugin as opt-in reference
  • /steer <prompt> injects a mid-run note the agent sees after its next tool call, without interrupting the turn or breaking prompt cache
  • Shell scripts usable as lifecycle hooks (pre_tool_call, post_tool_call, on_session_start, …) without a Python plugin
  • Webhook direct-delivery mode forwards payloads straight to a platform chat with no LLM in the path
  • Subagents get an explicit orchestrator role that can spawn workers (max_spawn_depth, flat by default); concurrent siblings share a file-coordination layer
  • hermes model gains a Configure auxiliary models screen (compression, vision, session_search, title_generation); auto now defaults side tasks to the main model for all users
  • Folds in the highlights deferred from v0.10.0; 1,556 commits, 761 merged PRs since v0.9.0
  • Note: source text truncated mid-sentence in the auxiliary-models item
featurecapabilitymodel-supportapi-change

COMMUNITY

No curated reactions recorded for this event. Facts and takes are kept in separate layers — community context is added by hand, never blended into the record above.