Enforces Snow Gloves connector-gate on catalog-managed MCP servers: hold and refuse cards never run, a server must be enabled for the tenant, and high-risk or…

<img src="docs/assets/snowgloves-hero.png" width="100%" alt="Snow Gloves OS — Hand-In-Glove Business Operations Platform" />
Hand-In-Glove Business Operations Platform
<!-- og:image: docs/assets/snowgloves-og.png (upload via repo Settings → Social preview) -->
<!-- readme-gen:start:badges -->
<!-- readme-gen:end:badges -->
<!-- readme-gen:start:tech-stack --> <img src="https://skillicons.dev/icons?i=py,bash,yaml,rust,tauri,github,githubactions&theme=dark" alt="Tech Stack" /> <!-- readme-gen:end:tech-stack -->
<!-- readme-gen:start:social -->
<!-- readme-gen:end:social -->
Run a business as if a senior team were watching it 24/7. Snow Gloves OS is a reusable, tenant-scoped operations platform. It wraps your tools (G-Stack connectors), your knowledge (NVIDIA embeddings), your judgement (an interpretation layer), and your actions (Hermes + Paperclip orchestration), so events don't get dropped, decisions are auditable, and risky actions are gated. It is agent-agnostic: you pick the runtime (Claude Code, Codex, Cursor, OpenCode, Grok, Hermes, OpenClaw, or any other agent), and Snow Gloves renders its skills, MCP servers, and rules into that runtime's own format.
Docs for new users: snow-gloves-wiki — how to use the app and the ecosystem (install, pick a runtime, onboard, modules, a quiet week). Slides, audio, and video live there. The modules dashboard is the catalog, not the human guide.
The Infrastructure cockpit connects the playable Infra Block city to catalog, agent routing, runtime adapters, connector gates, scoped tenant metadata, wing profiles, activity records and acceptance evidence. Run python3 scripts/infra_cockpit.py and npm --prefix apps/infra-block run dev, then open http://127.0.0.1:18760/. The workspace supports source inspection and gate-checked proposal exports. Private instance metadata requires an explicit --data-root; execution and enablement remain outside this read-only interface.
<img src="https://capsule-render.vercel.app/api?type=rect&color=gradient&customColorList=12,13,14&height=2" width="100%" />
<table> <tr> <td width="50%" valign="top">
Seven specialized agents (CEO, CTO, Chief of Staff, Librarian, Interpreter, Dispatcher, Sentinel) work as one operations team, each with IDENTITY · SOUL · TOOLS · SKILLS · HEARTBEAT · MEMORY · EVOLUTION.
</td> <td width="50%" valign="top">
A dedicated Chief of Staff routes four native snowgloves: skills via glob-matched hooks so CEO/CTO stay strategic. inference-sh and Explee are catalog packs, not core. Enable a module and it is callable across agents.
</td> </tr> <tr> <td width="50%" valign="top">
132 catalog cards (skills, MCP servers, plugins, playbooks, connectors) harvested from X, Instagram, Field Theory, founders-kit, and marketingskills — including optional inference-sh and Explee packs. Every card is a pointer with a disposition: add, pointer, hold, or refuse.
</td> <td width="50%" valign="top">
One adapter.yaml per runtime says where skills, MCP config, and rules live and which question tool drives plan mode. The founder picks the runtime; nothing is hard-wired to one agent.
</td> </tr> <tr> <td width="50%" valign="top">
Composio-like scoped connectors (Gmail, Calendar, Drive, Slack, PMS, Accounting, HRIS). Per-capability risk + approval flags. Webhooks signed and verified.
</td> <td width="50%" valign="top">
Tenant-isolated vector index (nv-embedqa-e5-v5, 1024-dim). Plug into NVIDIA NIM with NVIDIA_API_KEY, or run offline with the deterministic stub backend.
</td> </tr> <tr> <td width="50%" valign="top">
Minimal HTTP listener on port 4100 (scripts/hermes.py). POST /publish, GET /events, POST /test/e2e for end-to-end smoke. Every event is append-logged to _audit/hermes-events.jsonl.
</td> <td width="50%" valign="top">
Daily aggregation of hook usage, fallbacks, escalations, and top skills per agent. Auto-appended to each agent's EVOLUTION.md for self-improving loops.
