Subagents
Ark ships three subagents that the main session can dispatch at specific phases. They are read-mostly audit and research roles with tight write scopes — they never touch code, never run git, and never spawn each other.
| Subagent | Phase | Writes only to | Role |
|---|---|---|---|
ark-researcher | DESIGN / PLAN, research tier | .ark/tasks/<slug>/research/*.md | Gather third-party / prior-art / codebase-pattern findings |
ark-reviewer | REVIEW (deep) | REVIEW.md | Judge the PLAN; write verdict + findings |
ark-verifier | VERIFY | VERIFY.md | Audit shipped work against PLAN/PRD; run build/test/lint |
All three ship for every installed platform — Claude Code (.claude/agents/*.md), Codex (.codex/agents/*.toml), and OpenCode (.opencode/agents/*.md). The prompt body is identical across platforms; only the per-platform frontmatter differs.
Reserved stems
The filenames ark-researcher, ark-reviewer, and ark-verifier (with each platform's extension) are reserved. Ark re-applies them on init, upgrade, and load — a user-authored file at one of those stems is overwritten. Sibling agents at any other stem are preserved untouched.
Who picks them
The slash commands don't auto-dispatch. At REVIEW and VERIFY entry, /ark:design stops and asks you which to use — the Ark subagent, a different model, or do it yourself. The choice is yours per phase per task.
After dispatching any subagent, the main session checks git status and reverts (via git restore) anything written outside the agent's documented write-scope before incorporating its findings.
Research tier vs. embedded researcher
Both gather knowledge, but they answer different questions:
- Research tier (
/ark:research) — you cannot yet write a PRD. The open question is what to build, or whether to build at all. The corpus underresearch/is the deliverable; follow-up implementation is optional. See Tiers → Research. - Embedded
ark-researcher— you can already write the PRD's What / Why / Outcome; you just need to fill technical unknowns inside a tiered task (a third-party library's shape, a prior-art comparison, a codebase pattern map). It stays inside the parent task and persists findings to that task's.ark/tasks/<slug>/research/<topic>.md.
Rule of thumb: if the research changes whether there's a task, use the research tier. If it changes what goes in the PRD/PLAN of a task you already know you're doing, dispatch the embedded researcher.
What they don't do
- No subagent self-fixes.
ark-verifierreports FAIL items; the main session resolves them. - No subagent runs git or edits code, SPECs, or the PLAN it's reviewing.
- No subagent spawns another subagent — every prompt carries a recursion guard. Only the main session dispatches.