feat: Hermes Agent integration — memory provider, skills, crons

[benslockedin]

Summary

• Memory provider for NousResearch Hermes Agent — 3 tools (alive_world, alive_load, alive_search), smart prefetch, stash persistence, squirrel YAML on session end
• 11 Hermes skills in agentskills.io SKILL.md format (load, save, world, capture, search, create, bundle, daily, history, mine, cleanup)
• 8 cron templates with setup script (morning rhythm, project pulse, inbox triage, health audit, stash router, mining, pruning, people refresh)
• Runtime files — soul-patch.md (personality layer) + agents.md (squirrel runtime rules)
• 3-path onboarding installer — detects ALIVE/Hermes/both and guides setup

Built from the 14-page alive-hermes-spec.pdf (LaTeX). Memory provider tested: 138 walnuts discovered, smart prefetch working, stash persistence confirmed.

Companion PR to NousResearch/hermes-agent will add just the memory provider (3 files) following the pattern of their existing providers (Honcho, Mem0, etc.).

Test plan

• Run memory provider import test: python -c "from hermes.memory_provider import ALIVEProvider; p = ALIVEProvider(); print(p.is_available())"
• Test alive_world tool returns walnut list
• Test alive_load tool returns briefing for a known walnut
• Run full Hermes Agent session with ALIVE as memory provider
• Test setup-crons.sh registers all 8 crons
• Verify install.sh detects correct onboarding path

🤖 Generated with Claude Code

[patrickSupernormal]

Merge safety review (advisory) — Patrick Brosnan, non-collaborator on alivecontext/alive. Comment weight only, not an approving review.

Verdict: YELLOW — 6 mechanical fixes needed in a follow-up commit before merge. Architecturally clean.

Headline: Architecturally this PR is the safest possible shape for a new-subsystem integration — pure additive (+2961/-0, zero deletions, zero modifications of existing files), everything under a new hermes/ top-level directory, stdlib-only Python (no PyYAML / requests / cryptography), zero file overlap with any other open PR, no hook registration inside plugins/alive/hooks/hooks.json, no changes to plugin.json. On merge, the worst-case impact is "a new directory exists" — no runtime interaction with the rest of the plugin. Path handling uses Path.relative_to(world_root) with a ValueError guard which correctly prevents escaping the world. Credit where due: this is how to ship a major new integration safely.

But 6 real bugs need a follow-up commit before merge:

1. hermes/install.sh:2025-2032 — quoted heredoc silently disables shell expansion. cat > ... << 'WORLDKEY' uses a single-quoted delimiter, so line 2028 created: $(date +%Y-%m-%d) lands in the world key as the literal string $(date +%Y-%m-%d). Every Path B onboarding run produces a malformed .alive/key.md. Fix: change << 'WORLDKEY' to << WORLDKEY (unquoted) or compute the date into a variable first.

2. hermes/hermes-skills/alive-search/SKILL.md:34 — typo: missing underscore in glob. First glob is "$ALIVE_WORLD_ROOT"/*/kernel/log.md (missing the _ prefix). Top-level walnuts (the common case, e.g. 04_Ventures/stellarforge/_kernel/log.md) are silently excluded from log search because the typo'd glob doesn't match them and the second glob needs 2 levels of nesting. Fix: */kernel/log.md → */_kernel/log.md.

3. hermes/memory-provider/__init__.py:2838-2871 — started and ended timestamps both equal session-end time. on_session_end computes a single now_iso and writes it to both fields. initialize() never captures the start time. Every squirrel YAML entry shows zero-duration sessions. Fix: capture self._started_at in initialize(), use it for started: in on_session_end.

4. hermes/memory-provider/__init__.py:2682-2694 — on_turn_start walks the entire world on every turn. for key_file in self._world_root.rglob("_kernel/key.md"): iterates every walnut on every user message, then only uses the result to emit a logger.debug(...) call — no state change. At 138 walnuts (per PR body), that's 138 filesystem stat calls per turn for a debug log line. Contradicts the provider's headline "smart prefetch saves tokens" value prop. Fix: cache at initialize() time, or remove the loop entirely since it doesn't influence state.

5. hermes/setup-crons.sh:3050-3104 — silent error hiding on every hermes cron create. 2>/dev/null && echo "[+]" || echo "[!] failed" discards the stderr that would tell the operator WHY the create failed (hermes not installed, flag unknown, schedule invalid), and the || echo rescue prevents set -euo pipefail from aborting. Operator sees 8 "[!] failed" lines with no diagnostic. Fix: write stderr to a log file or capture into a variable.

6. hermes/setup-crons.sh:3090 vs hermes/cron-templates/alive-mine/SKILL.md:1 — skill name mismatch. Cron registers --skill alive-mine-cron, but the folder is alive-mine/ and the SKILL.md frontmatter declares name: alive-mine-cron. Whether this works depends on Hermes' resolver — if it uses folder name, this crons to a non-existent skill. Fix: pick one name, apply consistently.

Minor (not blockers): walnut-name match uses substring instead of word boundary (daily matches "write a daily summary"); on_memory_write stashes every write unconditionally as insight_candidate (over-capture); hand-rolled YAML parser won't handle block scalars or lists; YAML writes are non-atomic (no tmp-rename); no tests for 838 lines of provider code. I would NOT block merge on adding tests given the isolation — ship the provider, get real Hermes users, add tests as the API stabilises.

Merge order: Middle (position 3-4). Fully isolated, can technically merge at any position, but size (27 files, +2961 LOC) deserves fresh reviewer attention — not late when fatigued by conflict resolution. Zero inter-PR conflicts with #28, #29, #31, #32, #34, #35.

Full cross-PR synthesis (conflict matrix, recommended merge order for all 7 open PRs, per-PR verdicts, blocker list): see the cover comment on #29 — #29 (comment)

Generated via flow-next epic fn-8-vly (stackwalnuts walnut). 🐿️