</td> </tr> </table>
<img src="https://capsule-render.vercel.app/api?type=rect&color=gradient&customColorList=12,13,14&height=2" width="100%" />
git clone https://github.com/Sheshiyer/snow-gloves-os.git
cd snow-gloves-os
./scripts/install.sh # deps + print next steps (does not go live)
make doctor # pre-flight
make smoke # bus up, test event, stub embed, Sentinel
make walk # machine-check CoS graph + native inner loops
make graph-upgrade # learning-edge dry-run (constraints; hooks gated)
make onboard-prompt R=claude # print the plan-mode interview for your runtime
make test # pytest
Snow Gloves runs on three office Mac minis by function (marketing, design, coding) over Tailscale. The Coding Mac hosts OmniRoute :20128 and Hermes :4100 for the whole fleet. A brand (tenant) is enabled on a wing (node), and make fleet-render W=<wing> T=<tenant> R=<runtime> writes that brand's modules and the wing's MCPs into the runtime's own files. Start with docs/fleet/README.md; make fleet-doctor tells a mini whether it is wired as its wing says.
New Mac Mini (use the product, do not compile Tauri): docs/MAC-MINI-SETUP.md. Emulate the desktop wizard on a dev Mac with make app-install then make app-dev — apps/onboarding/README.md. The fleet inventory is fleet.yaml in your private data checkout (template: fleet.example.yaml), and each wing's machine profile lives in nodes/ (nodes/<wing>/node.yaml; templates here, the instance copy in the data checkout wins).
This repository is the platform. Instance data (brand tenants, the fleet inventory and wing profiles, receipts, decisions, planning state) lives in a separate private checkout, snow-gloves-ops. Point the scripts at it:
export SNOWGLOVES_DATA=/path/to/snow-gloves-ops # tenants/, fleet.yaml, nodes/, _audit/
Unset, everything resolves to this checkout and its public fixtures (_demo, acme, tryambakam-noesis), which the local validation commands use; make smoke and make walk always use the fixtures. The resolver is scripts/lib/paths.py; scripts/onboard.py, node_profile.py, graph_upgrade.py and upgrade.py also take --data DIR. Never commit tenants, fleet.yaml, receipts or planning state here.
Planned fresh-node CLI onboarding: layered setup plan and Claude start prompt. This covers an always-on Mac mini workspace, developer tooling, Temperance/OmniRoute, secrets, and selected brand capabilities. It is a planning candidate; the node bootstrap CLI is not implemented yet.
| Target | Purpose |
|---|---|
make install | Bootstrap (paperclipai + python deps) |
make doctor | Pre-flight diagnostic |
make onboard | Interactive tenant + sources prompt (legacy flow) |
make onboard-prompt R=<runtime> | Print the plan-mode interview prompt for a runtime |
make catalog / make catalog-check | Rebuild catalog/modules.json + registry.yaml / fail if stale |
make legacy-check | Fail if pre-Hermes bus names remain (skips .bak-*) |
make site | Build the static dashboard site (npm run build:site) |
make hermes | Foreground Hermes listener on :4100 |
make smoke | Hermes → e2e → bridge dry-run → embed stub → sentinel |
make walk | Route fixtures through CoS; native produce/check/correct; receipt JSON |
make tui | Onboarding TUI (agent vs manual; --headless for agents) |
make graph-upgrade [T=<slug>] [WRITE=1] | Learning edge dry-run (or apply constraints; hook diffs need approval) |
make embed T=<tenant> | Run NVIDIA embed worker (QUIET=1 for cron) |
make sentinel | Daily drift sweep |
make release-dry V=x.y.z | Preview a platform version bump |
make release V=x.y.z | Bump every version file, rebuild catalog, commit, tag (no push) |
make release-push V=x.y.z | Push HEAD + tag; signing and publication require explicit manual steps |
make release-check | Verify all version files agree with VERSION |
make upgrade [T=<slug>] [WRITE=1] | Dry-run (or apply) tenant platform migrations |
make kill-hermes | Free port 4100 |
make fleet-doctor [W=<wing>] / fleet-render W= T= R= / fleet-enable W= T= / fleet-connect W= / fleet-kit-export / fleet-remote-access W= | Three-mini fleet: wing diagnostics, brand x wing x runtime render, enable, Tailscale shell, gateway kit, remote-access plan (docs/fleet/README.md) |
