[sangyeT]

Why

Tickets ship well through the delivery loop, but demos paid the full cost of that loop: Jira, controller, frozen evals, independent review, CI and merge. This PR adds a separate demo lane, plus a UI checker that both demos and delivery can use. With the checker, an agent drives the UI, then looks at every screenshot, with computer use or by opening the images, before the UI counts as passing.

What changed

sstack-demo skill (tools/sstack/skills/sstack-demo/SKILL.md)

• The lane runs script → build and self-check → honesty check → two rehearsals, then either promotes the demo to tickets or archives it. It skips the delivery gates.
• The script has 3–7 steps, each labeled real or faked. A real step needs proof, and every mock in the branch diff must appear on the faked list.
• Demo work stays on demo/* branches, never merges, and never counts as ticket evidence (AGENTS.md).
• Kizen writes defer to platform.md approval. Record writes go only to a sandbox approved for this demo, and resets use a fresh run marker instead of deleting records.
• Promotion maps real steps to schema-valid eval cases: business values, phase and surface. Faked steps become scope with their own acceptance criteria.
• Routing: /sstack and sstack-mode send demo requests here. The skill is registered in both host skill folders, install.py, the contract checker, CLAUDE.md, the README and the installer's managed instructions.

tools/sstack/ui_eval.py

• run SCRIPT.json --approved-origin <host> drives a JSON step script in Chromium. Steps are goto, click, fill, press, wait_for, expect_text and capture.
• Evidence per run:
  • a screenshot for every step, and optionally a video
  • console errors, page errors, 5xx and network failures, and top-level 4xx
  • a {run_marker} value unique to each run
  • captured values in the eval result-case shape
• Origin guard:
  • top-level navigation elsewhere is aborted before any request is sent; this covers links, delayed scripts, popups and every hop of a server redirect chain
  • goto paths must stay under base_url, so file:, scheme-relative and backslash paths are rejected
  • the saved login (--storage-state) must live outside the repository
  • recorded URLs drop query strings and fragments
• review REPORT VERDICTS holds the visual verdicts. It binds each pass/fail verdict to its screenshot's SHA256, blocks on missing verdicts or reports with nothing reviewable, and fails on any fail verdict or failed automated check. Verdicts are the reviewer's declarations and are not authenticated.
• Docs: tools/sstack/skills/sstack-verify/references/ui-eval.md. sstack-demo and sstack-verify use this checker for UI steps.

Dependencies and CI

• playwright==1.56.0 is added to requirements-dev.txt. This is the stack's first non-stdlib runtime tool; it is imported lazily, so the other tools stay stdlib-only.
• CI installs Chromium and sets SSTACK_REQUIRE_BROWSER=1, so browser tests fail rather than skip there. Locally they skip only if Chromium can't launch.
• test_install.py counts host links from SKILLS instead of a hard-coded 8.

Validation

• The full stack gate passes: control.py verify changed --base origin/main --jobs 2 ran 124 tests, lint and docs checks, with the browser tests running in real Chromium.
• Mutation checks: each of these makes a test fail:
  • removing the route guard
  • replacing checked redirects with direct fulfilment
  • removing the post-run settle
  • disabling the 5xx or page-error listeners
  • review ignoring automated failures
  • dropping the sign from number capture
    One mutation isn't caught at run time: letting a screenshot failure pass, because tests can't force a screenshot error. Review still blocks that step.
• One independent reviewer covered five rounds: the skill, the checker, and three re-reviews. Its exploits for local files, off-origin hosts, popups, form posts, meta refresh, single and multi-hop redirects, late navigation and empty reviews are all blocked now. It also confirmed that cookies across redirect hops, POST→303→GET, same-origin chains, timing, history and screenshots behave correctly.

Known limits

• The checker has been tested only against a local fixture, not against a live Kizen login (SSO, MFA, session expiry). A login redirect to another host fails the run by design, so real runs need a saved session made with npx playwright codegen --save-storage=….
• Redirect hops are now script navigations: Referer is the previous hop's URL, and user activation isn't carried over. 307/308 redirects of form posts fail the run.
• CI downloads Chromium on every run with no cache, and the post-run settle raises stack-test time from about 20s to about 30s.
• Upgrades: existing consumer installs will be told "Managed instructions changed" or "Existing file differs" and need a manual reconciliation, like any release that changes tool files.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HSV4zUy6rF1vgHFSyUmuva

---

Generated by Claude Code