LIQUID BRAIN AI HARNESS

Description

Most “second brain for AI” setups are a folder of notes and a naming convention. Liquid Brain tracks work. It is a project-memory and work-management system for AI coding agents such as Claude Code and Codex, built on plain Markdown files and git. No service, no database, no account.

The three quiet failures it fixes

• Work ships and is never recorded.
• Decisions get re-argued because the reasoning scrolled away in a chat window.
• Tracking rots — cards sit “in progress” for months and nobody notices.

And the fourth one: an agent saying “tests pass” does not count. The harness runs the checks itself and keeps the results.

What is in the box

• Kanban cards with full history — BACKLOG → TODO → IN-PROGRESS → BLOCKED → VERIFY → DONE, plus DECISION-OWED, WON’T-DO and ON-HOLD. Cards are archived, never deleted.
• Generated views — board, manifest, timeline, artifact index, graph links and a single-file dashboard that needs no server. Malformed cards make generation refuse rather than produce a wrong board.
• Drift detection — cards sitting past a per-column threshold surface as STALE and escalate.
• A write lock with a queue — one writer at a time, others wait in order and are told where they stand. It heals itself.
• Evidence freshness — claims are hashed against the files they cite and marked current, stale, unverified or invalid. “Current” means the bytes still match, not that the claim is true.
• Bounded context — a task brief capped at a few thousand characters and a computed read scope per session, so the agent does not load the whole vault.
• Decision records, trust fields and reports — ADRs with rationale; every number carries a basis label (actual, estimate, projection); reports never add a guess to a signed contract.
• The harness — plan and run agent sessions gated by schema, tests, credentials, destructive-action, diff, scope and budget checks. It stops at VERIFY and never marks its own work DONE. It rejects sessions that changed nothing and logs tokens and cost per card.
• Safety hooks for Claude Code — a destructive-action guard (rm -rf, force-push, SQL DROP and friends are blocked unless dry-run or approved) and a scope fence that limits edits to what the card declares.
• Liquid Brain Desk (Windows) — one window with the board, card editing, Claude and Codex terminals tied to a card, scheduled agent tasks, graph and impact views and run history. Localhost only.

The rules it runs on

No step may depend on a human remembering to do it. A model cannot forget a rule it was never asked to remember. Strong models think, cheap models file, scripts do the rest. It is not magic: it is a set of scripts and a set of rules, enforced by something other than good intentions.

Honest limits

Locking across machines means taking turns. The read scope is a contract, not a sandbox. The guards are best-effort pattern matching. The harness works one card at a time. The Desk app is Windows-only and still a development candidate. Multi-user collaboration is not part of this release.

What you get

The Liquid Brain skill and vault tooling (Python 3.10+, standard library only, git), the harness, the Claude Code hooks, templates, a demo vault to learn from, and the Desk app for Windows. Pay what you can, from $0.