Start here

What OAK is

This is the two-minute picture of what the observatory records and why.

OAK is an observatory for agentic development: it observes every edit an agent makes through local hooks that cost no tokens, lets you audit what actually changed in your codebase, and lets you keep β€” or undo β€” each change at the scope that fits the work.

OAK is built for established, mission-critical codebases rather than throwaway prototypes, where a change accepted without being understood is a real liability.

πŸ”¬
The premise

Coding agents change code faster than anyone can read it, and a change accepted without being understood is a liability for you, not the agent's provider. OAK exists to make review keep pace with generation.

How it works

OAK derives everything from two sources that cost no tokens and are invisible to the agent. The transcript the agent already writes β€” prompts, tool calls, results, reasoning β€” is read as evidence, never modified. And a local capture hook snapshots each file before and after a change and appends a record to the observatory's own store. From those two, OAK reconstructs what you actually review.

β„Ή
No tokens, no interference

Reading happens outside the model loop. The hook prints nothing and returns immediately, so the agent's behavior and context are unchanged β€” and observing an agent is the default, not something you switch on.

What you review

The record is the same for every agent: prompts (what you asked), tasks (the to-dos the agent set itself), actions (every tool call), and edits (each captured file change). The unit of decision is the review unit β€” what an agent changed for a given prompt β€” so revising the same lines five times before settling asks you for one decision on the net change, not five.

Every edit keeps its own diff and its own conflict-guarded undo, so reverting one leaves the others standing.

One agent, or a fleet

Everything hangs off a session β€” one conversation with a coding agent. But work rarely stays in one session: an agent spawns subagents, a workflow fans out a dozen at once, and siblings run in parallel git worktrees of the same repo. OAK correlates all of them, so a fan-out reads as a single account of who did what β€” path-only, never mixing one agent's file contents into another's.

Where you review it

The same numbers appear in three places: the terminal app, VS Code, and JetBrains IDEs. A number shown on one surface is the same number acted on everywhere β€” because all three render one CLI backend.

β†’
Next steps

Install OAK, then walk through your first review. To understand the model first, start with Sessions & agents.