RuView/harness/ruview/README.md

npx @ruvnet/ruview — RuView WiFi-sensing operator harness

An AI agent harness that knows how to operate RuView (WiFi-DensePose): onboard a newcomer, provision an ESP32 CSI node, calibrate a room, train pose models, and — crucially — refuse to overstate accuracy. Minted from the RuView monorepo via metaharness and hardened per ADR-182.

WiFi sensing infers coarse pose/presence/breathing from Channel State Information. It is not a camera. Every accuracy number this harness emits must be MEASURED against a baseline — that rule is enforced in code (ruview_claim_check).

Quick start

npx @ruvnet/ruview                       # onboard — pick a setup path
npx @ruvnet/ruview claim-check --file REPORT.md   # the honesty guardrail (non-zero exit on untagged claims)
npx @ruvnet/ruview verify                # run the deterministic proof (VERDICT: PASS)
npx @ruvnet/ruview doctor                # structured diagnostics with fixes (ADR-372)
npx @ruvnet/ruview ports                 # find the ESP32 serial port
npx @ruvnet/ruview flash --port COM7 --bundle ./esp32-csi-node-s3-8mb   # plan only
npx @ruvnet/ruview train-plan --mode pose-smoke
npx @ruvnet/ruview train-gate --file eval-report.json
npx @ruvnet/ruview devices               # ESP32 / mmWave / LiDAR on this machine
npx @ruvnet/ruview hosts run --host lab-pi --tool ruview_devices_scan
npx @ruvnet/ruview guidance --topic homecore --query "Wasmtime plugins"
npx @ruvnet/ruview spaces --resource spaces
npx @ruvnet/ruview spaces --resource events --limit 25
npx @ruvnet/ruview --help

The operator tools are pure Node and the published package has no runtime dependencies (ADR-263 O3). MetaHarness, Darwin and Flywheel are exact-pinned development dependencies used only for scoring, evolution proposals and replay verification.

Tools (ruview_*)

Exposed both as CLI verbs and as an MCP server (npx @ruvnet/ruview mcp start):

Tool	What it does
ruview_onboard	Pick docker-demo / repo-build / live-esp32; print the next command
ruview_claim_check	Lint text for untagged / overstated accuracy claims (guardrail)
ruview_verify	Run verify.py deterministic proof → VERDICT
ruview_node_monitor	Assert CSI is flowing on an ESP32 (read-only)
ruview_calibrate	ADR-151 room pipeline (baseline→enroll→train-room→room-watch)
ruview_node_flash	Cross-platform esptool flash with checksum + chip preflight and boot-log evidence (mutating, guarded)
ruview_firmware_plan	Verify a firmware bundle and return the exact esptool command (read-only)
ruview_firmware_ports	Enumerate serial ports via pyserial (read-only)
ruview_doctor	Structured diagnostics: runtime, integrity, hosts, repo, Rust, Python/esptool, bundles, serial, kernel
ruview_train	Run pose-smoke / pose / room training from the checkout (writes checkpoints, guarded)
ruview_train_plan	Resolve a training command without running it (read-only)
ruview_train_gate	Mean-pose-baseline + leakage gate that returns the only publishable claim wording
ruview_devices_scan / ruview_esp32_capture / ruview_mmwave_read / ruview_lidar_read	Host device access: USB discovery, ESP32 UDP stream, 60/24 GHz mmWave, RPLIDAR/iPhone LiDAR (device-access grant)
ruview_host_list / ruview_host_run	SSH remote hosts, read-only tools only (remote-host grant)
ruview_guidance	Source-cited code map, capability maturity, validation commands, and limitations
ruview_spaces_list	OAuth-only paging for sites/buildings/floors/spaces/zones/entities/events/alerts (guarded over MCP)
ruview_memory_search	Search the reviewed, source-cited contributor brain
ruview_kernel_selftest	SYNTHETIC self-test of the optional @ruvnet/ruview-kernel WASM/napi-rs compute package (ADR-368)

Every tool is fail-closed: missing repo / python / binary / port → an honest negative, never a fabricated success.

MCP authority model

Read-only tools need no grant. Over MCP, ruview_calibrate and ruview_train need the workspace-write grant plus confirm: true; ruview_node_flash needs hardware-write plus confirm: true; ruview_spaces_list needs credential-use. The plan tools (ruview_firmware_plan, ruview_train_plan) cannot take confirm, so a read-only client can review exactly what would run. The MCP doctor cannot probe a board or contact a network host; those are CLI/SDK options (--probe, --url).

