Run performance benchmarks for a Homeboy component and surface regression deltas against a stored baseline.

Synopsis

homeboy bench <component> [options] [-- <runner-args>]
homeboy bench matrix [<component>] --setting-matrix <key=value[,value...]> [options] [-- <runner-args>]
homeboy bench list <component> [options] [-- <runner-args>]
homeboy bench history <component> [--scenario <id>] [--rig <id>] [--limit 20]
homeboy bench distribution <component> --field <metadata.path> [--scenario <id>] [--rig <id>] [--status <status>] [--limit 20]
homeboy bench compare --from-run <run-id> --to-run <run-id> [--metric <name>]

Description

The bench command invokes the extension’s bench runner, which measures one or more scenarios over N iterations and emits a structured JSON results file. Homeboy parses the results, compares declared numeric metrics against a saved baseline, and returns a structured report plus an exit code suitable for CI gates.

Bench evidence surfaces artifact access explicitly. Entries that can be opened directly include public_url; bundle-relative or internal artifacts include relative_to and fetch_command so agents can fetch them with homeboy runs artifact get instead of guessing public URL shapes.

Output

By default homeboy bench prints a compact human summary for run and comparison invocations: pass/fail, the persisted run ID, runner, component SHA, selected scenarios, key metrics, generic hotspots, generic coverage summaries when present, and artifact pointers, followed by the homeboy runs show <run-id> / homeboy runs artifacts <run-id> commands to see everything. No data is dropped — the full structured payload is always written to --output <file>, and --json prints it to stdout instead of the summary.

For successful full-surface runs, the hotspot section ranks the slowest numeric timing metrics and the hottest query/count-style metric families directly from the bench payload. The extraction is schema-blind and component-agnostic: it uses metric names such as duration, elapsed, *_ms, *_ms_per_item, *_queries_per_*, and grouped count metrics rather than product-specific scenario names.

When bench/fuzzer runners provide coverage metadata, the compact summary also renders a schema-blind coverage block. Homeboy reads generic fields such as coverage_summary, coverage_gaps, surface_count, exercised_count, skipped_count, and failed_count from result metadata or artifact records. It reports discovered, exercised, skipped-unsafe, and failed surface counts plus top uncovered groups when those fields are available; missing fields are simply omitted.

Bench runners that mutate runtime state can report fuzz-safe lifecycle evidence under run_metadata.lifecycle. The lifecycle result uses homeboy/lifecycle-result/v1 and includes phase statuses plus snapshot_refs for prepared/seeded runtime snapshots. Core stores the metadata but concrete prepare, seed, snapshot, reset, rollback, and teardown implementations remain extension-owned. See docs/architecture/lifecycle-contracts.md.

bench is a sibling of test, lint, and build under homeboy’s extension capability model. The runner contract, manifest shape, and baseline primitive (homeboy.json → baselines.bench) are shared with the other capabilities.

Arguments

• <component>: Component to benchmark. Auto-detected from the current working directory if omitted. The component must have a linked extension that declares a bench capability.

