Skip to content
Kindgi

Try a version on the test cases

View as Markdown

This guide shows whether people would rate a new agent or flow version higher than the runs they corrected, before anyone uses it.

A run on the test cases tries an agent version on each case of the test cases, as a replay of the past run. It rates the version's answers with the corrections people already made, beside the recorded runs.

Before you start

Section titled “Before you start”
  • Test cases: the agent's test set, and the new version of their agent or flow. Under kindgi dev, saving the agent's file with its new version is enough. On a runtime you deploy to, deploy the new version first.
  • Kindgi 0.4's CLI, and the project's id.
  • The right to start one: write on the project the run lands in, and the right to run its agent or flow. A test set other than the agent's own also takes the right to execute it, which its project's editors have.
  • For the app's half: Node 22.12 or later, or Python 3.11 or later, with Kindgi 0.4's client, and KINDGI_API_URL and KINDGI_API_TOKEN set.

Try a version on the test cases

Section titled “Try a version on the test cases”

A run on the test cases replays the version on each case. Version 0.1.0 of acme-ops.article-picker answered with three articles each time, and people marked most of them wrong. Version 0.2.0 changes its instructions. Ask your coding agent: "Make a new version of my agent that answers with only the articles that answer the question." For reference, 0.2.0's instructions look like this:

instructions:
  "Call acme-ops.search-articles with the customer's question. Answer with only the articles that answer it, most relevant first: one is often enough.",
Excerpt

The test cases keep their copies of the 0.1.0 runs. Try 0.2.0 on them:

Terminal window
pnpm exec kindgi test-runs start agent.acme-ops.article-picker.<project-id> --project=<project-id> --agent=acme-ops.article-picker --agent-version=0.2.0 --wait
  • agent.acme-ops.article-picker.<project-id> is the agent's test set in the project: agent., the agent's id, a dot, and the project's id.
  • --project=<project-id> is the project the replay runs land in. It's the test set's project: another agent or project is refused with 400 bad-input, and the message names the test set id this run would take.
  • --agent and --agent-version are the version to try.
  • --wait waits for the run to finish.

It prints the run, with its metrics:

{
  "testRunId": "b0b9f27a-14d9-4a4d-bc46-31aea6948d5f",
  "testSetId": "agent.acme-ops.article-picker.<project-id>",
  "testSetVersion": "1.0.3",
  "status": "completed",
  "result": {
    "summary": {
      "cases": 3,
      "candidate": { "kind": "agent", "agentId": "acme-ops.article-picker", "version": "0.2.0" },
      "baseline": { "kind": "recorded", "versions": [{ "agentId": "acme-ops.article-picker", "version": "0.1.0", "cases": 3 }] },
      "metrics": {
        "passRate": { "baseline": 0.3333333333333333, "candidate": 1, "delta": 0.6666666666666667, "n": 3, … },
        "caseCoverage": { "baseline": 0.75, "candidate": 0.5, "delta": -0.25, "n": 3, … },
        "weightedPrecisionAtK": { "k": 10, "baseline": 0.3333333333333333, "candidate": 1, "delta": 0.6666666666666667, "n": 3, … }
      },
      "diverged": 1,
      "refusedWrites": 0,
      "errors": 0,
      "sampling": { "models": [{ "providerId": "anthropic", "model": "claude-sonnet-5-5", "runs": 3 }] },
      …
    },
    "perCase": [ … ]
  },
  …
}
Excerpt

Without --wait, the command returns the run's id at once, and kindgi test-runs show <test-run-id> reads it later.

The run freezes the test set first: its cases as they are when it starts, as a version of the set (1.0.<n>), which the run records as testSetVersion. A new version is made only when the cases changed since the newest; otherwise that one runs again. So two runs of different versions compare on the same cases, unless a case was added or retired between them. With no cases yet, the run is refused with 409 suggested-change-nothing-to-check, and a runtime that doesn't freeze a project's test set answers 501 suggested-change-check-unsupported.

A suggested change's check runs the same two test runs for you: its version and the version it's built on, on one frozen version of the set.

From your code, ask your coding agent: "I want a script that tries my new agent version on my test cases, and shows me the scores." For reference, the script looks like this, with the project's id as its argument:

compare.ts
import { comparisonOf, createClient } from '@kindgi/sdk/client';


const kindgi = createClient(); // KINDGI_API_URL, KINDGI_API_TOKEN, or the running kindgi dev
const projectId = process.argv[2];


