Skip to project
Qingping Champ
All projects
Agent RuntimeIndependent2026

Napier

A local-first work ledger for inspectable, recoverable agent runs

How can multi-step agent work resume from durable evidence instead of chat history?

An ordered Napier event ledger feeding conversation, terminal, and evidence projections

Project context

Napier treats an ordered evidence ledger, not chat history, as system state. Messages, model calls, tool outcomes, goals, branches, artifacts, and runtime decisions are recorded before they are presented, so recorded runs can be inspected, resumed, branched, reconstructed, compared, and evaluated.

SYSTEM PATH

Evidence before interface

  1. 01RunEvent contracts
  2. 02SQLite WAL ledger
  3. 03Governed runtime
  4. 04Web, CLI, SDK, and RPC

Key decisions

The ledger comes before chat

Messages, model calls, tool outcomes, goals, branches, artifacts, and runtime decisions enter one ordered event stream. Chat and the inspection UI are projections of that durable record.

Interrupted work keeps a continuation path

Leased runs, persistent goals, budgets, operator decisions, sequence-accurate branches, explicit resume, and safe read-only recovery let eligible interrupted work continue without reconstructing state from chat.

Governance stays in the runtime

Workspace confinement, execution-time tool-policy checks, preview-bound changes, sandboxed execution, loop guards, and verification remain in the main execution path, with their outcomes recorded in the same ledger.

DESIGN CONTRACT

Principles and honest boundaries

  • Ordered evidence is authoritative
  • Interrupted runs have explicit recovery paths
  • Tool policy is checked at execution time
  • Replay is hash-verified evidence, not re-execution
Next project

Tessmora