Catalog Detector

Read and follow references/workflow.md.

• Archive previous run: Before starting, check if docs/catalog-detector/{slug}/ exists. If it does, archive it: mv docs/catalog-detector/{slug} docs/catalog-detector/{slug}-archived-$(date -u +%Y%m%dT%H%M%S). Then create the fresh directory: mkdir -p docs/catalog-detector/{slug}. Every invocation generates from scratch — never read from or make decisions based on archived directories ({slug}-archived-*).
• Before starting the workflow, run the catalog probe script: uv run .claude/skills/catalog-detector/scripts/catalog_probe.py --url {site_url} --slug {slug} --knowledgebase-dir docs/platform-knowledgebase --output-dir docs/catalog-detector/{slug} Parse stdout as JSON. Individual probe results are saved as JSON files in the output directory. Read docs/catalog-detector/{slug}/probe.log for diagnostics if needed.
  • Exit code 0 or 1: use the probe results to inform Steps 1-3. Routing logic:
    • If recipe_match is "full" AND transport_health is "healthy" AND anti_bot shows no blocking AND js_rendering_signals show no SPA framework: → Skip Steps 1-3, proceed to Step 4 using the probe data.
    • If recipe_match is "full" BUT anti_bot shows issues OR js_rendering_signals.text_to_html_ratio < 0.05: → Skip Steps 1-2, proceed to Step 3 for deeper investigation of the flagged concern.
    • If recipe_match is "partial" or "poor": → Use probe data as a starting point, proceed to Step 3 (starting at 3b, the probe already covers 3a).
    • If recipe_match is "untested" (no knowledgebase or all URLs failed): → Proceed to Step 3 from 3a, using probe's anti_bot and platform signals as context.
    • If platform_guess is "unknown" but signals suggest a recognizable CMS, classify as "custom" — the script never returns "custom" because that requires LLM judgment.
    • Check the errors array: any entry means the corresponding section's data is incomplete. Supplement with manual investigation for those sections.
  • Exit code 2: ignore the script output, fall back to manual Steps 1-3 entirely.
  • DO NOT modify the probe script. If it produces unexpected results, use manual reasoning to verify or override — the script output is evidence, not a verdict.
• After writing the catalog assessment, run the validation script: uv run .claude/skills/catalog-detector/scripts/validate_assessment.py docs/catalog-detector/{slug}/assessment.md --knowledgebase-dir docs/platform-knowledgebase --output-json docs/catalog-detector/{slug}/validate.json 2>docs/catalog-detector/{slug}/validate.log Fix any failing gates. DO NOT modify the validation script.
• These scripts are infrastructure, not generated code. Never edit, patch, or fix them during a pipeline run — even if you can identify the bug. If a script fails (exit code 2) or produces unexpected results, fall back to manual reasoning for those steps. Record the issue in the catalog assessment's Platform-Specific Notes so it can be fixed in a dedicated maintenance pass. Editing scripts mid-pipeline risks introducing regressions that affect other companies.
• Read the company report first for the company URL, categories, and business model context.
• Provide the file paths from the table above when the workflow references logical resources (e.g., "the company report", "write the catalog assessment report", "the platform knowledgebase").
• Use web search, web fetch, and Playwright browser tools to investigate catalog pages as the workflow directs.
• This skill does not escalate — all decisions are autonomous stops or track switches. If the workflow determines scraping is not viable (no_public_catalog, auth_required, anti_bot_severe, js_only, or attributes_not_extractable), it writes a minimal catalog assessment and stops. If the fast-path recipe verification fails (platform_recipe_failed), the workflow autonomously switches to the deep investigation path. Report the outcome to the user.
• The platform knowledgebase may not exist for platforms not yet encountered. On the fast path, a missing knowledgebase file means the platform is treated as unknown (deep investigation). On the deep investigation path, a new knowledgebase file is created if the platform is identifiable.

Notes

Owns the platform knowledgebase — reads on every run, writes on every successful run.