Compile pipeline + phases/gates prompts + full eval framework (mcp 0.7.0)

[milstan]

Summary

Introduces @leadbay/promptforge — a build-time workspace package that compiles .md.tmpl source files into prompts.generated.ts / tool-descriptions.generated.ts. Authors edit prose in one place; bundles inline the generated string constants. The MCP wire format is unchanged: every consumer (Claude Desktop, Cursor, Claude Code, OpenClaw) sees the same tools/list / prompts/list / prompts/get shapes.

Backwards-compat preserved: every tool name, every inputSchema, every annotation set, every outputSchema is byte-for-byte identical. Only the description prose changes — and it's tighter, more structured, and easier to evolve.

Highlights

• All 6 MCP prompts restructured into PHASES + IRON LAWS + GATES with byproduct outputs (COLUMN PRESERVATION PLAN, DECISION LOG, FINAL REPORT, STOP — awaiting user decision).
• All 60 tool descriptions migrated to .md.tmpl (21 composite + 37 granular + 2 file-import companions) using a shared snippet library of 11 partials.
• Daily check-in overhaul: top-10 motivational triage; auto-top-up via leadbay_bulk_qualify_leads; research every promising lead; ask before consuming contact-enrichment quota.
• Import-file overhaul: explicit GOAL ("maximize how many rows the Leadbay system can ingest and match"); two deliverables named (max-coverage column mapping + augmented file with LEADBAY_ID column); WHAT GOOD LOOKS LIKE section; per-CRM EXTERNAL_ID guidance extended from HubSpot to Salesforce / Pipedrive / Close / Attio.
• Full eval framework (packages/mcp/test/eval/): in-process MCP server driver, Evidence-pyramid pass rule, mission-match judge with rubric-from-frontmatter, drift meta-judge, replay-by-default backend HTTP, worktree-based drift detector, 70-fixture tool-routing classifier. Gated by EVAL=1; default pnpm test ignores it.
• Audit tests (packages/mcp/test/audit/): eval-coverage stubs, no-inline-descriptions-in-server.ts, per-tool ≤3500-char budget, naming convention, snippet orphans, freshness gate.
• CI gate: root pnpm test runs pnpm prompts:check first; PR fails if .md.tmpl edits don't have matching regenerated .generated.ts updates.

Version bumps

• @leadbay/mcp: 0.6.3 → 0.7.0 (publishable)
• @leadbay/core: 0.4.0 → 0.5.0 (private)
• @leadbay/promptforge: 0.1.0 (new, private)
• dxt/manifest.template.json login-command string: @leadbay/mcp@0.6 → @leadbay/mcp@0.7

CHANGELOG: rolled the previous 0.6.4 UNRELEASED import-flow entry forward into 0.7.0 (the 0.6.4 work landed via PR #33 but never published).

Test plan

• pnpm prompts:check — generated files in sync with .md.tmpl sources
• pnpm test — 430/430 pass (240 core + 12 leadclaw + 169 mcp + 9 promptforge)
• pnpm build — all packages build cleanly; published artifact is leadbay-0.7.0.mcpb (275.7 KB)
• Live MCP server smoke: tools/list returns 58 tools at full exposure with migrated descriptions; prompts/get(leadbay_import_file) returns a 14k-char body with 5 PHASE blocks, 3 GATE blocks, the GOAL section, the WHAT GOOD LOOKS LIKE section, and the multi-CRM record-link guidance
• Snapshot regression test: pre-migration body of leadbay_import_file (PR Add agentic Leadbay file import flow #33 prose) captured under packages/promptforge/test/snapshots/ so accidental drift surfaces as a test failure
• Audit-compliance test feat(mcp): install/login subcommands + npm publish prep + LLM-hallucination-proof docs #6 confirms each tool name resolves to exactly one Tool object (cross-tier sharing for leadbay_list_mappable_fields / leadbay_add_note / leadbay_create_custom_field preserved as intended)

Post-merge release steps

Per RELEASE.md:

1. git checkout main && git pull
2. git tag mcp-v0.7.0
3. git push origin mcp-v0.7.0
4. Watch the release workflow in GitHub Actions. It runs preflight-npm → publish-mcp (build + test + tag/version check + npm publish --access public --provenance).

🤖 Generated with Claude Code