setup

分类: 工具与效率 | 上传者: SponsioLabs | 下载: 0 | 版本: v1.0（最新）
Configure the sponsio-claude-code runtime end-to-end — bootstrap the per-plugin contract library tree at ~/.sponsio/plugins/, install bundled starter libraries for popular MCP servers (github, filesystem, playwright), generate fresh starter libraries for plugins / MCP servers that don't ship one (via `sponsio plugin scan`), tune the shipped rules to the user's actual environment (workspace path, expected call volume, dev/CI/prod 配置文件），并通过冒烟测试进行验证。当用户说出“设置 sponsio-claude-code”、“配置 sponsio-claude-code”、“安装插件”、“首次设置 sponsio”、“在此 Claude 代码会话中连接 sponsio”、“为我的 MCP 工具添加 Sponsio 护栏”、“针对我的环境调整插件”、“插件太严格/太松”、“校准 sponsio 规则”、“扫描此插件”时使用“为 X 生成 sponsio 规则”、“为我的插件创建合约库”、“我刚刚安装了 Y 插件/MCP 服务器，sponsio 应该阻止什么”，或者询问如何使 sponsio-claude-code 实际上阻止某些内容。
更新日志: Source: GitHub https://github.com/SponsioLabs/Sponsio
目录结构

当前层级: tree/main/plugins/sponsio-claude-code/skills/setup/
📄 SKILL.md 15.8 KB
SKILL.md

---
name: setup
description: Configure the sponsio-claude-code runtime end-to-end — bootstrap the per-plugin contract library tree at ~/.sponsio/plugins/, install bundled starter libraries for popular MCP servers (github, filesystem, playwright), generate fresh starter libraries for plugins / MCP servers that don't ship one (via `sponsio plugin scan`), tune the shipped rules to the user's actual environment (workspace path, expected call volume, dev/CI/prod profile), and verify with a smoke test. Use when the user says any of "set up sponsio-claude-code", "configure sponsio-claude-code", "install the plugin", "first-time setup of sponsio", "wire up sponsio in this Claude Code session", "add Sponsio guardrails for my MCP tools", "tune the plugin for my environment", "the plugin is too strict / too loose", "calibrate sponsio rules", "scan this plugin", "generate sponsio rules for X", "create a contract library for my plugin", "I just installed Y plugin / MCP server, what should sponsio block", or asks how to make sponsio-claude-code actually block things.
---

# sponsio-claude-code — configure the host shield

You are walking the user through configuring the **sponsio-claude-code**
plugin so it actually wraps tool calls in this Claude Code session.
Without these steps, the plugin loads but every per-plugin library
is empty and every tool call passes through unguarded.

The flow has two halves: **setup** (steps 1–2, 4–6) covers the common
case where the user runs popular MCP servers we ship starters for;
**scan** (step 3) covers plugins / MCP servers without a bundled
starter and lives inline because most sessions touch both halves.

## Prerequisites (check silently first)

Run:

```bash
sponsio --version
```

* Not found → tell the user to install: `pip install sponsio` (or
  `pip install -e ".[all]"` from a local clone). Stop here until they
  confirm.
* Present → continue.

If the user owns the agent code and wants to wire Sponsio into their
own framework integration (LangGraph, OpenAI Agents, Claude Agent
SDK, …) **instead of** running it as a Claude Code host plugin,
delegate to the `sponsio` skill (`sponsio onboard .`). This skill is
only for the host-plugin case where the user is gating tool calls
inside their Claude Code session.

## Step 1 — bootstrap the library root

Run:

```bash
sponsio plugin init
```

This creates `~/.sponsio/plugins/_host/sponsio.yaml` (covers Claude
Code's first-party tools — Bash, Edit, Write, …) and runs a built-in
allow + block smoke test. Show the user the output verbatim; if the
smoke test fails, **stop** and surface the error — the install is
broken and we shouldn't pretend otherwise.

If the file already exists (the user re-ran setup), the command
prints `…already exists. Re-run with --force to overwrite.` That's
informational, not an error.

## Step 2 — enumerate the user's MCP servers

You need to know which MCP servers / plugins the user actually has
installed before deciding which libraries to ship. Run
`claude mcp list`, or have the user open `~/.claude/settings.json`
and read `mcpServers`. Don't guess. If the user doesn't know, that's
the answer — they have none beyond the built-ins, and you can skip
straight to step 4.