Options

• --iterations <N>: Iterations per scenario (default 10). Forwarded to the runner via $HOMEBOY_BENCH_ITERATIONS. Extensions may clamp.
• --baseline: Save the current run as the new baseline under homeboy.json → baselines.bench.
• --ignore-baseline: Run without comparing to any saved baseline.
• --ratchet: When scenarios improve, auto-update the saved baseline so the improvement "sticks". Ignored when the run regresses.
• --regression-threshold <PERCENT>: Legacy p95 regression tolerance (default 5.0) used when the runner does not declare metric_policies. A p95 scenario regresses when its current p95_ms exceeds baseline.p95_ms * (1 + threshold/100).
• --shared-state <DIR>: Directory shared across iterations and concurrent runner instances. Forwarded to workloads via $HOMEBOY_BENCH_SHARED_STATE.
• --concurrency <N>: Number of parallel bench runner instances to spawn (default 1). Values greater than 1 require --shared-state. With --matrix, this controls generic scheduler task concurrency instead.
• --setting <key=value>: Override component settings (may be repeated).
• --setting-json <key=json>: Override component settings with typed JSON values for arrays, objects, numbers, booleans, or null.
• --path <PATH>: Override the component’s local_path for this run.
• --json: Print the full JSON output envelope to stdout instead of the compact human summary. The full payload is also always written to --output <file>; this flag only changes the default stdout presentation.
• --json-summary: Include a compact machine-readable summary in the JSON output envelope (for CI wrappers).
• --report side-by-side: Select the combined side-by-side comparison report for a multi-rig bench envelope. The report is emitted under reports.side_by_side and includes each rig’s status, elapsed time, key metrics, artifact paths/URLs, and failure reason.
• --rig <RIG_ID[,RIG_ID...]>: Pin the run to one or more rigs. Single rig pins the rig and stores its baseline under a rig-scoped key. If that rig declares bench.components, the command fans out across those components under one rig-state snapshot. Multiple rigs (comma-separated) run the same component + workload + iteration count against each rig in sequence and emit a cross-rig comparison envelope. See "Cross-rig comparison" below.
• --rig-concurrency <N>: For multi-rig comparisons, run up to N rigs concurrently. Default 1 preserves sequential CI behavior. Values greater than 1 are opt-in and preserve output ordering by the selected rig order.
• --matrix <axis=value[,value...]>: Add an agent-task matrix axis. Repeat for multiple axes. When present, homeboy bench builds a generic homeboy/agent-task-matrix-plan/v1, dispatches cells through the generic scheduler, and returns both scheduler and matrix aggregate metadata.
• --setting-matrix <key=value[,value...]>: For homeboy bench matrix, add a local settings axis. The command expands the Cartesian product, runs one normal child bench per cell, and aggregates child statuses, run IDs, and numeric metric samples.
• --runner-pool <BACKEND>: Executor backend string for matrix cells. Core keeps this as data so concrete backends such as Sample Runtime remain extension-owned. Omit to use the local no-op executor for plan/scheduler smoke checks.
• --max-tasks <N> / --max-queue-depth <N>: Scheduler backpressure limits for matrix cells.
• --expect-artifact <NAME>: Record an expected artifact name on each matrix task request. Repeat to declare multiple expected artifacts.
• --scenario <SCENARIO_ID>: Run or list only the exact scenario id. May be repeated. Homeboy validates selected ids against discovery before execution and forwards the comma-separated selector to runners via $HOMEBOY_BENCH_SCENARIOS.
• --ci-profile <ID>: Run using env and passthrough args from a single extension-declared CI profile whose only job declares command: "bench". This keeps bench parity on the normal Homeboy bench workflow instead of parsing arbitrary provider YAML into runnable commands.
• --ignore-default-baseline: Skip automatic single-rig expansion when the rig declares bench.default_baseline_rig.

Arguments after -- are passed verbatim to the extension’s bench runner script.

Matrix Fan-Out

For local benchmark scale sweeps, homeboy bench matrix is the narrow reusable runner over existing bench execution. It supports zero or one rig, rejects baseline writes, and varies string settings via --setting-matrix:

homeboy bench matrix 
  --rig gutenberg-rtc 
  --scenario gutenberg-rtc-protocol-load 
  --setting-matrix clients=10,100,500,1000 rounds=3 batch_size=1,10,25,100 
  --runs 3

The JSON output uses the settings_matrix bench variant and includes:

• cells: one entry per expanded settings combination with status, exit code, extracted child run ID, hints, and numeric metric samples.
• summary: cell totals and collected child run IDs.
• follow_ups: intentionally deferred pieces such as matrix-level observation records, typed JSON axes, and cross-rig matrix aggregation.

Agent-Task Matrix Fan-Out

Matrix fan-out is the operator-facing path over Homeboy’s generic agent-task substrate. It is intentionally executor-neutral: the CLI turns product inputs into AgentTaskRequest cells, the scheduler owns concurrency/backpressure, and backend-specific execution remains outside Homeboy core.

homeboy bench studio-web 
  --rig studio-web-evals 
  --matrix model=gpt-5.5,kimi,claude 
  --matrix prompt=site-a,site-b,site-c 
  --runner-pool sample-runtime 
  --concurrency 8 
  --max-queue-depth 16 
  --expect-artifact bench-results 
  --report side-by-side

The JSON output uses the matrix_fanout bench variant and includes:

• scheduler: homeboy/agent-task-aggregate/v1 scheduler totals, events, queue state, outcomes, and backpressure.
• matrix: homeboy/agent-task-matrix-aggregate/v1 cells with axes, status, artifacts, evidence refs, diagnostics, and metadata.
• report: compact product metadata for the selected report format plus succeeded, blocked, failed, cancelled-cell, and timed-out-cell totals.

When --ci-profile <ID> is used, args declared on that profile’s bench job are forwarded before explicit CLI passthrough arguments, and job env is passed to the bench runner. Profiles with zero jobs, multiple jobs, or a non-bench job are rejected for command-native bench reproduction; use homeboy ci run --profile <ID> when you need to execute a multi-job CI profile directly.

homeboy bench is resource-policy aware. If the current machine is already warm or hot according to homeboy doctor resources, Homeboy prints a stderr warning before running because the extra load can skew benchmark results. Use global --force-hot when running under load is intentional. When a default Lab runner is available for portable bench runs, --force-hot is not a local bypass by itself; add --allow-local-hot only when controller-machine execution is intentional:

homeboy --force-hot --allow-local-hot bench my-component

Memory Timeline Evidence

Each bench runner invocation samples the runner process tree while the workload is active. Successful runs add memory metrics to results.metric_groups.memory and each scenario’s metrics:

• peak_rss_mb: peak sampled RSS for the runner process tree.
• peak_child_count: number of descendant processes at the peak RSS sample.
• memory_sample_count: number of retained timeline samples.