// Replay version 0.2.0 on each case of the agent's test set in the project.
const { testRunId } = await kindgi.testRuns.start(
  `agent.acme-ops.article-picker.${projectId}`,
  { agentRef: { agentId: 'acme-ops.article-picker', version: '0.2.0' } },
  { projectId },
);


let run = await kindgi.testRuns.get(testRunId);
while (run.status === 'pending' || run.status === 'running') {
  await new Promise((resolve) => setTimeout(resolve, 2000));
  run = await kindgi.testRuns.get(testRunId);
}


const comparison = comparisonOf(run); // undefined until it's finished
for (const [name, metric] of Object.entries(comparison?.summary.metrics ?? {})) {
  console.log(`${name}: ${metric.baseline} → ${metric.candidate}`);
}
caseCoverage: 0.75 → 0.5
passRate: 0.3333333333333333 → 1
weightedPrecisionAtK: 0.3333333333333333 → 1

comparisonOf reads a finished run's result, typed: its summary and each case. It's comparison_of in Python, and Comparisons.comparisonOf in Java and Scala. For a run that hasn't finished, or a dry run, it returns undefined: None in Python, an empty Optional in Java and Scala.

A model doesn't answer the same way every time: try the same version on the test cases again, and the scores may differ. --repetitions=<n> runs each case several times, up to 10, and each metric gets a spread: its highest value minus its lowest. In the clients, it's repetitions.

Read the result

Section titled “Read the result”

Each metric has its value for the recorded runs (baseline) and for the version (candidate), the delta, and n, the cases it's based on.

  • passRate, "Share marked right" in the console: of what people marked, the share marked right. Each item of the answer they corrected counts, weighted by corrector group, and so does each part of a case's graded expect, its decisions and its answer, as one check of weight 1. The example's cases expect items alone: 0.1.0's answers had 3 marked right of 9 corrected articles. 0.2.0's had 3 of 3: it kept only the right article each time.
  • caseCoverage, "Corrected items": the share of the answer's items that people corrected. Items no one corrected bring it down. Here, each answer's text is an item too, and no one corrected it.
  • weightedPrecisionAtK, "Top 10 marked right": marked right among the first k items of a ranked list. k is 10 unless you set --k.

How a correction carries over. An item with its own id, an id or key field, is the same item when the id matches, at any place in the list. So the yes on kb-301 counts for 0.2.0 too. The answer's text, and a list element without an id, carry a correction only when they're unchanged. A reworded answer is a new item, with no corrections yet.

Count only experts' corrections

Section titled “Count only experts' corrections”

--group-weights=experts-only counts only the corrections made under an expert group. An item none of them corrected counts as not corrected. In TypeScript, it's groupWeights: 'experts-only', and in Python group_weights="experts-only". The summary says which it used, in groupWeights, as-recorded by default.

On an agent's test set, it grades only experts' cases, which have expert: true: a correction's case when the correction was made under a restricted expert group, and every case added by hand, since adding one takes publish on the agent (or write on its project). No other case's expect is graded.

On three test cases, one of them corrected only under a group restricted afterwards:

as-recorded experts-only
passRate 0.556 (5 of 9) 0.5 (3 of 6)
caseCoverage 1 0.667

Case by case, perCase has, for each case:

{
  "caseId": "240c0a0b-542b-4cb1-9156-fe7153a4377a",
  "runIds": ["732235ae-2321-46fe-876e-9e9818671cbf"],
  "changes": {
    "kept": [{ "key": "kb-101", "rankBefore": 0, "rank": 0 }],
    "dropped": [{ "key": "kb-102", "rankBefore": 1 }, { "key": "kb-103", "rankBefore": 2 }],
    "new": [{ "key": "answer", "pointer": "/appended/3/content" }]
  },
  "tools": [
    {
      "toolId": "acme-ops.search-articles",
      "toolVersion": "0.1.0",
      "arguments": { "query": "I forgot my password and can't sign in" },
      "source": "live",
      …
    }
  ],
  "diverged": true,
  "refusedWrites": 0,
  …
}
Excerpt
  • changes: the corrected items the version kept, with their place before and now, the ones it dropped, and the new ones, for people to correct.
  • tools: each tool call, and what happened to it: live when it ran, recorded when it got the past run's result, or refused.
  • diverged: the version read something the past run didn't, so the read ran live. Here, 0.2.0 left the final period off the customer's question when it searched, so the call wasn't the same as 0.1.0's.

What it should have done

Section titled “What it should have done”

When the test cases' runs have corrections that say what the run should have done (runtime 0.2.0 or later), each one is checked against what the version did on that case:

CAPTURE: one perCase entry with "corrections": [{ "correctionId": …, "pass": false, "repetitionsPassed": 0, "repetitions": 1 }] and "correctionDifferences": [{ "correctionId": …, "repetition": 1, "differences": [ … ] }]
Excerpt
  • corrections: for each one, whether it passed, and in how many of the case's repetitions (repetitionsPassed of repetitions). It passes only when every repetition did.
  • correctionDifferences: for one that failed, what differed in the first repetition that failed it, and that repetition's number. A difference is a call count (count), a call that shouldn't have been made (must-not-call), or an argument that's missing (missing-field) or has another value (field-value). refused marks a call the replay refused.
  • The summary's corrections counts the ones checked, and how many passed:
CAPTURE: "corrections": { "checked": …, "passed": … } from the same run's summary
Excerpt

With --group-weights=experts-only, only the ones made under an expert group are checked. A release gate checks them only with cases, so read them before you pin the version.

