Five-harness development framework (v2, 27 Oct)¶

This page is the as-designed view of how Money Marche engineering works. It merges two pictures: - Atul's concept (Gemini image, 26 Oct). It is kept as a reference, not as the as-built state: concept - MMDEP-PM's as-built view (v1, 27 Oct).

The decision (TD-072): take Atul's five harnesses as the frame (Design, Code, Testing, Research, Work), with one Orchestration core underneath. Each component carries its real status from the tech-stack register (adopt / trial / hold), so the picture never claims more than what runs. The editable source is the Mermaid diagram below. It is a text file in the repository, so it is versioned, reviewable and drawn on GitHub and on the docs site. Image files are for concepts only.

flowchart LR
  A([Atul and Ajay: goals, business answers])
  subgraph O[Orchestration core]
    O1[MMDEP Actions: for-atul issues and panel]
    O2[Routing and allocation: who does what, escalation ladder]
    O3[Execution: laptop runner now, server from S3]
    O4[Access isolation: data classes, keys per environment]
  end
  subgraph D[1 · Design harness: pre-flight]
    D1[Claude Design, ChatGPT projects, MMDEP Assistant] --> D2[Overall, phase and sprint designs; LLD with V/S/C/D/L]
    D2 --> D3[Ready issues: routing block and scope paths]
  end
  subgraph C[2 · Code harness: unattended]
    C1[Aider with free/cheap models, Ollama for client data] --> C2[Change limited to scope paths]
  end
  subgraph T[3 · Testing harness: quality gates]
    T1[CI: pytest with coverage, ruff, bandit, pip-audit, design check]
    T2[Live checks: API, NAV duplicates and gaps, staging parity, SEO crawl]
  end
  subgraph R[4 · Research harness: knowledge]
    R1[Knowledge extraction from chats] --> R3[(docs/knowledge)]
    R2[GitLab snapshots, AMFI/AMC sources] --> R3
  end
  subgraph W[5 · Work harness: operations]
    W1[Business flows: renewal cadence, lead routing]
    W2[Systems of record: CRM, website, accounting]
  end
  A <--> O1
  D3 --> C1
  C2 --> T1
  T1 -->|green| M[(Repositories, merge by risk)]
  T1 -->|red twice| O2
  M --> T2
  T2 --> D1
  R3 --> D1
  M -->|release with Atul's approval| W
  W --> T2
  D2 -->|business question| O1
  O --> D & C & T & R & W

What each harness holds: Atul's concept against what we run¶

Status: adopt = in use or decided; trial = being tried this sprint or next; hold = not now, with the reason; legal = legal check before build (TD-030).

Harness	Component (concept)	Status	Notes
Design	Claude, ChatGPT projects	adopt	Claude Design and MMDEP-PM; ChatGPT project agents; MMDEP Assistant for low/medium-risk design
Design	Deterministic state machines	adopt (practice)	Every workflow with states (lead, renewal, NAV ingest run) gets a Mermaid `stateDiagram` in its LLD
Design	DBML schema (dbdiagram.io)	hold	Mermaid `erDiagram` in the LLD does the same inside the repository, at ₹0, with no extra tool
Design	OpenAPI spec	adopt	The ingest-routes LLD already writes the routes into the OpenAPI file
Design	Excalidraw	trial	For rough sketches in discussion. Anything that becomes a design is redrawn in Mermaid
Design	Penpot	hold	UI design tool; revisit when a UI designer joins. WPDev works from the live site
Code	Aider + Ollama	adopt	Aider runs in the code harness. Ollama is on the LAN, firewalled, for client-data tasks
Code	LiteLLM proxy	trial (S3)	Goes on the server with SRV-1; one place for model keys and limits (B42)
Code	Repo slicing gateway	adopt (as built)	This is our routing block `scope.paths` plus the context pack: the model sees only what the issue allows
Code	Python, JavaScript	adopt	Plus PHP for WordPress (WPDev)
Testing	GitLab runner (host network)	adopt	Laptop runner; tests run where the code lives
Testing	ruff, pytest	adopt
Testing	`pytest --cov`, bandit, pip-audit	adopt (add)	₹0. Added to the CI templates as a backlog item; reported first, not blocking, for 2 weeks
Testing	SEOnaut crawler	trial	WPDev's SEO report already uses a crawler; Lighthouse CI on the laptop runner
Testing	(missing in concept) live production checks	adopt	The collector: API routes, NAV duplicates and gaps, staging status
Research	"Air-gapped" knowledge bus	adopt, renamed	We aren't air-gapped (we use cloud models). What exists: knowledge extraction, GitLab snapshots, `docs/knowledge`. Client data stays local by data class
Research	Local vector DB (Chroma/Qdrant)	hold	When RAG is needed (medium horizon), pgvector in the server's Postgres: one database, not two
Research	RAG indexing	hold (medium horizon)	Facts for content and support; after the server
Research	Bhashini (Indian languages)	hold, legal	Useful for regional-language content and client messages; needs a use case and a terms check
Research	DPI open-banking feeds, Account Aggregator	hold, legal	Client financial data (data class `client`). Joining the AA network as a data user (FIU) needs a regulated status and consent flows: legal opinion first
Work	n8n workflow bus	trial (S3)	On the server; flows exported to git. Until then, Python jobs on the laptop
Work	Zoho CRM	adopt (current)	Connected; MCRM is the CRM project
Work	SuiteCRM	business decision	A second CRM would conflict with Zoho/MCRM. Which is the CRM of record? (raised to Atul)
Work	Strapi headless CMS	hold	The website is WordPress; no second CMS without a need
Work	ERPNext (accounting, rewards)	hold	No accounting project yet; a business decision when rewards or commission accounting starts
Core	Orchestration core, execution engine	adopt	MMDEP: routing, allocation, Actions, runner on the laptop, server from S3
Core	Access isolation gateway	adopt (as built)	Data classes (public/indicative/internal/client), keys per environment (a staging key is refused on production), segregation (DR-008)
Core	India DPI connections (UPI, AA)	hold, legal	Not in the current phase; see Research

Corrections to the concept image: the image text has errors that came from the image generator ("Eirmenprent and interrspace", "Brower session", the garbled subtitle). The Mermaid diagram above is the version to use.

Where it is weak today, and the fix in hand¶

Orchestration core: inbox → panel copy happens only when Atul messages MMDEP-PM. Proposed fix: a 4-hourly panel sync (panel item SYNC-1).
Code, testing and research harnesses depend on the laptop being on. SRV-1 (the server) removes that.
Design → code: the hand-off is manual for ChatGPT LLDs. Fix: B36, LLD → ready issues.
Testing: no coverage or security scans in CI yet. Fix: add pytest coverage, bandit and pip-audit (TD-072).
Work harness: nothing automated yet; n8n arrives with the server. This is where revenue flows (renewals, leads) live, so it comes first after the server.