hermes-sci v0.1.0 — Hermes-native autonomous paper writer

First public release of hermes-sci, a Hermes-native skill that turns
a research idea + author-supplied results into a LaTeX paper with a
built-in numerical-claim audit.

Highlights

• Hermes-native config — reads ~/.hermes/config.yaml to inherit
  whatever LLM provider you've selected (MiniMax / OpenAI / DeepSeek /
  Moonshot / Anthropic / Groq / Together / xAI / Zhipu / Gemini).
• Dual backend — pure provider API or hybrid with claude -p
  subprocess shim (useful when you have a Claude Max subscription).
• Anti-hallucination audit — every numeric claim in the generated
  LaTeX is matched against an author-supplied `results.json`
  registry with tolerance; unverified numbers can be highlighted.
• Apple Silicon aware — tier-detects MPS, injects honest compute
  budget into prompts so the LLM doesn't claim 8×A100 on an M2.
• Modular sanitize pipeline — LaTeX cleaning is a list of passes
  (reasoning / fences / CJK / packages / markdown / bad-cmds / items /
  escape); most rules live in `data/*.yaml` and new rules can be
  contributed by editing YAML, no Python change.
• Async parallel sections — provider-aware concurrency (drops to
  1 during MiniMax peak hour 15:00-17:30 Asia/Shanghai).

Install

```bash
hermes skills tap add easyvibecoding/hermes-sci
hermes skills install easyvibecoding/hermes-sci/hermes-sci --force --yes
```

`--force` is needed because the skill intentionally reads Hermes
config and sets provider env vars — both of which look "dangerous"
to the generic audit but are core features, not exfiltration.

Validated

End-to-end on a real 30-run delegation benchmark:

• MiniMax backend → 6-page PDF, 91-100% numerical-claim verification
• Hybrid (Claude Opus 4.6) backend → 6-page PDF, 94% verification
• 0 true hallucinations across both runs

See `skills/hermes-sci/SKILL.md` for full usage.