make app-install / app-dev / app-build | Tauri onboarding app: npm install, tauri dev (GUI), release bundle |
<img src="https://capsule-render.vercel.app/api?type=rect&color=gradient&customColorList=12,13,14&height=2" width="100%" />
<!-- readme-gen:start:architecture -->
graph TD
subgraph Runtime layer
Ext[3rd-party app] -- signed webhook --> GS[G-Stack /webhook]
GS -- normalized envelope --> Her[Hermes :4100]
Her --> CoS[Chief of Staff<br/>skill-hooks.yaml]
CoS -->|strategy| CEO
CoS -->|technical| CTO
CoS --> Lib[Librarian]
CoS --> Int[Interpreter]
CoS --> Dis[Dispatcher]
CoS --> Sen[Sentinel]
Dis --> PB[Paperclip Bridge<br/>:3100/api/tasks]
Lib --> EW[NVIDIA Embed Worker]
EW --> VI[(tenant vector-index.jsonl)]
Her -- audit --> Log[(_audit/hermes-events.jsonl)]
Log --> SS[Sentinel Sweep<br/>daily]
SS --> Evo[(agents/*/EVOLUTION.md)]
end
subgraph Catalog and adapter layer
Cards[catalog/cards/*.md] --> BC[build_catalog.py]
Agents[agents/*/MANIFEST.yaml] --> BC
Ad[adapters/*/adapter.yaml] --> BC
BC --> MJ[(catalog/modules.json)]
MJ --> Prompt[onboard.py --prompt runtime]
Prompt --> Harvest[snowgloves-harvest.md]
Harvest --> Apply[onboard.py --apply-harvest]
Apply --> En[(tenants/slug/enabled.yaml<br/>runtime.yaml)]
En --> Render[onboard.py --render-adapter]
Ad --> Render
Render --> RT[Runtime files:<br/>skills, MCP config, rules block]
MJ --> Dash[Tauri dashboard + Pages site]
En --> Gate[connector-gate]
end
| Engine | Owns | Files |
|---|---|---|
| Connector | G-Stack fabric, scopes, webhooks | connectors/g-stack/, skills/connector-gate/ |
| Knowledge | Ingest, chunk, embed, retrieve | scripts/ingest.py, scripts/embed_worker.py |
| Interpretation | Skill routing, escalation | workflows/skill-hooks.yaml, agents/chief-of-staff/ |
| Orchestration | Event bus + Paperclip bridge | scripts/hermes.py, scripts/paperclip_bridge.py |
More detail: docs/architecture-overview.md.
<!-- readme-gen:end:architecture -->
catalog/cards/<id>.md holds one card per third-party skill, MCP server, plugin, connector, or playbook. A card is a pointer, never a copy of upstream code. scripts/build_catalog.py compiles the cards, the agent manifests, the adapters, and the G-Stack connectors into catalog/registry.yaml and catalog/modules.json (schema snowgloves.modules.v1). Run make catalog-check locally to reject stale generated output.
| Disposition | Count | Meaning |
|---|---|---|
add | 61 | Tenant can enable it |
pointer | 45 | Tenant can enable it as a reference; the host already provides it |
hold | 14 | Visible, not enableable until the founder picks it and it passes review |
refuse | 12 | Visible, never enableable |
By category: skills 79, playbook 36, plugin 13, mcp 3, connector 1. Optional packs inference-sh-agent-skills and explee-skills are add cards, not core hooks. Card schema: docs/catalog.md. Harvest notes: docs/research/2026-09-29-ecosystem-review.md.
Snow Gloves does not assume an agent. Each runtime has adapters/<runtime>/adapter.yaml (schema snowgloves.adapter.v1): skill, MCP, and rules paths; file formats; the runtime's question tool; and how to enter plan mode. Shipped adapters: hermes, claude, codex, cursor, opencode, grok, openclaw, muse, generic. Fields nobody has confirmed against the real runtime are marked verify: true and reported on every render. The Muse adapter is a guess throughout.
Onboarding is a plan-mode interview run inside the project the founder has been building:
make onboard-prompt R=cursor # 1. interview → snowgloves-harvest.md
python3 scripts/onboard.py --apply-harvest snowgloves-harvest.md --tenant acme # 2. apply
python3 scripts/onboard.py --list --category skills # 3. adjust
python3 scripts/onboard.py --enable ms-copywriting --tenant acme
python3 scripts/onboard.py --render-adapter cursor --tenant acme # 4. render (dry run)
python3 scripts/onboard.py --render-adapter cursor --tenant acme --write
The agent asks one decision at a time, takes options only from modules.json, and writes FILL: instead of guessing. See docs/onboarding.md and docs/adapters.md.
