Libra — Product North Star

Never start work you are unlikely to afford to finish.

Supporting invariant

Estimate the full Definition of Done before admission, reserve enough resources to finish, and re-plan only when the expected benefit of replanning exceeds its own cost, switching cost, and delay.

Every architectural and product decision in this repository is judged against this North Star and its supporting invariant. If a feature does not serve estimation, admission, resourcing, or disciplined replanning, it is out of scope for Libra.

What Libra is

Libra is a local Governor — a data-plane daemon that sits between agentic work and the LLM provider it talks to (currently integrated with agentic coding tools — Claude Code, Codex, etc.). It estimates the cost and time-to-complete of a task before admitting it, tracks spend against that estimate as the task runs, and makes replanning a deliberate, auditable decision rather than an implicit one.

Three Golden Journeys

1. Preflight worker

A developer stays inside Claude Code or Codex. They describe a real task. Before any tokens are spent, Libra plans the task and shows a probabilistic cost and time-to-complete estimate, with a confidence band. The developer decides whether to proceed — and then executes without ever switching to a separate app or dashboard.

2. Auto-mode power user

A power user configures a policy that prioritizes time over cost, enforces only a hard quality floor, and allows controlled cost elasticity within that policy. Normal replans — the kind the policy already anticipates — happen automatically, with no interruption. Only a material boundary crossing (a decision the policy did not already authorize) interrupts the user.

3. Engineering management

An engineering manager or team lead looks at Libra’s reporting and sees cost per successful outcome, on-budget/on-time completion rates, the split between rational and wasteful overruns, avoidable spend, and estimator calibration over time. They do not see raw token counts as the primary signal — token counts are an input to the metrics that matter, not the metric itself.

Non-goals

Libra is explicitly not:

• A generic token counter. Token counts are an input to cost estimation and ledger accounting, not the product.
• A generic LLM router or gateway. Libra’s optional gateway component exists solely to enforce hard provider-spend limits where traffic can be routed through it — it is not a general-purpose multi-provider router or a way to compare model quality/pricing across vendors.
• Another coding-agent chat UI. Claude Code, Codex, and other existing agent UIs remain the host user experience. Libra does not build or compete with a chat interface — it plans, estimates, meters, and gates the work happening inside those tools.