graceful-boundaries（优美的界限）

Graceful Boundaries Conformance Audit

What This Skill Does

Assesses a URL's Graceful Boundaries conformance level through direct HTTP inspection, then provides a concrete implementation plan for reaching the next level. The output is an actionable document with code examples the user can implement immediately. No special tooling or dependencies required — the skill works with any HTTP client.

When To Use This Skill

• User provides a URL and asks about its rate limit communication
• User asks to check Graceful Boundaries conformance for a service
• User wants to know what level an API is at
• User asks how to improve their API's 429 responses
• User wants to elevate from one conformance level to the next
• User says "audit this API" in the context of rate limits or boundaries

Assessment Process

Follow these phases in order. Each phase builds on the previous one.

Phase 1: Discovery Fetch

Fetch the limits discovery endpoint directly. Try both standard paths:

GET <url>/api/limits
GET <url>/.well-known/limits

Use curl, fetch, or any HTTP client available in the current environment. No special tooling is required.

If either path returns a JSON response, record:

• Whether the response contains a service field
• Whether the response contains a limits object
• Whether limit entries are well-formed (each has type, maxRequests,

windowSeconds, description)

• Whether a conformance field is present (self-declared level)
• Whether the response includes a Cache-Control header with s-maxage
• Whether changelog or feed URLs are present (v1.1 change discovery)
• Whether resource-dedup entries include returnsCached: true (v1.1)

If neither path returns a valid response, the service has no discovery endpoint and cannot be Level 2 or above.

Optional accelerator: If the graceful-boundaries repo is cloned locally, the automated checker provides a structured report:

node evals/check.js <url> --json

This is a convenience, not a requirement. The skill works entirely through direct HTTP inspection.

Phase 2: Proactive Header Check

If the limits endpoint documents specific API endpoints, fetch one of them and check for proactive headers on the success response:

• RateLimit: limit=N, remaining=N, reset=N
• RateLimit-Policy: N;w=N

These headers indicate Level 4 conformance.

Do NOT attempt to trigger 429s. That would require hammering the service and is not appropriate for an audit. Level 1 and Level 3 conformance cannot be verified without observing an actual refusal response — note these as unverifiable and explain why.

Phase 3: Level Assessment

Map findings to the conformance levels defined in spec.md:

If the service self-declares a conformance level via the conformance field, compare declared vs. validated. Flag any discrepancy.

Report the assessment as:

• Confirmed level: what the evidence supports
• Declared level: what the service claims (if any)
• Likely level: best estimate including unverifiable aspects

Phase 4: Gap Analysis

For each level above the current confirmed level, list exactly what is missing. Reference specific sections of spec.md:

To reach Level 1 (spec sections 2 and 6):

• Do ALL non-success responses (400, 401, 403, 404, 429, 500, 503) include

error, detail, and why? (v1.1: why is MUST for all error classes)

• Are 429 responses JSON with the 5 required fields (error, detail,

limit, retryAfterSeconds, why)?

• Does error use a stable machine-parseable string (snake_case)?
• Does detail include a specific retry time in human-readable form?
• Does why explain the purpose, not restate the error?
• Is retryAfterSeconds a non-negative integer?
• Does the HTTP response include a Retry-After header?
• For HTML 429 pages: is there a <meta name="retry-after" content="N">

tag or a <link rel="alternate" type="application/json" href="...">? (v1.1)

To reach Level 2 (spec section 1):

• Does a limits endpoint exist at /api/limits or /.well-known/limits?
• Does it return JSON with a limits object?
• Are limit entries well-formed (type, maxRequests, windowSeconds, description)?
• Is the endpoint cacheable (Cache-Control header)?
• Does it include changelog or feed URLs for change discovery? (v1.1, optional but recommended)

To reach Level 3 (spec sections 3 and 5):

• Do refusal responses include constructive guidance fields?
• Which guidance categories apply? (cachedResultUrl, alternativeEndpoint,

upgradeUrl, humanUrl, cached)

• Does the service prefer guidance in the recommended order:

use cached > try alternative > upgrade > wait > human handoff?

• For resource-dedup limits: does the service return cached results as a

200 instead of a 429? If so, does the discovery endpoint include returnsCached: true so agents skip retry logic? (v1.1)

To reach Level 4 (spec section 4):

• Are RateLimit headers present on success responses?
• Do they include all three components: limit, remaining, reset?
• Is a RateLimit-Policy header present?
• Does the policy format match N;w=N?

Phase 5: Implementation Guidance

Provide concrete, copy-pasteable code for each gap. Use the service's actual domain and endpoints in examples.

Limits discovery endpoint skeleton:

{
  "service": "<service name>",
  "description": "<what the service does>",
  "conformance": "level-2",
  "changelog": "https://<domain>/api/changelog.json",
  "feed": "https://<domain>/feed.json",
  "limits": {
    "<endpoint-key>": {
      "endpoint": "<path>",
      "method": "<HTTP method>",
      "limits": [
        {
          "type": "ip-rate",
          "maxRequests": 100,
          "windowSeconds": 3600,
          "description": "100 requests per IP per hour."
        },
        {
          "type": "resource-dedup",
          "maxRequests": 1,
          "windowSeconds": 86400,
          "returnsCached": true,
          "description": "One operation per resource per day. Repeat requests return the cached result."
        }
      ]
    }
  }
}

Structured refusal body:

{
  "error": "rate_limit_exceeded",
  "detail": "You have exceeded the limit of 100 requests per hour. Try again in <N> seconds.",
  "limit": "100 requests per IP per hour",
  "retryAfterSeconds": 1234,
  "why": "<one sentence explaining why this limit exists — not just restating the error>"
}

Constructive guidance fields (add to the refusal body):

{
  "cachedResultUrl": "/api/result?id=<resource>",
  "alternativeEndpoint": "/api/<alternative>",
  "upgradeUrl": "https://<domain>/pricing",
  "humanUrl": "https://<domain>/contact"
}

Proactive headers (add to success responses):

RateLimit: limit=100, remaining=99, reset=3600
RateLimit-Policy: 100;w=3600

Reference security considerations where relevant:

• SC-1: Published limits may be higher than enforced limits
• SC-2: why must describe the category of protection, not the mechanism
• SC-3: expected must use positive descriptions
• SC-6: Guidance URLs must be relative or same-origin

Phase 6: Generate the Assessment Document

Output a structured markdown document:

# Graceful Boundaries Assessment: <domain>

## Summary
- Confirmed level: <N>
- Declared level: <N or "not declared">
- Likely level: <N>

## What was checked
- Limits endpoint: <path> — <found/not found>
- Proactive headers: <present/absent>
- Refusal format: <not verifiable without triggering a 429>

## Gaps to next level
<prioritized list with spec section references>

## Implementation plan
<concrete code examples using the service's actual domain>

## Security notes
<relevant SC-* considerations>

What This Skill Does NOT Do

• Does not implement changes on the target service
• Does not deliberately trigger rate limits or 429 responses
• Does not require access to the service's source code
• Does not assess general API design quality beyond limit communication
• Is distinct from the agent-readiness-audit skill (which assesses

overall AI discoverability, not rate limit conformance specifically)