ReferenceAgents

Agents Reference

This is the complete, no-omissions list of every agent in ai-core-kit. An agent is a specialized subagent Claude delegates to (via Task) when its description trigger matches. Agents split along the META vs CHILD boundary:

  • META build team.claude/agents/. Phase specialists that /ack-build spawns to author the kit itself. Never rendered into a fork.
  • CHILD reusable fleettemplates/agents/. The everyday delivery and review agents that /ack-init renders into a fork, including the RPI-flow agents.
META agents (.claude/agents/) author the kit during /ack-build; CHILD agents (templates/agents/) are rendered into a fork by /ack-init — the RPI-flow agents plus the engineering review fleet.

META — build team (.claude/agents/)

/ack-build spawns these per phase in a ground-truth → author → adversarial-QA team. They write only META artifacts; a worker that would write a concrete project.manifest.yaml or contract instance at the META root is a layer violation and is refused. Not rendered into forks.

AgentLayerWhat it doesTrigger / when
contractMETAAuthors the CHILD-payload methodology — the contract template and the 3-mode manifest-driven contract-gate hook.Contract template, contract gate, gate modes, approved-oracle, design-contract-first scaffolding.
design-systemMETAAuthors the fullstack archetype’s design-system payload — frontend-design and brand-guidelines-derived skills plus a design-tokens scaffold.Design system, design tokens, frontend-design skill, brand-guidelines, fullstack UI scaffold.
infraMETAAuthors META-layer platform plumbing — the .claude/ tree, the project .mcp.json, the offline telemetry aggregator + pricing.json.settings.json, .claude layout, MCP wiring, telemetry aggregator, pricing.json, budgets, linter/validator script.
qaMETAValidates phase artifacts against acceptance tests and adversarially verifies the two-layer boundary at the end of every phase.Validate, verify against acceptance tests, QA this phase, test-fork the archetypes.
researchMETAExtracts ground truth from reference repositories or the Anthropic docs — exact conventions and per-repo license.Extract conventions, verify against docs, what does <repo> actually do, build the license ledger.
templateMETAAuthors or revises the CHILD-payload template sets — the archetype trees, interview questions.yaml, render map.Archetype templates, scaffold the <archetype> tree, interview branching, questions.yaml, render.map.

CHILD — reusable fleet (templates/agents/)

The RPI + review fleet that /ack-init renders into a fork. The model column is the agent’s declared model tier.

RPI-flow agents

The two agents that drive the RPI loop (Research → Plan → Implement).

AgentLayerModelWhat it doesTrigger / when
requirement-parserCHILDsonnetTurns an unstructured feature request into structured requirements, goals, constraints, complexity, and clarifying questions.Parse this request, what exactly are we building. RPI research step 1; rendered when features.rpi == true.
constitutional-validatorCHILDopusValidates roadmap items, features, and technical decisions against the project’s constitution, principles, and chosen archetype.Does this fit our principles, validate against the constitution, is this in scope. RPI research→plan gate.

Engineering & review agents

AgentLayerModelWhat it doesTrigger / when
architectCHILDopusSoftware-architecture specialist for system design, scalability, and technical trade-off decisions — produces high-level designs, component responsibilities, data flows, API contracts, ADRs.Design this, how should we structure, what are the trade-offs, write an ADR. RPI plan phase.
code-explorerCHILDsonnetCodebase-analysis specialist that traces execution paths, maps architecture layers, and documents dependencies so new work starts grounded in code reality.How does X work, trace this flow, map the architecture, what touches this module. Discovery step of RPI.
code-reviewerCHILDsonnetSenior code-review specialist that reviews a diff for correctness, security, and maintainability and returns severity-ranked, evidence-backed findings.Review this, check my changes, is this PR ready.
refactor-cleanerCHILDsonnetDead-code cleanup and consolidation specialist that finds unused code, exports, dependencies, and duplicates, then removes them safely in small verified batches.Remove dead code, clean up unused imports/deps, consolidate duplicates, tidy the codebase.
security-reviewerCHILDsonnetSecurity-vulnerability detection and remediation specialist that scans changed code for OWASP Top 10 issues, hardcoded secrets, injection, SSRF, and unsafe crypto.Security review, is this safe, check for vulnerabilities.
silent-failure-hunterCHILDsonnetError-handling auditor that hunts silent failures — swallowed exceptions, empty catch blocks, dangerous fallbacks, lost stack traces, and missing error propagation.Find swallowed errors, why is this failing silently, audit error handling.

Subagents that are not kit agents. skill-creator/agents/{analyzer, comparator,grader}.md are internal sub-agent prompts of the skill-creator skill — they carry no kit agent frontmatter and are not linted or rendered as standalone agents.

See also: Skills Reference, Commands Reference, Skills Catalog (provenance + license).