AI Assistants
CLI assistants share operating rules, skills, read-only review profiles, and local memory tools. Agents work inline by default and query memory on demand. OMP's native advisor adds main-session notes; child advisors stay off.
What is installed
| Piece | Source | Deployed to | Page |
|---|---|---|---|
| Global instructions | home/readonly_AGENTS.md | ~/AGENTS.md, linked into every harness | Instructions |
| Skills | home/exact_dot_agents/exact_skills/ | ~/.agents/skills/ | Skills |
| Subagent profiles | k-agent-reviewer (Claude, Pi, OMP), k-agent-scout (Pi) | each harness's agents dir | Subagents |
| Knowledge base | ,ai-kb | ~/bin/,ai-kb, data in ~/.local/share/ai-kb | Memory and handoffs |
| Session handoffs | ,handoff | ~/bin/,handoff, notes in ~/.local/share/agent-handoffs | Memory and handoffs |
| Behavior maps | ,behavior-map | ~/bin/,behavior-map, maps in ~/.local/share/k-ai-behavior-map | Memory and handoffs |
| Intent cards | ,task and the intent-guard hooks | ~/bin/,task, cards in ~/.local/share/agent-tasks | Staying on the goal |
| MCP servers | home/.chezmoidata/mcp_servers.yaml | rendered per harness | MCP servers |
| Harness configs | per-harness files under home/ | ~/.claude, ~/.pi, ~/.omp, ~/.codex, … | Tool configs |
| Local inference | llama.cpp router and launchers | ~/bin/,llama-cpp, ,*-llama-cpp | llama.cpp |
Harnesses
| Harness | Default model | Notes |
|---|---|---|
| Claude Code | claude-opus-5-5, effort medium | ,claude-openrouter and ,claude-llama-cpp switch the backend |
| Pi | openai-codex/gpt-6.1-sol | pi-subagents with built-in agents disabled; repo profiles only |
| OMP | anthropic/claude-opus-5-5:medium | native scout; repo k-agent-reviewer |
| Codex | gpt-6.1-sol | ,codex, ,codex-openrouter, ,codex-llama-cpp |
| OpenCode, Crush, freebuff | per-harness config | basic config and MCP; OpenCode also runs the intent-guard plugin |
Principles
The global operating rules define the exact scope, approval, verification, and completion gates. Review and verification describes the review loop, including returned findings and review baselines.
-
Inline by default. Subagents handle broad read-only searches and independent reviews. Reviews required by a risky code or instruction change use two fresh reviewers in parallel.
-
Static verification. Run affected unit and static checks once on the finished change, under time limits. Use source reads and isolated local runs for evidence. Live checks require the user's request, a hard timeout, and a copy when they would change external state or overwrite user data.
-
Fix before finishing. Resolve every supported issue in the request's scope and every defect in the change's impact, including low-severity findings. Review fixes and continue until a full review round is clean. Before reporting, finish every doable step in scope or the change's impact, and cleanup; then report evidence, assumptions, proposals, and blockers.
-
On-demand memory.
,ai-kb searchwhen it can help;,ai-kb rememberonly for verified, reusable gotchas, or--supersedesto retire a stale capsule. -
Intent card. Before its first edit the agent writes the task's goal, reading,
Done whenitems, and scope; hooks in every harness check each iteration against it (Staying on the goal). -
Explicit handoffs. "handoff X" writes a short note with
,handoff; "continue X" reads it in any harness. -
Behavior maps. Each repo keeps a short map of critical behaviors seen from the user side; agents update only the entries their diff touches.
Related: Reviewing agent diffs.