Skip to content

AGENTS.md

Latest commit

 

History

History
58 lines (37 loc) · 3.22 KB

AGENTS.md

File metadata and controls

58 lines (37 loc) · 3.22 KB
Raw

AGENTS.md — hermes-prime

Guidance for autonomous agents and LLM tooling that may consume, extend, or be primed by this package.

What this is

A bash bootstrap (bin/hermes-session-init) plus a markdown fragment (CLAUDE-fragment.md). The bootstrap injects the fragment into a project's CLAUDE.md so the next session opens with conventions in scope. Four subcommands: --check, --print, --inject, --uninject. ~120 LOC of bash, no Python, no dependencies.

What this is not

  • Not a model, an API, or a judge.
  • Not an automatic grounding caller — the fragment teaches the orchestrator when to call hermes-ground; it does not call it for you.
  • Not a global config patcher — it never touches ~/CLAUDE.md. Per-project only.
  • Not a runtime daemon — it runs once per inject; the fragment lives in the project repo from then on.

When to invoke (as an agent)

Use hermes-session-init --inject <project-dir> when:

  • You are starting recursive or emergent work in a fresh project (taxonomy design, rubric construction, structure-discovery debugging) AND
  • The project's CLAUDE.md does not yet contain the marker <!-- session-init: BEGIN -->.

Skip if:

  • The session is a single-task fix (the fragment adds noise without payoff).
  • The project's CLAUDE.md already has the marker (--inject is a no-op anyway, but skip the call).
  • The work is generation-task only with no rubric or grounding loop.

Minimal invocation

hermes-session-init --check                          # verify prereqs
hermes-session-init --inject /path/to/project        # idempotent
# ... do work ...
hermes-session-init --uninject /path/to/project      # cleanup

What the fragment teaches

After injection, a session reading the project's CLAUDE.md will know:

  1. When to call hermes-ground — 7 specific trigger conditions (e.g. "3+ commits without a user-visible ship", "about to coin a noun-phrase label for something with no implementation at a path").
  2. Where the audit-stack tools live — paths for hermes-ground, hermes-rubric-blinded, the handbook.
  3. The four standing conventions — calibrate-before-ship, rubric pass-through, grounding-on-emergence, no-noun-phrase-before-file.

The fragment is a snapshot of conventions at injection time. If the underlying memory feedback files change, re-inject to refresh.

Honest scope

hermes-prime does not measure whether the conventions are followed after injection. The companion tool hermes-ground does that for grounding. hermes-rubric does that for ship-quality. hermes-prime only guarantees the conventions are in scope.

The empirical claim that injection improves convention-recall in a fresh session is the subject of the E1 eval in evals/. Read the eval transcripts in evals/runs/ for the actual numbers.

Don't

  • Don't auto-inject into every project on a hook. The fragment costs tokens; only inject where the conventions apply.
  • Don't modify the fragment in-place inside a project's CLAUDE.md. The marker block is the contract; edits there will be wiped on --uninject.
  • Don't ship a noun-phrase label for a session-priming concept that does not yet have a file. The repo name is hermes-prime because the files now exist; that earned the noun.