Buildkite Test Engine Client

Buildkite Test Engine Client (bktec) is a command-line tool for running test suites with Buildkite Test Engine.

Its main responsibilities are discovering runnable tests, requesting a test plan from Test Engine, running the tests assigned to the current parallel job, reading runner results, and uploading results when configured.

Use bktec to split test work across parallel Buildkite jobs using timing data from Test Engine. It can also retry failed tests and mute or skip known failures for runners that support those features, while leaving you in control of the command your test runner executes.

For runner-specific setup, see the runner guides. For the canonical setup guide, see Installing and using the Test Engine Client.

In a Buildkite parallel step, bktec run follows this general flow:

flowchart LR
    discover["Discover test targets"] --> plan["Request or reuse a Test Engine plan"]
    plan --> slice["Select this parallel job's tests"]
    slice --> run["Run the test framework"]
    run --> results["Read results<br/>Upload when configured"]
    results --> done(["Report and exit"])
    results -.->|"Retry eligible failures"| run

    classDef boundary fill:#0f172a,color:#fff,stroke:#334155
    classDef local fill:#eff6ff,color:#172554,stroke:#3b82f6
    classDef remote fill:#ecfdf5,color:#064e3b,stroke:#10b981
    class done boundary
    class discover,slice,run,results local
    class plan remote

flowchart TD
    subgraph job["Buildkite parallel job · bktec"]
        start(["bktec run"]) --> config["Load configuration<br/>and select runner"]
        config --> discover["Discover test targets<br/>Selector file, file list, or runner discovery"]
        discover --> request["Request a test plan"]
        fallback["Local fallback<br/>Alphabetical round-robin split"]
        slice["Take this job's assigned tests<br/>BUILDKITE_PARALLEL_JOB"]
        slice --> run["Run the test framework<br/>Stream output and parse results"]
        run --> upload["Upload raw results after each attempt<br/>Only when configured"]
        upload --> retry{"Retry eligible failures?"}
        retry -->|"Enabled, supported, budget remaining"| failed["Select only failed tests"]
        failed --> run
        retry -->|"No"| report["Print final report"]
        report --> done(["Exit with final status"])
        fallback --> slice
    end

    subgraph engine["Buildkite Test Engine"]
        plan["Return cached plan<br/>or create a new split"]
        results["Collect test results<br/>for future timing data"]
        metadata["Receive plan timing<br/>and statistics"]
    end

    request --> plan
    plan -->|"Plan available"| slice
    plan -.->|"Recoverable failure or empty plan"| fallback
    upload -.->|"Upload enabled"| results
    report -.->|"Unless using fallback"| metadata

    classDef boundary fill:#0f172a,color:#fff,stroke:#334155
    classDef local fill:#eff6ff,color:#172554,stroke:#3b82f6
    classDef remote fill:#ecfdf5,color:#064e3b,stroke:#10b981
    classDef optional fill:#fffbeb,color:#78350f,stroke:#d97706
    class start,done boundary
    class config,discover,request,slice,run,report local
    class plan,results,metadata remote
    class fallback,upload,retry,failed optional

bktec discovers tests using the configured runner, requests a cached plan or sends the discovered work and parallel job details to Test Engine to create one, and selects the current job's task. It then runs only the assigned tests by expanding runner-specific placeholders such as {{testExamples}} and {{resultPath}} in the test command.

Splitting quality depends on the timing data available for the suite. First runs, newly added tests, or large changes to the test suite may be less evenly balanced until Test Engine has more recent results. Different runners support different features, and Buildkite Pipelines provides BUILDKITE_PARALLEL_JOB and BUILDKITE_PARALLEL_JOB_COUNT when parallelism is configured on the step.

bktec supports multiple test runners. The table below shows the features supported by each runner:

The latest version of bktec can be downloaded from https://github.com/buildkite/test-engine-client/releases

ARM and AMD architecture for linux, darwin, and windows

The available Go binaries

• bktec-darwin-amd64
• bktec-darwin-arm64
• bktec-linux-amd64
• bktec-linux-arm64
• bktec-windows-amd64.exe
• bktec-windows-arm64.exe

Download bktec from the latest release, make it available in your test environment, and configure your Test Engine suite for authentication. From bktec 2.6.0, the default authentication path uses a Buildkite Agent OIDC token. See Authentication for details.

