Do the main SEO and MCP workbenches preserve their declared decisions?
One runner imports the public production core and uses the same downloadable fixtures as the workflows. It checks duplicate metadata, schema-aware recipes, crawl plans, robots rules, batch indexability, run manifests, and MCP sessions.
All 13 deterministic cases matched their declared outputs. The runner makes no live network request and does not treat a local pass as proof that a remote page, crawler, or MCP server behaves the same way.
The runner imports the production core.
The fixtures call the exact functions loaded by the complete workflow pages. Assertions cover the output and the failure boundary; they do not repeat a second implementation inside the test.
corepack pnpm field-test:workbench- Does keyword cleanup expose conflicting duplicate metadata and apply the saved decision?PASS
- Do line-based keyword and crawl inputs keep their first entry even when a previous table setting marked the first row as a header?PASS
- Does crawl preparation enforce host, protocol, include, and exclude rules while preserving rejection reasons?PASS
- Does a pasted sitemap index stay separate from page crawl targets?PASS
- Does the robots evaluator select the longest applicable rule and Allow on an equal tie?PASS
- Does the crawl plan preserve the winning robots rule and every non-target group?PASS
- Does a recipe identify a moved column by name instead of silently reusing its old position?PASS
- Does the debugger treat a 200 response with noindex as technically blocked?PASS
- Does pasted response evidence produce the declared status, canonical, robots, and index signals?PASS
- Does a crawler export separate technical eligibility from noindex, robots, canonical, and conflicting signals?PASS
- Does the MCP lab identify and repair core tools/call structure?PASS
- Does the MCP session analyzer pair messages and expose unanswered and orphaned IDs?PASS
- Does a run manifest bind input, settings, and output without embedding source or export contents?PASS
Inspect the files before trusting the verdict.
The inputs, expected outputs, and portable recipes are public. They contain synthetic test data, not customer exports.
- Keyword inputExpected keyword CSVKeyword recipe
- Schema-aware keyword recipeRecipe schema v2
- Crawl inputExpected exclusionsCrawl recipe
- Crawl-plan inputrobots.txt fixtureExpected crawl plan
- Indexability evidenceExpected signalsIndexability recipe
- Batch indexability inputExpected URL groups
- MCP expected JSONMCP actual JSONMCP recipe
- MCP session JSONLExpected session summary
- Verifier manifestVerifier inputChanged verifier output
A fixture is not a live crawl.
- These downloadable deterministic fixtures exercise parsing and decision logic. They do not crawl a live URL, query a search engine, or contact an MCP server.
- The keyword fixture proves that conflicting metadata becomes visible and requires an explicit resolution; the receiving importer still requires a user-reviewed column mapping.
- The MCP checks cover selected version-specific core fields and expected-shape comparison, not the complete normative schema or a real client/server exchange.
- Robots, canonical, and indexability results remain bounded to the values supplied to the local functions.
Run the saved contract, or recapture the dated JSON after intentionally reviewing a core change.
corepack pnpm field-test:workbench:capture