# Preapproval Agent

> Approval pre-review agent. Process webhook-driven approval items, analyze request reasonableness, and execute auditable approve/reject decisions.

- **Type:** Skill
- **Install:** `agentstack add skill-cloudchef-atlasclaw-providers-preapproval-agent`
- **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior
- **Seller:** [CloudChef](https://agentstack.voostack.com/s/cloudchef)
- **Installs:** 0
- **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills)
- **Latest version:** 0.1.0
- **License:** Apache-2.0
- **Upstream author:** [CloudChef](https://github.com/CloudChef)
- **Source:** https://github.com/CloudChef/atlasclaw-providers/tree/main/providers/SmartCMP-Provider/skills/preapproval-agent
- **Website:** https://atlasclaw.ai/en/

## Install

```sh
agentstack add skill-cloudchef-atlasclaw-providers-preapproval-agent
```

Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent.

## About

# Preapproval Agent

Autonomous backend agent for approval pre-review. **Not a human confirmation flow.**

## Purpose

When triggered by a webhook:
1. Fetch and analyze approval request details
2. Evaluate request reasonableness with the shared pre-approval analysis contract
3. Execute approve/reject via existing approval skills
4. Return structured decision summary

## Trigger Conditions

This skill activates when:
- Webhook payload targets approval pre-review
- `agent_identity` is `agent-approver`
- Valid `request_id` is provided

## Robot Admin Execution

For webhook-driven backend execution, run this agent against an explicitly selected SmartCMP provider instance with a robot/admin credential. Set `ATLASCLAW_PROVIDER_INSTANCE` to the intended instance name; if that instance is not configured, execution must fail closed rather than falling back to `prod` or another instance.

The robot provider instance should use a SmartCMP `cmp_tk_*` provider token when available. The shared scripts send those tokens as `Authorization: Bearer ` and keep non-`cmp_tk_*` session tokens on the existing `CloudChef-Authenticate` header.

Treat `ATLASCLAW_USER_ID=webhook-*` as the AtlasClaw trigger identity, not as a CMP actor. Approval execution uses `../approval/scripts/approve.py` and `../approval/scripts/reject.py`, which load the selected robot credential through `_common.require_config()`.

Use this mode only for robot profiles whose `allowed_skills` include `smartcmp:preapproval-agent`. The same SmartCMP robot profile may also allow `smartcmp:request-decomposition-agent` when the same robot/admin account is approved for both workflows.

## Inputs

| Input | Type | Required | Description |
|-------|------|----------|-------------|
| `provider_instance` | string | Yes | CMP provider instance name (e.g., `cmp-prod`) |
| `robot_profile` | string | For webhook robot mode | Robot profile configured on the selected provider instance |
| `agent_identity` | string | Yes | Must be `agent-approver` |
| `request_id` | string | Yes | SmartCMP Request ID for execution, e.g. `RES20260505000010`, `TIC20260502000003`, or `CHG20260413000011` |
| `trigger_source` | string | No | Source label (e.g., `cmp-webhook`) |
| `policy_mode` | string | No | Policy preset (default: `balanced`) |

**Validation Rules:**
- If `request_id` is missing → **Stop immediately**
- If `request_id` cannot resolve to a pending approval activity → **Fail closed**
- If `agent_identity` ≠ `agent-approver` → **Stop immediately**

## Orchestrated Skills

This agent does NOT access the platform directly. It orchestrates:

| Skill | Purpose |
|-------|---------|
| `smartcmp_preapproval_get_request_detail` | Fetch pending approval details |
| `smartcmp_preapproval_get_catalog_detail` | Fetch the service catalog/card Markdown by `catalogId` |
| `smartcmp_preapproval_analyze_request` | Run the shared read-only pre-approval evaluator |
| `smartcmp_preapproval_approve` | Execute approval with reason |
| `smartcmp_preapproval_reject` | Execute rejection with reason |

## Workflow

```
1. Validate Inputs
   ├── Check provider_instance, agent_identity
   └── Verify request_id exists
         ↓
2. Fetch Approval Context
   └── smartcmp_preapproval_get_request_detail → Verify request_id
         ↓
3. Build Review Summary
   ├── Service/request name
   ├── Requester notes
   ├── Full parameters
   ├── Cost estimate
   └── Approval history
         ↓
4. Evaluate Against Shared Analysis
   └── smartcmp_preapproval_analyze_request → apply catalog policy and default rubric
         ↓
5. Choose Outcome
   ├── approve
   ├── reject_with_guidance
   └── manual_review_required
         ↓
6. Execute Decision
   ├── approve → smartcmp_preapproval_approve  --reason ""
   ├── reject  → smartcmp_preapproval_reject  --reason ""
   └── manual  → stop without mutating SmartCMP and return manual-review guidance
         ↓
7. Return Structured Result
```

## Shared Analysis Contract

Use `smartcmp_preapproval_analyze_request` as the common evaluator before any automated decision. The same script powers user-facing read-only analysis, so approval reasoning does not drift between the manual UI flow and the backend preapproval agent.

