Skip to content

Checkpoint: MMDEP-PM (reviewer chat), 10 Oct 2026

This is the current-state record for the reviewer chat, kept in the repository so that the chat is never the only copy (STD-AI-CONT-001 v1.1 §5). Newer repository files win over this note. The MMDEP inbox wins on item status.

1. Who does what

Role Held by Notes
Spec owner, approver Atul Approves items, merges PRs, runs releases
MMDEP-PM: design advisor, independent reviewer Claude (chat) Documents and issues only (Atul, 10 Oct). No code, no main, no approvals
MMDEP-DEV: architect, developer Claude Code (Sonnet), run by Atul on the desktop Rules in CLAUDE.md; one session per story or iteration
Test Runner MMDEP, plain code Under mmdep-agent

2. How the reviewer chat works

  • For each Claude Code report: which items to approve, reject, answer or close (exact commands for the runtime C:\Projects\ai-sandbox\mmdep_runtime\Scripts\mmdep.exe), then the next Claude Code prompt.
  • Documents go on a docs/<date>-<slug> branch and reach main only through a PR that Atul merges.
  • Issues: may create, label and comment on issues in repositories attached to the chat.
  • Every commit carries the trailers Agent:, Role:, Seat:, Work-Item: (STD-AI-CONT-001 v1.1 §8), because commits and issues appear under Atul's GitHub account moneymarche-dev.

3. What the reviewer chat can and cannot reach (tested 10 Oct)

Works Does not work
GitHub REST (gh api): read repos, create, label, comment on and close issues GitHub GraphQL (so gh issue list / gh pr list fail; use gh api repos/...)
git push of new branches from its own clone Contents and refs write APIs (blocked by the proxy)
Claude project docs, Drive and Notion connectors The MMDEP MCP server and hub (they run on the desktop); item status comes from docs/status/inbox.md or pasted mmdep show output
Scheduled cloud sessions (with Atul's approval each time) Messaging Atul's Remote Control or other local sessions

Repositories attached so far: moneymarche-dev/MMDEP. MCRM and Portfolio Engine need to be added to the chat before use.

4. Decisions since the 3 Oct handover (pointers, not copies)

  • Spec 0.10 (#59) and 0.11 (#71): golden rules, objectives, §5C staffing rule, §15B online phases, §17.0 readiness gate.
  • docs/decisions/DR-002-agent-platforms-and-automation.md: forge as the bus, MMDEP as the conductor.
  • docs/decisions/DR-003-controller-hosting.md: hybrid now; a 4 GB Lightsail VM pre-approved in principle behind a trigger.
  • docs/decisions/gitlab-drafts/AI-Engineering-Continuity-and-Handover-v1.1-amendment.md: role transfer vs delegation, Work Request and Work Result, pointer prompts, session types and names (§5.1, added 10 Oct), context tiers, seats, trailers.
  • Session naming <project>-<role>-i<NN> (Atul, 10 Oct; amendment §5.1).
  • Work-request loop: proposal only, see docs/research/claude-code-sessions-and-work-request-loop.md.
  • Current design and state, experiments and proposed 0.12 changes: docs/design/MMDEP-current-design.md.

5. Open and pending (as of 10 Oct; check the inbox)

What Who Next step
#80 release main at e412297 Atul Run it; close with done and the actual values. Then reject #77 and #74 with "Superseded by #80."
#81 HO-1 design (Work Request/Result records) Reviewer Needs mmdep show 81 output or the design on a branch
#70 DR-002 stage 1 sandbox setup Atul After #80
#72 Lightsail and EC2 snapshot clean-up Atul Delete the June and 27 Aug snapshots and EC2 snap-0d43b7c74288eef19 (after its AMI); recommend deleting the Nov 2024 WordPress-1 snapshot; keep the automatic ones and 25 Sep
GitLab MR for the handover amendment and ADR-004–006 drafts Atul Submit; then git -C C:\Projects\MoneyMarche pull --ff-only
Test branch docs/connector-test Atul Delete when convenient
Work-request loop (LOOP-1) Atul Decide after the working-arrangement discussion
GitLab projects: a way for the reviewer chat to read them Atul and reviewer Open (Atul, 10 Oct)
Gemini free tier used with private PE content Atul Decide: paid key, public or synthetic only, or another engine (docs/research/pe-gemini-pilot-review.md §8)
Write access for the reviewer chat Atul Deferred; suggested GitHub app permissions: Issues read/write, Contents read-only

6. Do not repeat

  • Never close a manual item with a pre-written template; use placeholders Atul fills with actual values (lesson from #55).
  • manual items close with done, not approve.
  • Answers with <placeholder> text are refused by HU-1; text inside backticks is allowed.
  • Don't create scheduled tasks or message other sessions without Atul asking for it in that turn.

