basou: A New Open-Source Harness to Rein In AI Coding Agents

As AI coding agents take over more of the actual writing, the scarce resource for developers shifts from typing speed to control over process. basou, a new open-source project, targets three recurring pain points: the loss of intent and reasoning between sessions, the repeated cost of re-explaining context, and the inability to audit what an agent actually did. It stores this information locally in a .basou/ directory alongside the codebase, with nothing leaving the machine.

The tool is built on two concepts: a 'saddle,' a declarative manifest defining a repo's language, visibility, and instruction-file locations once; and 'reins,' which carry intent across sessions. The basou orient command reconstructs current state from the record rather than memory, while basou decision capture has the agent itself emit structured decision data—including rejected alternatives—while it still holds conversational context, rather than relying on human recollection after the fact.

Underneath, a hash-chained JSONL event log serves as the single source of truth, verifiable via basou verify, while human-readable Markdown files (handoff.md, decisions.md) are regenerated copies. The project ingests logs from tools like Claude Code and Codex either by wrapping processes or importing after the fact, avoiding dependency on any vendor-specific format.

Designed to be local-first, deterministic (no runtime LLM), reversible, and licensed under Apache-2.0, basou is still at version 0.x but has frozen its CLI surface and versioned its on-disk format with a JSON Schema, aiming for format and semver guarantees at 1.0. It offers a practical provenance layer for engineers running multiple AI-assisted projects in parallel or needing auditable records of agent decisions.