The analysis result is advisory. This agent is the only component in this workflow that may subsequently call `smartcmp_preapproval_approve` or `smartcmp_preapproval_reject`, and it must do so only after validating the analysis output and request context.

## Catalog Policy Override

For every pending approval, fetch the service catalog/card by the `catalogId` returned from approval metadata. If `catalogId` is missing, fail closed and do not approve.

If the catalog Markdown contains `# Pre Approval Instructions`, that section is the authoritative approval policy for this request. Also accept `# Preapproval Instructions` and `# Pre-Approval Instructions` as heading variants.

Only use the built-in Decision Rubric when the catalog/card was fetched successfully and no pre-approval section exists.

If the catalog/card cannot be fetched, fail closed and do not approve. If catalog pre-approval instructions exist but cannot be deterministically evaluated by the shared analyzer, require manual review instead of speculative approval.

## Decision Rubric

### Approve When (most satisfied):

| Factor | Criteria |
|--------|----------|
| **Business Purpose** | Requester explains what the resource is for |
| **Resource Fit** | Size, environment, options proportional to stated use |
| **Configuration** | Parameters don't conflict, technically plausible |
| **Least-Necessary** | No excessive CPU, memory, storage without justification |
| **Environment** | Production requests have stronger rationale |
| **Cost** | Proportionate to described scenario |
| **Actionable Notes** | Description concrete enough for approval |

### Reject When (any true):

- No meaningful business justification
- Resources obviously oversized for stated need
- Production resources for vague/low-risk scenarios
- Request incomplete, contradictory, or copy-pasted
- Unusual/expensive resources without explanation
- Material risk with insufficient data

## Decision Style

> Be strict, concise, and auditable.

- Do NOT invent facts missing from request
- Do NOT ask requester follow-up questions
- Prefer rejection with guidance over speculative approval
- Explain what would make request approvable

## Comment Templates

**Approval example:**
```
Approved by agent pre-review. Business purpose clear, resource specs reasonable.
```

**Rejection example:**
```
Rejected by agent pre-review. Missing business justification, resource specs, and target environment. Please resubmit with details.
```

## Output Contract

```json
{
  "decision": "approve",
  "confidence": "high",
  "reasoning": [
    "Business purpose is explicit.",
    "Requested capacity proportional to described workload."
  ],
  "improvement_suggestions": [],
  "provider_action": {
    "skill": "../approval/scripts/approve.py",
    "success": true
  }
}
```

For rejections, include `improvement_suggestions`.

## Failure Handling

| Scenario | Action |
|----------|--------|
| Detail retrieval fails | Return failure, do NOT approve |
| Approval execution fails | Return provider error as-is |
| Rejection execution fails | Return provider error as-is |
| Ambiguous/expensive/high-risk | Return manual-review guidance without mutating SmartCMP |

## References

- [review-guidelines.md](references/review-guidelines.md) — Detailed review criteria

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [CloudChef](https://github.com/CloudChef)
- **Source:** [CloudChef/atlasclaw-providers](https://github.com/CloudChef/atlasclaw-providers)
- **License:** Apache-2.0
- **Homepage:** https://atlasclaw.ai/en/

Install and usage instructions live in the source repository linked above.

## Pricing

- **Free** — Free

## Security capabilities

Automated source analysis of v0.1.0 — what this tool can access:

- **Network access:** no
- **Filesystem access:** no
- **Shell / process execution:** no
- **Environment & secrets:** no
- **Dynamic code execution:** no

*"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.*


## Versions

- **0.1.0** — security scan: passed — Imported from the upstream source.

## Links

- Listing page: https://agentstack.voostack.com/l/skill-cloudchef-atlasclaw-providers-preapproval-agent
- Seller: https://agentstack.voostack.com/s/cloudchef
- Browse the marketplace: https://agentstack.voostack.com/browse

---
Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.