MemoPop Connector API, Phase 7: The Plugin and Real Claude

A thin skill that keeps Claude on the method, packaged with the connector, and a test where real Claude writes the fixture memo.

Path
plans/MemoPop-Connector-API-Phase-7-Plugin-and-Real-Claude.md
Authors
Michael Staton
Augmented with
Claude Code on Claude Opus 5.5 (1M context)
Tags
Plan · MemoPop · Claude-Plugin · Agent-Skills · TDD

MemoPop Connector API, Phase 7: The Plugin and Real Claude

Read the spec (§Docs that aren’t generated, §The test suite, layer 11) and [[The-Server-Hands-Claude-the-Method-One-Step-at-a-Time]].

Steps

  1. Check current Claude plugin documentation before packaging: the plugin manifest format, how a plugin declares a remote MCP connector, and which surfaces load skills. Record the sources in the changelog.
  2. Failing tests for both owned IDs. CONN-PLUG-02 is gated on ANTHROPIC_API_KEY and a reachable server URL.
  3. plugin/: the plugin manifest, the connector entry pointing at https://memopop.didi.sh/mcp, and skills/memopop/SKILL.md. The skill is thin and states the rules: this is a multi-step pipeline with an artifact trail; start with list_deals; always call next_step and do only what it says; show the partner every research file and wait for approval before submitting it with partner_approved: true; on skipped, say so in one line and continue; on down, stop and tell the partner. CONN-PLUG-01 checks each rule is present.
  4. The real-Claude test (tests/connector/test_real_claude.py): the Claude API’s MCP connector support pointed at a server running fixture-co, with the skill as the system prompt and a scripted partner that approves research. Assert from the server’s step log: tools in order, research shown before drafting, one forced skip handled in a line, one forced down stopping cleanly. Use a current Claude model, and keep the fixture small so a run costs little.
  5. docs/operator/connect-claude.md: adding the connector in Claude Desktop and installing the plugin.
  6. Changelog, commit, push. Report the cost and duration of one real run.
  7. Follow-up from the first live runs (2026-10-11): the skip line. Real Claude saw research.sources in skipped_since_last_call and went on to drafting without telling the partner. next_step now writes the sentence itself (tell_partner, CONN-STEP-10, owned here because the finding came from this plan’s run), the instruction opens with it on that call, and the skill says to relay it word for word before the step’s work.