7. Next when resuming

Atul wants to discuss a few more points and then set up the working arrangement afresh (10 Oct). Start there; update §1–§2 of this file and PROJECT.md when it is agreed.

8. Decisions and direction from Atul (10 Oct, evening)

  • Role transfer: Atul wants the MMDEP Programme Manager role moved to the reviewer chat (STD-AI-CONT-001 kind A). Steps:
  • a handover from the current holder (DR-002 names ChatGPT tasks for triage);
  • a role-map change by an approve item (raised by Claude Code);
  • this checkpoint then becomes the PM's current-state record.

GitLab projects are a gap. Bridges: ChatGPT agents assist through GitLab, and the reviewer chat reads the local clones under C:\Projects through the device bridge (read only; as current as Atul's last pull). - First priority project: the Money Marche website, with dependencies on other projects to resolve first. Website and MCRM are the revenue path that funds hardware, tools and AI. - Gemini API key: personal Google AI Studio key, free tier. Atul has looked at the paid option (prepaid). Per Google's billing page, prompts are treated under paid terms, and not used to improve Google products, once at least one API project has billing enabled. - Gemini: meant for terminal work. In a manual trial, Gemini CLI did Portfolio Engine shell work well. A second Gemini session turned ChatGPT's unbounded prompts into safe bounded tasks, and that made the relay work. Proposed role: task shaper and terminal agent (spec §8 task shaping), taking over what was planned for local Ollama models. Data class still applies (free tier: public and synthetic only). - Antigravity: since 23 Sep 2026 the Antigravity SDK can run open models locally (Google's reference: Gemma 4 26B, needing more than 24 GB of GPU or unified memory). Fully local runs need no cloud connection; hybrid runs send some data to a cloud Gemini planner. It runs open models only, not Claude, GPT or Gemini. To evaluate when local hardware exists (the desktop's memory already paused local models). - Copilot: not kept (Atul, 10 Oct). A subscription only as a backup isn't worth it; the same money goes to Claude, ChatGPT or Gemini. - Local access: the desktop app links this chat to the folders C:\Projects\MoneyMarche, C:\Projects\ai-sandbox\mmdep and C:\Projects\testcop\pe_gemini_pilot_workspace. There is no shell on the computer. - Claude in Chrome (tested 10 Oct): connected to Atul's Chrome with his sign-ins. It reads GitLab pages (work items, projects) read-only, so the reviewer chat can see GitLab while Atul's PC and Chrome are on. Projects visible to Atul's GitLab account: - Money Marche Digital Engineering Program (29 open MRs, 20 open items); - money-marche-website-staging; - Money-Marche / Money Marche Website; - moneymarche/api; - moneymarche/mm-crons; - moneymarche/webapp; - moneymarche/webapp-sakai; - moneymarche/wordpress_plugins; - Money Marche Testing / Money Marche Sandbox.

Writing on GitLab through the browser is possible but would act as Atul. It needs explicit permission per action. - Vacation 16–30 Oct: readiness plan in docs/handovers/2026-10-16-vacation-readiness-plan.md (proposal). - WPDev and testing (Atul, 10 Oct): WPDev (the second-generation website) tests both production and WPDev sites through Opera with ChatGPT; ChatGPT drove the Opera Browser Connector (77 actions on 10 Oct) until the connector began asking for an update. A testing repository with Selenium also exists; production uses the same Selenium version. - Capability questions raised (10 Oct), answered in chat; to be confirmed against the governance repo's TEA tasks: - Can TEA audits be recreated with terminal tools and models? Yes, as deterministic collectors (read-only AWS profile, read-only SSH user with allowlisted commands), with models only writing scripts and summarising evidence. - Can TEA do manual deployments during forge outages, with traceability? A person deploys from a local mirror at a recorded commit, with preflight and postflight evidence and a deployment record, reconciled to the forge when it returns (spec §15.6–15.7: agents never deploy). - Portfolio Engine needs tests and Python reports run: on the desktop (Claude Code or Gemini CLI, or MMDEP's Test Runner), because cloud sessions can't reach the MoneyMarche API. - Relaying bounded prompts to ChatGPT in Chrome: possible, one message at a time, each with Atul's explicit approval. - Next (Atul): evaluate the MMDEP plans first; then go through the projects and the blockers against each one's working style.

