# hermes-blind

> Dependency-free prompt and session-recovery scaffolds for LLM workflows.

## Public surface

- `wrap(prompt, variant="v1")`: prepend an evidence-gating evaluation prefix.
- `extract_disclosure(response)`: parse an optional disclosure line.
- `hermes-blind apply --session ...`: extract a turn-one goal anchor from
  Claude Code or Codex JSONL.
- `compose_intent(...)`: add optional intent-debias and scope-class framing.

The package is deterministic. It makes no model calls and no network requests.

## Install

    pip install hermes-blind

## Recovery example

    hermes-blind apply --session rollout.jsonl --format auto \
      --anchor-mode goals --turn 9 --out recovery.md

`--turn` is metadata, not a detector threshold. Recovery output may contain
user text and should be reviewed before sharing.

## Find a supported session log

Blind reads an explicit local JSONL path and currently supports Claude Code
and Codex session logs:

    ls -t ~/.claude/projects/*/*.jsonl | head
    find ~/.codex/sessions -type f -name 'rollout-*.jsonl' -print 2>/dev/null | sort | tail

Gemini CLI installs expose the package's prompt and skill surfaces; Gemini
session-log recovery is not currently supported.

## Evidence boundary

The 0.1.3 release validates package mechanics, JSONL parsing, anchor modes,
build/install behavior, and CLI operation. In a frozen 66-goal extraction
audit, goal-set extraction retained 40 goals versus 7 for the prior
first-sentence heuristic.

That result does not establish downstream drift recovery. Bias reduction,
behavioral recovery, automatic drift detection, optimal timing, adversarial
robustness, and non-English behavior remain unproven.

0.1.3 is the first public 0.1.x release; public delta: 0.0.6 → 0.1.3.

Repository: https://github.com/hermes-labs-ai/hermes-blind
License: MIT