release aidev BREAKING NEEDS REVIEW SIG 4/5

Hermes Agent v0.15.0: agent-loop refactor, Kanban orchestration, rebuilt session_search

The conversation loop in run_agent.py drops from 16,083 to 3,821 lines across 14 agent/* modules with forwarders kept for compatibility, Kanban becomes a multi-agent orchestration platform, and session_search is rewritten without an aux-LLM (free, ~20ms discovery instead of ~90s and ~$0.30/call). Anyone scripting Kanban, relying on the session_search config knob or companion skill, or running Grok models retired on May 15 should read before upgrading.

PUBLISHED2026-05-28
OBSERVED2026-08-11
AGE2mo
SOURCES1
  • run_agent.py 16,083 → 3,821 lines (-76%), extracted into 14 agent/ modules; thin forwarders on AIAgent keep test patch paths and external callers working
  • Kanban (104 PRs): triage auto-decomposes a task into a sub-task tree; hermes kanban swarm builds a Swarm v1 graph (root, parallel workers, gated verifier, gated synthesizer, shared blackboard); per-task model overrides, board default workdirs, per-task worktree paths/branches, scheduled start times, configurable claim TTL, retry fingerprinting, stale-task detection, respawn guards, drag-to-delete; worker reporting via /workers/active, /runs/{id}, /inspect
  • Cold start: deferred openai._base_client import (-240ms / -17MB per CLI invocation), hot-path work cuts per-conversation function calls 47% (399k → 213k over 31 turns), deferred compression-feasibility check (-170–290ms per agent construction), adaptive subprocess polling (-195ms per tool call). Termux 2.9s → 0.8s; hermes --version 701ms → 258ms
  • session_search: no aux-LLM, no cost; three modes (discovery, scroll, browse) inferred from which args are set — the mode parameter, config knob and companion skill are gone. Discovery ~20ms, scroll ~1ms
  • Promptware defense against Brainworm-class prompt injection at three chokepoints (tool output, recalled memory, stored skills), with tools/threat_patterns.py as single source of truth
  • Bitwarden Secrets Manager: one bootstrap token replaces per-provider API keys
  • Skill bundles — one slash command loads a whole workflow
  • Ink TUI gains a multi-session orchestrator
  • image_gen: Krea 2 Medium + Large added; FAL ported to a plugin
  • Nous-approved MCP catalog with interactive picker; OpenHands orchestration skill; ntfy as 23rd messaging platform
  • xAI: Web Search plugin, xai-oauth upstreamed into hermes proxy, detection of models retired May 15 plus hermes migrate xai, TTS speech-tag pauses, base_url leak guard, OpenAI-style execution guidance for Grok
  • 1,302 commits, 747 merged PRs, 560+ issues closed (15 P0, 65 P1, 19 security-tagged)
  • Note: the source text ends mid-sentence in the Promptware section, so the defense details are partial
featurecapabilitysecuritybreakingdeprecationapi-changemodel-support

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.