AI review summary

A first read of every attempt. Evidence for every claim, and no score.

Q: Does it score candidates?

No. The summary has no score, no pass or fail and no hiring recommendation, and Kendor's prompt rules them out. Your extra instructions can change what it focuses on, not that rule. Reviewers rate and decide, and nothing in Kendor reads the summary to move a candidate.

Q: Which models can it use?

Claude Opus 5.5, Claude Sonnet 5.5 or Claude Haiku 5.5 on an Anthropic key, and GPT-6 Astra, GPT-6.1 Sol or GPT-6 Luna on an OpenAI key. An owner sets the default model under Integrations, and a reviewer can pick another each time they run it. Anthropic or OpenAI bills your account, and the tokens count toward your monthly AI cap.

Q: Does it work for live coding sessions?

Yes. A session is an attempt too, so the summary can run when it ends or by hand afterwards. In a session the code, terminals and agent are shared with the interviewers, so the summary credits work to the candidate only where the evidence supports it, and says when it can't tell.

Q: Which plans include it?

AI review summaries are part of the Pro and Enterprise plans. They aren't included in Starter.

When a screen is submitted or a live coding session ends, Kendor runs the challenge's tests on the final code and asks Claude or GPT for one summary: what the candidate built, how they worked, how they used the AI agent and what to ask next. It runs on your organization's own key. Reviewers read it before the code, not instead of it.

Start hiring See pricing

AI summaryAdvisoryHide

Maya built a fixed-window rate limiter on Valkey, with one key per client and route. 9 of 11 tests pass; both failures are bursts across a window boundary. She read the tests first and ran them after most changes. Claude Code wrote the Valkey client setup, and she rewrote its retry loop herself.

Outcome: 9 of 11 tests pass. The two failures send a burst across a window boundary and get twice the limit through.
How they worked: Read the tests before writing code, then ran them 14 times in 52 minutes. Added log lines to chase the boundary failure but didn't fix it.
AI agent use: Asked Claude Code for the Valkey client and a first Allow. Replaced its retry loop with one INCR and EXPIRE, and explained why in the chat.

Strengths

Keys are namespaced per client and route, so clients never share a window.limiter/window.go:12
Added her own tests for an empty client ID and a zero limit.limiter/window_test.go:64

Concerns

The window resets on the first request after expiry, so a burst at the boundary gets twice the limit through.limiter/window.go:41
Valkey errors are ignored and the request is allowed.limiter/window.go:52

Integrity notes

Left the tab for 3 minutes at minute 31, then pasted 12 lines into window.go.activity

Interview questions

Why a fixed window instead of a sliding window or a token bucket?
What should the limiter do when Valkey is down?
Walk us through the boundary failure. What would you change in Allow?

What the AI checked (7)

Claude Sonnet 5.5 · 2 hours ago · advisory, check the evidence · Ran automatically when submittedRegenerate

Every claim points at the evidence.

Each strength, concern and integrity note carries the file and line it rests on, or the terminal, agent chat or activity. Click one and the review page opens that file at that line, or that recording, so you check the claim yourself. When the model can't point to evidence, its instructions are to leave the claim out.

A link to something that wasn't recorded is greyed out instead of clickable
Integrity events are stated as facts with times, never as proof of cheating

Concerns

The window resets on the first request after expiry, so a burst at the boundary gets twice the limit through.limiter/window.go:41
Valkey errors are ignored and the request is allowed.limiter/window.go:52

Strengths

Ran the tests after each change to Allow.terminalNothing was recorded for this challenge

window.go

36func (l *Limiter) Allow(ctx context.Context, client, route string) bool {

37 key := l.key(client, route)

38 n := l.count(ctx, key)

40 if n == 1 {

41 l.store.Expire(ctx, key, l.window)

42 }

44 return n <= int64(l.limit)

45}

47func (l *Limiter) count(ctx context.Context, key string) int64 {

See what it read.

