gstack-hermes

YC-style review pipeline + threat-modeling for Hermes Agent.

Adapted from garrytan/gstack (Claude Code plugin) with minimal changes — full skill content preserved upstream, only the bits that genuinely break in Hermes are touched.

What you get

12 skills:

Skill	What it does
gstack	Entry-point menu. Two-level inline-keyboard (L1 categories → L2 concrete skill) via hermes clarify_tool. Run /gstack when you don't remember which /yc-* to call.
yc-office-hours	YC-style 6 forcing questions for a new idea. Saves a design doc to ~/.hermes/projects/<slug>/designs/.
yc-ceo-review	CEO-mode plan review: 4 scope modes (Expansion / Selective / Hold / Reduction).
yc-eng-review	Engineering-mode plan review: diagrams, edge cases, test plan.
yc-design-review	UX/UI plan review, AI-slop detection.
yc-devex-review	Time-to-hello-world DX audit.
autoplan	Batch orchestrator — runs CEO + Eng + Design + DevEx reviews sequentially with auto-decisions.
yc-retro	Weekly engineering retrospective.
yc-learn	Capture / search / prune project learnings. Writes ### sections to ~/.hermes/memories/lessons.md (compatible with file-size-guard hook).
sec-cso	Chief Security Officer mode: OWASP Top 10 + STRIDE threat-modeling on a plan.
document-generate	Generate missing documentation for a feature, module, or entire project.
ship	Release workflow: detect base branch, run tests, review diff, bump VERSION, update CHANGELOG, push.

Plus one hook:

Hook	What it does
pipeline-next-step	After agent:end, appends a ### Next Steps suggestion to MEMORY.md if you just ran a /yc-* or /sec-cso skill. Throttled to once per hour per suggestion.

Install

1. Add the tap

hermes skills tap add veeskelad/gstack-hermes

2. Install skills

Pick the ones you want:

hermes skills install veeskelad/gstack-hermes/yc-office-hours
hermes skills install veeskelad/gstack-hermes/yc-ceo-review
hermes skills install veeskelad/gstack-hermes/yc-eng-review
hermes skills install veeskelad/gstack-hermes/yc-design-review
hermes skills install veeskelad/gstack-hermes/yc-devex-review
hermes skills install veeskelad/gstack-hermes/autoplan
hermes skills install veeskelad/gstack-hermes/yc-retro
hermes skills install veeskelad/gstack-hermes/yc-learn
hermes skills install veeskelad/gstack-hermes/sec-cso
hermes skills install veeskelad/gstack-hermes/document-generate
hermes skills install veeskelad/gstack-hermes/ship

Note: hermes security scanner may flag the skills as dangerous (false positive on bash patterns like AGENTS.md, rm, sudo). If installs are blocked, sideload manually:

git clone https://github.com/veeskelad/gstack-hermes /tmp/gstack-hermes
mkdir -p ~/.hermes/skills/gstack-hermes
cp -r /tmp/gstack-hermes/skills/* ~/.hermes/skills/gstack-hermes/

How /gstack and auto-progression work

Discovery via /gstack:

/gstack
  ↓ (L1 clarify)
🧠 Review  ⚙️ Workflow  📝 Retro & Learn  🔒 Security
  ↓ (L2 clarify, per category)
/yc-office-hours  /yc-ceo-review  /yc-eng-review  /yc-design-review
  ↓
agent reads ~/.hermes/skills/gstack-hermes/<chosen>/SKILL.md
and executes it

If ~/.hermes/memories/MEMORY.md has a fresh ### Next Steps (gstack pipeline, ...) from the previous run, /gstack adds a 5th L1 option «↪ Продолжить с прошлого шага» that jumps straight to the suggested next skill.

Auto-progression after each skill:

Each skill's frontmatter declares its pipeline_successor:

metadata:
  hermes:
    pipeline_successor: yc-ceo-review

The HERMES_NOTE at the top of every SKILL.md instructs the agent: after completing the skill body, call clarify_tool with «Готово. Дальше — /?» and 3 buttons:

• Да → agent reads + executes the successor SKILL.md
• Пропустить → clean exit
• Сделать другое → free-form

Terminal skills (document-generate, yc-learn, gstack itself) have pipeline_successor: null and do not auto-suggest.