For each registered server, decide which branch of step 3 applies:

```bash
sponsio plugin install --list
#   _host (auto-installed by `plugin init`)
#   filesystem
#   github
#   playwright
```

* **Bundled** (`github`, `filesystem`, `playwright`) → step 3a.
* **Not bundled** (anything else) → step 3b.

A single session often hits both branches.

## Step 3a — bundled MCP servers (use `sponsio plugin install`)

Show the user what each bundled starter covers so they can pick:

* `_host` — already installed by step 1.
* `github` — github-mcp-server. Hard-deny on `delete_repository`,
  blocks deletes of `main` / `master` / production branches, blocks
  writes to `.env`, `.github/workflows/*.yml`, `CODEOWNERS`. Caps
  issue / PR / merge rates.
* `filesystem` — @modelcontextprotocol/server-filesystem. Blocks
  read/write/edit/move on `.env`, `.ssh/`, `.aws/credentials`,
  `/etc/`, browser cookie databases, etc.
* `playwright` — microsoft/playwright-mcp. Blocks navigation to
  internal hosts, `browser_evaluate` exfil patterns
  (`document.cookie`, `sendBeacon`, `fetch` to remote hosts),
  credit-card-shape typing.

Then install **only what the user has** — don't push the whole
bundle:

```bash
sponsio plugin install github filesystem
# or, if they confirm they want everything:
sponsio plugin install --all
```

## Step 3b — unbundled plugins (use `sponsio plugin scan`)

For every MCP server / plugin that isn't in the bundled set, generate
a starter library. Don't try to hand-author one — the heuristics in
`sponsio plugin scan` cover the common patterns (irreversible ops,
bash exec, sql, rate limits, loops) better than ad-hoc rules.

### 3b.1 — discover the plugin's tool inventory

The scanner can't currently introspect an MCP server's tool list at
rest, so collect names another way:

* If the plugin defines tools as **`skills/*/SKILL.md`** (slash
  commands), tool names are `<plugin-id>:<skill-name>`. Read
  `<dir>/skills/` to enumerate.
* If the plugin runs an **MCP server**, real tool names are
  `mcp__<server>__<tool>`. Run `claude mcp list` to see registered
  servers, or read the plugin's `.mcp.json` for server names. The
  actual tool list comes from running the MCP server — not parsable
  statically. Ask the user: "What tools does this MCP server expose?
  You can paste a list, or open `claude` and let me see." If they
  don't know, suggest they open a session, list any
  `mcp__<server>__*` tool names that appear in `/help`, and bring
  those back.
* If the user just says "everything", run scan with no `--tools` and
  explain the result is baseline-only (just `core/runaway`).

### 3b.2 — dry-run the scan

Always start with the dry-run. **Never** `--apply` until the user has
seen the output.

```bash
sponsio plugin scan <plugin-dir> --tools tool_a,tool_b,tool_c
```

The output is one rendered yaml per **routed group** (the same
partitioning that `sponsio.guard_stdin.derive_plugin_id` does at
runtime — `Bash` → `_host`, `mcp__github__X` → `github`, etc.). Show
the user every group.

### 3b.3 — review every contract

For each contract in each group, state:

* **What it blocks** — translate the regex / rate cap into plain
  English. Example: "blocks `delete_repository` outright; for one-off
  deletes the user has to add an override."
* **Why** — point at the `heuristic:` evidence: `starter_irreversible`
  / `starter_bash` / `starter_sql` / `starter_rate_limit` /
  `starter_loop`. Each comes from the rules in
  `sponsio/discovery/starter_pack.py:_per_tool_rules`.
* **What it doesn't catch** — heuristic rules are conservative
  baselines, not airtight. Tell the user explicitly when a rule feels
  generic so they don't assume it's covering everything.

If a rule is wrong (e.g. the heuristic flagged `list_users` as
external-send because the name contains `send`), tell the user to
either:

1. Drop that line from the rendered yaml before running with
   `--apply`.
2. Apply, then add an override:
   ```yaml
   overrides:
     - match: { desc: "list_users at most 10 times per session" }
       disabled: true
   ```

### 3b.4 — apply

Once the user is happy with the dry-run:

```bash
sponsio plugin scan <plugin-dir> --tools ... --apply
```

Show every path the apply writes to:

```
✓ wrote ~/.sponsio/plugins/_host/sponsio.yaml
✓ wrote ~/.sponsio/plugins/github/sponsio.yaml
```

If a target file already exists, scan refuses without `--force`.
**Don't pass `--force` for the user automatically** — surface the
warning and let them decide.

