memories/claude-tag-pilot-6cf1d324.md

memory

A multiplayer Slack integration where tagging Claude in a channel provides a shared session for all participants, rather than per-user contexts. It leverages ephemeral sandboxes tied to dedicated service accounts, currently piloted by the Zive application platform and AI co-pilot infrastructure teams.

Core Mechanics

  • Session Scope: Tied to individual Slack threads. No cross-thread session sharing.
  • Access Model: 1:1:1 (one service account : one access bundle : one channel) for auditability.
  • Authentication: Uses Anthropic enterprise workspace roles; users must belong to the 'claude tag user' group to trigger responses.
  • Credential Handling: Proxied; secrets are masked before hitting downstream services to prevent exposure through terminal/curl usage.

GitLab Integration

  • Requires custom header PRIVATE-TOKEN with raw token value (no Bearer prefix).

Governance

  • Passive monitoring is disabled; only explicit @-tags trigger activity.
  • No self-approval on protected branches (GitLab-enforced).
  • Audit trail records routines, memory, and egress, but currently lacks detailed logging for Git/MCP traffic.
  • Token rotation is managed by AI Ops (Next due: Sept 28, 2026).

Pilot Constraints

  • No web/link egress access; cannot read Slack canvases or external Google Docs.
  • Limited to two teams initially; metrics on efficiency, quality, and governance are tracked bi-weekly.

Pilot Status (as of Aug 31, 2026)

  • Live Status: Deployed to both pilot channels; monitored by senior leadership.
  • Operations: AI Ops service account requested via Jira (pending).
  • Key Connectors: Identified as priority targets: Drive, GitLab, Jira, Confluence, and Vercel (the latter specifically for monitoring citizen-dev deployments).