Subagents

Subagents are rare by design. A child starts without the main session's context, spends about 25k tokens before it reads anything, and returns a summary the main session must check.

When a subagent is worth it

• Broad read-only search across many files or unknown paths: use Claude Explore, OMP scout, Pi k-agent-scout, or a fresh Codex agent. The child returns a focused result.
• Independent review required by ~/AGENTS.md §3 step 5 or requested by the user: use the k-review Reviewer launch procedure.

Everything else stays inline: implementation, verification, memory, and routine reads of known files.

Rules

• A reviewer is read-only and returns at most 10 findings once. Verify mode launches two fresh reviewers in parallel, merges their findings, and runs the fix loop in Skills. The review skill defines the exact packet and scope; the fix procedure owns the loop.
• A wording-only fix round skips review. Findings from the last review still need fixes. Mark each fix after the last review that is not wording-only "(unreviewed)" in Checked:; a cause that returns twice, or a finding that cannot be fixed, goes under Open:.
• Reviewers check request coverage against source evidence and report every severity. They run static checks under a time limit, start no live checks, and put claims only a live check could settle under Not checked:. They never run a secret-printing command, even a fake named after a real command. Every review applies k-code-quality's De-slop section, which loads k-anti-slop.
• Reviewers must not start subagents. Claude enforces CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1, Pi sets maxSubagentDepth: 0, and OMP sets task.maxRecursionDepth: 1. Codex receives this restriction in the shared reviewer prompt; its native fresh-fork tool exposes no per-child tool or depth control.
• Claude reviewers use the session's effort (medium in settings.json). CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1 disables the advisor for the main session and its subagents, since an agent profile has no key to turn it off (2.1.304); OMP keeps task advisors off (task.agentAdvisor).

Profiles

The reviewer body is shared: home/.chezmoitemplates/reviewer-prompt.md. It collects candidates across correctness, removed safeguards, contracts, simplification, conflicting state operations, edge inputs, instruction text, project rules, and slop, then deduplicates and refutes them before reporting evidence-backed file:line findings. Its verdict follows the evidence: there is no finding quota, and the caller's or author's preferred outcome does not justify inventing, suppressing, or withdrawing a finding.

Harness	Reviewer	Search
Claude Code	home/dot_claude/exact_agents/k-agent-reviewer.md.tmpl (Opus, session effort, omitClaudeMd), launched in the foreground	built-in Explore
Pi	home/dot_pi/agent/exact_agents/k-agent-reviewer.md.tmpl (claude-opus-5-5, medium)	k-agent-scout (gpt-6-luna, low)
OMP	home/dot_omp/private_agent/exact_agents/k-agent-reviewer.md.tmpl	native scout, APPEND_SYSTEM.md overlay
Codex	native spawn_agent with fork_turns: "none"; shared prompt from ~/.agents/skills/k-review/references/reviewer_prompt.md	fresh native agent
OpenCode	none; review inline	native tools

Pi's pi-subagents built-in agents stay disabled (subagents.disableBuiltins), so only these profiles exist there. OMP disables its own reviewer and security-reviewer agents in task.disabledAgents.

Codex 0.163.0 spawn_agent exposes a role selector (agent_type) only when agent roles are defined ([agents.<name>] in config.toml or a ~/.codex/agents/ file); this repo defines none. k-review passes the shared prompt and review packet directly, keeping the parent's model and effort. Full reviews launch two fresh agents; fix-diff reviews launch one.

Claude, Codex, Pi, and OMP share the SOP, skills, ,ai-kb, ,handoff, ,task, and ,behavior-map workflow; the intent-guard hooks run in every harness (Staying on the goal). Only Claude runs the SOP, checked-answer, stop-review, and behavior-map mods. OMP's Gate advisor reviews the final answer against ~/AGENTS.md (Oh My Pi); the other harnesses follow the written rules without enforcement.