tdmcp submission orchestrator

Coordinates four agents to prepare (and re-prepare) tdmcp's submission to the Anthropic Connectors Directory via the Desktop Extension (MCPB) path. Defines who works when; each agent's skill defines how.

Execution mode — sub-agents (this environment has no TeamCreate)

TeamCreate/SendMessage are not available here, so run agents as sub-agents with the Agent tool. For every agent:

• subagent_type: "general-purpose", model: "opus".
• Prompt MUST tell it to read its definition file .claude/agents/<name>.md and its required skill(s) first, then do the task. The definition file is the source of truth (reusable next session); the prompt just points at it + passes the task and the relevant _workspace/ paths.
• Parallelize independent agents by sending multiple Agent calls in one message (or run_in_background: true).

Data passing — file-based via _workspace/

Create _workspace/ at the repo root for inter-agent artifacts:

• 00_submission-spec.md — architect → everyone
• 01_migration-notes.md — bundle-engineer → QA/human
• 02_form-answers.md — docs-author → QA/human (paste-ready form draft)
• 03_qa-report.md — QA → human

Final deliverables (privacy page, edited repo files, tdmcp.mcpb) land in the repo proper. _workspace/ is preserved for audit / re-runs.

Phase 0: context check (initial vs re-run)

1. If _workspace/ is absent → initial run: full pipeline below.
2. If _workspace/ exists and the user asks to fix/update one part → partial re-run: re-invoke only the affected agent(s), pass the prior artifacts + the new feedback (e.g. a directory rejection reason). E.g. "privacy page rejected" → docs-author only, then QA.
3. If _workspace/ exists and the user gives fresh inputs → move it to _workspace_prev/ and start a new initial run.

Pipeline (initial run)

1. architect (solo, first). Produces 00_submission-spec.md. Everything downstream depends on it — do not parallelize anything before it finishes.
2. docs-author ∥ bundle-engineer (parallel — disjoint file sets):
   • docs-author owns: privacy page(s), config.ts nav additions, 02_form-answers.md. (Prose lane.)
   • bundle-engineer owns: dxt/manifest.json, scripts/build-mcpb.mjs, package.json scripts, scripts/setup.mjs, .dxt→.mcpb text sweep in existing docs, .github/ release workflow, 01_migration-notes.md. (Code lane.)
   • They never edit the same file — that's why they can run together safely.
3. submission-qa (solo, last). Runs builds + checks, writes 03_qa-report.md. If it reports blocking issues, loop back to the owning agent (partial re-run), then re-run QA on the affected checks.

Error handling

• An agent failing once → retry once with the failure context. Still failing → proceed and record the gap in the final summary; never silently drop it.
• Never bypass a failing gate (--no-verify, deleting a check). A red QA gate is the signal to fix, not to suppress.
• Conflicting facts (e.g. spec says manifest_version X, packer wants Y) → trust the live tool/repo, and have the architect update the spec rather than forcing the stale value.

After the run

Summarize to the user (in Portuguese): what changed in the repo, the contents of 02_form-answers.md and the QA verdict, and the explicit Needs human before submit list (support email, privacy contact, cutting a new release so the .mcpb asset exists, and the actual form click — which only the user can do). Then offer to address any QA nits or feedback.

Test scenarios

• Happy path: no _workspace/ → architect spec → privacy page + form draft + .mcpb migration → npm run docs:build and build:mcpb pass → QA all-PASS except human-input items → summary lists those items.
• Rejection re-run: user says "directory rejected: privacy policy incomplete" → Phase 0 detects existing _workspace/ → re-invoke docs-author with that feedback → QA re-checks the privacy gate only → updated summary.