a nightly reviewer for Claude Code

It reads your day
while you sleep.

oneiro goes through the day's Claude Code sessions, finds where the agent kept misunderstanding you and where it made the same mistake twice, and queues concrete fixes to your CLAUDE.md and memory. Nothing lands until you approve it.

View on GitHub Release notes

$ oneiro run  then  oneiro serve

The review

A morning queue, not an autopilot.

Every proposal waits until you keep or deny it. Approving is what writes the memory note or the CLAUDE.md edit. Deny one and it takes a reason, so the next run does better.

127.0.0.1:8260
the review queue
The morning queue. Each proposal waits until you keep or deny it.
a proposal as a full file diff
Open one and it is a full file diff, scrolled to the change, so you see exactly what would land.
nothing to review
A quiet day, so iro just sleeps.
How it works

Five steps, every night.

01

Collect

Every transcript touched in the window. Default 24h, yours to change.

02

Distill

A 20mb session becomes a small digest the model can actually read.

03

Review

Pull out misunderstandings, repeated mistakes, gaps in your instructions.

04

Synthesize

A thing that bit you in five chats ranks above a one off.

05

Queue

Proposals wait for your keep or deny. Plus a short morning report.

Engines

Bring your own brain.

Pick who does the thinking in the config. Same prompt in, same structured findings out, so the rest of the pipeline never changes.

claude-cli · your subscription anthropic · Messages API openai · any compatible endpoint
Stays safe

Paranoid on purpose.

  • Nothing is written without your approval. Every proposal waits in the queue.
  • Memory notes are additive and prefixed oneiro-, easy to find and delete.
  • Zero runtime dependencies. The real config stays out of git and never holds a secret.

oneiro is Greek for dream.

The review runs at night, so iro does too. Open source, MIT, zero deps. Take it for a spin.

View on GitHub