Suede MCP QA

Gate policy: advisory, not blocking

Every claim-verification step, check, quality gate, and ship verdict in this skill is a recommendation to the user, not a control on the agent. This policy governs every gate, check, verdict, and "do not ship / publish / proceed" line elsewhere in this skill:

• Run every check and report the results honestly. Verdicts (ship, ship-with-caveats, hold, letter grades, BLOCKED or OPEN items) are advice attached to the work, not orders that change it.
• Never block, delay, skip, rewrite, or refuse the action the user asked for because a check failed or a gate said hold. Complete the requested action as asked, and deliver the gate output alongside it as a clearly labeled recommendation.
• A failed gate changes what you report, never what you do.
• Single exception: if a finding is extremely risky (data loss, security or credential exposure, legal or rights violations, payment mistakes, or irreversible public damage), pause, tell the user exactly what the risk is and what the options are, and let them pick. Their choice is final.

Use this skill when a Suede MCP server or MCP docs surface changes.

Core principle: a check that did not run against the live server did not happen.

Operating Stance

• Run against a live MCP server, not a spec document. If the server is not running, start it before checking.
• For each check, record the exact command run and the exact output received. Do not summarize.
• A check that cannot run (server unreachable, tool not implemented) is a FAIL, not a skip.
• Report failures immediately: do not wait until all checks complete to surface a blocker.
• Never mark a skill as present in the catalog unless its folder exists and its SKILL.md is readable.
• Never mark an install command as working unless you ran it from a temporary destination directory.

Checks

1. Run syntax checks and the repo's hermetic MCP protocol tests.
2. Parse catalog JSON and confirm every listed skill folder exists, then run scripts/mcp-surface-snapshot.sh to compare the catalog's mcp block against what the live server actually serves (exit 1 means drift; the server wins).
3. Exercise the full lifecycle in one process: initialize, the notifications/initialized notification, then ping, tools/list, tools/call, resources/list, resources/read, prompts/list, and prompts/get.
4. Verify supported protocol versions are echoed and an unsupported client version negotiates to the server's latest supported version.
5. Confirm every tool has a closed inputSchema, an outputSchema, and read-only/non-destructive/idempotent annotations.
6. Confirm every successful tool call returns structuredContent, a useful human-readable text block, and a serialized JSON text fallback for older clients.
7. Check pre-initialization calls, repeated initialization, bounded input, bounded arguments, invalid names and schemas, malformed JSON, and unknown methods.
8. Confirm healthy stderr is empty and stdout contains newline-delimited JSON only; logs and stack traces must never corrupt the transport.
9. Confirm install output leads with public GitHub skill installs, local plugin commands are labeled local-only, and README/docs/catalog language agrees with the live server.

This Server's Real Surface

mcp/suede-skills-mcp.mjs is the only server this skill QAs. Do not check it against a generic MCP checklist: check it against this exact surface. Read mcp/catalog.json first; the mcp block there must match what tools/list, resources/list, and prompts/list actually return.

Derive that surface, never recite it: scripts/mcp-surface-snapshot.sh runs one stdio session against the server, prints the tool names, resource URIs, and prompt names it actually serves with their counts, and diffs them against the catalog's mcp block. Exit 1 means drift: the server's own tools/resources/prompts arrays are ground truth, and mcp/catalog.json is what gets corrected.

Stdio Test Blocks

The copy-paste JSON-RPC blocks that exercise initialize, tools, resources, prompts, and the lifecycle and malformed-input probes are in references/stdio-test-blocks.md. Open it when you are actually running the checks, not when deciding which checks apply.

Failure Handling

Failure type	Severity	Action
Server fails to start	Critical	Stop. Report startup error verbatim.
tools/list returns empty	Critical	Stop. The MCP is non-functional.
Lifecycle or protocol negotiation fails	High	Hold. Capture the request/response transaction.
Tool schema, output schema, or read-only annotation missing	High	Hold. Repair the published contract and rerun the suite.
Structured result lacks either text fallback	High	Hold. Preserve structured and legacy-client output together.
Listed skill folder missing	High	Flag each missing folder. Continue checking others.
Malformed JSON-RPC response	High	Report the raw response. Flag as broken.
Install command leads with local-only path	High	Flag. Install output must lead with public GitHub route.
Docs/catalog language mismatch	Medium	List each mismatch. Flag as hold-with-caveat.
Tool implemented but not in catalog	Low	Flag as undocumented. Not a blocker.

Recommended ship gate rules (advice to the user, not a lock on any action):

• Any Critical or High failure → hold
• Medium failures only → ship-with-caveats (list each caveat)
• No failures → ship

Output

Server:
Commands run:
Tools checked:
Resources checked:
Prompts checked:
Install output:
Failures:
Fixes:
Ship gate: ship | ship-with-caveats | hold

Red Flags: Stop

• "The server ran fine last week; no need to restart it for this.": Run every check against the live server now.
• "The catalog parses, so the folders are surely there.": Open every listed folder and read its SKILL.md.
• "That check can't run, I'll mark it skipped.": A check that cannot run is a FAIL.
• "The output looked right, close enough.": Record the exact command and exact output, verbatim.

Boundaries

• Check and report only. Do not edit the server source, mcp/catalog.json, or the docs surface to make a check pass: hand each fix back through Routing and re-run.
• Do not publish, tag, or release anything; this skill clears an MCP release, it does not ship one.
• Never record a check as passed from a spec, a README, or a previous run. Only output captured from the live server in this session counts.
• Do not extend a verdict to a third-party MCP server: the surface above is this pack's, and a generic server has not been checked against it.

Routing

After QA:

• MCP source needs fixes → return to the MCP source file and fix, then re-run this skill
• MCP source changed to fix a QA failure → suede-code-review on that diff before re-running this skill, so the repair itself is not shipped unreviewed
• Catalog JSON needs updates → edit mcp/catalog.json and re-run steps 2 and 7
• Docs/README language mismatch → update the docs surface to match live MCP output (private Suede Labs companion, not in this pack: suede-docs), then re-run check 7
• Install command broken → suede-launch-packaging to fix and test the install path