9. Session summary, 10–11 Oct (for the next session)

Written this session (all on docs/2026-10-10-sync, PR #2): - Current design and state: docs/design/MMDEP-current-design.md. - Plan review step 1 (objectives, phases, risks, rules, repository protection, folders): docs/design/MMDEP-plan-review-2026-10.md. - MMDEP Cloud Layer (MCL) design: docs/design/MMDEP-cloud-layer-design.md. - Capacity planning engine design: docs/design/MMDEP-capacity-engine-design.md. - Cloud terminal and hosting estate proposal (tools box, runner controls, headless Chrome, server browser, trial plan): docs/design/cloud-terminal-and-estate-proposal.md. - MM-Cron agent augmentation: docs/design/mm-crons-agent-augmentation.md. - Research notes: agent modes and Lightsail; repositories, accounts and quotas; AI vendor data terms; PE Gemini pilot review; local folder inventory; Claude Code sessions and the loop. - Assumptions register (A-01 to A-34): docs/decisions/assumptions-register.md. - Atul's action tracker: docs/status/atul-actions.md. Ajay meeting brief: docs/handovers/2026-10-ajay-meeting-brief.md. Vacation plan: docs/handovers/2026-10-16-vacation-readiness-plan.md.

Decided by Atul: Copilot dropped; GitLab for code and GitHub for work (to confirm); company-owned organisations and groups; repositories per project; decide the server early; test first on a new AWS free-plan account (12 Oct); Portfolio Engine's dev deployment moves to AWS after the test; keep Bluehost (≈ 2.75 years paid; also company email); staging check by test copy; browser-on-server decision parked until the test; budgets at the end of the discussion.

Still to discuss: part 3 (browser and terminal on Atul's machine); the projects, verticals and their priorities, and blockers per working style; MCL decisions (weights, Claude reserve, dashboard); final budgets.

10. Project facts from Atul (11 Oct, 15:24) and low-usage mode

  • GitLab group moneymarche (legacy, managed by Ajay and a developer until about a year ago; Atul took it over in Sept 2026) holds the production repositories.
  • webapp-sakai, branch dashboard, is the current production website, including the customer dashboard. The dashboard calls SuiteCRM through the MoneyMarche API; the CRM runs at crm.moneymarche.com. One agent has worked there for the past few days.
  • wordpress_plugins: the plugins in use. api and mm-crons: the APIs and related scripts; the website depends on both.
  • webapp: old, apparently unused (a retirement candidate; the governance Portfolio Register still lists it as active, to correct).
  • Website staging: an old implementation with features never released to production.
  • Low-usage mode (Atul, 11 Oct, weekly limit near): only work that supports the projects; short replies; no web research unless asked; read existing summaries before raw sources; one commit per session; heavy reading delegated to ChatGPT or Gemini with results in repositories.

11. Portfolio facts from Atul (11 Oct, 15:55)

Project Repo Hosting Facts
Production website GitLab moneymarche/webapp-sakai, branch dashboard moneymarche.com Serves public pages. Dashboard pages exist but have no users and were never tested; preserve them. Streams by label: Maintenance, SEO, API, MCRM, WPDev. Agent: ChatGPT project "Money Marche Website Maintenance Convergence". SEO work needed
MMCLONE GitLab money-marche-testing/money-marche-sandbox (tests and the differences from production, to be documented there) mmclone.arevalifestyle.com (Bluehost) Exact replica of production; every change is tested here before production. No schema, configuration or install script existed; knowledge came from Ajay's notes, TEA audits and agents' experience
Staging (old) GitLab money-marche-group/money-marche-website-staging (extracted from the deployment by TEA) staging.moneymarche.com Old implementation with unreleased features; its database has a richer catalogue that was never moved to production (coverage and depth)
Governance / MMDEP standards GitLab money-marche-group/money-marche-digital-engineering-program — Now owned by the PM (reviewer chat), delegating work to other agents. Writes reach GitLab only through Atul or delegated agents
MCRM GitHub moneymarche-dev/MCRM — Owns the new dashboard framework, calculators, the catalogue (delinking categories and products from the WordPress database; served by the API at api.moneymarche.com), Mautic and user forms. Its audit may have found further items
WPDev GitLab money-marche/money-marche-website wpdev.moneymarche.com Parallel next-generation site: everything production does plus more, except MCRM's areas. Waits for MCRM to refactor production and give details. Goal: push low-risk code to production when production signals, then bulk or full replacement
API, MM-Cron, plugins GitLab moneymarche/api, moneymarche/mm-crons, moneymarche/wordpress_plugins AWS production server The website depends on API and MM-Cron
Portfolio Engine GitHub moneymarche-dev/portfolio-engine Bluehost (dev), moving to AWS after the test Important; marketing edge
Recommendation engine (future) — — Combines insurance, investments and finance to recommend solutions to customer problems
webapp GitLab moneymarche/webapp — Old, unused

Usage note (Atul): 26% of the weekly Claude limit used by 11 Oct with no coding; work in low-usage mode (§10).