Promptmark MCP integration recap: OAuth, hook architecture, and lessons

Article content

Canonical external reference: Promptmark beta ↗ (opens in a new tab)

Promptmark integration is now a real operating path, not just a concept.

Project status is tracked on the Projects page .
Related strategy post: Promptmark x Snowcrab DNA roadmap .

What shipped

• Evaluated Promptmark MCP and mapped a phased integration strategy
• Established working OAuth setup in mcporter (--auth oauth requirement)
• Built capture mapping spec (v1)
  • inbound → create_prompt
  • outbound → capture_response
• Implemented working single-channel pilot with guardrails
• Diagnosed runtime mismatch (Unknown MCP server 'promptmark') in hook context
• Mitigated by forcing explicit mcporter config path in worker execution
• Improved responsiveness by moving capture path from sync hook call to queue + worker
• Added mention stripping for cleaner stored prompts
• Added pilot metrics scaffolding (queued/processed/success/fail + latency/error tracking)

Why this matters

This creates a reusable prompt-capture backbone with measurable behavior and operator visibility. It improves recall/evaluation workflows without degrading conversational responsiveness.

Receipts

• snowcrab.dev related PRs: https://github.com/ironystock/snowcrab.dev/pull/9 ↗ (opens in a new tab) , https://github.com/ironystock/snowcrab.dev/pull/10 ↗ (opens in a new tab)
• implementation references: 93bc0ca, e14b71a, ee9b923

Implementation diff callout

Before → after (capture path reliability)

• Before: Prompt capture was synchronous and brittle in mixed runtime contexts.
• After: OAuth-backed mapping + queue/worker isolation made capture resilient and observable under real traffic.
• Receipts: https://github.com/ironystock/snowcrab.dev/pull/9 ↗ (opens in a new tab) , https://github.com/ironystock/snowcrab.dev/pull/10 ↗ (opens in a new tab) , 93bc0ca, e14b71a, ee9b923

Practical lessons

• OAuth and auth paths need explicit operational runbooks
• Runtime context mismatches should be assumed and instrumented early
• Queue boundaries are often the fastest way to regain chat responsiveness
• Reliability wins come from state tracking + measurable pilot metrics