Sketchi log analysis

Skill shpitdev/sketchi/.agents/skills/sketchi-log-analysis

Analyze Sketchi playground Cloudflare AI Gateway logs for prompt tuning, Gemini flowchart IR reliability, retained request/response payloads, and scenario regression evidence. Use when inspecting google-ai-studio Gateway logs, validating generated Sketchi IR, or turning model failures into scenario or prompt changes.From its SKILL.md

Install

npx -y skills add shpitdev/sketchi --skill sketchi-log-analysis

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

13 stars13 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

2.9 KB, as published. Nobody here has run it

Sketchi Log Analysis

Use this skill to inspect Sketchi playground AI Gateway logs through the Cloudflare API MCP. Keep the workflow read-only.

Cloudflare MCP

Use the Cloudflare API MCP server:

{
  "mcpServers": {
    "cloudflare-api": {
      "url": "https://mcp.cloudflare.com/mcp"
    }
  }
}

The target Gateway is usually google-ai-studio in the Sketchi Cloudflare account.

Workflow

Use the Cloudflare API MCP to find the AI Gateway log endpoints if needed.
List recent logs for google-ai-studio with a small page size first.
Prefer non-cached logs when evaluating model behavior.
Read log detail for useful samples.
Prefer retained detail fields such as request_head and response_head when they are marked complete.
Use separate request/response payload endpoints only when detail does not include enough retained payload data.
Parse the retained response into candidate Sketchi IR and classify it.
Report concrete follow-up: scenario assertion, prompt change, deterministic converter issue, or Gateway/provider issue.

Do not mutate Cloudflare resources. Do not delete logs. Do not print secrets, provider credentials, bearer tokens, or full unrelated payloads.

Report Fields

For each relevant request, report:

Timestamp, model, provider route, status, latency, and token usage.
Scenario ID or scenario title from request metadata or payload.
Whether request and response payloads were retained.
Whether the request has a true system/user split.
Whether the system prompt contains the flowchart IR rules.
Whether the user prompt avoids duplicating the flowchart IR rules.
Whether the model response was valid JSON.
Whether the response satisfied the flowchart IR contract.
Any recurring failure pattern that should become a scenario assertion or prompt change.

Flowchart IR Contract

Classify a candidate as pass only when all of these are true:

type is flowchart.
Exactly one node has kind: "start".
At least one node has kind: "end".
Every node has id, label, and kind.
Every edge uses existing source and target node IDs.
Every non-end node has at least one outgoing edge.
End nodes have zero outgoing edges.
Every decision node has at least two outgoing edges.
Every outgoing edge from a decision node has a non-empty unique label.

Use prompt-risk when JSON and contract shape are valid but scenario semantics are missing or ambiguous. Use contract-fail for typed IR violations. Use gateway-fail for auth, provider, route, timeout, or Gateway failures that prevent evaluation.