This example runs RSpec across 10 parallel Buildkite jobs and uploads results using bktec's built-in upload support, available in bktec 2.7.0 or later:

steps:
  - label: ":rspec: RSpec"
    command: bktec run
    parallelism: 10
    env:
      BUILDKITE_TEST_ENGINE_SUITE_SLUG: my-suite
      BUILDKITE_TEST_ENGINE_TEST_RUNNER: rspec
      BUILDKITE_TEST_ENGINE_RESULT_PATH: tmp/rspec-result.json
      BUILDKITE_TEST_ENGINE_UPLOAD_RESULTS: "true"

This setup does not require a Buildkite Test Collector. If your test suite already uses a collector, set BUILDKITE_TEST_ENGINE_UPLOAD_RESULTS: false to avoid double uploads. See Upload test results to Test Engine section below.

The RSpec runner defaults to:

bundle exec rspec --format progress --format json --out {{resultPath}} {{testExamples}}

bktec replaces {{testExamples}} with the tests assigned to the current parallel job and {{resultPath}} with the configured result path. Other runners may need different result output settings or commands; see the runner guides before adapting this example.

For complete Buildkite pipeline examples across supported runners, see the test-engine-client-examples repository.

After the step runs, use the Buildkite job log and Test Engine suite page to check that bktec:

• authenticated successfully, either with OIDC or an API access token
• created or reused a test plan for the current build step
• received tests for the current parallel job
• ran the configured test command with assigned test files, examples, packages, or selectors
• uploaded results to Test Engine when BUILDKITE_TEST_ENGINE_UPLOAD_RESULTS=true

If the first run is uneven, check whether the suite has recent timing data. You can inspect the generated plan with bktec plan --plan-out -, or enable debug output with BUILDKITE_TEST_ENGINE_DEBUG_ENABLED=true.

bktec uses the following Buildkite Pipeline provided environment variables.

From bktec 2.6.0, bktec automatically requests a Buildkite Agent OIDC token for authentication. You don't need to create or configure an API access token. You will need to configure an OIDC policy for your Test Engine suite to allow this.

If you're running bktec older than 2.6.0, or if you want to use an API access token instead, you can create a Buildkite API access token with read_suites, read_test_plan, and write_test_plan scopes from your Personal Settings in Buildkite, then set:

export BUILDKITE_TEST_ENGINE_API_ACCESS_TOKEN=token

To use bktec, you need to configure the BUILDKITE_TEST_ENGINE_SUITE_SLUG environment variable with your Test Engine suite slug. You can find the suite slug in the URL of your suite. For example, in the URL https://buildkite.com/organizations/my-organization/analytics/suites/my-suite, the slug is my-suite.

export BUILDKITE_TEST_ENGINE_SUITE_SLUG=my-slug

Test Engine needs test result data to provide timing data for intelligent test splitting and to support features such as retry and muting. Choose one method to upload results from each test run. A Buildkite Test Collector is not required to use bktec.

Option 1: Use bktec's built-in upload (requires bktec 2.7.0 or later)

Built-in upload is off by default. If your test suite already uses a collector, remove or turn off its reporter, plugin, or test integration. Then, set BUILDKITE_TEST_ENGINE_UPLOAD_RESULTS to true:

export BUILDKITE_TEST_ENGINE_UPLOAD_RESULTS=true

bktec uses BUILDKITE_ANALYTICS_TOKEN as the upload token when it is set. Otherwise, bktec requests an OIDC token. See Authentication for details.

You can attach key/value tags to each upload using --tag or BUILDKITE_TEST_ENGINE_TAGS. Tags are useful for filtering and grouping test results in Test Engine.

# As CLI flags (repeatable)
bktec run --tag env=production --tag region=us-east-1

# As an environment variable (comma-separated)
export BUILDKITE_TEST_ENGINE_TAGS="env=production,region=us-east-1"

Option 2: Let a Buildkite Test Collector upload results

Test collectors are available for many languages and frameworks. Some collectors also provide richer data collection such as execution-level tagging and span tracing. bktec can continue to discover, split, run, and retry tests while the collector uploads the results.

Leave BUILDKITE_TEST_ENGINE_UPLOAD_RESULTS unset, or set it to false if it is enabled elsewhere in your pipeline:

export BUILDKITE_TEST_ENGINE_UPLOAD_RESULTS=false

Some runner-specific features require an installed collector package. In those cases, let the collector upload the results and keep built-in upload off. See the runner guides and test collector docs for details on your framework.

For test collectors and OpenTelemetry SDKs that export OTLP/HTTP protobuf traces, bktec can receive those traces on loopback and relay them to Buildkite. Enable it on bktec run with:

export BUILDKITE_TESTS_OTLP_RELAY=true

The relay injects the standard OTEL_EXPORTER_OTLP_TRACES_* configuration into the test process, acknowledges local exports immediately, and retries delivery to Buildkite in the background. It obtains and refreshes the upstream credential through buildkite-agent oidc request-token, reusing bktec's own OIDC-minted collector upload token when one exists; --no-oidc cannot be used with the relay. Exporters authenticate to the relay with a random, per-run local credential; existing collector upload credentials are left unchanged.