Backup channel: the pipeline-next-step/handler.py hook still writes the suggestion to ~/.hermes/memories/MEMORY.md on every agent:end. This is a secondary signal — if the model skipped the auto-progression clarify call for any reason, the suggestion still surfaces in the next session via memory prefetch.

3. Install the pipeline-next-step hook (optional)

git clone https://github.com/veeskelad/gstack-hermes /tmp/gstack-hermes
cp -r /tmp/gstack-hermes/hooks/pipeline-next-step ~/.hermes/hooks/

Make sure hooks_auto_accept: true is set in ~/.hermes/config.yaml, or accept the hook on first use.

Pipeline

new idea ────────────► /yc-office-hours
                           │ (design doc)
                           ▼
scope unclear ───────► /yc-ceo-review
                           │ (4-mode scope review)
                           ▼
need task plan ──────► /writing-plans            (hermes-native)
                           │
              ┌────────────┴─────────────┐
              ▼                          ▼
  individual reviews            batch shortcut:
   /yc-eng-review                    /autoplan
   /yc-design-review                 (CEO+Eng+Design+DevEx in one shot)
   /yc-devex-review
   /sec-cso (if auth/payments)
              │                          │
              └────────────┬─────────────┘
                           ▼
                  /executing-plans            (hermes-native)
                           │
                           ▼
                  /requesting-code-review     (hermes-native)
                           │
                           ▼
                       /ship                  (tests + version bump + deploy)
                           │
                           ▼ (post-release)
                       /yc-retro
                           │
                           ▼
                       /yc-learn

orthogonal (any time):  /document-generate (scaffold docs for module/project)

After each pipeline skill, pipeline-next-step hook writes a suggestion to MEMORY.md — it'll be picked up by hermes memory prefetch on the next session. Pure recommendation; ignore if not relevant.

Why not just use upstream gstack?

garrytan/gstack is already in hermes DEFAULT_TAPS, so you can pull it directly. This fork differs in:

• Smaller scope — 8 skills (vs 35+). Only the review/retro/learn/CSO subset that complements hermes uniquely. We dropped yc-review because hermes has software-development/requesting-code-review v2.0.0 which is more mature (baseline-aware, auto-fix loop).
• No gstack-* binary dependencies — original skills call ~/.claude/skills/gstack/bin/gstack-config and similar that don't exist in hermes; we removed those touchpoints.
• Hermes adaptation note at the top of each skill — explicit guidance for the model about AskUserQuestion → plain chat replacement, path differences.
• yc-learn rewritten to write ### sections to ~/.hermes/memories/lessons.md (auto-archived by hermes's file-size-guard) instead of separate learnings.jsonl.
• Pipeline hook rewritten as native Python async handler on agent:end (was Claude-Code Stop hook).

Skill format

Each skill is a single SKILL.md with YAML frontmatter:

---
name: yc-office-hours
description: "..."
version: 1.0.0
metadata:
  hermes:
    tags: [yc, planning, design-doc, scope]
    related_skills: [yc-ceo-review]
---

Some skills have side-files (sections/, specialists/, ACKNOWLEDGEMENTS.md) — referenced from the body via ${HERMES_SKILL_DIR}.

Most skills contain Plan Mode / Voice / Telemetry / Continuous Checkpoint sections from upstream gstack that reference Claude-Code-only gstack-config binaries. These degrade silently in hermes — the model sees them but the bin calls fail gracefully (and most are wrapped in 2>/dev/null || true). We kept them as-is to stay close to upstream; if context size becomes an issue, run scripts/strip-boilerplate.py (not included here, but the original is in garrytan/gstack).

Updating

hermes skills check    # see which installed skills have upstream changes
hermes skills update   # pull and reinstall

Credits

• Upstream: garrytan/gstack (MIT)
• Hermes: NousResearch/hermes-agent
• This fork: minimal adapter for hermes — see scripts/gstack-to-hermes.py for the exact transformations.

License

MIT — see LICENSE.