Reference · Information

CLI commands

Every MCP tool has a CLI equivalent for shell use: senkani exec, senkani search, senkani fetch, senkani bench, senkani kb, senkani learn. Scriptable, pipeable, CI-friendly.

Command

What it does

Status

senkani exec

Run a command with the filter pipeline applied to output. Mutating commands pass through unchanged.

—

Live

senkani search

Symbol lookup from the tree-sitter index.

—

Live

senkani fetch

Read a symbol's source lines by name.

—

Live

senkani explore

Navigate the codebase via the import graph from an entry point.

—

Live

senkani index

Build or update the symbol index manually.

—

Live

senkani validate

Local syntax validation across 25 languages. Zero API calls.

—

Live

senkani stats

Lifetime metrics from the session database. FTS5 search over tool output history.

—

Live

senkani export

JSONL dump of sessions + commands + token_events. GDPR-adjacent data portability.

—

Live

senkani wipe

Destructive: delete session DB + socket-auth token.

—

Live

senkani bench

Run the benchmark suite: 10 tasks × 7 configs. Plus bench pii: headless PII-classifier cold/warm/shape bench (stub backend, no model load).

Diagnose and optionally repair hook registration, binary paths, grammar versions, config.

—

Live

senkani compare

Diff two files or git revisions through the filter pipeline.

—

Live

senkani grammars

Show installed tree-sitter grammar versions.

—

Live

senkani schedule

Schedule recurring commands via launchd.

—

Live

senkani uninstall

Clean removal of all senkani config, hooks, and MCP entries.

—

Live

senkani kb

Query and manage the project knowledge base.

—

Live

senkani eval

Quality gates: bench savings + KB health + regression detection.

—

Live

senkani ml-eval

Per-tier Gemma 4 output quality eval — drives the 20-task harness against each installed tier and writes ~/.senkani/ml-tier-eval.json.

—

Live

senkani learn

Compound learning CLI — review staged proposals, accept/reject, audit.

—

Live

senkani authorship

Bulk-tag legacy KB rows whose authorship column is NULL. Operator-triggered; never automatic.

—

Live

senkani skill

Lint and export HandManifest skill packages. Round-trips one canonical manifest into Claude Code, Cursor, Codex, OpenCode, or Senkani output.

—

Live

senkani pack

Install, uninstall, and list SkillPacks (multi-skill bundles + HookRouter policy fragment + context doc). Collision-diff on install; chained provenance via pack_audits.

—

Live

senkani models

List, pull, or verify Senkani's local ML models (vision, embeddings, Layer-3 PII classifier). Operator-explicit pulls; no auto-download.

—

Live

senkani onboarding milestones

Read-only verification surface for ~/.senkani/onboarding/milestones.json. Prints the live SENKANI_ONBOARDING_MILESTONES gate state, resolved path, file's JSON, and a derived summary.allComplete line — bypasses the env gate on reads so =off still surfaces on-disk truth.

—

Live

senkani walk rebuild-bundle

Rebuild SenkaniApp from current HEAD and refresh the wrapped .app bundle under tools/soak/runner/: swift build --product SenkaniApp → copy debug binary into Contents/MacOS/SenkaniApp → ad-hoc codesign --force --sign - → lsregister -f. Shared callable form used by senkani doctor check #20 and the walk-runner .command scripts (eliminating the inline rebuild dance). Fast swift build path — fine for the soak runner, which never exercises the crashing Bundle.module access.

—

Live

senkani walk rebuild-bundle-xcodebuild

GUI-Cowork-walk path. Builds SenkaniApp via xcodebuild -scheme SenkaniApp -destination 'platform=macOS' -configuration Debug build (NOT swift build), locates the product (build-log Ld parse → DerivedData glob fallback), then wraps it into an open-launchable .app: copy exe into Contents/MacOS/ → dynamically discover & copy every *.bundle resource into Contents/Resources/ → ad-hoc codesign --force --deep --sign - → lsregister -f. Only the xcodebuild product resolves Bundle.module correctly under an open-launch; a swift build-wrapped bundle crash-loops. Distinct from rebuild-bundle by design — the fast path and senkani doctor auto-rebuild stay untouched.

—

Live

senkani monitor

Read-only operator dashboard in the terminal — a live view over the session database (no mutations). --tui enters the interactive event loop (raw mode; q quit, j/k cursor nav, r refresh, / filter) and restores cooked mode on exit, including the SIGINT/SIGTERM paths. --single-frame prints a single ANSI render and exits — useful for scripts and snapshot tests. --app-version <string> sets the version shown in the header bar (default 0.4.0). --poll-interval=<Duration> sets the TUI refresh cadence with an ms/s/m/h suffix (e.g. 500ms, 5s, 30s, 5m) — default 5s, must be ≥ 1s (anything below is rejected with --poll-interval must be ≥ 1s), no upper bound. After the first paint each tick repaints only the changed cells (RenderFrame.diff delta painting), keeping steady-state updates byte-cheap over a remote/SSH session. Plain senkani monitor (no flag) prints use --tui and exits 0.

—

Live