A transient credential failure (endpoint outage, timeout) never fails the job: the relay starts anyway, buffers traces, and keeps retrying in the background, reporting any traces it had to drop when the run ends. The job only fails at startup when the Buildkite API positively refuses to issue the token (agent exit status 77, e.g. a disallowed audience), since that cannot succeed by retrying.

After the final test attempt, bktec spends at most 10 seconds draining queued requests and prints the number of forwarded and dropped requests. Requests over 900 KiB are rejected synchronously, and a full 64 MiB byte queue returns HTTP 429 so the SDK can apply backpressure.

The child resource also receives buildkite.otlp.endpoint=<loopback URL> through OTEL_RESOURCE_ATTRIBUTES. This is a Buildkite-specific attribute—OpenTelemetry does not define a semantic convention for the relay route—and lets stored spans show that their SDK exported through bktec without modifying the opaque OTLP payload in the relay.

--plan-identifier (or BUILDKITE_TEST_ENGINE_PLAN_IDENTIFIER) sets the identifier the plan command generates the plan under. The identifier is the plan's server-side cache key: distinct values produce distinct plans, while reusing a value returns the previously cached plan for that identifier.

Inside a Buildkite build you don't need to set this; the identifier defaults to ${BUILDKITE_BUILD_ID}/${BUILDKITE_STEP_ID}. Supply --plan-identifier explicitly when generating a plan off-agent (for example, running it in your local development environment or on your own machine). Doing so also removes the need to set BUILDKITE_BUILD_ID and BUILDKITE_STEP_ID, which are otherwise required.

# UUIDv7 via Python (3.14+); use uuid4() on older versions
./bktec plan --json --plan-identifier "$(python3 -c 'import uuid; print(uuid.uuid7())')"

# UUIDv4 via uuidgen (preinstalled on macOS and most Linux)
./bktec plan --json --plan-identifier "$(uuidgen)"

bktec run accepts the same flag. It fetches the plan cached under the identifier, or, on a cache miss, creates and caches one under it. A reused identifier therefore returns the previously cached plan even if the inputs have changed, so keep the value unique per distinct plan.

--plan-out makes bktec plan write the full test plan without running any tests. Unlike --json, which emits only the plan identifier and parallelism, --plan-out writes the whole plan: the tasks, the per-node breakdown of tests, the muted and skipped tests, and the timing metadata. It takes a destination: - for stdout, or a file path.

./bktec plan --plan-out -            # stdout
./bktec plan --plan-out plan.json    # a file

The human-readable planning summary and any warnings are written to stderr, so stdout carries only the plan. --json, --plan-out and --pipeline-upload are mutually exclusive; choose one.

Both plan and run distinguish this invocation's requested selection and split settings (Requested) from the returned Selection summary and Split summary. The returned plan may be cached: Using existing plan identifies a successful fetch, while a create response makes no claim about cache status. Returned selection parameters appear only when they differ from this invocation; matching values do not establish that the invocation created the plan. Ordinary values use key = value; whitespace/control characters are escaped, long values are bounded, and unknown request payloads are omitted. Returned strategy names are displayed as bounded, escaped text, including names introduced by newer servers. Applied/skipped status comes from the returned metadata, not the strategy name or this invocation's settings.

Selection counts use the backend's eligible denominator, labelled test selectors, even when it includes a mixture of file and example formats. Estimated compute compares selected and candidate cumulative mean durations using the same candidate-pool median/default fallbacks. It requires supported metadata and at least 50% candidate timing coverage. Its share is not the requested duration cutoff or a wall-clock saving; a zero candidate total has no defined share. It does not use the separate top-level duration estimates.

Estimated nodes needed is the pre-cap count and labels the returned sizing basis (mean or P90 durations, including median/default fallbacks). Node limit: capped at N appears only for an independently binding cap, not merely a reached limit; tied constraints are not independently binding. The longest-node estimate uses the actual P90-packed allocation. P90 durations; target exceeded compares that estimate with the target used by sizing, not the mean decision estimate or whole-build runtime. Sparse-history and unusable-timing sizing do not use the configured target. Estimates are not runtime guarantees, and zero estimates do not promise instant execution. Older plans and unknown estimators leave details unavailable rather than inferring them from current flags. Historical timing counts and missing-history median/default estimates remain visible separately.

