diff --git a/adr/20260514-benchmark-process-directive.md b/adr/20260514-benchmark-process-directive.md
new file mode 100644
index 0000000000..1fc8fe70e9
--- /dev/null
+++ b/adr/20260514-benchmark-process-directive.md
@@ -0,0 +1,272 @@
+# `benchmark` process directive for per-task performance metrics
+
+- Authors: Edmund Miller
+- Status: draft
+- Deciders:
+- Date: 2026-05-14
+- Tags: directive, trace, metrics, benchmark
+
+## Summary
+
+Introduce a `benchmark` process directive that emits per-task runtime metrics
+(wall time, peak memory, CPU usage, I/O) to a user-specified file, with an
+optional `repeats` option for noise-reduction via repeated execution.
+
+## Problem Statement
+
+Nextflow already collects rich per-task runtime metrics through `TraceRecord` —
+`realtime`, `peak_rss`, `peak_vmem`, `%cpu`, `rchar`, `wchar`, context switches,
+and roughly 40 other fields — and surfaces them through the workflow-wide
+`trace.txt`, the HTML `report.html`, and `timeline.html`. These artifacts are
+useful for whole-pipeline post-mortem analysis but are not designed for
+per-process benchmarking workflows:
+
+- The trace file mixes all tasks across the whole run; isolating measurements
+  for a single process requires post-processing.
+
+- There is no first-class way to capture a per-task metrics artifact at a
+  stable, predictable path that downstream processes can consume as input.
+
+- There is no built-in support for repeated execution of a task to obtain
+  multiple samples and reduce measurement noise — a common requirement when
+  benchmarking algorithms or hardware.
+
+- Users who want to benchmark a third-party pipeline (e.g. nf-core workflows)
+  currently have no way to enable metrics capture for a specific process
+  without forking the pipeline source.
+
+Snakemake addresses the same need with a `benchmark` rule keyword that records
+wall time, memory, I/O and CPU usage to a per-job TSV (or JSONL) file, with an
+optional `repeat(file, N)` wrapper for multiple samples. Nextflow has no
+equivalent.
+
+## Goals
+
+- Provide a declarative, per-process way to capture performance metrics as a
+  named pipeline artifact.
+
+- Make the directive settable from `nextflow.config` (via `withName` and
+  `withLabel` selectors) so users can benchmark an existing pipeline without
+  modifying its source.
+
+- Reuse the existing `TraceRecord` collection — no new probes, no changes to
+  the bash wrapper.
+
+- Support repeated execution by spawning N real task executions and aggregating
+  their metrics, not by looping inside a single task.
+
+- Make the benchmark file a regular workflow artifact with a deterministic
+  schema, usable as input to downstream processes.
+
+- Multi-format output driven by file extension (`.tsv` by default,
+  `.jsonl` opt-in).
+
+## Non-goals
+
+- Aggregating benchmarks across tasks — the existing `trace.txt` and
+  `report.html` already cover whole-workflow summaries.
+
+- Collecting metrics beyond what `TraceRecord` already records. Snakemake's
+  `max_uss` and `max_pss` are not exposed by Nextflow's `nxf_stat` probe and
+  are documented as gaps rather than added in this ADR.
+
+- Reliable benchmarking inside `exec:` blocks. As with Snakemake's `run:`
+  body, the Groovy-side execution time is intertwined with Nextflow's process
+  supervision and cannot be measured cleanly.
+
+- Changes to the trace, report, or timeline subsystems.
+
+## Decision
+
+Introduce a `benchmark` process directive that accepts either a path string or
+a map of options. Each task writes its own metrics file at the resolved path.
+When `repeats: N` is set, Nextflow submits N independent task executions and
+aggregates their `TraceRecord`s into a single output file.
+
+## Core Capabilities
+
+### Syntax
+
+The `benchmark` directive accepts a string shorthand or a map of options,
+mirroring the `publishDir` convention:
+
+```groovy
+// process definition — shorthand
+process align {
+    cpus 8
+    memory '16 GB'
+    benchmark "benchmarks/align/${sample}.tsv"
+
+    input:
+    tuple val(sample), path(reads)
+
+    script:
+    """
+    bwa mem -t ${task.cpus} ref.fa ${reads} > out.bam
+    """
+}
+```
+
+```groovy
+// process definition — map form
+process align {
+    benchmark file: "benchmarks/align/${sample}.jsonl", repeats: 3
+    // ...
+}
+```
+
+The path is templated with task variables (`task.process`, `task.tag`,
+`task.hash`, plus any input variables in scope) and resolved relative to
+`workflow.launchDir`, the same as `publishDir`.
+
+### Config-level usage
+
+Because `benchmark` is registered in the standard process directive list, it
+is settable from `nextflow.config` via the usual `withName` and `withLabel`
+selectors. This is a primary use case: benchmarking an existing pipeline (for
+example, an nf-core workflow) without editing its source.
