[sreejithraman]

The chief's command center fell behind its record. The chief kept two hand-made copies, record.md and program.json, and republished the page only when it judged a refresh worth it. This PR makes the page a mechanical result of one committed record.

What changes

• One record in git. program.json holds the whole program, organized by topic: program facts, decisions[], and work[] rows that carry their owner, scope, and sources. The program folder is a git repo, and HEAD:program.json is the record. Owners still write only reports/<ID>.md.
• A fixed write procedure. Settle any uncommitted change with the user, edit, run render_dashboard.py --check, read git diff, commit, publish. On a git or publish error the chief stops and reports.
• The command center is a private Claude artifact.
  • --page renders HEAD to a .next file and prints changed when it differs from the last published page.
  • The chief publishes on changed, then moves .next into place. A failed publish shows changed again on the next write.
  • Record-only commits leave the page alone.
  • The skill states only when and what to publish; agents already know the artifact tools.
• No hand-set timestamps. The page's Updated time comes from the commit that changed it. Each row keeps its own checkedAt.
• Stricter --check.
  • It rejects unknown keys, which catches half-migrated files.
  • It reads --revision from HEAD.
  • It refuses to overwrite the state file.
  • It keeps every coordination rule, including scrap-or-rebase holds on dead bases.
• Chief moves. Only a Claude session with artifact tools can be chief. A new chief confirms it can read the artifact. The old chief hands off right after its commit, and the new chief publishes.
• Removed:
  • record.md and the separate snapshot;
  • the static server, staging, and swap steps;
  • the refresh judgment;
  • page polling;
  • the Tailscale hosting steps.
• Browser tests use Playwright's Chromium when it is cached, so check.py no longer leaks a branded-Chrome code-sign clone on every run.

Verification

• python3 scripts/check.py passes. The command-center browser tests now load pages from a loopback server, which fixed a reload test that failed about 1 run in 8. Full runs went from 12 failures in 40 to 0 in 40.
• The real publish loop ran against a private example artifact:
  • A display commit published.
  • A record-only commit didn't publish.
  • A skipped publish stayed changed.
  • The Updated time matched the display commit.
• In claude.ai's viewer:
  • The page renders, and its icons load.
  • An open tab moved to the new version without a reload and kept its selected tab.
• Ten review passes (Primary, Thermo, Ponytail, Adversarial) plus fresh-session dry runs. The last pass found nothing material.
• Not checked: the Claude iPhone app.

After merge

The live seatcommons program still uses the old format: record.md plus the old program.json. It needs a one-time migration into the new file in a git repo, then its first artifact publish.

🤖 Generated with Claude Code