adversarial-squad

Adversarial Squad Deployment (harness)

The operational HOW. The doctrinal WHEN/WHY is the procedure adversarial-squad-deployment (packs/built-in/procedures/adversarial-squad-deployment.procedure.yaml), which sits under the brownfield-onboarding paradigm. This skill changes no mission type or guard; it is a technique the orchestrator opts into.

When to use

A squad is worth its tokens at a high-leverage point-cut where one reviewer's blind spot is expensive:

• after /spec-kitty.specify → pre-spec investigation (scope, prior art, live repros)
• after /spec-kitty.plan → post-planning brownfield check (foldable issues, split-brain, deprecations)
• after /spec-kitty.tasks → post-tasks anti-laziness pass (fakeable DoDs, decomposition realism)
• before merge → architectural-gate / cross-base sweep
• ad-hoc decision → proponent + adversaries + synthesizer (e.g. delete-vs-migrate)

Do NOT use it as a rubber stamp, and do NOT wire it as a mandatory gate.

The recipe

1. Frame one sharp question and pick the point-cut. A squad answers a question; it is not a vibe check.
2. Select 3–4 distinct profiles by lens (bounded). Complementary, not redundant:
   • architect-alphonso — structure / seams / topology
   • debugger-debbie — live-evidence, coverage, "would this catch the regression?"
   • reviewer-renata — anti-laziness, contract-vs-implementation, fakeable assertions
   • randy-reducer — duplication / dead code (⚠ duct-tape bias — read critically)
   • paula-patterns — decomposition, boundaries, second-opinion adjudication
   • planner-priti — scope, sequencing, tracker hygiene
   • python-pedro — implementer feasibility
   • doctrine-daphne — doctrine integrity / DRG wiring Scale past 4 only for an explicit "audit / comprehensive" ask.
3. Dispatch in parallel, profile-LOADED. Each delegate's prompt MUST begin with: "FIRST run spec-kitty agent profile show <id> and spec-kitty charter context --action <action> --json; apply the resolved initialization, boundaries, directives, and tactics, then state which you applied." Loading the profile — not naming a persona — is the point. Only a read-only harness that cannot invoke the CLI may read packs/built-in/agent_profiles/<id>.agent.yaml; that degraded fallback can diverge because overlays, specializes_from lineage, and enhances/overrides semantics are not applied. Keep delegates read-only unless the task is an isolated implementation in its own worktree.
4. Require structured, non-fakeable output. Each returns findings as [SEVERITY] file:line — issue — recommendation, ending in a verdict, grounded in cited evidence, with honest concession of where its lens does not apply. A steelman that over-claims is weak; an adversary that concedes nothing is noise.
5. Match model tier to difficulty. Strong tier for analytical/adversarial lenses; lighter tier for mechanical/tracker delegates.
6. Synthesize; second-opinion on divergence. Aggregate. Where delegates disagree on a consequential point, do NOT average — adjudicate from the source, or dispatch one focused second-opinion delegate. Be critical of any delegate with a known bias. If irreconcilable, escalate to the human with both positions.
7. Record the convergent evidence and act. Capture confirmed findings (artifact, findings doc, or memory). The value is convergent evidence that survived independent scrutiny — not a single opinion.