Skill Claude CodeCodex
Trigger when the user asks to check an existing plugin for conformance: "audit this plugin", "check plugin conformance", "does this plugin follow the ecosystem rules?". Given a plugin directory path, checks plugin.json required fields, subagent file presence for every declared agent, YAML frontmatter completeness…
Skill Claude CodeCodex
Trigger when the user wants to design a new inter-agent artifact schema: "design a schema for X", "add a new schema to shared/schemas", "what should this handoff look like?". Given a description of what the artifact represents, who produces it, and who consumes it, produces a JSON Schema (draft 2020-12) following…
Skill Claude CodeCodex
Trigger when the user asks to build a new plugin from scratch: "build a new plugin", "create a plugin for X", "scaffold a plugin directory". Given a plugin ID and a description of what it should do, generates a complete, installable plugin directory — plugin.json, skills/ /SKILL.md, one stub agent file per declared…
Skill Claude CodeCodex
Trigger when the user wants one new agent added to an existing plugin: "write an agent for X", "add a subagent to this plugin", "create a subagent for this task". Given an existing plugin's directory, a task description, and a model/effort tier, generates a single conformant agent file at plugins/ /agents/ .md using…
Agent
Delegate to this subagent when you have a completed requirement@1 (or free-text design intent) and need a formal component spec@1 drafted. Input is a requirement@1 and optionally design references (mockup description, existing component family, design-system tokens). Output is a spec@1 conforming to…
Agent
Delegate to this subagent when a component spec@1 needs a definitive pass/fail judgment before entering the planning phase. Input is a spec@1 JSON object describing a UI component. This is an adversarial gatekeeper — the default disposition is fail, and the spec must earn a pass. The spec passes only if all four…
Skill Claude CodeCodex
Trigger when the user asks to spec a UI component: "spec this component", "draft a component spec", "write a spec for this design", "audit this component spec", "is this component spec ready?", "gate this component spec". Given a requirement@1 or free-text design intent, produces a spec@1 covering props, variants…
Agent
Delegate to this subagent when a plan@1 needs adversarial review before any implementation begins. Input is a plan@1 JSON object and its source spec@1 (read from specfilepath if set). The agent checks eight dimensions: acceptance criteria in the spec with no corresponding task (missing), tasks referencing symbols not…
Agent
Delegate to this subagent when a plan@1 exists and you need a rough effort estimate before committing to execution. Input is a plan@1 JSON object. The agent produces a per-task time estimate in minutes, marks each task as parallelizable or sequential, lists blocking dependencies by task ID, and produces a totalminutes…
Agent
Delegate to this subagent when you have a completed spec@1 and need a sequenced, executable implementation plan. Input is a spec@1 (or its specfilepath, read from disk if provided). Output is a plan@1 conforming to shared/schemas/[email protected], with specfilepath and spechash propagated from the spec. Decomposes into…
Skill Claude CodeCodex
Trigger this skill when the user says "write a plan for X", "how do we implement this spec", "break this down into tasks", "what are the implementation steps", "create an implementation plan", or "plan this out". Also activate when a spec (spec@1) exists and the user is ready to move to implementation but hasn't…
Agent
Invoke once per candidate after scanner (and boundary-tracer for T7 or T10 candidates). Never batch multiple candidates in a single invocation. Input is one candidate@1 entry and the recon manifest; for T7 or T10 candidates, also receives the field survival map from boundary-tracer. The agent reads actual code at the…
Agent
Conditional agent invoked only when scanner produces candidates classified as T7 (write-only fields or intent-capture discard) or T10 (error downgrade). Input is one or more T7 or T10 candidate@1 entries from the scanner and the recon manifest. For Rust workspaces, the agent loads shared/references/rust-hazards.md…
Agent
Invoke before any ranger scanning begins. Input is a workspace root path. This agent detects the primary language by inspecting Cargo.toml (rust) or package.json (typescript or javascript) at the workspace root. It identifies entry points such as binary crates, main.ts, or exported index files, then traces imports and…
Agent
Invoke after recon has produced a workspace manifest. Input is the manifest from recon (including livefiles and language) and optionally a specific hazard focus category. For Rust, the agent loads shared/references/rust-hazards.md; for TypeScript or JavaScript, it loads shared/references/typescript-hazards.md. It…
Skill Claude CodeCodex
Trigger this skill when the user asks to perform a bug hunt, code audit, find bugs, security audit, defect scan, vulnerability scan, code review for bugs, or asks "what's wrong with this code." Activate for any codebase language — Rust, TypeScript, or JavaScript. 5-phase pipeline: (1) recon builds a verified module…
Agent
Delegate to this subagent when ranger has produced a finding-report@1 and the defect class requires a structural fix — a change to the architecture, type system, interface boundary, or abstraction layer — rather than a patch of individual bug instances. Input is a finding-report@1. Output is a spec@1 conforming to…
Agent
On-demand diagnostic: given a specfilepath and a workspace root, verifies the current code still implements every acceptance criterion in the spec. Reads the spec from disk, not from context, and reads implementation/tests from the workspace. When a prior changeset@2's criteriaevidence is available, uses its pointers…
Agent
Delegate to this subagent when you need to verify that a draft spec@1 is grounded in its source artifacts — the requirement@1 and optional research-report@1 — before it proceeds to auditing or gating. Input is a spec@1 draft plus the originating requirement@1 and any research-report@1 used to produce it. For each…
Skill Claude CodeCodex
Trigger when ranger has returned a finding-report@1 and the confirmed defect class needs a structural fix — a change to the architecture, type system, interface boundary, or abstraction layer — rather than patches to individual instances. Given a finding-report@1, produces a spec@1 for the structural change that makes…
Skill Claude CodeCodex
Trigger when the user asks to review a spec's quality: "review this spec", "is this spec complete?", "audit this spec". Given a spec@1, adversarially reviews it for untestable criteria, ambiguous language, missing error cases, scope overlap with other specs, and incomplete sections. Returns a structured issue list…
Skill Claude CodeCodex
Trigger when implementer reports that a spec criterion contradicts what the system actually does: a speccontradiction status from smith. Given a specfilepath, the criterionid of the affected acceptance criterion, and a contradiction report (what the spec says vs. what the system actually does), revises the affected…
Skill Claude CodeCodex
Trigger when the user asks to write or draft a spec: "spec this out", "draft a spec", "write a spec for this requirement". Given a requirement@1 and optionally a research-report@1, produces a spec@1 draft with purpose, scope, nongoals, apisurface (if applicable), and acceptance criteria. Every criterion is a testable…
Skill Claude CodeCodex
Trigger when a spec@1 needs a definitive pass/fail judgment before it enters planning: "gate this spec", "is this spec ready?", "check this spec before we plan it". Default disposition is fail — the spec must earn a pass. Passes only if every acceptance criterion is testable with no vague language, no TBDs remain…