Readiness plan for Atul's absence (16–30 Oct 2026)¶
Goal. For two weeks, work keeps moving without Atul. Nothing reaches main, production or a release without him, and he returns to a short, ordered list of decisions, not a pile of unreviewed work.
Principle. While he's away, only work that is read-only or advisory runs unattended: independent reviews, digests and evidence. Production work by agents (PRs) continues only under a WIP limit, because nobody merges for two weeks.
1. Before 16 Oct: decisions and small actions for Atul (about 1 hour in total)¶
| # | Action | Time |
|---|---|---|
| A1 | Run release #80; close #77 and #74 as "Superseded by #80." | 15 min |
| A2 | Review and merge PR #2 (documentation) | 15 min |
| A3 | Gemini: enable billing on the API project with a small prepayment, so prompts fall under paid terms (not used for training) | 5 min |
| A4 | Approve the common label scheme (below), so the reviewer chat can create the labels in the three GitHub repos | 5 min |
| A5 | Approve the Programme Manager role transfer item (Claude Code raises it) | 5 min |
| A6 | Choose what runs while away (§2) and approve the scheduled tasks | 10 min |
| A6b | The PC is a laptop and travels with Atul (10 Oct). He checks it twice a day; in the daytime it stays open in the hotel room. Settings: sleep and hibernate off when plugged in; closing the lid does nothing when plugged in; Windows updates paused for 3 weeks; screen locked (Win+L), not signed out; device encryption (BitLocker) on and a strong PIN, because it holds sign-ins to GitHub, GitLab, AWS (read only) and Google; the Claude desktop app and Chrome left open. A test with the screen locked before leaving | 15 min |
| A7 | Name an emergency contact for production (website, API, MM-Cron), for example the legacy developer, with what he may do | 5 min |
2. While away: what runs (proposal)¶
| Process | Who | How | Output | Needs Atul's PC? |
|---|---|---|---|---|
| Independent review of MCRM documentation PRs (the current backlog) | Claude, in a scheduled cloud session (a different vendor from the author, ChatGPT) | Weekdays, up to 3 PRs per run, oldest first, labelled state:review + role:review; Atul's standing question per PR (to be written by Atul or taken from the PR) |
A review report comment on each PR answering the question, with a recommendation; label moves to state:needs-atul |
No |
| Programme digest | Claude, in a scheduled cloud session | Mondays and Thursdays: reads the GitHub repos (MMDEP, Portfolio Engine, MCRM) | One comment on a pinned "Status digest" issue in MMDEP: what moved, what is blocked, waiting for Atul, review debt. GitHub notifies Atul's phone | No |
| Project work by ChatGPT (Portfolio Engine, MCRM) | ChatGPT, as now | Continues under a WIP limit: at most 5 open PRs per project; beyond that, only reviews and evidence | PRs labelled for review | No |
| MMDEP development (Claude Code) | Paused | Nothing merges for two weeks; unfinished branches wait | — | — |
| Gemini audit controller (PE pilot) | Paused | Needs the desktop and the fixes in docs/research/pe-gemini-pilot-review.md |
— | Yes |
| GitLab reading (governance, website, API, MM-Cron) | Reviewer chat, best effort | Through Claude in Chrome on Atul's PC (kept on) | Included in the digest when reachable | Yes |
Not done while Atul is away: merges, releases, deployments, MMDEP item approvals, anything on production, new spend.
3. Build before 16 Oct (small; only if A6 approves it)¶
- Label scheme in the GitHub repos (reviewer chat, 11–12 Oct):
state:*,role:*,type:*,work-request,cross-project; Portfolio Engine keepsbatch-*for its controller. - Review task prompt (reviewer chat, 12–13 Oct): a pointer prompt (STD-AI-CONT-001 §4) for the scheduled review session: what to read, the report format (question, answer, findings, recommendation, evidence, trailers), limits (3 PRs, no approval or merge, comment and label only).
- Dry run (13–14 Oct): one scheduled run on two MCRM PRs, which Atul checks before leaving.
- Digest task (14 Oct): the first digest posted; Atul checks it on his phone.
4. On return (31 Oct)¶
Atul starts from the digest issue: PRs reviewed and waiting for him (with recommendations), blocked items, and the decisions list. Then the working-arrangement discussion continues with two weeks of evidence on review quality and the WIP limit.
5. Risks¶
- Claude usage: scheduled sessions draw on Atul's Claude plan. Each run is capped (3 PRs; the digest only reads).
- Quality: reviews are advisory. Atul still decides each PR.
- Agent identity: comments appear as
moneymarche-dev; each report starts with a header naming the agent and role, and ends with trailers. - The laptop travels: hotel Wi-Fi and moves mean gaps. The cloud processes (reviews, digest) don't depend on it. If the laptop is lost, revoke its GitHub, GitLab and Google sessions from another device.
- GitLab is covered only while the PC is awake and Chrome is open. The reviewer chat can't wake a sleeping PC, and after a restart (an update or a power cut) nothing runs until Atul signs in again.