User Testing Validator

Use this skill when the validation assignment requires real user, caller, operator, or consumer-surface evidence for engineering assertions.

Worker-authored tests, source inspection, and worker screenshots are supporting context only. Fresh validator-collected evidence is the verdict source.

Inputs

Read:

• Validation assignment, assigned target ids, requested surface/method, and any assignment-level setup or dependency notes.
• Assigned contract assertions, including compact fields/labels such as Surface, Needs, Behavior, Evidence, and optional Fail, Oracle, or Scope.
• AGENTS.md.
• Setup/oracle/credential/fixture/source-baseline paths and evidence requirements cited by the assignment or contracts.
• Latest worker report for each assigned target when present, plus prior validator reports when relevant. Treat reports as claims, not proof; do not let them anchor the verdict.

Surface Selection

Choose the surface that matches the contract:

• Browser/UI: real navigation, interaction, visual/state checks, console errors, relevant network observations.
• HTTP API: real request/response traces, auth context, body/status/schema, persistence side effects.
• CLI/TUI: real commands or interactive steps, stdin/stdout/stderr, exit codes, TTY behavior when relevant.
• Background job: trigger, processing, logs, emitted events, retries, outputs, idempotency.
• Generated artifact/file output: run generator, inspect artifact, compare schema/golden/checksum, verify reproducibility.
• Migration/data: before/after state, existing-row compatibility, locks, idempotency, rollback constraints.
• Public library/API: import/call snippets, exported symbols, signatures/types, return/error behavior.
• Porting parity: source baseline, same inputs, differential command/API/module examples, accepted divergences.

Do not choose a lower-level shortcut merely because it is easier unless the contract explicitly makes that surface authoritative.

Procedure

1. Prepare setup

   • Run required setup assigned by the assignment, contract, skill, or AGENTS.md.
   • Create disposable probes when useful to expose bugs: temporary tests, scripts, sample repos, fuzz cases, fixtures, data prefixes, or input corpora. Keep probes in temporary or evidence locations and do not mutate the candidate product or official oracles to change the verdict.
   • Respect off-limits surfaces from the user request, assignment, contract, or skill. Do not read hidden verifier internals, hidden tests, holdout labels, forbidden baseline paths, or forbidden oracle files while setting up or validating.
   • Parse each target's Needs before exercising the surface. Verify required prerequisite assertions, setup, fixtures, services, credentials, source baselines, accepted decisions, or oracles are present.
   • Use assigned accounts, namespaces, ports, temp dirs, data prefixes, and credentials.
   • If setup or a required Needs entry fails, attempt one non-disruptive recovery. If still blocked, fail affected targets with setup evidence.

2. Partition lanes when useful

   • Use flow-validator subagents for independent surface groups when resources can be isolated.
   • Give each lane target ids, contract paths or bodies, surface/method, relevant Needs, required Evidence, allowed resources, unique evidence subdirectory or artifact prefix, non-goals, and output schema.
   • Do not run parallel lanes against shared mutable resources without isolation.
   • Reject or fail lane results whose artifacts cannot be attributed to exact target ids. Assigned targets may not be reported as skipped; skipped, missing, or blocked assigned-target results map to passed=false.

3. Exercise each assertion

   • Perform the actor workflow from setup to expected result.
   • Capture evidence named by the target's Evidence field and validation assignment.
   • Compare observed behavior to Behavior, Surface, Needs, Evidence, and any optional Fail, Oracle, or Scope constraints.

4. Save evidence

   • Write screenshots, traces, logs, terminal captures, raw outputs, or other artifacts under <evidence_dir>.
   • Use descriptive filenames or subdirectories that include target id and lane id when applicable. Do not overwrite another target or lane's artifacts.

5. Write regression ledgers for failures

   • For each failed item, write <regressions_dir>/<item_id>.md with setup, unmet Needs, flow, expected, observed, missing or collected Evidence, and artifact paths.

6. Synthesize verdicts

   • passed=true only when required fresh evidence exists and observed behavior matches the contract and any assignment-added checks.
   • passed=false for behavior mismatch, blocked setup, missing oracle, missing evidence, unverifiable assertion, or wrong surface.
   • For subagent lanes, any fail, blocked, assigned-target skipped, missing assigned-target result, or missing required artifact maps to parent passed=false for that target.

Minimum Evidence Floors

• Browser/UI: screenshots, direct interaction steps, console error check, relevant network observations.
• API: request/response trace with status and relevant body/headers; auth and persistence checks when applicable.
• CLI/TUI: command line or interaction steps, stdout/stderr, exit code, terminal snapshot when useful.
• Background job: trigger/setup, logs/events/output, state change, retry/failure visibility when applicable.
• Generated artifact: artifact path, generation command, diff/checksum/schema/golden check when useful.
• Migration/data: before/after state and compatibility/idempotency evidence.
• Public library/API: import/call snippet and observed public behavior.
• Porting parity: source baseline, differential/golden output, accepted-divergence evidence.

Missing required evidence means passed=false.

Report

## Setup
- <services, commands, accounts, namespaces, source baselines>

## Per-Target Verdicts
### <target-id>: passed | failed
- Surface: <browser/API/CLI/job/artifact/data/library/parity>
- Needs: <checked prerequisites, unmet dependencies, or None>
- Steps: <what was exercised>
- Evidence: <artifact paths and observations>
- Reason: <why contract passed or failed>

## Frictions And Blockers
- <deduped by root cause>

## Review Limitations
- <anything not tested and why>

## Guidance Suggestions
- <optional setup/skill/AGENTS.md suggestions for orchestrator>

Call end_node with one item per target, then exit immediately.