Homeboy also writes bench-memory-timeline.json and bench-memory-timeline.csv in the run directory. Concurrent bench instances use bench-memory-timeline-i<N>.json and .csv. Observation runs record these as bench_memory_timeline artifacts, including failure runs that exit after samples were captured but before benchmark JSON is complete.

Observation History

Every homeboy bench run is persisted to the local observation store. The normal bench output includes hints with the persisted run ID, the observation DB path, and follow-up commands such as:

homeboy runs show <run-id>
homeboy runs list --kind bench --component <component> --rig <rig>

Omit --rig <rig> from the list command for unpinned bench runs.

homeboy runs is the canonical reader for persisted observation-store records. Use these commands for new automation:

homeboy runs list --kind bench --component <component> [--scenario <id>] [--rig <id>] [--limit 20]
homeboy runs distribution --kind bench --component <component> --field <metadata.path> [--scenario <id>] [--rig <id>] [--status <status>] [--limit 20]
homeboy runs bench-compare --from-run <baseline-run-id> --to-run <candidate-run-id> [--metric <name>]

The legacy homeboy bench history <component> reader is a compatibility wrapper over runs list and returns the same runs.list payload. --scenario filters to runs whose stored metadata includes the scenario, and --rig narrows to rig-pinned runs.

The legacy homeboy bench distribution <component> --field <metadata.path> reader delegates to runs distribution and returns the same runs.distribution payload. It is generic over metadata shape: scalar string, number, and boolean values are counted directly, and arrays are flattened. Use --scenario, --rig, --status, and --limit to narrow the persisted run window before aggregation.

The legacy homeboy bench compare --from-run <baseline-run-id> --to-run <candidate-run-id> reader delegates to runs bench-compare and returns the same runs.bench-compare payload. It compares numeric metrics recorded in two persisted benchmark runs, captures both run IDs, component state, shared bench context, selected metric deltas, and includes a Markdown table under reports.markdown in the JSON payload.

Bench runners may attach generic tracker provenance to the top-level results envelope or to individual scenarios/workloads. Homeboy persists that metadata in the run record, keeps it in JSON artifacts/exports, and renders a Provenance section with clickable links in compare Markdown when present:

{
  "provenance": {
    "labels": ["source: zendesk", "privacy: internal"],
    "links": [
      {
        "url": "https://wordpress.org/support/topic/checkout-is-very-slow/",
        "label": "WordPress.org support thread",
        "source": "wordpress.org"
      }
    ]
  }
}

Use labels for non-public or freeform context that should stay as text, such as source: zendesk, scenario: shortcode checkout place-order latency, or privacy: internal. Use links for reviewer-clickable references.

Bench artifacts can attach a viewer descriptor when a file artifact has a browser-based reviewer view. Homeboy keeps link generation generic: descriptors name the viewer kind, base URL, and query parameter whose value is the public artifact URL. Built-in viewer descriptors live in core::artifact_links so new viewer families, such as Codebox artifact viewers, can be added without duplicating URL construction at call sites.

The command is read-only and exits successfully for a valid comparison even when the numbers regress. Repeat --metric <name> to keep the comparison focused; omit it to compare all shared numeric scenario metrics.

Homeboy rejects comparisons when the stored shared benchmark context differs for settings, selected scenarios, workload fingerprints, iteration count, run count, or concurrency. This keeps manually-created baseline/candidate runs from being compared when they were not actually measuring the same scenario contract.

This slice compares two existing persisted run IDs. A future execution mode should wrap two homeboy bench invocations in one command, using the same rig/scenario/settings and separate component paths or refs, then feed the resulting persisted run IDs through this comparison layer.

Scenario Discovery

homeboy bench list <component> asks the extension runner for its scenario inventory without executing any workload code. The runner receives $HOMEBOY_BENCH_LIST_ONLY=1 and writes the normal BenchResults envelope with iterations: 0, empty per-scenario metrics, and optional discovery metadata such as file, source, default_iterations, and tags.

This is the safe first step for agent-driven or CI-driven perf work: inspect what can be measured before deciding which full bench run is worth paying for.