+
+```groovy
+// nextflow.config — benchmark one named process, 5 repeats
+process {
+    withName: 'ALIGN' {
+        benchmark = [file: "benchmarks/align/${task.tag}.tsv", repeats: 5]
+    }
+}
+```
+
+```groovy
+// nextflow.config — benchmark every process matching a label
+process {
+    withLabel: 'heavy' {
+        benchmark = "benchmarks/${task.process}/${task.tag}.tsv"
+    }
+}
+```
+
+```groovy
+// nextflow.config — benchmark all processes (high cost; documented caveat)
+process {
+    benchmark = [file: "benchmarks/${task.process}/${task.tag}.jsonl", repeats: 3]
+}
+```
+
+### Options
+
+| Option | Type | Default | Description |
+|--|--|--|--|
+| `file` | String (templated) | _required_ | Destination path. Templated with task variables; resolved against `workflow.launchDir`. |
+| `repeats` | Integer | `1` | Number of independent task executions. See _Repeat semantics_ below. |
+| `fields` | List\<String\> | sensible default subset of `TraceRecord.FIELDS` | Which columns to emit. Unknown names are rejected at parse time. |
+| `format` | `'tsv'` \| `'jsonl'` | inferred from `file` extension | Explicit override when the extension is ambiguous. |
+
+### Schema
+
+Column names match `TraceRecord.FIELDS` directly. This keeps benchmark files
+consistent with `trace.txt` and `report.html` so users can correlate them
+without translation.
+
+Default columns:
+
+`task_id`, `hash`, `process`, `tag`, `attempt`, `realtime`, `%cpu`,
+`peak_rss`, `peak_vmem`, `rchar`, `wchar`, `vol_ctxt`, `inv_ctxt`
+
+Users can override via `fields:`. Any name in `TraceRecord.FIELDS` is
+accepted.
+
+TSV files have a header row followed by one data row per repeat. JSONL files
+contain one JSON object per repeat, with keys matching the column names.
+
+### Repeat semantics
+
+`repeats: N` does not loop inside the wrapper script. Nextflow submits N
+independent task executions (each with its own workdir, its own
+`.command.trace`, its own `TraceRecord`), and aggregates them into a single
+benchmark file. This stays within Nextflow's existing execution model and
+avoids introducing new wrapper-script behavior.
+
+Implications:
+
+- Each repeat is a full, isolated task run — independent scheduling, retry,
+  and failure handling under the configured `errorStrategy`.
+
+- The benchmark file is written exactly once per logical task invocation,
+  with N rows (TSV) or N JSON records (JSONL), one per repeat. Each row's
+  `task_id` and `hash` identifies the underlying execution so users can
+  cross-reference with `trace.txt`.
+
+- For output channels, the first successful repeat's outputs are emitted
+  downstream; other repeats' workdirs are retained on disk for inspection but
+  their outputs are not emitted. This avoids inventing new semantics around
+  duplicate outputs and matches Snakemake's behavior where repeated runs do
+  not change the rule's output graph.
+
+- All N repeats must reach a terminal state (success or terminal failure)
+  before the benchmark file is written.
+
+### Field coverage compared to Snakemake
+
+Snakemake's columns map cleanly onto Nextflow's native fields. The ADR
+chooses to expose Nextflow names directly rather than aliasing, but the
+mapping is recorded here for users porting workflows:
+
+| Snakemake | Nextflow `TraceRecord` |
+|--|--|
+| `s` | `realtime` |
+| `h:m:s` | derived from `realtime` |
+| `max_rss` | `peak_rss` |
+| `max_vms` | `peak_vmem` |
+| `max_uss`, `max_pss` | not collected — gap, documented |
+| `io_in` | `rchar` |
+| `io_out` | `wchar` |
+| `mean_load`, `cpu_time` | derived from `%cpu` and `realtime` |
+
+### Cloud and remote executors
+
+`TraceRecord` collection already works for AWS Batch, Google Batch, and
+Kubernetes through the existing bash wrapper, so no executor-specific work
+is required. The benchmark file is materialized to the resolved path using
+the same path-handling code that `publishDir` uses, which supports remote
+launch directories (S3, GCS, Azure Blob) without special-casing.
+
+## Implementation surface
+
+The change is small and self-contained:
+
+- `modules/nextflow/src/main/groovy/nextflow/script/dsl/ProcessBuilder.groovy` —
+  add `'benchmark'` to `DIRECTIVES`.
+
+- `modules/nextflow/src/main/groovy/nextflow/script/dsl/ProcessConfigBuilder.groovy` —
+  parse the string-shorthand and map forms.
+
+- New `modules/nextflow/src/main/groovy/nextflow/processor/Benchmark.groovy` —
+  mirrors `PublishDir`; fields: `file`, `repeats`, `fields`, `format`.
+
+- `modules/nextflow/src/main/groovy/nextflow/processor/TaskProcessor.groovy` —
+  after all repeats complete, serialize the collected `TraceRecord`s to
+  TSV/JSONL at the resolved path.
+
+- `modules/nextflow/src/main/groovy/nextflow/trace/TraceRecord.groovy` — no
+  schema changes; `FIELDS` is reused.
+
+- `docs/reference/process.md` — new `benchmark` section in the directive
+  reference.
+
+No edits to `command-trace.txt` or any wrapper template.
+
+## Open questions
+
+- Exact mechanism for fanning out N repeats: reuse the existing
+  attempt/retry path (treating repeats as forced retries with success) or a
+  dedicated fan-out scheduler hook. The ADR commits to the user-visible
+  contract; the mechanism is left to follow-up implementation design.
+
+- Whether to warn when a workflow-level `benchmark` directive is combined
+  with caching (`cache true`) — repeated runs against a cached task will
+  return cached metrics rather than fresh measurements. Tentative default:
+  emit a warning the first time a benchmarked task hits a cache.
+
+## Links
+
+- Snakemake benchmark rules: <https://snakemake.readthedocs.io/en/stable/snakefiles/rules.html#benchmark-rules>
+- Related directive ADR: [hints process directive](20260323-hints-process-directive.md)