Skip to content

Phone set, parallel harness work and actions

1. Phone set: at most three chats on the go

The ChatGPT phone app shows one account at a time, so on the go Atul sends PROCEED to three chats only:

Phone chat Why
API (Lane P, account A) Business-critical, and the API's high-risk work needs him most
MMDEP Coding — Sonnet (Claude app) Builds the platform that removes work from everyone else
MMDEP Assistant (account A) Digest, actions and the fallback for Claude

Everything else (MWMC, MCRM, PE, WPDev, API Lane S on its new account) runs in attended laptop sessions (flow A), or through the local harness (flow T) from issues. Their queues live in their repositories, so nothing is lost between sessions.

2. What the local harness runs in parallel (flow T, free models)

# Work Source Starts
1 MMDEP cx 1–2 items GitHub MMDEP Now (batch 2: #47–#51 done 19 Oct)
2 AMC factsheet parsers (§3) New GitHub repo moneymarche-dev/mm-factsheet-parsers When OWP-1b is approved
3 API staging cx 1–2 items (lane:staging, risk: low) GitLab moneymarche/api After #56 (harness on GitLab)
4 MM-Cron cx 1–2 items: AMFI factsheet-link list job, NAV duplicate/gap report script GitLab moneymarche/mm-crons After #56
5 Actions sync after every collector run (§4) MMDEP After #57

Capacity: the harness runs hourly, two items a run, on Gemini's free tier first (spend this month: $0.18). Its gates (scope, tests, lint, exact strings, risk low) keep it safe without review before merge (DR-012).

3. AMC factsheet parsers (DR-009 OWP, made concrete)

  • Repository: a new satellite repo moneymarche-dev/mm-factsheet-parsers. Public data only, trust level T0, no estate code.
  • First job (built by the harness): read AMFI's monthly factsheet link list and download the PDFs for the top 5 AMCs by our clients' holdings for one month. They go into samples/<amc>/<yyyy-mm>.pdf; they are public documents.
  • One issue per AMC: the harness writes parsers/<amc>.py. Output is fixed JSON: TER, AUM, inception date, benchmark, fund manager, holdings, and ratios as reported. Acceptance tests check that schema plus spot values transcribed by hand from the sample PDF (the spot values live in the issue, as exact strings).
  • Months of test data: once a parser passes, a registered job runs it over the past 12 months of PDFs and commits JSON to data/. MM-Cron (API lanes) later loads it through the API.
  • Free agents (a ChatGPT Go account, Gemini chat) can take the same issues when the harness fails twice. Their output passes the same tests (DR-009 intake).

4. Atul's actions without Claude (ACT-1)

Today: - The Actions panel is a Claude page, and only a working Claude check-in can write it. The check-in hasn't written anything since 18 Oct. - The Assistant writes panel proposals only in ONLINE MODE. - Result: the panel goes stale when Claude is unavailable.

Proposal: actions live in the repository (rules §19), and the local harness processes them after every collector run: 1. Sources, collected by script: open question-atul issues in every repository (the collector lists them already), plus the Assistant's newest panel-updates file on branch reports. 2. Output: reports branch actions/ACTIONS.md, a phone-friendly list. Each item has an ID, a one-line question, options, a recommendation, the age, and a link. 3. Atul answers from his phone in one GitHub issue, Atul answers (MMDEP), as comments like PE329 B or all recommended. The GitHub app works for this. 4. After each collector run, mmdep actions sync runs: - Gemini Flash-Lite turns the comments into decisions; - each decision becomes a row in docs/relay/answers.md and a comment on the source issue; - ACTIONS.md is refreshed. - Ambiguous replies stay open and are flagged. 5. The Claude panel becomes a mirror, updated by the check-in when it works.

5. Decisions

  • ACT-1: move Atul's actions to the repository as in §4. Sonnet builds mmdep actions sync (#57). The panel becomes a mirror.
  • OWP-1b: create moneymarche-dev/mm-factsheet-parsers, and let the harness build the AMFI link job and the top-5 AMC parsers.