Device and host access (ADR-373, ADR-374)

Run these on the machine the hardware is plugged into. A cloud agent cannot reach your USB ports; use a local session or an SSH host.

npx @ruvnet/ruview devices                                   # classify USB serial devices by VID:PID
npx @ruvnet/ruview esp32 --seconds 10                         # ESP32 / Realtek RAC1 node UDP stream (default :5005)
npx @ruvnet/ruview esp32 --seconds 45 --analyze               # ...and run live CSI through @ruvnet/ruview-kernel
npx @ruvnet/ruview monitor --port COM10 --baud 1500000        # serial console without resetting the node
npx @ruvnet/ruview mmwave --port /dev/ttyUSB0                 # MR60BHA2 60 GHz / LD2410 24 GHz, auto-detected
npx @ruvnet/ruview lidar --source rplidar --port /dev/ttyUSB1 # RPLIDAR scan summary
RUVIEW_LIDAR_TOKEN=… npx @ruvnet/ruview lidar --source iphone --url ws://HOST:8787/ws/lidar

Tool	Access	MCP grant
ruview_devices_scan	pyserial port list + VID:PID roles (opens no port)	device-access
ruview_esp32_capture	receive-only UDP; CSI/vitals/feature packets per node, loss, RSSI	device-access
ruview_mmwave_read	serial; firmware-identical MR60BHA2/LD2410 parsers	device-access
ruview_lidar_read	RPLIDAR SCAN over serial, or iPhone relay WebSocket (stats only)	device-access
ruview_host_list	configured SSH hosts	none
ruview_host_run	one read-only tool on an SSH host	remote-host + the tool's grant

Remote hosts live in ~/.config/ruview/hosts.json (mode 0600) and are added only from the CLI:

npx @ruvnet/ruview hosts add --name lab-pi --ssh ruv@lab-pi.local
ssh ruv@lab-pi.local true      # pin the host key once; BatchMode + StrictHostKeyChecking are enforced
npx @ruvnet/ruview hosts run --host lab-pi --tool ruview_esp32_capture --args-json '{"seconds":10}'

The remote side runs npx -y @ruvnet/ruview@<this version> call <tool> --read-only. Every argument is validated against the tool schema on both ends and single-quoted for the remote shell. Flash, train, calibrate, credentialed reads, and host hopping are never forwarded.

Firmware flashing (ADR-370)

Works on Windows, macOS, and Linux through python -m esptool (pip install esptool pyserial). The bundle directory is an extracted release flash bundle, firmware/esp32-csi-node/release_bins/<variant>, or an ESP-IDF build/.

npx @ruvnet/ruview ports
npx @ruvnet/ruview flash-plan --port /dev/ttyUSB0 --bundle ./bundle --variant s3-8mb
npx @ruvnet/ruview flash --port /dev/ttyUSB0 --bundle ./bundle --variant s3-8mb --confirm

Before writing, every image is checked against SHA256SUMS(.txt) (a mismatch is always refused), and esptool chip_id must report the chip the variant targets. After writing, a 15 s serial capture is summarized. The result sets hardwareValidated: true only when that log shows CSI callbacks without a panic. A successful write alone is not hardware validation. WiFi provisioning stays with provision.py (see the provision-node skill) so credentials never pass through the harness.

Training (ADR-371)

npx @ruvnet/ruview train-plan --mode pose-smoke        # review the command
npx @ruvnet/ruview train --mode pose-smoke --confirm   # SYNTHETIC smoke run (needs libtorch)
npx @ruvnet/ruview train --mode pose --data-dir data/mmfi --confirm
npx @ruvnet/ruview train --mode room --enrollment enrollment.json --confirm
npx @ruvnet/ruview train-gate --file eval-report.json  # PASS/FAIL + publishable wording

Pose training builds wifi-densepose-train with tch-backend. It needs libtorch: tch 0.24 expects torch 2.11.0, via LIBTORCH or LIBTORCH_USE_PYTORCH=1. Checkpoints and trainer logs go under v2/target/ruview-train/<mode> unless --checkpoint-dir names another path inside the checkout. The gate refuses any score that has no mean-pose baseline, uses a random-frame split, shares subjects across a grouped-subject split, overlaps in time, or fails to beat the baseline.

Debugging doctor (ADR-372)

npx @ruvnet/ruview doctor                          # all groups, human output
npx @ruvnet/ruview doctor --json --group python,serial,firmware
npx @ruvnet/ruview doctor --port COM7 --probe      # esptool chip_id (resets the board, writes nothing)
npx @ruvnet/ruview doctor --url http://localhost:3000   # sensing-server /health

Each check is pass, warn, fail, or skip, and every non-passing check carries a fix:. The exit code is non-zero only for failures.