Semantic Gates

Bench runners can attach scenario-level correctness gates to non-timing metrics. Gates are evaluated after metrics are parsed and aggregated. Any failed gate marks the scenario and run failed, even when timing metrics improve, and the failure details are emitted before baseline comparison data in the JSON output.

Supported operators are eq, gte, and lte:

{
  "id": "studio-agent-loop",
  "iterations": 10,
  "metrics": {
    "p95_ms": 1200.0,
    "assistant_message_count": 1,
    "identifies_studio_rate": 1.0
  },
  "gates": [
    { "metric": "assistant_message_count", "op": "gte", "value": 1 },
    { "metric": "identifies_studio_rate", "op": "eq", "value": 1.0 }
  ]
}

Evaluated gates add scenario-level gate_results. Failed gates set the scenario’s passed field to false and add top-level gate_failures plus budget_findings to the bench output.

Bench output also exposes a top-level gate_results array using the shared homeboy/gate-result/v1 schema consumed by cook loops and PR finalization. IDs include the scenario id and metric name, for example bench.gate.studio-agent-loop.success_rate, so multiple scenarios can gate on the same metric without collisions. Failed normalized gates are marked retryable: true and include agent feedback plus metric evidence for the next candidate iteration.

Budget Findings

Benchmark and profile workloads can emit top-level budget_findings for fixed threshold failures that should report and gate consistently across extensions. The common shape is:

{
  "category": "budget",
  "code": "rest.max_response_bytes",
  "severity": "error",
  "file": null,
  "context_label": "profile:wordpress-rest",
  "message": "REST response exceeded 250 KB budget",
  "actual": 4378195,
  "expected": 250000,
  "unit": "bytes",
  "subject": "/wp-json/sampleplugin/v1/pipelines?per_page=100",
  "passed": false
}

severity: "error" or passed: false fails the bench run. Lower severities are report-only. code is the stable grouping key, subject identifies the endpoint/resource/phase being measured, and actual / expected / unit are rendered by homeboy report failure-digest.

Metric Policy Presets

Bench runners can declare generic metric_policy_presets when they want common performance policy semantics without hand-writing full metric_policies or gates. Homeboy expands presets during result parsing into the existing policy, gate, and budget_findings paths.

{
  "metric_policy_presets": {
    "agent_loop_ms": {
      "preset": "latency_regression",
      "regression_threshold_percent": 7.5,
      "phase": "warm"
    },
    "peak_rss_bytes": {
      "preset": "memory_regression"
    },
    "rest_response_bytes": {
      "preset": "absolute_budget",
      "max": 250000
    },
    "coverage_ratio": {
      "preset": "min_coverage",
      "min": 0.8
    },
    "failure_rate": {
      "preset": "max_failure_rate",
      "max": 0.02
    },
    "blocked_rate": {
      "preset": "max_blocked_rate",
      "max": 0.05
    },
    "critical_findings": {
      "preset": "max_critical_findings",
      "max": 0
    }
  }
}

Supported presets are latency_regression, memory_regression, cold_warm_delta, flake_noise_threshold, absolute_budget, min_coverage, max_failure_rate, max_blocked_rate, and max_critical_findings. Regression and outcome presets expand to metric_policies so saved baselines and ratchets compare the same generic numeric metrics across runs. Outcome presets also accept min or max threshold values and expand those into normal gates, so absolute threshold failures render through the existing budget finding and failure digest surfaces.

Rigs can also declare gates for scenario metrics when the workload output is owned elsewhere. The keys under metric_gates are exact scenario ids:

{
  "bench": {
    "default_component": "studio",
    "metric_gates": {
      "wordpress-is-dead": {
        "native_block_quality_pass": { "equals": 1 },
        "tool_error_count": { "equals": 0 },
        "success_rate": { "equals": 1 }
      }
    }
  }
}

Rig-declared metric gates are merged with workload-emitted gates before Homeboy evaluates scenario status. Supported rig operators are equals, gte, and lte.

Invocation Isolation

Every bench child process receives generic invocation-scoped environment variables, independent of runner implementation:

• HOMEBOY_INVOCATION_CONTEXT_JSON: structured JSON contract for the same invocation context. Prefer this for new extension helpers so they consume one audited Homeboy-owned shape instead of reconstructing context from individual env names.
• HOMEBOY_INVOCATION_ID: stable identifier for this child workload invocation.
• HOMEBOY_INVOCATION_STATE_DIR: private state directory for files that should outlive one internal iteration but remain scoped to this invocation.
• HOMEBOY_INVOCATION_ARTIFACT_DIR: private artifact directory for logs, screenshots, traces, and downloaded/build artifacts.
• HOMEBOY_INVOCATION_TMP_DIR: private temporary directory for project copies, browser profiles, wasm caches, and other scratch state.

