From kensai
Local multi-agent code review with adversarial falsification. Orchestrated by the kensai MCP server — call `start`, follow returned instructions through grounding → surfacing → proving → filing. Supports committed, uncommitted, or all changes. Optionally pass a path.
How this skill is triggered — by the user, by Claude, or both
Slash command
/kensai:review [uncommitted|committed|all] [project-path][uncommitted|committed|all] [project-path]modepathThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
**Read-only** — do not edit, write, or modify files.
Read-only — do not edit, write, or modify files.
The kensai MCP server holds all session state. Agents self-serve context via mcp__kensai__session_priming — no data
threading from lead to agents.
Agent teams: !echo ${CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS:-disabled}
If agent teams shows disabled above, stop and tell the user:
Agent teams required. Add to settings:
{ "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" } }Then restart.
Call mcp__kensai__session_start with resolved mode and path. On error, inform the user and stop.
Mode: $mode (default: uncommitted) Path: $path (default: current directory)
If mode looks like a path, treat as $path with mode uncommitted.
committed — HEAD~1..HEAD. Review the topmost commit.uncommitted (default) — HEAD..working tree. Review work-in-progress.all — HEAD~1..working tree. Review commit + uncommitted (amendment scenario).After session_start, check session_state for changed_files total. 20+ files → spawn teammates with
model: "<current-model>[1m]". Otherwise default model.
review-grounder teammate. No data in spawn prompt — grounder self-serves via
session_priming.mcp__kensai__grounding_complete.Read grounding from mcp__kensai__grounding_get — summary, integration surface, hotspots, blind spots, intent.
Derive surfacing dimensions from grounding — each dimension names a specific risk area this change introduces, not a generic review category. The dimension tells the surfacer WHERE to focus; the surfacer decides WHAT to find.
Derivation process:
Self-check before spawning: if you could copy the same dimension name to a different change and it would still make sense, the dimension is too generic. Rewrite it using terms from the grounding.
Create a task per dimension. Spawn review-surfacer teammates in parallel — one per dimension. Spawn prompt:
dimension name + quality criteria from grounding hotspots (risk patterns to watch for, not methodology).
Always spawn one additional general-review surfacer alongside the dimension surfacers. No prescribed dimension or quality criteria. Spawn prompt: "General review -- no prescribed dimension. Follow any thread. Focus on behavioral regressions: semantic contracts, lifecycle guarantees, concurrency, and anything the dimension surfacers might miss."
Wait for all surfacers. They record findings and call mcp__kensai__surface_clean or
mcp__kensai__finding_surface via MCP.
Shut down all surfacers.
mcp__kensai__surfacing_complete. If 0 findings -> skips to filing.review-prover teammate. Prover self-serves findings and context via MCP, applies falsification gates,
verdicts each finding.mcp__kensai__findings_list with status: "pending". If any remain,
message the prover with the pending finding IDs and wait. Repeat until all are handled.mcp__kensai__findings_list with status: "confirmed".file:line.mcp__kensai__filing_complete.mcp__kensai__session_end.The MCP enforces phase order. Tools reject calls outside their valid phase.
Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.
npx claudepluginhub gaijinentertainment/yatai --plugin kensai