SDK (ADR-369)

import { createRuView } from '@ruvnet/ruview/sdk';   // typed: src/sdk.d.ts

const ruview = createRuView();                  // { strict: true } throws RuViewError on ok:false
const report = await ruview.doctor({ groups: ['python', 'serial'] });
const plan = await ruview.firmware.plan({ port: 'COM7', bundle: './bundle', variant: 'c6' });
const flashed = await ruview.firmware.flash({ port: 'COM7', bundle: './bundle', variant: 'c6', confirm: true });
const gate = await ruview.training.gate({ model_score: 0.595, baseline_score: 0.501, split: 'chronological',
  train_end: '2026-03-01', test_start: '2026-03-02', n_test: 800, reproducer: 'python eval.py' });
const kernel = await ruview.kernel.load();      // optional @ruvnet/ruview-kernel
const radar = await ruview.devices.mmwave({ port: '/dev/ttyUSB0', model: 'auto' });
const nodes = await ruview.devices.esp32({ seconds: 10 });
const remote = await ruview.hosts.run('lab-pi', 'ruview_lidar_read', { source: 'rplidar', port: '/dev/ttyUSB1' });

The SDK calls the same policy-checked registry as the CLI and MCP server. Mutating calls still need confirm: true.

Cognitum Spaces OAuth

Activate the additional read scope through the Rust CLI, then use the same validated client through the metaharness:

wifi-densepose login --spaces
wifi-densepose whoami
npx @ruvnet/ruview spaces
npx @ruvnet/ruview spaces --resource sites --limit 50
npx @ruvnet/ruview spaces --resource events --cursor '<opaque-next-cursor>'

The metaharness never accepts a bearer token or API key and removes COGNITUM_SPACES_API from the child environment, so this surface cannot silently fall back to the compatibility API-key path. The API origin is fixed to https://api.cognitum.one, and the credentialed adapter requires an installed wifi-densepose binary rather than running Cargo build scripts from an auto-detected checkout. It returns only the bounded P2/P3 semantic projection. --resource selects one of sites, buildings, floors, spaces, zones, entities, events, or alerts; --limit is 1–100 and --cursor is the opaque value from the prior page. An empty list is a valid authenticated result, not sensing-quality evidence. An expired session may rotate the stored refresh credential before the read completes.

MCP use is denied unless the server operator starts it with RUVIEW_MCP_GRANTS=credential-use. Set RUVIEW_CREDENTIALS_PATH in the MCP server environment when a non-default store is needed; MCP calls cannot choose an arbitrary credential file or URL. spaces:read grants no write, pairing, command, policy-approval, spending, or actuator authority. See the bundled cognitum-spaces skill for the full playbook.

Codebase guidance

ruview_guidance is the read-only starting point for unfamiliar work. Filter by architecture, sensing, hardware, training, homecore, integrations, deployment, community, or testing, and optionally add a free-text query:

npx @ruvnet/ruview guidance --topic sensing --query "UDP CSI ingestion"
npx @ruvnet/ruview guidance --topic homecore --query "restore migration voice"

Each result separates implementation maturity from evidence, cites current repository paths, names focused validation commands, and states known limitations. In a RuView checkout, cited paths are checked before the result passes. Outside a checkout, the tool labels them as a reviewed packaged catalog. Related shared-brain records are bounded, reviewed, and treated only as evidence.

Skills

Host-neutral playbooks in skills/ (onboard, provision-node, calibrate-room, train-pose, verify, cognitum-spaces). npx @ruvnet/ruview skill <name> prints one.

Use as a Claude Code MCP server

The bundled .claude/settings.json registers the ruview MCP server (npx -y @ruvnet/ruview mcp start). Drop this package's .claude/ into a repo, or run npx @ruvnet/ruview install --host claude-code.

Claude Code mod: live sensing pane (ADR-377)

This package ships ruview-live, a Claude Code mod: a plugin whose behaviour is a function-hooks module. /ruview opens a pane beside the transcript, refreshed on a timer, with:

• your CSI nodes (ESP32, Realtek): rate, loss, RSSI and CSI shape;
• an optional ESPHome radar kit: presence, distance, and device-reported heart/breathing;
• alerts, such as a board that sends heartbeats but no CSI.

It also keeps a one-line status: RuView · 2 nodes · radar present · 1 alert.

