Agent Catalog

13 specialized agents with defined roles, models, tools, and routing triggers.

Agent Table

Note: For /discuss sessions any agent may be summoned as a teammate (SendMessage-capable) via TeamCreate — researcher is the most common, but architecture debates summon planner/coder etc. (see canonical examples in rules/agent-routing.md). Outside discussions, all run subagent-only.

Agent Detail

searcher

• Trigger: "검색해", "찾아봐", internet resource needed, paper lookup
• Constraints: Always called by other agents — never directly performs code edits
• Special: All agents may freely summon searcher for search sub-tasks
• Tavily: Uses search for general queries, extract for full-page content

scout

• Trigger: "파일 찾아", "코드 어디", project structure unknown, before coder/qa
• Modes:
  • Standard: symbol/file search via Glob + Grep
  • deepinit: full project structure analysis for new projects
• Output: File list + dependency graph + impact scope

planner

• Trigger: Architecture design, experiment planning, "설계해", workflow creation
• Constraints: Delegates all search to searcher; no direct internet access
• Output: Phase plan with gate conditions, team composition, artifact list

coder

• Trigger: Code write, modify, bug fix, new module
• Constraints: No execution — writes to file only; user runs
• Special: Calls scout before writing to understand impact scope
• Refactor guard: No refactoring during active experiments

qa

• Trigger: After code changes, test execution, "검증해", "테스트해"
• Modes:
  • Standard: syntax/import check, test run, PASS/FAIL report
  • tracer: causal bug tracing — follows call chain to root cause
  • ralph-loop: auto-retry loop until tests pass (5-Gate guarded)
• Ralph Wiggum Loop: Max N retry rounds; escalates to main on convergence failure

researcher

• Trigger: Paper writing, experiment analysis, "논문", "비교 분석", critical evaluation
• Teammate mode: Participates in /discuss sessions with SendMessage
• Constraints: Delegates search to searcher; does not execute code

designer

• Trigger: Figure generation, TikZ diagrams, matplotlib plots, "그림 그려", "시각화"
• Standards: TikZ for structural diagrams; matplotlib for data plots
• Output: Saved to docs/paper/figures/ (or project-specified path)

docs

• Trigger: Confluence page create/update, Jira issue, "문서화", "컨플루언스"
• Protocol: 3-step confirmation before any Confluence update (see CLAUDE.md)
• MCP: Uses Confluence MCP + Jira MCP; Jira comments only via scripts/jira_comment.py (preview → user approval → post·update), never the MCP comment tool

gitops

• Trigger: "커밋해", "브랜치", "diff", "로그", git operations
• Constraints: push/pull are user-only — gitops only handles local git ops
• Output: Commit message following {issue} {type}: {Subject} format

citation-verifier

• Trigger: right before citing a paper in text/experiments; SOTA baseline or domain-mechanism references; pre-2020 citations (routing 3.6)
• Modes: advocate + skeptic in parallel → synthesis debate; retraction gate (Crossref + OpenAlex) runs first
• Constraints: returns verdict only — no file writes (untrusted isolation); caller (main) stores results (mem0 T21 + detail log per rules-lazy/citation-verification-log.md)
• Verdict: 허용 / 경고 / 금지 with supporting/contrasting/mentioning counts and confidence

tutor

• Trigger: "이해 안 돼", "다시 설명해줘", "퀴즈 내줘", "가르쳐줘" on an existing artifact; /paper-read Step 7 and /research close-out after conclusions are confirmed by the key-paper deep read (mandatory); /discuss and /experiment reports (offered)
• Modes: quiz / diagnose (misconception type: term·formula·causality·number·context) / reteach (explanation ladder, one rung lower per round) / verify (re-quiz wrong items only) / explain (math·science concept note + ELI5 Q&A, no quiz)
• Constraints: no web access — emits NEEDS_SEARCH, main runs searcher and feeds back the summary; writes only quiz.md (questions, round log) and quiz-key.md (answers, opened by main only at grading); questions and answers must be grounded in the artifact (notes/review) — external material is auxiliary and labeled
• Exit: all questions answered correctly (+ integration check), or user stops, or an item stuck at ladder rung 6 is handed off as "tutor limit" and excluded from the loop; main records rounds and final score in Mem0

Routing Priority

Canonical: rules/agent-routing.md § 라우팅 테이블 (24 rows, first-match wins). The table below is a condensed mirror for quick lookup — on any conflict the canonical file wins, and edits must land there first.

Invocation Modes

Subagent (Single Task)

Agent(subagent_type="coder", description="sonnet: Modify loss function")

Teammate (Discussion)

TeamCreate(team_name="arch-review")
Agent(team_name="arch-review", name="researcher", ...)
Agent(team_name="arch-review", name="planner", ...)

Teammates use SendMessage for cross-agent communication. Required for /discuss sessions.

Skill Auto-Call (4-Gate)

Agents may autonomously invoke whitelisted skills after passing all 4 gates:

SKILL_DEPTH=1 in spawn prompt disables all skill auto-calls (recursion guard).