Intendant

Use to operate a running Intendant daemon from the CLI — sessions, approvals, displays and screenshots, computer-use input, browser workspaces, audio, and federated peers (message or delegate to another machine's agent, or drive its screen directly with --peer). Prefer `intendant ctl` over broad MCP tools to keep model context small.

Keyless end-to-end smoke for operator goals on a NATIVE session — drives a mock-provider intendant over the control socket and proves set/get/notice-injection/clear/honest-failure

Keyless end-to-end smoke for the session-vitals git + cache segments — mock-provider daemon in a scripted git repo, asserts the merged session_vitals event on the control socket (branch/dirty/ahead/parity plus cache hit/TTL)

Use when investigating Intendant run history, session logs, dashboard replay/search behavior, external-agent wrapper logs, Codex/Claude backend logs, context rewind/fission sidecars, display frames, recordings, or task artifacts under ~/.intendant/logs. Teaches how to find sessions, search session.jsonl and sidecars, follow turn-file byte spans, map wrapper sessions to backend ids, and interpret what each log artifact provides.

E2E test the display recording and replay system. Launches intendant --web with recording enabled, triggers Xvfb display creation, verifies ffmpeg recording starts, segments are created and serveable, and the replay UI loads in the browser. Asserts via /recordings and /debug HTTP endpoints. Human monitors via VNC on port 5950.

E2E test the live audio pipeline with real audio. Uses espeak-ng TTS + PulseAudio virtual mic to feed synthesized speech to Gemini Live or OpenAI Realtime through the browser's getUserMedia on /app. Asserts via /debug endpoint JSON. Human monitors via VNC on port 5950. Tests the full audio path: TTS -> virtual mic -> Firefox -> AudioWorklet -> WASM -> live model -> tool calls -> agent.

E2E test the --web app UI. Launches Xvfb, runs intendant --web as a background process, opens Firefox on /app, and asserts via /debug endpoint and WebSocket JSON. Human monitors via VNC on port 5950.

Live end-to-end test of the Claude Code external-agent integration: spawn a supervised Claude Code session in a toy repo, exercise the approval protocol (allow + deny), native session-id capture, mid-turn steering, interrupt, thread actions (/compact, fork via --fork-session), the async Agent-tool sub-agent surfacing (ephemeral task-* child sessions), and resume-by-native-id, driving everything through the Unix control socket.

Live two-daemon test of the Files-tab editor against a federated peer: pair two isolated daemons on localhost, scope the inbound peer identity to file-operator with a single write root, then drive a real browser through daemon A's dashboard to browse/edit/save files on daemon B — including cross-daemon conflict detection and both IAM denial paths (write outside write_roots, read outside all roots), verified on disk and in B's [peer-fs] audit trail.

Live single-daemon proof that browser mTLS certificates bind to scoped IAM principals end-to-end: mint extra client certs from the rig CA, bind them to role:files-write (with fs roots) and role:operator via the IAM API from the root cert, then hit the HTTPS dashboard with Playwright client certificates and watch the daemon enforce role ceilings, fs scoping on write/rename/delete (both rename legs), and the peer.use gate on peer signaling relays and quick controls.

Use when the user should see an agent-owned display through Intendant's shared view: demoing a finished result, letting the user watch live GUI/browser work, focusing attention on a region, capturing a display frame, or asking the user to take input authority.

Fission-shaped evaluation runner. Launches a managed Codex session on the jobline service quartet (a REST API job store, a worker, a CLI client, and a read-only metrics service — all Python stdlib, in disjoint directories), then scores the result behaviorally with verify.sh (per-component partial credit + a live integration bonus, max 5.0) and records whether the model spontaneously used model-driven fission. The four components own disjoint write scopes and share only the HTTP protocol, so fission into chartered worktree branches is a genuine win — but it is never required and never mentioned to the agent. Whether it fissions is the measurement.

Fission-shaped evaluation runner. Launches a managed Codex session on the salestream ETL task (four independent components — a Python CSV→JSONL normalizer, a Python business-rule quarantine stage, a Rust merge/dedupe tool, a bash+jq report generator — plus an end-to-end Makefile pipeline), then scores the result behaviorally with verify.sh (per-component partial credit + an integration bonus, max 5.0) and records whether the model spontaneously used model-driven fission. The four work streams own disjoint write scopes, so fission into chartered worktree branches is a genuine wall-clock/context win — but it is never required and never mentioned to the agent. Whether it fissions is the measurement.

E2E test the managed Codex long-context feature against installed vanilla Codex. Uses a deterministic large-output task to force vanilla auto-compaction, then verifies managed Codex avoids hidden compaction, enters rewind-only recovery where ordinary tools are gated, rewinds to an exact tool-call anchor chosen from the anchor catalog, and returns to a safe context-pressure state.

Live smoke test of model-driven fission on a managed Codex session: spawn two chartered branches in a toy git repo (one read-only, one write-scoped with an isolated worktree), verify the fission ledger and the dashboard endpoint key the group at the real spawn tool-call anchor, wait on and import a completed branch into the parent, claim the canonical branch, then rewind the parent past the spawn anchor and assert the group detaches.

Use when validating the Station canvas dashboard end to end — renderer health (WebGPU/fps/frame pacing), agentic canvas interaction via station.debug_json()/station.activate(), and the perf regression gate — with scripts/validate-dashboard.cjs on a headless CI-style box or a headed GPU host. Covers environment setup, the dual-GPU Chromium recipe, headless gate vs headed acceptance invocations, the --station-perf-eval before/after workflow, and portal-grant caveats for display streams.

Use when running Intendant Wayland display-capture E2E tests against a remote GNOME Wayland VM where the XDG Desktop Portal screen-sharing dialog must be approved programmatically or through a remote desktop window. Covers GNOME Remote Desktop setup, FreeRDP wrapping for Computer Use, portal approval, verification, cleanup, and safety boundaries.

Make a voice call through any app (Element, FaceTime, WhatsApp, etc.) using computer use to navigate the UI and spawn_live_audio for the AI voice conversation. Returns typed structured data.

Make an outbound phone call via SIP and conduct a voice conversation using spawn_live_audio. The AI model talks through the Vortex Audio virtual device, which pjsua routes to the SIP call. Returns typed structured data from the conversation.