General · summary
Final retry failure is silently discarded
The exhausted path converts the final failure into apparent success.
retry.ts L29
A fresh context for the final pass
The agent wrote it.
EV reviews it.
EV is a review workflow for your coding agent. Run it entirely through the CLI, or open Desktop to inspect the evidence and make the final decisions yourself.
APPLE SILICON · INSTALLS DESKTOP + CLIStart here
From install to first review.- 01 · INSTALL Install EV. Desktop and CLI arrive together.
- 02 · CONNECT Connect your agent.
ev agent setup codex - 03 · REVIEW Ask for a fresh read.
/ev-review
FIG. 01 / ONE REVIEW, TWO WAYS THROUGH IT
24let lastError: unknown;
25for (let attempt = 0; attempt < 3; attempt++) {
26try { return await run(); }
27catch (error) { lastError = error; }
28}
29return undefined as T;
> /ev-review
EV review · general + security-audit
✓ checked linked fp issue requirements
7 findings returned to agent
Independent context
A second context,
on purpose.
The reviewer sees the work with fresh eyes. It can fetch the issue, inspect the change, and gather evidence without inheriting the implementing agent’s assumptions.
- IMPLEMENTATION CONTEXT The work as it was built
- EV REVIEW Fresh session · chosen skills
- DURABLE REVIEW Evidence, findings, decisions
What EV gives you
A review surface built for agent-written changes.
- 01 · SCOPE
- Review a working copy, branch, commit, or ref.
- 02 · SKILLS
- Choose general, security, performance, or your own review lens.
- 03 · EVIDENCE
- Keep every finding tied to the files, lines, and reasoning behind it.
- 04 · STORY
- Turn a large change into focused sections you can review in order.
- 05 · TWO LOOPS
- Let the agent act automatically, or make the final calls in Desktop.
- 06 · MEMORY
- Return to local review history with its evidence and decisions intact.
How it works
One review.
Two ways through it.
Every review starts in your coding agent with /ev-review. Let the agent complete the loop, or pause in Desktop so you can make the final decisions.
Automated
The shortest path from review to fix.
Ask the same skill for Desktop when you want to make the final calls yourself.
Ask for a review.
Invoke the skill; your agent starts EV with the right review skills.
/ev-reviewGet a fresh read.
EV checks the change in a separate context.
7 findingsFix and repeat.
Findings return to the agent to address and re-run.
7 → 2 → 0 openWith Desktop
Take the final pass.
Stay in the agent loop. Story turns even a giant PR into focused sections before you decide.
Ask for the full review.
Invoke the skill and tell your agent to open the review in Desktop.
/ev-review · with DesktopRead it in sections.
Move through the story, inspect its evidence, then curate the findings and add comments.
42 files → 6 sectionsReturn one review.
Submit once; EV sends Markdown back to the waiting agent.
Markdown → agentGrounded findings
See every finding
in context.
Every finding stays anchored to its file, lines, and reasoning. Resolve what you addressed, dismiss what does not apply, or add your own comment before returning the review.
24let lastError: unknown;
25for (let attempt = 0; attempt < 3; attempt++) {
26try { return await run(); }
27catch (error) { lastError = error; }
28}
29return undefined as T;
General · finding 1 of 2
Final retry failure is silently discarded
The final error is stored but never returned or thrown, so callers observe apparent success.
return undefined as T
DismissResolve
Review memory
A review you can
come back to.
EV’s Home screen and navigation keep recent reviews with each project. Reopen one to recover its scope, findings, evidence, comments, and decisions.
- RECENT IN PROJECT
- Return to recent reviews from the navigation rail.
- ALL REVIEWS
- Browse the full history for the active project.
Give agent-written code
a proper review.
Download for macOS Read the documentation APPLE SILICON · CLI INCLUDED