QuietHarness system architecture

中文 · English

This document describes Leo's current AI work system. It is not a universal architecture invented from a blank page.

The public version preserves the business topology, task lifecycle, owner/worker split, readout/writeback boundaries, single-writer rule, and delivery receipts. It removes private paths, thread IDs, live tasks, accounts, positions, health records, and production controls.

How it evolved

Claude Code Workflow v1/v2 accumulated always-loaded rules, skill routing, memory flush, model routing, Morning, Today, Session End, hooks, and completion gates. These mechanisms solved real model limitations at the time.

As model capability improved, loading every mechanism into every request began to create repeated planning and configuration maintenance. v3 does not erase the system; it changes where each part lives and when it loads.

Current map

Leo's real work
        │
        ▼
Claude Code · Codex · Cursor          replaceable clients and models
        │
        ▼
QuietHarness Core                     work · evidence · verify · hard gates
        │
        ├── Business Lines            stable workspaces and owners
        ├── Task Continuity           per-task SSOT and receipts
        ├── Background Systems        reports, sync, monitors
        └── Domain Workflows          PM, data, content, video, health

Complexity remains in the system boundaries instead of occupying every prompt.

Three clients, one behavior boundary

Codex reads the shared 1,604-byte AGENTS.md; Claude Code imports it through a thin entrypoint; Cursor mirrors the same reliability boundaries in a project .mdc rule. Those boundaries keep direct execution, truthful evidence, proportional verification, and confirmation before irreversible actions.

The client, model, and conversation may change without changing ownership or task truth.

Nine stable business lines

Leo's current topology is public because it is part of the design:

Key	Line	Responsibility
personal_ops	Personal Ops	system, agents, gates, recovery
pm_strategy	Strategy Lab	strategy, validation, scale/kill
pm_intel	Data Platform	pipelines, indexes, profiles, quality
portfolio	Portfolio	allocation, risk, execution decisions
content_writing	Content Studio	X, articles, SEO, media assets
product_distribution	Products & Growth	products, monetization, distribution
health_ops	Health Ops	records, cadence, health research
daily_rhythm	Daily Rhythm	intake, routing, attention audit
research_dd	Research Desk	research, opportunities, knowledge

Users may rename, merge, or remove these lines. Their value is the pattern: conversations are replaceable entrypoints, while responsibility and workspace ownership remain stable.

Startup loads a bounded summary for the current line only. An exact task record is read only when the task is named.

Per-task truth

Each durable task has one canonical record. Dashboards and aggregate lists are derived views.

• status is the lifecycle authority.
• owner_thread is stable ownership.
• primary_worker and claimed_by describe the current batch.
• next_action, artifacts, and updated_at remain explicit.
• state_revision lets verified new facts supersede old narrative.
• A worker claim is not durable until an artifact, verification, and owner writeback exist.

readout → source/freshness/claim → work → verify
        → artifact/receipt → owner writeback → read again

See Task continuity.

One writer per mutable surface

Agents may research and review in parallel. Writes to the same repository or task surface are serialized by default.

Acquisition records the repository, worktree, writer, allowed paths, validation, and next gate. Release records the artifact, validation, writeback, rollback, outcome, and remaining gate.

This contract came from real cross-model write collisions.

Background systems

Reports, sync jobs, and monitors run independently. Success leaves a quiet artifact; missing, stale, or failed output raises an exception. An unrelated request does not need a Morning or closeout pipeline.

Open-source branches

Parts of the private system became independently useful repositories:

• prediction and research: polymarket-toolkit, asset-dd-and-opportunity-evaluation;
• content intake: x-reader, tg-reader-mcp, long-media-cli;
• video: claude-video-kit;
• health: ai-health-vault;
• system audit: claude-skill-audit.

QuietHarness shows how these branches connect; it does not duplicate their implementations.

Public and private boundary

Public: business-line structure, task contracts, owner/worker/writer rules, sanitized incidents, and the open-source map.

Private: thread UUIDs, absolute machine paths, live tasks, accounts, credentials, positions, health data, customers, schedulers, VPS, and production state.

Abstraction removes storage and private coupling. It does not remove the system's history or design choices.