## Step 4 — tune the shipped rules to the user's actual environment

The shipped libraries are templates with conservative defaults.
Without this step, half the rules are either too strict (blocking
legitimate work) or too loose (giving the user false confidence).
Walk through the four parameter classes below and write any
agreed-upon adjustments as `overrides:` blocks into the relevant
library. **Don't skip this step on the assumption defaults are
fine** — defaults are a starting point, not a fit.

### 4.1 — workspace path

Several rules in `_host` and `filesystem` use `<workspace>/` as the
path-allowlist root. Until that's substituted with the user's actual
project root, those rules either match nothing (silent no-op) or
match too broadly. Ask:

> "Where's your main working directory for this session?
>  (`pwd` or your project root.)"

Then write to the relevant `agents:<id>` block:

```yaml
agents:
  _host:
    workspace: /Users/<them>/projects/<repo>
```

If multiple plugins want different workspaces, add `workspace:` under
each agent block separately.

### 4.2 — expected call volume

The shipped `rate_limit` defaults are tuned for a single interactive
session (50 Bash calls / 200k tokens / 5 PRs). Operators running CI
scripts, batch jobs, or recurring agents often hit these
legitimately. Ask:

> "How chatty is this agent? Is this an interactive session, a CI
>  run, a long-running operator, or a one-shot script?"

Use the answer to override:

| Scenario | Adjustment |
|---|---|
| Interactive (default) | leave as-is |
| Heavy CI / batch | bump exec/Bash rate to 200, token budget to 500k |
| Read-only research | tighten Bash rate to 10, drop exec rate cap entirely |
| Long-running (>1h) | drop session-bounded counts, switch to time-window pacing (note: needs daemon mode for time-window — surface as a future-work caveat) |

Example override:

```yaml
agents:
  _host:
    overrides:
      - match: { desc: "Cap exec calls per session" }
        args: [Bash, 200]
```

### 4.3 — environment profile

Different blast radius means different default tightness. Ask:

> "What kind of environment is this — local dev, staging, production,
>  or a customer-data context?"

Apply this matrix:

| Profile | What changes |
|---|---|
| Local dev | leave defaults |
| Staging | enable `audit_after` on destructive tools (logs every action); keep delete rules permissive |
| Production | move `delete_*` from `rate_limit 0` to **assumption-gated** — require an explicit `confirm_reconfirmed` tool emission (see existing pattern in `capability/shell` §4) |
| Regulated / PII | tighten sto rules — `core/universal`'s β from 0.95 → 0.99; force `semantic_pii_free` even on agents that don't currently include it |

### 4.4 — known-false-positive overrides

Walk the user through each shipped rule that's commonly tripped by
legitimate workflows. For every "Yes, that's a problem for me"
answer, write a targeted `overrides:` entry. Common cases:

| Rule | When it false-positives | Override |
|---|---|---|
| `_host` "Each exec call needs its own confirm_reconfirmed" | Any agent that doesn't emit `confirm_reconfirmed` markers | `disabled: true` (until the integration ships markers) |
| `github` "delete_repository is blocked outright" | Cleanup bots, automated repo lifecycle | `disabled: true` + add a custom rule with a tighter pattern (only allow deletion of repos matching `^test-`) |
| `filesystem` "read_file must not exfiltrate dotenv" | dotenv rotators, secret-rotation agents | `disabled: true` only for `read_file` (keep `write_file` denied) |
| `playwright` "browser_navigate must not target internal hosts" | Anyone testing their own internal app | replace with a narrower allowlist of the user's actual internal hostnames |

### 4.5 — write the overrides + verify

After the walkthrough, write all agreed-upon overrides into the
relevant `~/.sponsio/plugins/<id>/sponsio.yaml` files. **Do not edit
the shipped library inline** — always add overrides under
`overrides:` so future `sponsio plugin install --force` doesn't
clobber the user's customisations.

Run `sponsio validate --config ~/.sponsio/plugins/<id>/sponsio.yaml`
on every file you touched. Any error means you wrote a malformed
override; fix before continuing.