The structured context currently has this substrate-agnostic shape:

{
  "id": "inv-0123456789",
  "state_dir": "/tmp/hb-501/0123456789",
  "artifact_dir": "/tmp/hb-501/0123456789.a",
  "tmp_dir": "/tmp/hb-501/0123456789.t",
  "port_range": { "base": 20000, "max": 20007 },
  "named_leases": ["playground-browser-profile"]
}

port_range is omitted when the workload did not request ports. named_leases is omitted when no named leases were requested. The legacy HOMEBOY_INVOCATION_* env vars remain part of the contract for existing extensions and scripts.

Runtime path contract

Invocation directories are placed under a short, platform-aware root so any workload can put a UNIX domain socket or other path-length-sensitive primitive under HOMEBOY_INVOCATION_STATE_DIR without bespoke defense.

Layout. The invocation owns three sibling directories under the runtime root:

• <root>/<short-id> → HOMEBOY_INVOCATION_STATE_DIR (the leaf the workload owns; downstream sockets bind directly here).
• <root>/<short-id>.a → HOMEBOY_INVOCATION_ARTIFACT_DIR
• <root>/<short-id>.t → HOMEBOY_INVOCATION_TMP_DIR

There is no s/a/t subdir layer underneath the short id. The invocation is 1:1 with a single workload run, so an extra workload-id segment under STATE_DIR would burn sockaddr_un budget for no isolation gain. Workloads that need internal subdirs under STATE_DIR can still create them, but they own the path-length budget at that point.

Root selection. In priority order:

• HOMEBOY_INVOCATION_RUNTIME_DIR env override (tests and unusual host configurations).
• /tmp/hb-<uid> on every Unix host when /tmp is a writable directory. The UID suffix keeps root and unprivileged invocations from colliding on one runtime root. macOS apps that respect $TMPDIR get per-user isolation under /var/folders/<14>/T/... (~50 bytes), which leaves no realistic sockaddr_un budget. Anchoring to /tmp saves ~35 bytes of headroom and is writable on every standard macOS / Linux configuration.
• Linux fallback: $XDG_RUNTIME_DIR/hb when set and /tmp is unusable.
• macOS fallback: $TMPDIR/hb when /tmp is unusable.
• Generic fallback: ~/.cache/homeboy/inv (or $XDG_CACHE_HOME/homeboy/inv).

Path components use a short opaque id (~10 hex chars) instead of a full UUID v4. The full UUID is retained inside the on-disk invocation lease for traceability, but is never embedded in path components.

Path budget contract. Homeboy guarantees that HOMEBOY_INVOCATION_STATE_DIR, HOMEBOY_INVOCATION_ARTIFACT_DIR, and HOMEBOY_INVOCATION_TMP_DIR leave at least:

• 48 bytes of headroom under the macOS sockaddr_un sun_path capacity (104 bytes), and
• 32 bytes of headroom under the Linux capacity (108 bytes).

Workloads can append a realistic workload-relative socket name like <workload-id>/daemon/daemon.sock (≈40 bytes) directly under STATE_DIR and bind to it without EINVAL. If $HOME or the configured root are unusually long, Homeboy fails fast at invocation acquisition with a clear error naming sockaddr_un, the platform-specific limit, the available headroom, and the HOMEBOY_INVOCATION_RUNTIME_DIR override — instead of letting a downstream workload hit the limit at bind time.

Rig workload object entries can request optional shared-machine primitives:

{
  "bench_workloads": {
    "nodejs": [
      {
        "path": "${package.root}/bench/playground-server.bench.mjs",
        "port_range_size": 8,
        "named_leases": ["playground-browser-profile"]
      }
    ]
  }
}

When port_range_size is set, Homeboy allocates a non-overlapping local port range for each child invocation and exports HOMEBOY_INVOCATION_PORT_BASE and HOMEBOY_INVOCATION_PORT_MAX. Leases are persisted under Homeboy’s local config directory while the child is running, guarded by a local index lock, and stale PID leases are pruned on the next allocation.

named_leases are for truly shared machine-local resources that cannot be namespaced with dirs or ports. Conflicts fail before the workload starts and name the held lease and holder invocation when available.

These primitives are intentionally generic enough for WordPress Playground-style benchmarks: one invocation can run browser/server/wasm/node services, consume multiple local ports, keep downloaded or built artifacts in the artifact dir, and use private temp project dirs without Studio-specific namespace code.

