An agentic review system

Code reviews that think before they block.

Give a fake person the context, standards, and hard-won instincts of a real one. It finds problems, attacks its own findings, and grounds every decisive claim before it touches your PR.

Create your fake person See how it reasons

No generic lint
Every claim at file:line
Human rails stay on

epoch-ml / zergPR #496

fake-idanreviewed 14s ago

COMMENT

[fake idan]Comment — the lease is fenced correctly.

I went deep on publication, expiry, and the takeover path. I did not line-review the generated clients.

REFUTED

The apparent stale-writer race is closed by the token compare at sync/lease.py:184.

SOLID

Heartbeat ownership is checked again immediately before publish. sync/publish.py:92

7 files scoped4 candidates killed$0.84

The method

Most findings should die.

Detection is easy. Trust comes from trying hard to prove yourself wrong—and knowing when a real defect still does not justify a block.

01
Scope
Map the changed surface and owner directives.
02
Find
Fan out across dimensions with a PR-specific rubric.
03
Attack
Independent agents try to refute every candidate.
04
Groundmandatory
The lead re-derives every decisive claim in the code.
05
Calibrate
Choose the honest severity, state, and scope note.

Three non-negotiables

01 / REFUTE

Adversarial verification

A second agent is paid to kill each finding. No concrete failure scenario, no finding.

02 / RE-DERIVE

Lead grounding

Anything decisive is independently re-read in the code. The lead never repeats a subagent’s confidence.

03 / CALIBRATE

Judgment over volume

Severity follows direction and blast radius—not how clever a bug sounds. Over-blocking counts as failure.

A platform for points of view

Your fake person remembers what the generic agent never knew.

Every person lives in your private workspace, with their own knowledge base, review doctrine, voice, traps, repository access, and audit trail.

Versioned knowledge and recurring failure traps
Repository-specific scope and severity policies
Private by default, publishable when you choose
Explicit identity on every GitHub artifact

Build a point of view

IPRIVATE · ACTIVE

Fake Idan

Systems, architecture, and calibrated review.

38 sources17 traps94% grounded

Start with one real opinion

Clone the judgment.
Keep the accountability.

Create your fake person