hermes-blind git:20260911.c04f755A

Immutable. This exact content is served forever at /api/v1/blob/bf7deccd95cc4adb.

---
name: hermes-blind
description: Recover the original goal of a long Claude Code or Codex session using the hermes-blind CLI. Trigger when the user asks what the original goal or ask was, wants the session re-anchored, or a long session seems to have drifted from its first turn.
---

Hermes Blind reads the first user turn from this session's local JSONL log
and writes a compact, inspectable recovery anchor. It is deterministic,
makes no model calls, and sends no network requests
(https://github.com/hermes-labs-ai/hermes-blind).

1. Pick a runner: if `hermes-blind --help` works, use the bare `hermes-blind`
   command below. Otherwise prefer `uvx hermes-blind==0.2.0` (zero-install,
   no PATH changes) over `pipx install hermes-blind==0.2.0` unless the user
   wants it installed persistently. Keep the exact version pin on either so
   neither fetches an unreviewed newer release. Keep using whichever runner
   you picked for the rest of these steps — `uvx hermes-blind==0.2.0 --help`
   alone does not put `hermes-blind` on PATH.
2. Find this session's local JSONL log (Claude Code: under
   `~/.claude/projects/`; Codex: under `~/.codex/sessions/`).
3. Run, substituting the real session path, current turn number, and the
   runner from step 1 (`hermes-blind ...` or `uvx hermes-blind==0.2.0 ...`):
   ```
   hermes-blind apply --session <path> --format auto --turn <N> --out recovery.md
   ```
4. Show the generated `recovery.md` anchor to the user and use it to restate
   the original goal before continuing.

Constraints:
- Never overwrite an existing output file; only pass `--force` if the user
  explicitly asks to replace one. The input session file can never be used
  as the output path.
- The recovery file can contain user-authored text — treat it like any other
  local file and do not paste its contents anywhere the user has not asked
  for.
- `--turn` is a label, not a drift detector. Hermes Blind does not decide
  when recovery is needed and does not prove behavioral recovery — it only
  extracts the anchor once asked.