Examples

# Benchmark a component with defaults (10 iterations, 5% regression threshold)
homeboy bench my-component

# Inspect the persisted observation from a completed bench run
homeboy runs show <run-id>

# List related persisted bench runs for comparison loops
homeboy runs list --kind bench --component my-component --rig studio-trunk

# List declared scenarios without executing them
homeboy bench list my-component

# 50 iterations, stricter 2% regression threshold
homeboy bench my-component --iterations 50 --regression-threshold 2.0

# Save a new baseline
homeboy bench my-component --baseline

# Run with auto-ratchet on improvement
homeboy bench my-component --ratchet

# Select a single scenario
homeboy bench my-component --scenario hot_path

# Select two scenarios in a cross-rig comparison
homeboy bench studio --rig studio-trunk,studio-branch 
    --scenario studio-agent-runtime 
    --scenario wp-admin-load

# Share warm state across invocations and run four instances in parallel
homeboy bench my-component --shared-state /tmp/homeboy-bench --concurrency 4

# Pin to a single rig — preflight + rig-scoped baseline
homeboy bench studio --rig studio-trunk

# Pin to one rig and run every component declared in bench.components
homeboy bench --rig mdi-substrates --shared-state /tmp/mdi-bench

# Cross-rig comparison: same workload, two rigs, side-by-side report.
# First rig (`studio-trunk`) is the reference; the diff table expresses
# every other rig's metrics as percent deltas vs the reference.
homeboy bench studio --rig studio-trunk,studio-combined-fixes 
    --iterations 10 
    --report side-by-side

# Opt into parallel cross-rig execution for side-by-side exploratory runs.
homeboy bench studio --rig studio-agent-sdk,studio-bfb 
    --scenario studio-agent-site-build 
    --rig-concurrency 2

# Three-rig comparison to isolate one PR's contribution.
homeboy bench studio 
    --rig trunk,combined-fixes,combined-fixes-without-3120 
    --iterations 20

Cross-rig comparison

--rig <a>,<b>[,<c>...] runs the same component + workload + iteration count against each rig and emits a single comparison envelope. By default, rigs run in sequence for stable CI behavior. Pass --rig-concurrency <N> to opt into bounded parallel rig execution for exploratory or product-demo workflows where fresh isolated sites should be built in the same wall-clock window.

How it runs

For each rig, in input order by default, or in bounded parallel batches when --rig-concurrency is greater than 1:

1. Load the rig spec and run rig check. Failure aborts the entire comparison — comparing against an unhealthy rig would produce garbage numbers.
2. Snapshot rig state (each component’s git SHA + branch) into the per-rig output entry.
3. Run bench against the resolved component with the rig pinned.

After every rig finishes, results are aggregated into a BenchComparisonOutput envelope with comparison: "cross_rig". The first rig in the list is the reference: per-metric percent deltas in the diff table express each subsequent rig as (current - reference) / reference * 100.

Multi-rig comparison envelopes include reports.side_by_side, a compact report artifact for demo and product harnesses. It references every rig result and surfaces:

• per-rig status, exit code, and failure reason
• elapsed time when elapsed_ms or duration_ms metrics are present
• flattened key metrics, including grouped metrics like prompt.hash_match
• artifact paths and URLs, including URL-looking artifact paths

Parallel execution preserves the selected rig order in the output envelope, so the reference rig and diff interpretation do not change when concurrency is enabled.

What’s intentionally not done

• No baseline writes. --baseline and --ratchet are rejected on cross-rig invocations. Baselines are per-rig; writing one from a comparison would silently bless one rig over the others. Run homeboy bench --rig <id> --baseline once per rig to ratchet individually.
• No statistical-significance gating. Two rigs with overlapping p95_ms distributions still produce a numeric delta. Treat single-digit percent moves with skepticism.

Rig bench defaults

Rig specs can reduce repeated CLI arguments for common main-vs-branch bench workflows:

{
  "bench": {
    "default_component": "studio",
    "components": ["studio", "playground"],
    "default_baseline_rig": "studio-trunk"
  },
  "bench_workloads": {
    "wordpress": [
      { "path": "${package.root}/bench/studio-admin.php" }
    ]
  }
}

• bench.default_component lets homeboy bench --rig <id> omit the positional component. With multiple rigs, every rig must agree on the default unless the component is provided explicitly.
• bench.components lets homeboy bench --rig <id> fan out across a list of components from one rig spec. Scenarios are merged into the standard single-run envelope with :c<component> suffixes (for example cold-boot:cstudio). When --shared-state <dir> is provided, each component gets its own <dir>/<component> subdirectory.
• bench.default_baseline_rig upgrades homeboy bench --rig <candidate> into homeboy bench --rig <baseline>,<candidate> unless the invocation already lists multiple rigs, writes a baseline (--baseline / --ratchet), passes --ignore-default-baseline, or the candidate rig declares a multi-component bench.components matrix.
• bench_workloads supplies rig-owned workload files keyed by extension ID. Entries must be objects with a path field. Paths support ~, ${env.NAME}, ${components.<id>.path}, and ${package.root} expansion. ${package.root} resolves to the installed rig package root, so portable rig packages can keep workload files next to the rig spec without hardcoded machine paths.

