Skip to content

Current implementation capture bundles

← Figma implementation

Visual acceptance in runs using --figma-file, an accepted UI handoff, or --figma-manifest needs implementation images bound to the current workspace. A PNG from an earlier build cannot establish acceptance for a later build, even when a reviewer gives its report the latest source revision.

Run from the task workspace after the implementation and capture fixture are ready:

autocode visual-capture --config capture.json --timeout 120

The command uses Node and the project's existing Playwright Chromium installation. It does not call a model or install dependencies. The project supplies a fixture that starts its server, selects the requested state and tears down resources. The collector owns the fresh browser context and screenshot operation. Worker permissions still apply; a denied browser/server operation must be addressed before visual acceptance can proceed.

Copy reference_hash from implementation_captures in the stage context. For an exported inventory, copy the exact case ID, route, state and viewport too. For a native Figma URL, use stable case IDs for the approved visual requirements. Prepare all fixtures and separate config files for each case before capturing; editing a shared config for the next case would invalidate the previous capture. Write those configs, any fixture created for the review, and comparison/review artifacts under .autocode/evidence/ so they do not change the captured source snapshot, and leave existing capture bundles intact. A later repair re-verifies every file a report cites, so under .autocode/ the runner accepts citations only from .autocode/evidence/, the run directory, capture bundles (a manifest and the artifacts it lists, for the run's design reference) and the retained design inputs. A Validator report citing anything else there is rejected when it is accepted.

{
  "reference_hash": "<64-character reference hash from the stage context>",
  "case": {
    "id": "workspace.desktop.empty",
    "route": "/workspace",
    "state": "empty conversation",
    "viewport": {"width": 1440, "height": 900, "device_scale_factor": 1}
  },
  "fixture": "tests/capture-workspace.cjs",
  "inputs": [],
  "assets": [{"url": "/workspace", "path": "index.html"}],
  "ready": [{"selector": "body", "attribute": "data-state", "equals": "empty"}],
  "build_command": []
}

Here is a minimal fixture for a static index.html with data-state="empty" on its body. Application fixtures should establish the actual approved state and await its readiness before returning from setup.

const http = require('node:http');
const fs = require('node:fs');
let server;
exports.setup = async ({page}) => {
  server = http.createServer((req, res) => {
    if (req.url !== '/workspace') {
      res.writeHead(404); res.end(); return;
    }
    res.writeHead(200, {'content-type': 'text/html'});
    res.end(fs.readFileSync('index.html'));
  });
  await new Promise(resolve => server.listen(0, '127.0.0.1', resolve));
  await page.goto(`http://127.0.0.1:${server.address().port}/workspace`);
};
exports.teardown = async () => {
  if (server) {
    server.closeAllConnections();
    await new Promise(resolve => server.close(resolve));
  }
};

Declare every loaded document, stylesheet, script, image, font and data response in assets, with an origin-relative URL (including its query string) or an exact HTTP(S) URL and the workspace file whose bytes it serves. The collector hashes the response bodies actually received by the browser. A separate HTTP fetch cannot substitute for those observations. Redirects, failed requests, undeclared responses and mismatched bytes refuse capture. Pending requests at readiness or new requests afterward also refuse capture; stop polling and await fixture data before setup returns. Service workers are blocked in the fresh browser context.

Use a production build/preview for dev servers that transform files in transit. Set build_command to its argv, for example ["npm", "run", "build"], and point the fixture and assets to those outputs. Ignored/generated outputs require a successful build during this capture operation. Its command, result, source revision and output hashes are retained. Keep generated outputs Git-ignored so the build does not change the source snapshot. Files named in inputs add exact hashes for other dependencies, including ignored configuration. All file paths must be normalized paths relative to the workspace; symlink inputs are refused. Optional playwright_module and executable select existing local installations.

Successful output contains capture_ref, capture_sha256, candidate_ref and visual_acceptance: null. Each attempt gets a new directory under .autocode/captures/. The manifest records source, reference, fixture, collector, input and artifact identities, readiness, route and viewport. Changes during capture refuse publication. Failures retain their available logs and artifacts; previous captures are never overwritten. Do not commit these generated bundles.

Stage contexts explicitly list current bundles, rejected captures with reasons, and unavailable inventory cases. The independent reviewer opens those candidates and compares them with the design. Exported-inventory reports copy the receipt into each design_results row alongside candidate_ref and comparison_ref. Native Figma reports use implementation_captures entries containing only capture_ref and capture_sha256, and cite their candidate images in criterion evidence. Partial nonvisual milestones can leave this array empty; whole-task visual acceptance cannot. An additional historical image in passing criterion evidence cannot be hidden beside a fresh receipt.

Report decoding and completion recheck provenance and pinned evidence. Source changes after capture, missing manifests, changed images/fixtures/build outputs, wrong reference/case/state/viewport and substituted candidates refuse acceptance. Older runs with unbound screenshots need fresh captures and independent review. Freshness never changes a FAIL or NOT_VERIFIED judgment into PASS.

This protocol assumes the project's fixture, build command and local tools are honest. It detects stale or inconsistent evidence; it is not an attestation against a malicious process fabricating every input and receipt. A native Figma reference hash binds the supplied URL, not the remote file's current contents; reviewers must still inspect that design. Selecting all required states and judging visual fidelity remain independent obligations.

Verification without model spend

The regular suite includes scripted-provider completion tests and provenance guards. To also exercise real Chromium acquisition with an installed browser:

AUTOCODE_TEST_PLAYWRIGHT=/absolute/path/to/node_modules/playwright \
AUTOCODE_TEST_CHROMIUM=/absolute/path/to/chromium \
  .venv/bin/python -m unittest -v tests.test_visual_capture

Omit AUTOCODE_TEST_CHROMIUM to use Playwright's default installed browser. Set AUTOCODE_TEST_CAPTURE_EVIDENCE to a fresh ignored directory to retain the disposable browser test projects. Without the Playwright opt-in, these real browser tests are explicitly skipped; offline guard tests still run. These checks do not establish live Figma fidelity or model review quality.