The model gets the brief, the candidate's changes against the starter, the test results, the terminal, the agent chat, integrity events and written answers. To check a claim it can list, read and search the starter or the final code, and read a terminal or the agent chat in full. The card lists every lookup it made, and the ones you can open are links.

When the challenge has a test command, Kendor runs it on the final code first, so pass counts aren't guessed
One model call with read-only tools. It can't run code or reach the network
The card says when an attempt was too large to read whole, or changed after the summary

What the AI checked (7)

Listed the files
Read limiter/window.go · lines 30–60
Searched the code for “EXPIRE”
Read limiter/window_test.go in the starter
Read cmd/server/main.go
Read the terminal (terminal) · from line 120
Read the AI agent chat

Claude Sonnet 5.5 · 2 hours ago · advisory, check the evidence · Ran automatically when submitted

Your key, your settings, and candidates are told.

Summaries run on your organization's Anthropic or OpenAI key, the same one behind the AI agents. Turn on automatic summaries per screen or live coding session, or leave them off and run one by hand from the review page. Extra instructions per screen shift what it focuses on.

An owner sets the default model, and a reviewer can pick another for each run
Before they start, candidates read that AI may summarise their work and that people make every decision
Tokens count toward the monthly AI cap you set in Integrations

Screen builderAI review summarySummarise each attempt with AI when it's submitted. Reviewers can also run it by hand.Extra summary instructionsFocus on error handling and how they tested the limiter.Added after Kendor's prompt. They can shift the focus, not its rules: no scores, and evidence for every claim.56 / 2000

IntegrationsDefault modelSummaries run on this model unless the reviewer picks another when they start one.Claude Sonnet 5.5

Candidate's start pageAI may help review your workAcme may use AI, on its own Anthropic or OpenAI account, to summarise your submission for its reviewers: your code, terminal, answers and any assistant conversation. It gives no score; people make every decision.

How it works

From your key to the first summary.

Connect a key
An owner connects an Anthropic or OpenAI key under Integrations and picks a default summary model.
Choose when it runs
Switch on the AI review summary for a screen or live coding session to run it when each attempt ends. Reviewers can also choose Summarise on any finished attempt.
Read it, then the code
The summary sits at the top of the review page. Follow its evidence, use its interview questions, and make the decision yourselves.

Questions

Does it score candidates?

No. The summary has no score, no pass or fail and no hiring recommendation, and Kendor's prompt rules them out. Your extra instructions can change what it focuses on, not that rule. Reviewers rate and decide, and nothing in Kendor reads the summary to move a candidate.

Which models can it use?

Claude Opus 5.5, Claude Sonnet 5.5 or Claude Haiku 5.5 on an Anthropic key, and GPT-6 Astra, GPT-6.1 Sol or GPT-6 Luna on an OpenAI key. An owner sets the default model under Integrations, and a reviewer can pick another each time they run it. Anthropic or OpenAI bills your account, and the tokens count toward your monthly AI cap.

Does it work for live coding sessions?

Yes. A session is an attempt too, so the summary can run when it ends or by hand afterwards. In a session the code, terminals and agent are shared with the interviewers, so the summary credits work to the candidate only where the evidence supports it, and says when it can't tell.

Which plans include it?

AI review summaries are part of the Pro and Enterprise plans. They aren't included in Starter.

Read the docs

More of Kendor

All features

Run your next screen on Kendor. Read what the code says.

Create your workspace, import a repository or start from a template, and send your first screen today.

Start hiring See pricing

Hiring at scale? Talk to us

A first read of every attempt. Evidence for every claim, and no score.

Strengths

Concerns

Integrity notes

Interview questions

Every claim points at the evidence.

Concerns

Strengths

See what it read.

Your key, your settings, and candidates are told.

From your key to the first summary.

Connect a key

Choose when it runs

Read it, then the code

More of Kendor

Run your next screen on Kendor. Read what the code says.