Output shape (cross-rig)

{
  "comparison": "cross_rig",
  "passed": true,
  "component": "studio",
  "exit_code": 0,
  "iterations": 10,
  "rigs": [
    {
      "rig_id": "studio-trunk",
      "passed": true,
      "status": "passed",
      "exit_code": 0,
      "artifacts": [
        {
          "scenario_id": "agent_boot",
          "run_index": 0,
          "name": "raw_result",
          "path": "bench-artifacts/agent_boot/run-0/raw-result.json",
          "kind": "json",
          "label": "Raw result"
        }
      ],
      "results": { ... },
      "rig_state": { "rig_id": "studio-trunk", "captured_at": "...", "components": { ... } }
    },
    {
      "rig_id": "studio-combined-fixes",
      "passed": true,
      "status": "passed",
      "exit_code": 0,
      "results": { ... },
      "rig_state": { ... }
    }
  ],
  "diff": {
    "by_scenario": {
      "agent_boot": {
        "p95_ms": {
          "studio-combined-fixes": {
            "reference": 31200.0,
            "current": 19400.0,
            "delta_percent": -37.82
          }
        }
      }
    }
  },
  "hints": [ ... ]
}

The reference rig is omitted from the inner diff.by_scenario.<id>.<metric> map — its delta against itself would always be zero. A scenario or metric missing from a non-reference rig is silently skipped (no synthetic zeros).

Each rig entry also includes an artifacts index when workloads emit artifact pointers. The full-fidelity data remains nested under results.scenarios[].artifacts and results.scenarios[].runs[].artifacts, but the index makes proof artifacts easy to find in cross-rig output. run_index is zero-based and omitted for scenario-level artifacts that are not tied to a specific --runs iteration.

Exit code

exit_code is 0 only when every rig passed. The first non-zero rig exit code wins. passed is true only when every rig passed.

Baseline Ratchet Semantics

The bench baseline is a list of per-scenario snapshots stored in homeboy.json under the baselines.bench key. Each snapshot records { id, metrics } plus the iteration count at capture time.

On every run without --baseline or --ignore-baseline:

1. Each current scenario is matched against the baseline by id.
2. If the runner declares metric_policies, only those metrics are compared. Each policy declares whether lower or higher values are better and optional percent/absolute tolerances.
3. If a policy declares variance_aware: true, Homeboy compares the metric’s raw sample distributions instead of only the summary value. The summary value still appears under metrics.<name> for reports; the per-iteration samples live under metrics.distributions.<name>.
4. If the runner omits metric_policies, Homeboy keeps the historical default: compare p95_ms as lower-is-better with the CLI threshold.
5. A scenario improves when any compared metric moves in the better direction.
6. Scenarios present in one run but not the other are flagged as new_scenario_ids / removed_scenario_ids. Neither state triggers a regression by itself — they’re informational.
7. If any scenario regressed, the command exits 1 regardless of the runner’s own exit code.
8. If any scenario improved and --ratchet is set, the baseline is overwritten with the current snapshot.

p95 remains the default for legacy latency benchmarks because it is less sensitive than mean to one-off GC pauses but more sensitive than p99 to genuine regressions. Runners that care about non-latency signals should declare metric_policies instead.

Runner Contract

The extension’s bench script must:

1. Read $HOMEBOY_BENCH_ITERATIONS to determine iteration count.
2. Write its JSON output to $HOMEBOY_BENCH_RESULTS_FILE.
3. Exit with a non-zero status only on runner-level failure (script error, workload crash) — regressions are homeboy’s domain.

JSON output schema

{
  "component_id": "string",
  "iterations": 10,
  "metric_policies": {
    "error_rate": {
      "direction": "lower_is_better",
      "regression_threshold_absolute": 0.01
    },
    "requests_per_second": {
      "direction": "higher_is_better",
      "regression_threshold_percent": 5.0
    },
    "agent_loop_ms": {
      "direction": "lower_is_better",
      "regression_threshold_percent": 10.0,
      "variance_aware": true,
      "min_iterations_for_variance": 20,
      "regression_test": "mann_whitney_u"
    }
  },
  "scenarios": [
    {
      "id": "scenario_slug",
      "file": "tests/bench/some-workload.ext",
      "iterations": 10,
      "metrics": {
        "mean_ms": 120.3,
        "p50_ms": 118.0,
        "p95_ms": 145.0,
        "p99_ms": 160.0,
        "min_ms": 110.0,
        "max_ms": 172.0,
        "error_rate": 0.0,
        "requests_per_second": 180.5,
        "status_500_count": 0,
        "agent_loop_ms": 1200.0,
        "distributions": {
          "agent_loop_ms": [1100.0, 1200.0, 1300.0]
        }
      },
      "memory": { "peak_bytes": 41943040 },
      "artifacts": {
        "raw_result": {
          "path": "bench-artifacts/scenario_slug/raw-result.json",
          "kind": "json",
          "label": "Raw result"
        }
      }
    }
  ]
}

