deepagent-hermes
Copyright (c) 2026 Kedar Dabhadkar

This project is a faithful reproduction of the design ideas in Nous Research's
Hermes Agent (https://github.com/nousresearch/hermes-agent), released by Nous
Research under the MIT License.

The following design decisions, prompt structures, configuration defaults, file
formats, and architectural patterns are adapted directly from the Hermes Agent
codebase (v0.15.1):

- The closed reflection loop with 10-iteration / 10-turn nudge intervals.
- The three-layer (stable / context / volatile) system prompt assembly with
  byte-stable prefix-cache discipline.
- The "system_and_3" Anthropic prompt-caching strategy.
- The 13-section context-compression summary template.
- The 90 / 50 parent / subagent iteration budget defaults.
- The frozen-snapshot pattern for MEMORY.md / USER.md files (loaded once at
  session start, mid-session writes do not change the system prompt).
- The agentskills.io–compatible SKILL.md format and progressive-disclosure
  loading via skill_view.
- The SQLite FTS5 schema with messages_fts (unicode61) and messages_fts_trigram
  (CJK) virtual tables, BM25 ranking, and the DISCOVERY / SCROLL / BROWSE
  session_search modes.
- The cron job JSON shape including the skills / script / no_agent /
  context_from / enabled_toolsets fields, the 60-second tick interval, and the
  [SILENT] suppression marker.
- The 4-source plugin discovery order (bundled, user, project, pip entry-points)
  and the 17 lifecycle-hook names.
- The curator skill-lifecycle state machine (active → stale at 30 days →
  archived at 90 days; pinned skills immune).

Where prompt text is reproduced verbatim, the source file in the Hermes Agent
repository is cited inline in the corresponding .md file under prompts/.

Nous Research is the originator of these ideas; this project's contribution is
the reproduction on a different framework substrate (LangGraph + deepagents +
langgraph-stream-parser instead of Hermes's native runtime).

In addition to the design adaptations above, the v0.1.0a1 release bundles 26
SKILL.md files copied verbatim from the Hermes Agent ``skills/`` directory
under the following categories: software-development, github, research,
data-science, mlops, productivity, and note-taking. Each SKILL.md retains its
original frontmatter, including the ``author`` field crediting the original
contributor(s) and the ``license: MIT`` field. The bundled set is a curated
subset chosen to align with this project owner's stated use cases; the full
Hermes skill catalog (90+ skills in ``skills/`` plus 86+ in
``optional-skills/``) is available at
https://github.com/nousresearch/hermes-agent/tree/main/skills.

Some bundled skills were originally adapted by Hermes Agent from upstream
sources (e.g. the ``systematic-debugging`` skill is credited as "adapted from
obra/superpowers" in its own frontmatter). Those upstream attributions are
preserved verbatim in the corresponding SKILL.md files.
