Agent Claude Code
Security auditor (read-only), OWASP-informed and threat-model-first. Finds ATTACKER-exploitable vulnerabilities: auth, injection, data protection, access control, real-time/transport, multi-tenant/IDOR. Normal-use breakage is Mack's, not Cipher's. Use proactively for auth, input handling, data access, secrets, network…
Agent Claude Code
Behavioral QA breaker (read-only). Finds how code breaks under NORMAL or careless use — races, state-machine violations, data-integrity bugs, boundary/numeric abuse, broken invariants. Attacker-exploitable issues are Cipher's, not Mack's. Use proactively when changes touch concurrency, persistence, critical state…
Agent Claude Code
Implementation agent (generator): coding, exploration, research, builds, tests, and file modifications. Returns Summary/Changes/Verification/Concerns — does not commit. Use proactively when writing or editing code, running builds/tests, exploring the codebase, or researching APIs/docs. Samantha evaluates; Monk…
Agent Claude Code
UX and accessibility reviewer (read-only, Haiku). Reviews user-facing code STRUCTURE — markup, labels, error strings, a11y attributes, empty/loading states — not rendered pixels. Use proactively when changes touch UI components, dialogs, forms, flows, or user-facing copy.
Agent Claude Code
Skeptical architect (read-only, Opus). Challenges the principal's decisions — not line-by-line code — on scope expansion, new abstractions, and architectural choices. Verdicts: PROCEED / SIMPLIFY / REDESIGN / DEFER. Use proactively when scope grows, a new abstraction appears, architecture is being chosen, or before…
Agent Claude Code
Translation and i18n specialist (Haiku). Bulk translation, quality review, format-specifier validation, pluralization, locale formatting, and hardcoded-string extraction. Use proactively for all translation, localization, locale-file, or i18n coverage work.
Settings file Claude Code
Samantha oversight and system prompt.
Skill Claude CodeCodex
Structured multi-agent adversarial review and challenge cycle. Use when a high-stakes change, major architectural decision, or important output needs verification beyond a single reviewer — or when the user asks for an adversarial review.
Skill Claude CodeCodex
Full codebase-vs-docs/spec discovery audit: finds missing features, half-built work, code/doc divergence, dead code, and doc gaps, then emits a prioritized backlog and work orders. Use when the user wants a queue-feeding sweep, 'what's missing vs the spec', or 'bring code up to spec'. For audit-and-fix-now in one…
Skill Claude CodeCodex
Broad post-change review across behavior, security, and UX — dispatches fitting specialists and synthesizes a verdict. Use AFTER a change when the user says review this, how does this look, or wants a multi-dimensional review. For a security-only deep audit, use threat-audit.
Skill Claude CodeCodex
Lightweight git stage-and-commit with a clear message, skipping the full build/test/review pipeline. Use when the user says commit this, save this, or wants a quick commit without shipping.
Skill Claude CodeCodex
Arm solo coordination mode: one agent session, in-session subagents as workers, optional local idle scheduler. Use when staying in one context window, or when the human says solo mode / in-session workers / no peer seats.
Skill Claude CodeCodex
Arm star coordination mode (formerly dual): orchestrator + implementer seats on a file mailbox STAR bus. Use when multi-process seats, durable work orders, multi-project hub, or the human says star mode / dual mode / arm the coord monitor.
Skill Claude CodeCodex
Common substrate for Samantha coordination modes (solo and star): seat identity, mailbox grammar, disaster rules that apply in both modes, harness wake bridging, and idle-kick/scheduler concepts. Use when arming coordination, choosing solo vs star, or when coordinate-solo / coordinate-star need shared protocol.
Skill Claude CodeCodex
Diagnostic triage for regressions and unclear breakage: launches parallel read-only investigation tracks across subsystems and synthesizes a severity-ranked verdict. Use when something is broken, regressed, was working before, or is not working as expected (and there is no single pasted stack trace — that is fix).
Skill Claude CodeCodex
Codebase orientation and explanation of how a component, flow, or subsystem works. Use when the user asks what does this do, how does X work, explain this, or wants to understand part of the codebase.
Skill Claude CodeCodex
Targeted diagnosis and fix for a known error — not a full diagnostic sweep. Use when the user pastes a stack trace, error message, failing assertion, or says this returns the wrong value.
Skill Claude CodeCodex
Skill router for ambiguous intents: decides whether something is broken (diagnose) or missing (build) and routes to the right protocol. Use when the situation is ambiguous, the user has not named a skill, or triage is needed before acting.
Skill Claude CodeCodex
Translation and internationalization quality: coverage gaps, locale issues, and cultural appropriateness (Rosetta does the translation work). Use for missing languages, locale bugs, i18n coverage, or translation quality reviews.
Skill Claude CodeCodex
Full GitHub issue resolution pipeline: recon, plan, skeptical review, implement, verify, communicate. Use when the user references a GitHub issue, pastes an issue link/URL, or says fix issue N.
Skill Claude CodeCodex
Applies Open Knowledge Format (OKF): author a new knowledge concept, validate a concept against code, or migrate a legacy .aispec doc to OKF. Canon create/delete/rename is gated on human go-ahead. Use when writing or validating knowledge docs, OKF concepts, or migrating .aispec files.
Skill Claude CodeCodex
Proactive code-quality cleanup in zone-partitioned waves (analyze/fix plus specialist categories). Use after major features, or when the user asks to clean up, polish, refactor debt, or improve code quality.
Skill Claude CodeCodex
Full pre-commit verification pipeline: build, test, review, then commit if clean. Use when work is done and ready to ship, or the user says ship it / ready to commit with the full pipeline (for a lightweight save only, use commit).
Skill Claude CodeCodex
Spec↔code compliance audit that builds what's missing until alignment. Use when the user asks about spec compliance, feature completeness, design-doc alignment, or whether the code matches the spec.