QA Sweep

QA & Independent Review Protocol

Follow this protocol on every implementation, fix, debugging, or review task.

Proportionality

Scale the depth of each phase to risk — never skip a gate for changed work, but match its weight to the task:

• Read-only or answer-only work (explaining code, reading logs, answering a question) with no worktree edit: Phases 2–3 are not required; verify claims against the actual source before answering.
• Any worktree edit (code, tests, docs, package metadata, release notes, skill files): Phases 2 and 3 are mandatory. For a small, low-risk edit, one fresh review agent covering both the adversarial and completeness checklists is acceptable; for high-risk or cross-file work, keep them separate.
• High-risk work (security, auth, isolation, IPC contracts, payments, migrations, concurrency): full protocol, no consolidation.

This proportionality applies only to qa-sweep's own Phase 2/3 passes. When swarm mode is enabled, the swarm-mode contract's separate independent implementation reviewer and final critic gates apply unreduced to any changed work — consolidation here never merges or replaces those gates.

For agent-type, model, and effort selection when spawning these sub-agents, load the orchestrating-subagents skill: economize on explorers, never on reviewers.

If no subagent tool is available

If this protocol executes in a context without a subagent tool (Agent or Task) — check your actual tool list rather than assuming — perform the Phase 2/3 checklists yourself as a clearly labeled fallback self-review and disclose in your report that independent review was unavailable, so the orchestrator can re-run the gate with a real fresh agent. Never present self-review as independent review.

Phase 1 — Parallel Implementation

• Use parallel sub-agents to speed up independent units of work wherever possible.
• Each sub-agent must read relevant source code end-to-end before making changes.
• Reference official documentation to verify whether any behavior is intended before treating it as a bug.
• Do not trust assumptions — prove every behavior against actual code.

Pre-implementation static-gate baseline

Before the first coder dispatch in any phase, capture a static-gate baseline so findings are scoped to the phase, not to pre-existing code:

• Defining the pre-phase file set: the canonical source is the file list passed to declare_scope({ taskId, files }) (see the swarm-implement skill for the discipline). If declare_scope was called with the task's exact file list, that IS the pre-phase file set. If declare_scope was not called (e.g., the file list is not 100% obvious), use the containing directories declared in the previous call, OR fall back to the task's files_touched array in the plan (save_plan writes this). Never use the entire repository as the pre-phase file set — that defeats the point of phase scoping.
• SAST: if sast_scan is available with capture_baseline:true, invoke it once at phase open against the pre-phase file set. Subsequent phase-scoped scans only fail on NEW findings. If swarm tools are unavailable, document the gap in the closure report and skip the baseline capture (do not silently fall back to a project-wide scan that fires on every historical finding).
• placeholder_scan: no capture-baseline mode, but it is diff-aware — pass added_lines (file path → phase-added line numbers) so only newly added lines drive the verdict; a file omitted from the map is scanned unfiltered (fail-closed), so cross-check its findings against the changed lines. For intentionally retained TODO/FIXME/HACK comments, pre-configure allow_globs (technical debt tracked elsewhere). Do not delete real placeholders to silence the gate; add them to the allowlist.
• Re-baseline if co-change detection expands the file list after the initial baseline — a missed re-baseline will mis-attribute pre-existing findings to the phase.

Acceptance tests first

Before running any broader test suite, run the smallest test set that covers the task's acceptance criteria. Full-suite passes can mask targeted failures when ordering or concurrency changes — exercise each FR's acceptance criterion explicitly first.

Phase 2 — Independent Adversarial Review (Mandatory)

After implementation, spawn a FRESH sub-agent that has not participated in any prior work. Give it this directive verbatim:

"Assume all work done by the implementing agent is incorrect until you can prove otherwise with absolute evidence from the actual code. The implementing agent makes frequent mistakes and tends to miss edge cases. Do not trust any claim without tracing it yourself. Review every change, test, and edge case end-to-end through the real source."

The review agent must:

• Independently trace each change end-to-end through the codebase
• Search for related issues and regressions the implementing agent may have introduced
• Verify documented behavior vs. actual code behavior
• Surface every edge case not explicitly covered

Timing requirement: Phase 2 must complete and all confirmed findings must be addressed before the commit you intend as the final substantive push. Do not defer this to "after CI passes" — CI passing on a buggy commit does not retroactively make the review optional. For high-risk work (security, isolation, IPC contracts, auth, payments), this is a hard gate with no exceptions.