The native Tauri v2 app in apps/onboarding/ gains a Modules dashboard and a guided onboarding flow. Both read the same catalog/modules.json plus each tenant's enabled.yaml. The same data is published as a static site, built with make site (npm run build:site) previously published by GitHub Pages. Automatic publishing is retired; this URL retains its last published content until a separate local publisher is selected:
https://sheshiyer.github.io/snow-gloves-os/
make app-install # one-time: npm install
make app-dev # launch in dev mode (hot-reload frontend)
make app-build # produce a signable release bundle
make site # build the static Pages site locally
See its README for the IPC surface and capability list.
The root VERSION file (currently 0.2.1) is the platform version. The app's package.json, tauri.conf.json, Cargo.toml, both lockfiles, and distribution.yaml are kept in lockstep with it, and catalog/modules.json embeds it.
make release-dry V=0.2.1 # show every file that would change
make release V=0.2.1 # bump, rebuild catalog, finalize CHANGELOG, commit, tag v0.2.1
make release-push V=0.2.1 # push only; no automatic signing or publication
make upgrade # dry-run tenant migrations to VERSION
make upgrade T=acme WRITE=1 # apply, rebuild catalog, re-render acme's adapters
The retired hosted release built signed Tauri installers and attached source, catalog, adapter, and checksum assets. Those publication jobs are no longer active. make app-build remains the local app build entry point; signing, cross-platform builds, updater latest.json, and publication require a separately reviewed local procedure. v0.2.0 is a reinstall for 0.1.x (new bundle id). See docs/RELEASING.md and docs/UPGRADING.md.
<!-- readme-gen:start:tree -->
📦 snow-gloves-os
├── 📄 VERSION # platform version (0.2.1)
├── 📄 distribution.yaml # what the platform owns vs what tenants own
├── 📂 .local-jobs/ # explicit local checks; schedules disabled
├── 📂 .specify/ # Spec-Kit templates + workflows
├── 📂 .planning/ # GSD state: local symlink into the private data checkout (gitignored)
├── 📂 agents/ # 7 agents, each with 8+ md files + MANIFEST
├── 📂 adapters/ # one adapter.yaml per runtime (9)
├── 📂 catalog/ # SCHEMA.md · cards/ (132) · registry.yaml · modules.json (generated)
├── 📂 prompts/ # onboard-interview.md (plan-mode interview template)
├── 📂 migrations/ # v0_1_to_v0_2.py (bus rename to Hermes)
├── 📂 apps/onboarding/ # Tauri v2 app: onboarding + Modules dashboard, Pages site
├── 📂 connectors/g-stack/ # Capability registry + auth + webhooks
├── 📂 scripts/ # hermes · ingest · embed · bridge · sentinel · build_catalog · onboard · release · upgrade
├── 📂 skills/ # registry.yaml (4 native skills) · connector-gate · sg-onboard · tn-seed · gtm-brief-synthesis
├── 📂 workflows/ # skill-hooks.yaml — Chief of Staff graph
├── 📂 specs/ # Spec-Kit features
├── 📂 docs/ # architecture, catalog, adapters, onboarding, releasing, upgrading, research, runbooks
├── 📂 tenants/ # Fixture tenants only; real tenants live in $SNOWGLOVES_DATA/tenants
├── 📂 nodes/ # Wing profile templates (marketing, design, coding)
├── 📄 fleet.example.yaml # Fleet inventory template (the real fleet.yaml is private)
├── 📂 tests/ # pytest suite
├── 📂 _audit/ # Append-only event log
├── 📄 config/snowgloves.yaml # Runtime config (ports, embeddings, scopes)
└── 📄 Makefile # make smoke runs the whole loop
<!-- readme-gen:end:tree -->
<img src="https://capsule-render.vercel.app/api?type=rect&color=gradient&customColorList=12,13,14&height=2" width="100%" />
<!-- readme-gen:start:health -->
| Category | Status | Score |
|---|---|---|
| Spec coverage | ████████████████████ | 100% |
| Agents wired | ████████████████████ | 100% |
| End-to-end smoke | ████████████████████ | 100% |
Local validation (pytest, version + catalog, make walk, smoke; execution is opt-in) | ████████████████████ | 100% |
Catalog (132 cards, --check in CI) | ████████████████████ | 100% |
Runtime adapters (9 shipped; several fields still verify: true) | ██████████░░░░░░░░░░ | 50% |
| Release + upgrade tooling (v0.2.1 stable) | ████████████████░░░░ | 80% |
| Real NVIDIA NIM integration | ████░░░░░░░░░░░░░░░░ | 20% |
| Live Paperclip wiring | ████░░░░░░░░░░░░░░░░ | 20% |
| Production hardening | ████░░░░░░░░░░░░░░░░ | 20% |
Overall: 70% — Stable v0.2.1 platform. Before a live pilot: confirm unverified adapter fields, and wire real connectors, NIM, and Paperclip.
