Skip to content

The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.

6.3 KB
Raw Blame

Experience verification

The seven authoring rebuilds use source inside each package's template/studio/ and their own builders. build-experiences.mjs delegates to those builders. The original models remain under store/tools/experiences/ as retained history; their tests do not stand in for the rebuilt tools. See current release status. These checks establish technical behavior and delivered output, not a subjective product response or an installed-agent trial.

Game Master and Lab Bench provide package-local browser/acceptance suites. Lab Bench additionally reopens all six authored deliveries, exercises numeric/category/block controls, and independently reconstructs its design matrices with Python statsmodels/SciPy. See each package's test/ACCEPTANCE.md.

Creative Direction's evidence includes three distinct briefs, targeted revisions, actual canvas/source saves and independently read print deliveries. Run its test/acceptance.mjs, then test/verify-delivery.py PATH_TO_RUN with Poppler available.

From the repository root:

node store/tools/build-experiences.mjs
node store/tools/build-experiences.mjs --check
node store/tools/build-experience-branding.mjs --check
node --test store/tools/experience-tests/models.test.mjs
node --test store/agents/generative-art/test/*.test.mjs store/agents/music-studio/test/*.test.mjs
node --test store/viewers/web-viewer/test/*.test.mjs

Legacy workspaces include a pure-model checker. Run it in their workspace:

node tools/check.mjs --seeds 100

It extracts the pure model from the actual edited artifact, exercises domain invariants over the requested seeds, repeats each result and writes .harness/model-check.json. It does not claim browser, visual or listening verification and does not mark the workspace ready.

The rebuilt Art and Music workspaces use node tools/check.mjs for source/preview agreement and node tools/export.mjs for real browser delivery. Art supports --seeds N --scale 1|2|4; Music exports the score, mix, MIDI and aligned stems. Their setup scripts install pinned tools locally. Neither exporter sets a ready verdict from technical measurements.

To reproduce all six targeted revisions, including the saved-project import/backup workflow:

node store/tools/experience-tests/revisions.mjs /tmp/authoring-revisions --export
python3 -m venv /tmp/music-delivery-reader
/tmp/music-delivery-reader/bin/pip install mido==1.3.3
/tmp/music-delivery-reader/bin/python store/tools/experience-tests/verify-music-delivery.py /tmp/authoring-revisions/keepsake/after

The independent reader checks WAVs, MIDI and stem summation with Python wave/zipfile and Mido. It deliberately reports listeningReviewed:false.

Real browser suite

Install the pinned development dependency in this directory, with system Chrome available:

npm install
npm run browser

BROWSER_EXECUTABLE selects another Chromium executable. PLAYWRIGHT_MODULE can select an existing Playwright Core module by absolute path. EXPERIENCE_OUTPUT selects the evidence directory (default: work/experience-evidence). Set EXPERIENCE_IDS=generative-art,music-studio to select the rebuilt tools. The test dependency is local to this directory; production export dependencies live in the two packages' toolchain directories, outside user workspaces. Set SOFTWARE_WEBGL=1 to exercise SwiftShader instead of the machine's graphics driver. Run node performance.mjs for 30 warmed artifact requests and five save-to-ready-and-paint samples per experience. It waits for the changed document and two animation frames, not just an iframe URL change. Timing is local measurement, not a performance guarantee.

The suite creates temporary workspaces, copies each template, runs its init script, expands the manifest's actual viewer URL, and starts this checkout's viewer. It checks real controls, downloads and their contents, seed changes and reload preservation, paused file edits, 390px layouts, browser exceptions and screenshots. A separate fixture verifies sibling fetch, localStorage, ES modules and recovery when a missing artifact is created.

The retained original-model suite checks 100 art/brand/data/music-score/arena/autopilot seeds, 16 rendered audio arrangements and 32 walkable voxel worlds. These are finite samples, not a claim about every seed or every browser. Browser tests are separate from Node coverage; a Node coverage number must never be reported as coverage of the UI or shell scripts.

Running app integration

With the local daemon running and these seven agents plus autonomous/web-viewer installed from this checkout, run HARNESS_MACHINE_ID=<local machine id> node daemon.mjs. It creates temporary idle Claude sessions through the desktop's loopback protocol, checks engine readiness, framework materialization, instructions, initial verdicts, installed viewer routing and a browser file reload. It sends no prompts and deletes only its own temporary sessions and workspaces in cleanup. EXPERIENCE_IDS also selects packages here. A passing idle-session check does not establish that the coding engine can complete a real brief. That is a separate acceptance gate.

Original harness identities

Each package's brand/icon.svg is the source for its mark. The experience builder embeds it into the workspace header and a data-URL favicon so both work offline. The desktop's shared EngineIdentity mapping uses the corresponding assets/engine-icons/<id>.png for Store cards, the picker, pane headers and native tabs. Those bundled icons arrive with a desktop build; publishing a Store package cannot replace an already-running desktop binary's assets.

After editing a source vector, use the same Chrome/Playwright setup as the browser suite:

node store/tools/build-experience-branding.mjs
node store/tools/build-experiences.mjs

The branding builder renders 256px PNGs, creates light/dark SVG logo lockups and a visual proof sheet at 16/24/48px. --check needs no browser: it verifies source hashes, logo generation, PNG dimensions and byte-for-byte agreement between package and desktop assets. The original marks are MIT-licensed work by OpenHarness contributors; they represent these harnesses, not third-party products. The brand assets are available from each package README.