### 4.6 — observe-mode dial for tuning runs

If the user has *no idea* what numbers to use, suggest:

```bash
export SPONSIO_GUARD_MODE=observe
```

Run their normal workflow for a day, then come back and:

```bash
sponsio report --since 24h
```

Surface the would-have-blocked rules. For every cluster of
legitimate-looking violations, tighten the matching `overrides:`.
This is the data-driven counterpart to the questionnaire above — the
questionnaire is the cold-start prior, the report is the posterior.

## Step 5 — verify the deny path actually works

Run a synthetic event through the hook command:

```bash
echo '{"hook_event_name":"PreToolUse","tool_name":"Bash","tool_input":{"command":"rm -rf /"}}' \
  | sponsio plugin guard --stdin
```

Expect: a JSON deny payload on stdout. If it's empty:

1. The library at `~/.sponsio/plugins/_host/sponsio.yaml` is wrong or
   wasn't written. Re-run `sponsio plugin init --force`.
2. The Sponsio CLI version is older than the libraries' rule shapes.
   Update with `pip install -U sponsio`.

For an unbundled plugin you scanned in step 3b, do the same with one
of its tool names plus a pattern you expect to block.

## Step 6 — tell the user to reload the plugin

If the user is in a live Claude Code session, the plugin needs to
re-pick up the manifest. Tell them to run `/reload-plugins` and
confirm it shows `sponsio-claude-code`.

## Common configuration adjustments

**Operator wants observe mode (log, don't block) — pilot rollout:**

```bash
export SPONSIO_GUARD_MODE=observe
```

This dial only affects this plugin; other Sponsio integrations in the
same shell still respect `SPONSIO_MODE` independently.

**Operator wants per-plugin overrides instead of editing the library:**

Add an `overrides:` block under the agent in
`~/.sponsio/plugins/<plugin>/sponsio.yaml`:

```yaml
agents:
  github:
    contracts: [...shipped...]
    overrides:
      - match: { desc: "delete_repository is blocked outright (overrides: disabled: true to allow)" }
        disabled: true
```

## Troubleshooting

**"My rule looks right but the deny doesn't fire."**

Most common causes:

1. **Wrong tool name.** The contract uses `mcp__github__delete_repo`
   but the actual tool is `mcp__github__delete_repository`. The
   stateless hook silently no-ops on non-matches (correct behaviour
   — we can't deny on a tool we have no rule for). Confirm the real
   name from `claude mcp list` or by triggering the tool once and
   reading the PreToolUse event from
   `~/.claude/projects/.../*.jsonl`.
2. **Wrong routing.** Check the file actually lives at
   `~/.sponsio/plugins/<routed-id>/sponsio.yaml` where the routed id
   matches what `derive_plugin_id` would return for the tool name.
   `mcp__github__X` → `github`; `acme:fetch` → `acme`; bare `Bash` →
   `_host`.
3. **Rate / count rules don't fire on the first call.** The
   stateless prototype gets a fresh empty trace per hook. Until the
   daemon mode lands, only argument-level rules (`arg_blacklist`,
   `arg_value_range`, `scope_limit`, `arg_length_limit`,
   `dangerous_*`, `tool_allowlist`) reliably fire on a single call.
   `rate_limit`, `loop_detection`, `must_precede`, `cooldown` are all
   daemon-future.

## What you must not do

* **Do not** auto-install everything without asking. The user picks
  which starter libraries they want — don't push the whole bundle.
* **Do not** apply scan output without showing the dry-run first.
* **Do not** invent tool names. If the user can't list them, tell
  them you can't safely scan and suggest they list MCP tools from a
  live Claude session first.
* **Do not** use `--force` to overwrite a user-edited library without
  explicit consent.
* **Do not** edit files under `sponsio/contracts/*.yaml` inside the
  installed Sponsio package. Those are read-only shipped packs.
  User-level adjustments go in
  `~/.sponsio/plugins/<plugin>/sponsio.yaml`.
* **Do not** flip `SPONSIO_GUARD_MODE=observe` "to make the smoke
  test pass". A failing smoke test means something is actually
  broken; silencing it hides the bug.
登录后下载/点赞/收藏 ❤ 10 | ★ 0
setup

目录结构

SKILL.md

举报内容

提示
