Hermes Blind

Recover the original goal of a long Claude Code or Codex session—and add evidence constraints to evaluation prompts.

Long agent sessions can lose the shape of the request that started them.
Hermes Blind reads the first user turn from a local Claude Code or Codex JSONL
log and writes a compact recovery anchor you can inspect and paste back into
the session. It also provides a small prompt wrapper for evaluations that asks
the model to disclose prior exposure, quote its evidence, and hedge when the
evidence is thin.

The package is deterministic, dependency-free at runtime, and local: it makes
no model calls and sends no network requests.

Install

For the isolated command-line app:

pipx install hermes-blind

Or install it into your current Python environment:

python -m pip install hermes-blind

Requires Python 3.10+.

Recover a long agent session

The lowest-friction path is to give your coding agent this instruction:

Install hermes-blind. Find the JSONL log for this Claude Code or Codex
session, then run hermes-blind apply --session <path> --format auto --turn <current-turn-number> --out recovery.md. Show me the generated
anchor and use it to restate my original goals before continuing. Do not
overwrite files or share the session text.

Or run it directly:

hermes-blind apply \
  --session /path/to/session.jsonl \
  --format auto \
  --turn 9 \
  --out recovery.md

The generated markdown starts like this:

# Recovery scaffold (anchor-extracted from turn 1, applied at turn 9)

## Original anchor
- stated_goal: "Ship the onboarding flow and verify the clean install"

## Session state
- session file: rollout.jsonl
- user turns observed: 9

--format auto recognizes Claude Code and Codex JSONL shapes. The default
goals mode preserves up to 12 goal-carrying sentences from the first user
turn; first-sentence keeps the compact legacy behavior and full includes
up to 4,000 characters.

The --turn value is only a label in the output. Hermes Blind does not detect
drift or decide when recovery is needed. Existing output files are preserved
unless --force is explicit, and the input session file can never be used as
the output path.

Recovery files include user-authored text. Inspect them before sharing.

Add evidence constraints to an evaluation prompt

From the CLI:

hermes-blind apply \
  --variant v1 \
  --prompt "Score this release from quoted evidence."

This prints a wrapped prompt without calling a model:

[HERMES-BLIND]
If you have prior exposure to this target or its author, state it in one line.
Score using only quoted evidence from the target text below.
Unknown or thin evidence = hedge; do not confabulate.
[/HERMES-BLIND]

Score this release from quoted evidence.

Or use the Python API:

from hermes_blind import wrap

prompt = wrap(
    "Rate this paper on novelty from 0 to 10 and cite the target text.",
    variant="v1",
)

Available variants are null, micro, short, v1, full, placebo, and
gate-only. The null variant is an exact no-op for controlled comparisons.
The package also exposes the dependency-free intent and scope preambles used
by Hermes Rubric.

Evidence and limits

The repository tests and CI cover deterministic wrapping, Claude Code and
Codex JSONL parsing, recovery modes, safe output handling, package
installation, and CLI invocation.

A frozen nine-session extraction audit found that the default goal-set anchor
represented 40 of 66 pre-listed goals, compared with 7 of 66 for the earlier
first-sentence heuristic. That supports better mission representation in the
generated artifact for the evaluated sessions. It does not establish that
reinserting the artifact changes model behavior or improves task outcomes.
See the evaluation report
for the method, limitations, sanitized results, and receipt hashes.

Not established:

• reliable bias reduction from the evaluation prefix;
• successful behavioral recovery after inserting an anchor;
• automatic drift detection or an optimal intervention turn;
• adversarial prompt-injection resistance; or
• non-English behavior.

Treat the output as a transparent scaffold for a human or agent to inspect,
not as a security boundary or independent evaluator.

Development

python -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[dev]"
ruff check src tests
pytest -q
python -m build
twine check dist/*

See the changelog
for release history and the
contribution guide
for contribution guidance.

License

MIT. See the license.

Built by Hermes Labs.