Role

Socratic interview → generates <run-path>/connector.md tailored to the user's source. Use when templates in ${CLAUDE_PLUGIN_ROOT}/templates/connectors/ don't match.

Invocation

/ultra-analyzer:connector-init <run-path>

Protocol

Step 1: Pre-flight

• Verify <run-path>/state.json exists (run is initialized).
• If <run-path>/connector.md already exists, ask: overwrite, or keep existing?

Step 2: Interview (use AskUserQuestion)

Ask these questions one at a time, in order. Each answer shapes subsequent questions.

Q1: Source category

• What kind of data source are you analyzing?

Options: database (Mongo/Postgres/SQLite/etc.) | filesystem (files on disk) | HTTP API (REST/GraphQL) | browser (scraping/automation) | streaming (logs, Kafka) | archive (tarball, zip of structured files) | custom/other

Q2: Access mechanism

• How does the pipeline connect to this source?
• If database: connection string (env var name?) or MCP tool name?
• If API: base URL, auth mechanism (bearer token? OAuth? API key header?)
• If filesystem: root path, file globs?
• If browser: playwright-persistent MCP, or browsermcp MCP, or CLI tool?
• If custom: describe the CLI / library / process that reads it

Q3: Unit of analysis

• What is a single "unit" in this source?
• Database: a collection/table
• Filesystem: a file (or a directory)
• API: an endpoint or a resource collection
• Browser: a page or a DOM subtree
• Streaming: a log file or a time window
• Listen for: is there a natural iteration that produces discrete units?

Q4: Schema introspection

• How do you discover the shape of a unit without loading it fully?
• Database: db.<coll>.findOne() or get_schema MCP tool
• Filesystem: read first N bytes, detect format by extension
• API: call the endpoint once and inspect response shape
• Browser: navigate + DOM sample
• Custom: the user's introspection approach

Q5: Query execution

• Given a topic's query spec (a JSON blob describing what to look for), how do you run it?
• This is the 80% of the work. Probe for: what query DSL does the source speak? Aggregation pipeline? SQL? regex? CSS selectors? XPath?

Q6: Cross-references

• Does a query result typically contain references (IDs, URLs, paths) that need to be followed to get the full picture?
• Yes → how? ($lookup, follow-up HTTP calls, symlink resolution)
• No → resolve_refs is a pass-through.

Q7: Citation format

• When the final report quotes a datum, what's the canonical "where this came from" anchor?
• Examples: [DOC:users._id=abc123], [FILE:src/auth.ts:42], [URL:api.github.com/repos/foo], [ROW:log.jsonl:1234].
• The anchor must be unambiguous AND parseable (validator checks it).

Q8: Sensitive/forbidden data

• What fields, patterns, or values must NEVER leak into findings?
• Examples: passwords, API keys, PII (emails, phone), PCI data.
• Static list? Regex patterns? Config-driven?

Q9: Rate limits / budget

• Any per-query or per-run limits?
• API rate limits (requests per second/minute)
• DB query cost ceilings
• File size limits
• Browser page-load ceilings

Q10: Auth secrets

• Confirm: are secrets (tokens, passwords) stored in env vars and NEVER in committed files?
• If the user hasn't set this up yet, stop and guide them to do so before proceeding.

Step 3: Synthesize connector.md

Based on answers, write <run-path>/connector.md with:

• All 6 operation sections filled in with concrete, runnable instructions
• Source type, auth mechanism, budget constraints declared upfront
• A "Known limitations" section listing any answers that were uncertain

Reference the templates in ${CLAUDE_PLUGIN_ROOT}/templates/connectors/ as style guides — follow their section structure and specificity level.

Step 4: Smoke-test

Immediately after writing, attempt enumerate:

bash ${CLAUDE_PLUGIN_ROOT}/bin/adapter.sh <run-path> enumerate

If it returns a non-empty list → connector is viable. If it errors, show the error and offer to revise the spec.

Step 5: Print summary

✓ Generated connector.md at <run-path>/connector.md
  Source: <one-line description>
  Smoke test (enumerate): <PASS/FAIL with detail>
  
Next: proceed to /ultra-analyzer:run (Gate 1 will review the connector).

Hard rules

• NEVER write auth tokens into connector.md. Always reference env vars (e.g. $MY_API_TOKEN).
• NEVER guess answers. If the user is unsure about a question, mark the section "TODO: clarify before first real run" and flag it to the user.
• NEVER skip the smoke test. A connector that can't enumerate is useless downstream.
• If the user describes a source so exotic none of the 6 operations cleanly apply, be honest: "This source may not fit the pipeline's contract. Consider preprocessing to a supported shape (e.g. dump to JSONL first), or extending the contract via docs/EXTENDING-SOURCES.md."