v1.3 — install actually works now (PR #3 merged)

v1.3 — doc correction.

Merged PR #3 (hendrixfreire): fix(install): use real hermes cron create CLI signature + doc consistency

The biggest deal here is install actually works now. v1.1's install.sh was calling hermes cron create with --schedule, --prompt, --prompt-file, --model, --provider — none of which the CLI accepts. bash install.sh was aborting before creating any cron jobs. Hendrixfreire validated against the live CLI and rewrote create_or_skip() to use positional schedule prompt, match the real cron list output format for idempotency, and feature-detect --model for forward compatibility.

README rename: SELF_SUPERVISED → PROPOSAL_ONLY_AMEND in the env table and the self-supervised section. (install.sh was always PROPOSAL_ONLY_AMEND; the README was the inconsistent piece.)

Constitutional gate alignment (prompts/critic.md): in proposal-only mode, GPG signature is now a warning, not a halt — aligned with the v3 design intent.

v1.3 follow-up: corrects the v1.1 changelog entry that incorrectly said PROPOSAL_ONLY_AMEND rename was "intentionally not changed". The rename was applied in PR #3; the changelog should have reflected it.

No code changes since v1.1 except the install.sh fix from PR #3.

v1.2 — clarify (no code changes)

v1.2 — clarify.

Adds a 5-row side-by-side comparison table to README explaining how hermes-self-evolution differs from Hermes' built-in learning loop.

The two mechanisms are complementary, not redundant:

• Hermes' built-in loop improves the agent in-session (skill creation + improvement during use, persistent user modelling).
• hermes-self-evolution audits and adjusts the agent between sessions under a formal governance protocol (Constitution + diverse-model Critic + falsifiable predictions).

Triggered by a recurring reader question on the launch tweet — first asked by @nahkhsinad. Embedding the answer in the README means future readers don't have to ask again.

Also softens the safety posture framing for the built-in loop from "None" to "No formal amendment protocol (in-session, LLM-curated)" — it has implicit in-session safety; it just lacks the formal amendment protocol this framework provides.

No code changes — README only.

v1.1 — review-driven hardening

v1.1 — hardening pass in response to hendrixfreire/hermes-self-evolution-review.

Fixes

• F1 (§1.5): SECURITY comment in scripts/execution-loop.sh expanded — the no-eval-of-markdown design choice (already in v2) is now explicit
• F2 (§1.3): install.sh now touches every facts file referenced by prompts or scripts on first run (was missing enforcement-log.md, learning-log.md)
• F4 (§1.4): scripts/execution-loop.sh drops the hardcoded known_arms list. _print_unused_arms() now reports only arms that have actually been dispatched
• F5 (§1.8): portable date parser — GNU date -d, BSD date -j -f, then stat mtime fallback. Silent-failure path removed
• F6 (§1.7): README §"Why this exists" softened — skill loader claim is a portability hedge, not a Hermes bug

Intentionally kept

• Gardener cron (review §4 redesign proposed deleting; we disagree — agent entropy is real)
• PROPOSAL_ONLY_AMEND=true default (constitution is human-only)
• 4-job architecture (Evolution + Critic + Verifier + Gardener)

Thanks to @hendrixfreire for the thorough audit.