<!-- readme-gen:end:health -->
com.tryambakam.snowgloves.onboarding (was com.thoughtseed.snowgloves.onboarding). v0.2.0 is a new app: existing 0.1.x users must reinstall from the DMG/MSI; OTA will not find 0.2.0.distribution.yaml lacks hermes_requires. Nothing declares which Hermes version or port a distribution needs yet. Do not set it in this release./plan, are unconfirmed. See docs/adapters.md.This repo uses GitHub Spec-Kit. Available slash commands (in Copilot/Codex chat):
/speckit.constitution /speckit.specify /speckit.clarify
/speckit.plan /speckit.tasks /speckit.analyze
/speckit.checklist /speckit.implement /speckit.taskstoissues
Active feature: specs/001-hand-in-glove-platform.
7 principles govern every change:
See .specify/memory/constitution.md.
<!-- readme-gen:start:footer -->
<img src="https://capsule-render.vercel.app/api?type=waving&color=gradient&customColorList=12,13,14&height=110§ion=footer" width="100%" />
Built with ❤️ by Mage Narayan · Tryambakam Noesis
<!-- readme-gen:end:footer -->
hooks/register.ts 123 lines1// sg-connector-gate: skills/connector-gate as code. Before Claude calls a tool on an MCP server
2// that the Snow Gloves catalog manages, it checks the card's disposition, the tenant's
3// enabled.yaml, and the approvals queue, and refuses the call when they say no.
4//
5// It only ever passes a call on with next(e) or refuses it with { deny }. It never answers a
6// call with a result, so the settings PreToolUse hooks (erp-read-only.py) still run after it.
7import type { EngineInterface, PluginOptions, Register } from 'claude-code'
8
9import type { GateTable } from '../types'
10import { approvalReason, asGateTable, decide, platformRootFrom, splitTool, statusText } from './lib'
11
12const RELOAD_MS = 60_000
13const TABLE_TIMEOUT_MS = 15_000
14
15// The gate decides from this copy, so a call never waits on a process start.
16// session.start loads it again after every reload of the module.
17let table: GateTable | null = null
18let loadError: string | null = null
19// register sets these from the plugin's options
20let pluginOptions: PluginOptions = {}
21let mode: 'enforce' | 'observe' = 'enforce'
22
23async function settingsOf($: EngineInterface, options: PluginOptions) {
24 const root = String(options.platformRoot || '') || platformRootFrom($.plugin.root)
25 const tenant = String(options.tenant || '') || (await $.env.get('SNOWGLOVES_TENANT')) || ''
26 const dataRoot = String(options.dataRoot || '') || (await $.env.get('SNOWGLOVES_DATA')) || ''
27 const venv = root + '/.venv/bin/python'
28 const python = (await $.fs.exists(venv)) ? venv : 'python3'
29 const mode = options.gateMode === 'observe' ? 'observe' : 'enforce'
30 const ttlHours = Number(options.ttlHours) > 0 ? Number(options.ttlHours) : 24
31 return { root, tenant, dataRoot, python, mode, ttlHours }
32}
33
34function bridge(cfg: { root: string; python: string; dataRoot: string }, args: string[]): string[] {
35 const argv = [cfg.python, cfg.root + '/scripts/sg_mods.py', ...args]
36 if (cfg.dataRoot) argv.push('--data-root', cfg.dataRoot)
37 return argv
38}
39
40// Never rejects: it runs from a timer. A failed reload keeps the last good table.