What a case expects

Section titled “What a case expects”

A case of an agent's test set is graded on its expect, part by part, on each repetition:

The case expects A repetition passes it when
decisions: "wrong" It calls every shouldCall tool (with its args where given) and none of mustNotCall, checked exactly
decisions: "right" It calls the same tools the recorded run called, by id, in any order. A refused call counts as called
answer: "wrong" A judging model finds that its answer does what shouldSay or why asks
answer: "right" A judging model finds that its answer says what the recorded answer said
items Nothing: the items count toward the metrics, as above, with no pass

The case has one entry in corrections, under the case's id, and it's counted in the summary's corrections. A repetition passes when every part the case expects passed on it, and the case passes when every repetition did. Its parts says whether each part, its decisions and its answer, passed on every repetition. Each graded part is one check in passRate, beside the case's items, so a case that expects both is two checks, and a release gate on passRate counts them. The recorded run passes a part when people marked it right, and each repetition that ran passes it when the part passed on every repetition. For decisions: "wrong", correctionDifferences says what differed. When an answer fails, why has the judging model's reason:

{
  "correctionId": "8f2c4e1a-6b0d-4c3e-9a57-2d1f0b7e4c93",
  "pass": false,
  "repetitionsPassed": 1,
  "repetitions": 3,
  "why": "It never says how long the refund takes.",
  "parts": { "decisions": true, "answer": false }
}
  • Which model judges. The version you try judges its own answers, with the model its model-settings block names. A suggested change's check judges both of its runs with the version the change is built on, and an Improve run judges with the version it starts from. So one model judges both sides. The model an Improve run names only writes its new wording.
  • What it costs. Each answer checked is a call to the judging model, which costs money. It's counted in its repetition's costUsd, so it's in the test run's cost, even when the reply couldn't be read. Answers are checked in order, and checking stops at the first answer that fails. An answer is checked on a repetition whose decisions failed too, since it's a check of its own.
  • No pass. A case that expects only items has no entry, and adds no check to passRate: its items count toward the metrics. Neither does a case whose answer couldn't be checked, say because the version's model settings name no model or the call to the judging model failed, or one that expects decisions: "right" when the recorded run's tool calls can't be read. The case's unchecked says why, and the summary's unchecked counts those cases and lists their reasons. A failed call doesn't stop the test run.
  • A repetition that errored or stopped fails every part, and its answer isn't checked, nor any after it.

What a replay does

Section titled “What a replay does”

A replay does nothing the past run didn't do. "The same call" means the same tool with the same arguments.

  • A read-only tool gets the past run's result for the same call. A read-only tool is mutating: false, has no effect that changes things, and no approval to ask for. It runs live when the past run didn't make the call, and the case is then diverged.
  • A read-only tool also runs live when the version you try pins other settings blocks than the past run's version. So a tool that reads its settings shows the change. That counts as diverged too. With --reads=live, or reads: 'live' in the clients, it always runs live.
  • A read that reaches outside, with a network or sensitive-data-egress effect, only ever gets the past run's result for the same call. Without one, it's refused, as below.
  • Any other tool gets the past run's result for the same call, or doesn't run. The model is told replay: this call changes things and has no recorded result, or that the call needs an approval, or reaches outside. The case shows what it would have done. refusedWrites counts every call a replay refuses, reads included, despite its name.
  • A case with a correction that says what the run should have done replays strictly (runtime 0.2.0 or later): no tool runs live, even with --reads=live, or when the version pins other settings blocks. A read gets the past run's result for the same call, or doesn't run: the model is told replay: this read has no recorded result, and refusedWrites counts it. So the version is checked on the results the expert saw.
  • A replay never waits for an approval. Where the past run's conversation was approved or rejected, the replay follows that decision. Where there's no decision, the case says approvalSkipped.
  • A replay runs for the past turn's end user. An agent whose memory is per end user, same-user, is offered the same tools as the past turn, kindgi_remember included, which stores nothing in a replay. A replay's conversation is never recalled later as an earlier conversation.
  • A replay of a run a webhook or an event started reads its message as data (it needs runtime 0.1.9 or later). Its model reads the turn's message, or a flow's agent step's, as an <event> block, data and not instructions, even when the past run was on an earlier runtime and read it as it came: its answer can differ for that alone. The replay's own run records no startedBy, as no replay does. A delivery earlier in the conversation reads that way too, unless an earlier runtime kept it: that one reads as it came.
  • A recorded result that was data from outside stays data (runtime 0.3.0 or later). Where the past run's model read a tool's result in an <event> block, as another system's answer to a call that paused, or as the result of a tool whose results are from outside, the replay's model reads the past run's result the same way. On an earlier runtime it reads as it came. In a flow run first corrected on an earlier runtime, an agent step's answer to a paused call reads as it came.