--plan-out writes what the server returns, unmodified. If the server cannot generate a plan it returns an empty plan, which is emitted as-is (a warning is printed to stderr). Only when the server cannot be reached at all does bktec fall back to a minimal locally-generated plan; this carries the identifier and parallelism but no tasks (it is not a computed split), and is noted on stderr.

bktec run --plan-out <path> writes the full plan before running tests, so verification tools can compare results against the actual split and selection without fetching the plan again or managing another API token:

bktec run --plan-out tmp/test-engine/plan.json
# Or configure the output path through the environment:
BUILDKITE_TEST_ENGINE_PLAN_OUT=tmp/test-engine/plan.json bktec run

The flag takes precedence over BUILDKITE_TEST_ENGINE_PLAN_OUT. Parent directories are created and existing files are overwritten. If the file cannot be created, written, or closed, the run stops before executing tests (exit 16). Otherwise test execution and exit statuses are unchanged. Unlike bktec plan, run --plan-out accepts only file paths: - names a literal file, not stdout.

For cached and freshly created plans, the file is the API's full test_plan JSON response, indented with a trailing newline. All server fields are retained, including identifier, parallelism, all nodes in tasks, test format, path, identifier, and value, and selection, skipped_reason, and timing metadata when present. It is saved before node-specific location-prefix adjustments or retries. Every parallel node writes the entire plan without any additional network requests; consumers can upload the file from node 0.

When bktec falls back to local splitting (including an empty server error plan), the file instead records the actual local tasks with an explicit fallback: true:

{
  "identifier": "build-id/step-id",
  "parallelism": 2,
  "tasks": {
    "0": {"node_number": 0, "tests": [{"path": "spec/apple_spec.rb"}]},
    "1": {"node_number": 1, "tests": [{"path": "spec/banana_spec.rb"}]}
  },
  "fallback": true
}

This is a local full-suite split, not a server selection. Verification tools should check fallback before interpreting the document as a server plan.

By default, bktec discovers tests and requests a plan by sending runner-specific selectors, the values Test Engine looks up when computing a split for the current job. For every runner except gotest, the selector is the same file path that bktec discovers, so selector splitting doesn't change which tests run where, only how that path is reported and matched. gotest is the exception: its selector is a Go package import path from go list, and selector splitting replaces the legacy even-count package split with duration-aware splitting, so Go users may see a different (and better balanced) split once historical timing data is available.

See Migrating from bktec v2 to v3 for collector requirements, runner-specific changes, and upgrade verification steps.

This is supported for RSpec, Jest, Vitest, Cypress, Playwright, pytest, gotest, Cucumber, and the custom runner.

By default, bktec discovers selectors itself using file discovery for every runner except gotest (which uses go list output) and custom (which falls back to its configured file pattern). You can instead provide a fixed list of selectors with --selector-file (or BUILDKITE_TEST_ENGINE_SELECTOR_FILE), a path to a newline-delimited file of selector values:

export BUILDKITE_TEST_ENGINE_SELECTOR_FILE=selectors.txt

When Test Engine computes a plan, it matches the selectors bktec sends against historical executions tagged with test.selector.primary. That tag is attributed automatically when results are uploaded via bktec's built-in upload, or via the Ruby, JavaScript, or Python collector at the minimum version noted in the runner guides.

If an execution has no test.selector.primary (for example an older collector, or the custom runner), Test Engine falls back to the file name for matching. If no selector history can be matched, Test Engine uses default duration estimates rather than legacy file timings, and bktec prints a warning. Run the tests once to record selector timings, or update to the minimum collector version noted in the runner guides if historical executions should already be available. The file-name fallback strips a configured location prefix on a best-effort basis, so updating matters most if you use a location prefix.

The custom runner is the main case that needs attention, since its selector might not be a file path. See the custom runner guide.

See the runner guides for runner-specific selector details and collector requirements.

You can pass test selection strategy configuration and additional change context to the test plan API request. This preview is enabled only when BKTEC_PREVIEW_SELECTION is truthy (1, true, yes, or on). This functionality is under development, and these flags currently have undefined behavior.

Environment variables:

export BKTEC_PREVIEW_SELECTION=true
export BUILDKITE_TEST_ENGINE_SELECTION_STRATEGY=percent

Command-line flags:

BKTEC_PREVIEW_SELECTION=true ./bktec plan --json --selection-strategy percent \
  --selection-param percent=40

When --selection-strategy is set, the plan command automatically collects git metadata from the current repository and sends it with the API request. This includes commit information (SHA, author, committer, message), diff data (files changed, numstat, full diff), and context fields (branch name, base branch, pipeline slug, build UUID).

