SysEngineering

SCSP AI Initiative: handoff

Operational state, so work continues on any machine (work laptop, home PC) without re-deriving context. Update the State and Next sections and append to the Log at the end of every working session.

Resuming on another machine

  1. git pull in sysengscripts. This file, CLAUDE.md and the root CLAUDE.md section “SCSP AI Initiative” are the durable context. Claude’s local memory does not travel between machines.
  2. Confluence is the shared record: IPO → 3.M SCSP Initiative. Claude can read and update it through the Atlassian connector as aschembri@hillspire.com; page IDs are in CLAUDE.md.
  3. Jira: the SCSP AI Initiative board in ITPROJ, built on epic ITPROJ-1401 (see Jira below).
  4. The personal prototype lives in cos_lab/, inside this folder. Its README, CLAUDE.md and HANDOFF.md carry its state.

State (2026-10-15)

Phase: discovery. No SCSP code or infrastructure exists yet. A personal prototype of the SCSP connector is under way in cos_lab/.

Jira

The prototype: cos_lab (ITPROJ-1421)

Adam’s own chief of staff, built the way SCSP plans to build one for everyone. Claude connects to one remote MCP connector on Cloud Run with Google sign-in, which reads Adam’s own GitHub, Jira and Confluence, Calendar, Gmail and Okta, all read-only. It proves the SCSP design on one person: the sign-in, per-person tools, permissions with two of Adam’s identities, Gmail by consent, and the record of use.

Now: Google sign-in is configured, and the connector is built and reviewed (2026-10-15), on branch feat/cos-lab-connector, in its own PR since #193 merged. It works with Claude Code only for now; claude.ai waits for step 7. Then: run it against real Google on Adam’s machine, and deploy it to Cloud Run. Details, the branch order and the session plan are in cos_lab/HANDOFF.md.

What exists that the build will reuse

Thing Where Why it matters
Gwen’s Claude connector: OAuth 2.1 + DCR, stateless HMAC tokens, allowlist integrations/gwen_grant_memo_agent/mcp_oauth.py, mcp_connector.py, test_mcp.py (on main since PR #181) The SCSP connector is this pattern with Google sign-in replacing the access code and the permissions model replacing the allowlist. Gwen’s HANDOFF, “Claude connector”, has the operational lessons.
cos_lab cos_lab/ in this folder The prototype: hand-written Claude loop, read-only tools, a local MCP server, and the remote connector (lab/connector/): its own OAuth server with a consent page and Google sign-in, sessions, tools per person, a record of use. 164 tests. The SCSP connector’s first draft should start from it.
Harry’s deploy pipeline .github/workflows/harry_deploy.yml The staged deploy (no-traffic revision → health check → promote) proposed for SCSP
SCSP’s Workspace tenant route scsp-prod, customer 101475582153570045240 Already in the helpdesk module’s routes

Facts learned (keep)

From building the prototype’s connector (2026-10-15), and binding on SCSP’s:

Open decisions (full list: Confluence 3.M.4)

  1. Which Google Cloud organization and project host SCSP Intelligence; who administers it.
  2. Google or Okta sign-in for SCSP staff; do tier and base groups exist.
  3. Which AI products SCSP licenses (Claude Team/Enterprise, ChatGPT) and who the org Owner is.
  4. What data SCSP approves for AI processing.
  5. Where SCSP’s own code lives: this repo or an SCSP repo. (The personal prototype is settled: it lives here, in cos_lab/.)
  6. Confluence access for SCSP staff: IPO is a restricted space, so adding them to it shows them all of IPO. Options are a dedicated space or page restrictions. Decide before inviting anyone.

Next

Log