• Top-level keys are strict — unknown top-level fields are rejected to keep the contract honest.
• metrics is an arbitrary map of numeric values. Homeboy core does not attach domain meaning to metric names.
• metric_policies is optional. If omitted, Homeboy compares p95_ms using the legacy lower-is-better latency policy.
• Policy direction accepts lower_is_better / lower and higher_is_better / higher.
• Policy thresholds are optional. regression_threshold_percent compares relative movement; regression_threshold_absolute compares raw numeric movement. If both are present, a metric must exceed both tolerances to regress.
• Policy variance_aware: true requires a matching metrics.distributions.<metric> array on every scenario that emits the metric. If min_iterations_for_variance is set and the sample array is smaller, parsing fails before baseline comparison.
• Policy regression_test accepts point_delta, mann_whitney_u, and kolmogorov_smirnov. point_delta is the legacy summary-value check. Variance-aware metrics default to mann_whitney_u when the field is omitted. Mann-Whitney uses a one-sided 95% normal approximation; Kolmogorov-Smirnov uses the standard 5% two-sample critical value.
• Scenario-level unknown keys are tolerated, so extensions can emit additional metadata (tags, environment info, warmup counts) without breaking parsing.
• Scenario id values must be unique within one bench results envelope. Workload-discovering runners should derive ids from paths relative to the bench root (for example, reads/heavy.php → reads-heavy) instead of file basenames alone.
• memory is optional. Extensions that can’t measure peak memory omit it.
• file is optional but recommended for diagnostics.
• artifacts is optional. Values are local paths plus optional kind and label metadata. Homeboy preserves and indexes these pointers but does not upload, retain, or diff artifact contents.

Artifact kind conventions

kind is an open string, not a closed enum. Homeboy preserves unknown kinds so extensions can add new evidence types without requiring a core release. Use these conventional values when they fit:

• json — structured machine-readable data, summaries, or transcripts.
• directory — a directory containing multiple related files.
• playwright-trace — Playwright trace archive, usually a .zip file.
• screenshot — browser or UI screenshot evidence.
• network-log — captured request/response metadata such as HAR or JSONL.
• console-log — browser console output or page-level JavaScript logs.

For browser benchmarks, prefer specific labels such as Playwright trace, Final screenshot, Network log, or Console log. The JSON result keeps the full artifact map under each scenario, and Homeboy also promotes artifact pointers into the top-level artifacts index. Markdown reports render those labels and paths so users can find trace and screenshot files without opening the raw JSON payload.

Environment variables injected

Bench scripts receive the standard runner contract plus bench-specific variables:

• HOMEBOY_BENCH_RESULTS_FILE — where to write JSON output.
• HOMEBOY_BENCH_ITERATIONS — iteration count to use.
• HOMEBOY_BENCH_RESPONSIVENESS_FILE — optional JSONL file for workload responsiveness pings. Workloads opt in by appending pings such as { "at": "2026-06-08T00:00:00Z", "t_ms": 1200 }; Homeboy summarizes missed_ping_count, max_ping_gap_ms, and last_ping_at in bench output.
• HOMEBOY_BENCH_RESPONSIVENESS_MISSED_MS — ping gap threshold used to classify missed responsiveness windows. Defaults to 10000.
• HOMEBOY_RUN_DIR — per-run directory (shared with test/lint/build).
• HOMEBOY_EXTENSION_ID, HOMEBOY_COMPONENT_ID, HOMEBOY_COMPONENT_PATH, and the usual execution-context vars.
• HOMEBOY_SETTINGS_JSON — component settings as JSON.

Component Requirements

For a component to be benchmarkable, it must have:

• A linked extension whose manifest declares a bench capability.
• A bench-runner script provided by the extension.

Extension manifest:

{
  "bench": {
    "extension_script": "scripts/bench/bench-runner.sh"
  }
}

Exit Codes

• 0 — All scenarios passed, no regressions detected (or no baseline exists yet).
• 1 — At least one scenario regressed beyond the threshold, or the runner itself failed.
• Other non-zero — Runner exit code passthrough (extension-specific).

Related

• test — Test sibling capability; bench mirrors its flag conventions for --baseline, --ignore-baseline, and --ratchet.
• lint — Lint sibling capability.
• build — Build sibling capability.