name: vool-cumulative-testing description: "Run and interpret the cumulative test pack: the full suite plus the pack's own bounded selection, attributing every failure to its change or to a pre-existing base failure. Use when the user asks to run the tests, the pack, or cumulative verification." version: 1.0.0 license: MIT author: sls_0x status: p1 type: capability_skill triggers: [cumulative, test pack, suite, run the tests, verification pack] allowed-tools: [workspace.run_tests, workspace.run_lint, workspace.run_formatter, workspace.read_file] capabilities: [workspace.read, workspace.validate] permissions: [read_files] effects: [validation_command, read_only] inputs: [workspace_root, selection optional] outputs: [test_results, failure_attribution] stop-conditions: [pack executed and every failure attributed or reported unattributed, the harness itself will not run - report that as the finding] recovery: [if a selection is too large, bound it and say what was skipped, if results are flaky, re-run once sequentially and report both outcomes] cumulative-test-law: [tests/native_skills must pass before this package changes, selection phrase "run the cumulative test pack" must keep selecting this package, a green claim without a real executed run is the defect this package exists to prevent] id: cumulative-testing risk-class: read_only task-families: [debugging, workspace_audit] capability-families: [workspace] tool-intents: [workspace.run_tests, workspace.run_lint, workspace.run_formatter, workspace.read_file] permitted-tools: [workspace.run_tests, workspace.run_lint, workspace.run_formatter, workspace.read_file] prerequisites: [] expected-outputs: [test_results, failure_attribution] verification: [cumulative_suite_green, deterministic_evidence] stopping-conditions: ["stop when every failure is attributed: pre-existing, in-scope, or flaky — with the evidence for the label"] incompatible-with: [] priority: 35

Cumulative Testing

You run the pack and ATTRIBUTE the results. A green claim without an executed run is the exact defect this package exists to prevent; so is a red suite reported as "mostly passing".

Scope

IN: run the bounded pack (full suite or named selection), attribute each failure (caused by the change / pre-existing at base / flaky with evidence), summarize pass/fail counts exactly. OUT: fixing failures (hand to root-cause-repair), re-running until green (selective re-runs are allowed once, reported as such).

Required capabilities (declaring is not granting)

workspace.read, workspace.validate. Declaring is not permission; the run goes through the runtime permission lane like any execution.

Procedure

1. workspace.read_file the harness config to learn the pack's bounds.
2. workspace.run_tests the named selection — bounded, never unbounded.
3. workspace.run_lint when the change touches style-sensitive surfaces.
4. Attribute every failure: caused / pre-existing / flake, with evidence.
5. Report counts exactly: N passed, M failed, K skipped, attribution list.

Stop conditions

Stop at the attribution report. A harness that will not start is reported as the finding, not worked around silently.

Recovery

Selection too large: bound it and say what was skipped. Flaky result: one sequential re-run, both outcomes reported.

Cumulative test law

tests/native_skills is the pack; testing-behaviour edits re-run it.

No direct execution bypass

Test runs go through the runtime tool door (execute_runtime_tool) and its validation lane with output caps. No direct runner invocation around it.