Evidence focus garden

Evidence: Focus Garden from idea to production

Summary

Note

Screenshots are real captures. The app screenshots come from scripted headless Chromium against the production build with seeded demo data (a fixed clock so the garden shows a full week). GitHub pages were captured while logged out, so no private data appears.

One brief went in. The orchestrator decomposed it into 21 tasks, dispatched 6 specialist agents, ran the gates, paused for a human, looped back once on changes requested, then deployed the approved commit.

How the evidence was produced

Item	Value
Tool	GitHub Copilot CLI 1.0.91, plugin ai-team-sdlc@ai-team-sdlc v1.0.2
Entry point	copilot -p "Use the ait-sdlc-orchestrate skill ..." --allow-all-tools --share <file>
Run id	2026-10-05-focus-garden (resumed from state.json across 4 invocations)
Stack produced	Vite, TypeScript, no runtime dependencies, offline PWA, GitHub Pages
Tasks	T-001 to T-021 (design, prototype, PRD, architecture, 7 build slices, QA, critic, security, fixes)
Tests at the end	159 unit tests, 16 Playwright scenarios, PWA and Lighthouse checks

Timeline

Step	Invocation	Duration	AI credits	Result
Design to QA	Run 1	2 h 47 min	2,097	Stopped at the QA gate (three manual checks)
Waiver, critic, security	Run 2	19 min	392	Stopped at human sign-off
Changes requested	Run 3	57 min	747	Three defects fixed, gates repeated, stopped at sign-off again
Sign-off and merge	Run 4	22 min	92	Approved commit merged; first release attempt cancelled by an Actions incident
Release fix and deploy	Manual	a few min	0	CI portability fix (PR 2), governed release, live smoke check

1. Install the plugin and bootstrap

The plugin is a marketplace in its own repo. Two commands install it, and ait-init wires a repository to it.

The plugin repo also publishes an executive presentation of the lifecycle through GitHub Pages.

The sample repository starts empty, then ait-init adds the plugin settings and a thin AGENTS.md.

2. The brief

The only human input was one page: the idea, the users, a v1 scope, non-functional requirements, the delivery target and the run rules ("never push, pause at sign-off").

3. Design and prototype

The designer agent produced experience flows, wireframes and three themes, then passed design-review.

A throwaway clickable prototype was built and driven in a real browser at desktop and 360 px widths: 7 of 7 checks passed with no console errors.

4. The first build, before the human review

The frontend developer built the app in seven slices (shell and themes, persistence, timer, rewards, garden, data controls, PWA). This is what the first signed-off candidate looked like.

5. Requirements and architecture

The product owner wrote a PRD that traces every item in the brief. The first draft failed spec-review (it had silently changed the timer presets) and was corrected. The architect then wrote the technical spec and four ADRs.

6. Human sign-off: changes requested

QA passed (with a recorded human waiver for three manual-only checks) and critic and security reviews passed. The human then reviewed the artifact and requested changes for three defects that every automated gate had missed: a malformed SVG path (console errors in the garden), plants that were small and similar in colour, and a heavy focus outline after navigation. The plugin re-ran the build, QA, critic and security gates on the fix.

The "before" captures are the section 4 images (04-*). The "after" captures below (06-*) use the same seed data, so they compare directly.

7. QA, critic review and security

See Gates and sign-off for what each gate caught.

8. Sign-off, merge and deploy

After the second sign-off review the human approved all three roles for commit 5d3c73f. The orchestrator recorded the approval, pushed the branch, opened a pull request and merged it without squashing, then verified that the merged tree equals the approved tree.

The first release attempts

The first release attempt did not deploy. It was cancelled before any step ran, during a GitHub Actions incident, and was re-dispatched once the platform recovered.

The second attempt ran for real and failed at the performance gate. tests/lighthouse-audit.mjs had a Windows Chrome path hardcoded as its default, so on the Ubuntu runner it died with spawn ... chrome.exe ENOENT. Every local gate had passed on a Windows machine; only the real deploy workflow could expose this.

The fix is a two-line change in a test script: default to Playwright's Chromium and keep CHROME_PATH as an override. It went through its own pull request, and the performance gate passed locally (score 1.00) before the merge. No product code changed.

Note on the approved tree. Pull request #2 changes one test file, so main is no longer byte-identical to the approved commit 5d3c73f. The deployed product code is the approved code; the difference is CI tooling only. A stricter team would ask for a fresh sign-off here. This run recorded the change and its reason instead.

The release

With the fix on main, the governed workflow ran with governance_approved=true. All three jobs passed: quality and release checks, the immutable Pages artifact, and the signed-off deploy.

9. Smoke check on the live site

The deploy gate checked the live URL, not the build output. A scripted headless browser loaded https://devopsabcs-engineering.github.io/ai-team-sdlc-sample-focus-garden/ and confirmed:

Check	Result
HTTP status of the home page	200
Web manifest served, start_url and scope under the repository base path	Yes
Service worker registered with the repository scope	Yes
Console errors	0
Requests to other origins	0
Reload while offline	200, page renders

Rollback. Each Pages deploy is an immutable artifact tied to a workflow run. To roll back, re-run the workflow on an earlier commit, or revert on main and dispatch again. The previous product is redeployed through the same governed path.

Lessons learned

• Gates catch what the author misses. The PRD, a state bug, a lockfile problem and an environment problem were all stopped by a gate, not by luck.
• Human review still finds things. Three visual and interaction defects passed every automated gate and were found by looking at the product.
• Long phases are expensive. The first run used about 2,100 credits and nearly three hours. Smaller scope and bounded invocations would cut that.
• Evidence needs a scripted capture. A fixed clock and seeded data make screenshots rep
• Run the deploy workflow early. Local gates on one OS missed a Windows-only path that the Ubuntu runner exposed on the first release.eatable and let console errors be counted.
• Windows consoles garble UTF-8. The --share export is clean; the console log is not.