41async function load($: EngineInterface, options: PluginOptions) {
42 try {
43 const cfg = await settingsOf($, options)
44 if (!cfg.tenant) {
45 table = null
46 loadError = 'no tenant: set SNOWGLOVES_TENANT or the tenant option'
47 return
48 }
49 const argv = bridge(cfg, ['gate-table', '--tenant', cfg.tenant, '--ttl-hours', String(cfg.ttlHours)])
50 const run = await $.process.run(argv, { cwd: cfg.root, timeoutMs: TABLE_TIMEOUT_MS })
51 const next = asGateTable(run.stdout)
52 if (next) {
53 table = next
54 loadError = null
55 } else {
56 loadError = run.stdout.trim() || run.stderr.trim().split('\n').pop() || 'exit ' + run.exitCode
57 }
58 } catch (err) {
59 loadError = String(err instanceof Error ? err.message : err)
60 }
61}
62
63async function requestApproval($: EngineInterface, options: PluginOptions, connector: string, capability: string, tool: string) {
64 const cfg = await settingsOf($, options)
65 const argv = bridge(cfg, [
66 'request-approval', '--tenant', cfg.tenant, '--connector', connector, '--capability', capability, '--tool', tool,
67 ])
68 const run = await $.process.run(argv, { cwd: cfg.root, timeoutMs: TABLE_TIMEOUT_MS })
69 const out = JSON.parse(run.stdout)
70 if (run.exitCode !== 0 || !out.ticket || !out.ticket.id) throw new Error(out.error || 'request-approval failed')
71 return String(out.ticket.id)
72}
73
74async function gate($: EngineInterface, e, next) {
75 const decision = decide(table, e.tool)
76 if (decision.kind === 'pass') return next(e)
77 if (mode === 'observe') {
78 $.ui.log('would refuse ' + e.tool + (decision.kind === 'deny' ? ': ' + decision.reason : ': needs approval'))
79 return next(e)
80 }
81 if (decision.kind === 'deny') return { deny: decision.reason }
82 const ticket = await requestApproval($, pluginOptions, decision.connector, decision.capability, e.tool)
83 $.ui.toast('Approval needed: ' + decision.connector + '.' + decision.capability)
84 return { deny: approvalReason(e.tool, table ? table.tenant : '', ticket) }
85}
86
87export const register: Register = (on, options) => {
88 pluginOptions = options
89 mode = options.gateMode === 'observe' ? 'observe' : 'enforce'
90
91 on('session.start', async ($, e, next) => {
92 // Waited for, so the table is in place before Claude's first tool call
93 await load($, options)
94 if (!table) $.ui.log('observing only: ' + loadError)
95 $.clock.every(RELOAD_MS, () => load($, options))
96 await $.command.register({
97 name: 'sg-gate',
98 description: 'Snow Gloves connector gate: which managed MCP servers may run',
99 argumentHint: '[status|reload]',
100 immediate: true,
101 })
102 return next(e)
103 })
104
105 // Fails closed for managed servers only: a failure on someone else's server lets it go on
106 on('tool.call', { tool: /^mcp__/ }, gate).catch(($, e, next) => {
107 if (next.called) return next(e)
108 const decision = decide(table, e.tool)
109 if (decision.kind === 'pass' || mode === 'observe') return next(e)
110 const parts = splitTool(e.tool)
111 return {
112 deny:
113 'sg-connector-gate could not check ' + (parts ? parts.server : e.tool) +
114 ' (' + next.error.kind + '), so the call did not run.',
115 }
116 })
117
118 on('command.run', { command: 'sg-gate' }, async ($, e) => {
119 if (e.args.trim() === 'reload') await load($, options)
120 return { text: statusText(table, mode, loadError) }
121 })
122}
123hooks/lib.ts 97 lines1// Pure rules for sg-connector-gate, mirroring skills/connector-gate/SKILL.md. No mods API calls here.