The replay runs are ordinary runs, marked with replayOf, the past run, and testRunId. Lists of runs leave them out. kindgi runs list --test-run=<test-run-id> lists those of one run on the test cases, and --replays=include or --replays=only adds them to any list. Conversation lists leave out their conversations the same way: kindgi conversations list --replays=only lists them. You can correct them like any run: that's how a version's new items get their corrections.

A run on the test cases stopped before it finished, because the runtime stopped, ends failed with the cases it got through: Interrupted: the test run's executor stopped before it finished. 2 of 3 cases finished. Start it again.

Try a flow version on the test cases

Section titled “Try a flow version on the test cases”

A corrections test set built from a flow's corrected runs tries a flow version. Each case runs the past run's input through that version, as a replay. The flow's whole output is rated: the output itself, the output item, and each element of a list in it.

To try some of its tools or agents at another version, without publishing a new flow version, add --with=<id>@<version> once for each. Here the flow acme-ops.suggest-articles 0.1.0 runs its acme-ops.find-articles tool at 0.2.0, which matches only the articles a question asks about:

Terminal window
pnpm exec kindgi test-runs start acme.flow-picks --project=<project-id> --flow=acme-ops.suggest-articles --flow-version=0.1.0 --with=acme-ops.find-articles@0.2.0 --reads=live --wait

--flow and --flow-version are the flow version to try, and --reads=live runs read-only tools live. Its summary names the swapped tool:

"summary": {
  "candidate": {
    "kind": "flow",
    "flowId": "acme-ops.suggest-articles",
    "version": "0.1.0",
    "versions": { "tools": { "acme-ops.find-articles": "0.2.0" } }
  },
  "metrics": {
    "passRate": { "baseline": 0.5, "candidate": 1, … },
    …
  },
  …
}
Excerpt
  • In the clients: flowRef: { flowId, version } in place of agentRef, and versions: { tools: { 'acme-ops.find-articles': '0.2.0' } }. Agents go under agents.
  • A read-only tool's new version shows only with live reads. With the default, recorded reads, its call gets the past run's result. So the same run on the test cases without --reads=live rates 50% before and after.
  • Under kindgi dev, a pack serves two versions of a tool side by side, so a tool swap works there. An agent has one version there, its file's, so swapping an agent needs a runtime you deploy to.
  • A --with the flow doesn't use is refused: flow acme-ops.suggest-articles 0.1.0 doesn't use acme-ops.article-picker.

A flow stops at a tool call the replay refuses. acme-ops.suggest-and-log also records the articles it suggested. With 0.2.0 of the tool, it would have recorded other articles than the past run did for one question, so that case stops there:

"stopped": {
  "toolId": "acme-ops.log-suggestion",
  "arguments": {
    "articles": [{ "id": "kb-201", "title": "Request a refund" }],
    "question": "I was charged twice for one order. How do I get my money back?"
  },
  "reason": "replay: this call changes things and has no recorded result"
}
Excerpt

A stopped case isn't an error: the refused call wasn't carried out, and the case has no output to rate, so it's left out of the metrics. The summary counts it in stopped, and the refused call in refusedWrites. The other cases are rated as usual. A step's agent turn doesn't stop the flow: its refused calls go back to the model, as they do when you try an agent version.

When a version does better, pin it where you choose, as Release an agent version shows.

In the console

Section titled “In the console”

The test cases' page has Try an agent version on the test cases: pick the version, how many runs per case, whether to read live data instead of the past runs' results, and Count only experts. Its Runs on the test cases list links each one to its page, which has:

  • a headline in words: "People's corrections rate acme-ops.article-picker 0.2.0 better than the recorded runs: 33% marked right before, 100% now, on the 3 of 3 cases with answers people corrected. In 1 case it read something the recording didn't have.";
  • the metrics, for the recorded runs and the version, with the change. A run that counted only expert groups says so: "Only experts' corrections counted; the others weighed 0.", and lists mark it "expert groups only";
  • Case by case: the recorded answer beside the version's, the corrected items it kept and the ones no longer there, the new ones, and each tool call;
  • The whole result, as JSON.

A flow's test cases have Try a flow version on the test cases instead, with a picker for each agent its steps run, to run it at another version. Tools are swapped from the CLI and the clients. The page of a flow's run on the test cases shows the recorded output beside the version's, and a stopped case says what it would have done: "It stopped at a tool call the replay refused: it would have called acme-ops.log-suggestion with {"articles":[…],"question":"…"}. That call wasn't carried out, and there's no output to rate."

Next

Section titled “Next”