For pipelines that use plan without --selection-strategy, you can opt in to metadata collection with the --collect-git-metadata flag (or BUILDKITE_TEST_ENGINE_COLLECT_GIT_METADATA=true). This collects the same git metadata without requiring selection to be configured:

BKTEC_PREVIEW_SELECTION=true ./bktec plan --json --collect-git-metadata

The base branch for diff computation is resolved using a fallback chain:

1. Explicit override via --metadata base_branch=<branch>
2. BUILDKITE_PULL_REQUEST_BASE_BRANCH (auto-set by Buildkite on PR builds)
3. Auto-detection via <remote>/HEAD, then <remote>/main, then <remote>/master

Most users don't need to configure anything. Override base_branch only if your repository uses a non-standard default branch (for example, develop or trunk) and <remote>/HEAD isn't configured.

The --remote flag (default origin) controls which git remote is used for base branch detection. You can also set BUILDKITE_TEST_ENGINE_REMOTE.

Auto-collected values are merged with any explicit --metadata flags you provide. Your explicit values always take precedence.

Use --metadata key=value to pass additional metadata or override auto-collected values. Use --selection-param key=value to pass strategy parameters. Both flags are repeatable. Values can be large and multiline.

BKTEC_PREVIEW_SELECTION=true ./bktec plan --json --selection-strategy percent \
  --selection-param percent=40 \
  --metadata base_branch=develop

--selection-param and --metadata are only supported as repeatable CLI flags.

bktec can collect historical git commit metadata from your repository and upload it to Buildkite for training test selection models. This is useful for bootstrapping models with historical changeset data so that test selection can identify which tests are relevant to your code changes.

The tools subcommands are hidden from bktec --help by default. Setting BKTEC_PREVIEW_SELECTION to a truthy value (1, true, yes, or on) makes them visible in help output. The commands can always be invoked directly regardless of this setting.

Two commands are available under bktec tools:

Collect and upload commit metadata:

bktec tools backfill-commit-metadata \
  --access-token "bkua_..." \
  --organization-slug "my-org" \
  --suite-slug "my-suite"

Generate the tarball locally for inspection before uploading:

bktec tools backfill-commit-metadata --output commit-metadata.tar.gz

# Inspect the contents
tar tzf commit-metadata.tar.gz
# commit-metadata.jsonl
# metadata.json

# Upload when ready
bktec tools backfill-commit-metadata \
  --upload commit-metadata.tar.gz \
  --suite-slug "my-suite"

The API access token requires read_suites and write_suites scopes.

For detailed usage, flags, and configuration options, see the Commit Metadata Backfill guide.

• Configure the runner-specific command and result output for your test framework.
• Inspect a generated plan with bktec plan --plan-out - when you need to understand how work was split.
• Choose one result upload method so future runs have timing data: enable built-in upload with BUILDKITE_TEST_ENGINE_UPLOAD_RESULTS=true, or let a Buildkite Test Collector upload the results and leave this environment variable unset or set to false.
• Configure retries, muting, or skipping where your runner supports them.
• Use BUILDKITE_TEST_ENGINE_DEBUG_ENABLED=true when troubleshooting authentication, uneven splits, or runner command issues.
• Run bktec --help, bktec run --help, or bktec plan --help for the current CLI options.

You can also find example configurations and usage instructions for each test runner in the examples repository.

• Jest
• Vitest
• Playwright
• Cypress
• pytest
• go test
• RSpec
• Cucumber
• Custom Test Runner

To enable debug mode, set the BUILDKITE_TEST_ENGINE_DEBUG_ENABLED environment variable to true. This will print detailed output to assist in debugging bktec.

bktec may exit with a variety of exit statuses, outlined below:

• If there is a configuration error, bktec will exit with status 16.
• If the test runner (e.g. RSpec) exits cleanly, the exit status of the runner is returned. This will likely be 0 for successful test runs, 1 for failing test runs, but may be any other error status returned by the runner.
• If the test runner is terminated by an OS level signal, such as SIGSEGV or SIGABRT, the exit status returned will be equal to 128 plus the signal number. For example, if the runner raises a SIGSEGV, the exit status will be (128 + 11) = 139.

Make sure you have Go, Ruby, and Node.js installed in your environment. You can follow the installation guides for each of these tools:

• Go Installation Guide
• Ruby Installation Guide
• Node.js Installation Guide

Once you have these dependencies installed, run bin/setup to install dependencies for the sample projects for testing purposes.

To test, run:

./bin/test