2import type { GateDecision, GateTable } from '../types'
3
4export const GATE_SCHEMA = 'snowgloves.mods-gate.v1'
5
6export function platformRootFrom(pluginRoot: string): string {
7 const parts = pluginRoot.replace(/\/+$/, '').split('/')
8 return parts.slice(0, -2).join('/') || '/'
9}
10
11export function asGateTable(text: string): GateTable | null {
12 try {
13 const value = JSON.parse(text)
14 return value && value.schema === GATE_SCHEMA && typeof value.servers === 'object' ? value : null
15 } catch {
16 return null
17 }
18}
19
20// mcp__<server>__<tool>; the adapter keys .mcp.json by card id, so <server> is the card id
21export function splitTool(tool: string): { server: string; name: string } | null {
22 if (!tool.startsWith('mcp__')) return null
23 const rest = tool.slice('mcp__'.length)
24 const cut = rest.indexOf('__')
25 if (cut <= 0) return null
26 return { server: rest.slice(0, cut), name: rest.slice(cut + 2) }
27}
28
29export function isManaged(table: GateTable | null, tool: string): boolean {
30 const parts = splitTool(tool)
31 return table !== null && parts !== null && parts.server in table.servers
32}
33
34export function decide(table: GateTable | null, tool: string): GateDecision {
35 const parts = splitTool(tool)
36 if (!table || !parts) return { kind: 'pass' }
37 const server = table.servers[parts.server]
38 // Servers the catalog doesn't list are not Snow Gloves'. They go on untouched, so the
39 // settings PreToolUse hooks that guard them (erp-read-only.py) still run.
40 if (!server) return { kind: 'pass' }
41 const id = parts.server
42 if (server.disposition === 'hold') {
43 return {
44 kind: 'deny',
45 reason: `${id} is on hold in the Snow Gloves catalog: it waits on a founder pick and a review, so it can't be called yet.`,
46 }
47 }
48 if (server.disposition === 'refuse') {
49 return { kind: 'deny', reason: `${id} is refused in the Snow Gloves catalog, so Snow Gloves will not run it.` }
50 }
51 if (!server.enabled) {
52 return {
53 kind: 'deny',
54 reason:
55 `${id} is not enabled for tenant ${table.tenant}. ` +
56 `Ask the user whether to enable it with: python3 scripts/onboard.py --tenant ${table.tenant} --enable ${id}`,
57 }
58 }
59 if (server.needs_approval) {
60 const granted = table.grants.some((g) => g.connector === id && g.capability === parts.name)
61 if (!granted) return { kind: 'approve', connector: id, capability: parts.name }
62 }
63 return { kind: 'pass' }
64}
65
66export function approvalReason(tool: string, tenant: string, ticket: string): string {
67 return (
68 `${tool} needs founder approval for tenant ${tenant}. Ticket ${ticket} is queued. ` +
69 'Ask the user to approve it in /approvals, then retry the call.'
70 )
71}
72
73export function statusText(table: GateTable | null, mode: string, error: string | null): string {
74 // Claude Code prints the plugin's name in front of a command's text
75 if (!table) return 'observing only (' + (error || 'no tenant') + '). Every MCP call goes on.'
76 const lines = [`${mode} for tenant ${table.tenant}`]
77 for (const id of table.managed) {
78 const s = table.servers[id]
79 const state =
80 s.disposition === 'hold' || s.disposition === 'refuse'
81 ? s.disposition
82 : !s.enabled
83 ? 'not enabled'
84 : s.needs_approval
85 ? 'enabled, approval per tool'
86 : 'enabled'
87 lines.push(` ${id}: ${state}`)
88 }
89 if (table.grants.length > 0) {
90 lines.push('approved: ' + table.grants.map((g) => g.connector + '.' + g.capability).join(', '))
91 }
92 if (table.pending.length > 0) {
93 lines.push('waiting: ' + table.pending.map((p) => p.connector + '.' + p.capability + ' ' + p.id).join(', '))
94 }
95 return lines.join('\n')
96}
97types/index.d.ts 27 lines1// What scripts/sg_mods.py gate-table prints (schema snowgloves.mods-gate.v1).
2export type GateServer = {
3 category: string
4 disposition: 'add' | 'pointer' | 'hold' | 'refuse'
5 risk: 'low' | 'medium' | 'high'
6 approval: 'yes' | 'no'
7 enabled: boolean
8 needs_approval: boolean
9}
10
11export type GateTable = {
12 schema: 'snowgloves.mods-gate.v1'
13 tenant: string
14 data_root: string
15 ttl_hours: number
16 managed: string[]
17 servers: Record<string, GateServer>
18 grants: { connector: string; capability: string; id: string; decided_at: number }[]
19 pending: { connector: string; capability: string; id: string }[]
20}
21
22// pass: let the call go on. deny: refuse it with a reason Claude reads. approve: queue a ticket, then refuse.
23export type GateDecision =
24 | { kind: 'pass' }
25 | { kind: 'deny'; reason: string }
26 | { kind: 'approve'; connector: string; capability: string }
27