Showcase views (ADR-378). The pane has three views, switched with keys 1, 2 and 3. The two live views are drawn in terminal cells (24-bit colour) and animated in place:

Key	View	What you see
1	Overview	the cards above
2	CSI waterfall	per-subcarrier amplitude over time (esp32 --spectrum), replayed at the frames' arrival rate and labelled MEASURED or SYNTHETIC from the packets' own flag; n cycles nodes
3	Radar	a 120° range fan with a sonar ping out to the measured distance; heart and breathing charts, and a ♥ and breathing gauge that pulse at the device-reported rates (a metronome, not a waveform)

The radar kit reports range, not bearing, so the arc covers every bearing at that range. Desktop and IDE surfaces show the overview, and a note in the live views.

npx @ruvnet/ruview mod                    # where the mod is in this install, and how to load it
claude --plugin-dir "<path printed above>"   # try it for one session
# or from the RuView marketplace:
#   /plugin marketplace add ruvnet/RuView
#   /plugin install ruview-live@ruview

In Claude Code:

Command	What it does
/ruview	Open or close the pane
/ruview refresh	Run one capture now; the result goes in the status line
/ruview off	Close the pane
/ruview waterfall, /ruview radar	Open the pane on that view

Settings (claude plugin configure ruview-live):

• udpPort (default 5005);
• radarHost, an ESPHome kit on your private network;
• refreshSeconds (default 15, minimum 5);
• liveRefreshSeconds for the waterfall and radar views (default 4, range 2–60);
• captureSeconds (default 3).

How it reads:

• The mod only runs this package's CLI, for read-only esp32 and mmwave --source esphome captures, with --json, so the pane shows exactly what those tools return.
• It never flashes, provisions or writes to a device.
• Like every mod, it runs inside Claude Code with Claude Code's access, so install it only from a source you trust.

Mods are early access in Claude Code. If /ruview is missing, start Claude Code with CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1.

ChatGPT and MCP Apps (ADR-375)

ruview_esp32_capture, ruview_devices_scan and ruview_doctor render in a self-contained console widget (ui://ruview/console-v1.html) in ChatGPT and other MCP Apps hosts, with a Refresh button that re-runs the tool. Results also carry structuredContent.

# MCP over HTTP: loopback, token-authenticated, read tools only
export RUVIEW_MCP_TOKEN="$(node -e "console.log(require('crypto').randomBytes(24).toString('base64url'))")"
RUVIEW_MCP_GRANTS=device-access npx @ruvnet/ruview mcp start --http --port 8790

Clients that can send headers use Authorization: Bearer <token> on http://127.0.0.1:8790/mcp. ChatGPT connectors cannot, so expose the port over HTTPS (a tunnel or reverse proxy) and add https://<host>/mcp/<token> as the connector URL. Write grants (hardware-write, workspace-write) are never honoured over HTTP: flashing and calibration stay on the CLI and stdio. Malformed request targets return HTTP 400 with invalid_request_target; the transport continues serving subsequent requests.

In a terminal, commands print a formatted view; pipes and --json print JSON. ruview esp32 --watch redraws live capture windows with per-node trends.

Hosts

Claude Code and Codex are implemented directly and tested with the local, non-interactive CLIs:

npx @ruvnet/ruview agent run --host claude-code --repo . --prompt "Map the sensing-server startup path"
npx @ruvnet/ruview agent run --host codex --repo . --prompt "Find the nearest tests for HomeCore restore state"

Prompts travel over stdin, never through a shell. Both adapters are read-only by default (claude -p --safe-mode in plan mode; codex exec in its read-only sandbox with user config and exec rules ignored), use a scrubbed environment, bound output/time, redact secrets, and require a trusted RuView checkout. Workspace writes require both --allow-write and --confirm; dangerous bypass flags are never emitted.

Shared contributor brain

The committed brain/corpus/core.jsonl is a small, reviewable source of repository facts. Every record has a source citation, evidence tier, tags, and review state:

npx @ruvnet/ruview brain search --query "darwin community memory"
npx @ruvnet/ruview brain verify --repo .
npx @ruvnet/ruview brain propose --id finding-id --title "Finding" \
  --content "Source-bound observation" --sourcePath README.md --sourceLine 1 \
  --tags onboarding,docs --contributor github-user

Proposals are unreviewed JSONL for a normal pull request. Local vector indexes, private overlays, raw agent transcripts, CSI/person data, and credentials are never part of the shared corpus. Retrieved text is quoted evidence, not an instruction or authority grant.

Ruflo + Darwin/Flywheel

Development tooling is exact-pinned in devDependencies: metaharness@0.4.1, @metaharness/darwin@0.8.0, and @metaharness/flywheel@0.1.7. Ruflo remains an optional contributor coordinator rather than cold-start weight for the dependency-free published MCP server:

claude mcp add --scope project ruflo -- npx -y ruflo@3.32.26 mcp start
codex mcp add ruflo -- npx -y ruflo@3.32.26 mcp start

npm run flywheel:plan is read-only. Darwin execution is human-triggered with node flywheel/run.mjs --confirm; it writes only an untrusted .metaharness/ proposal archive. The protected gate requires frozen-anchor retention, holdout lift, security and legacy-test success, verified provenance, and human approval. No contributor run can directly replace or publish the champion.

License

MIT © ruvnet