Save the state that play actually produces

Play Lakanoa, a small musical garden game built with agents in Tillandsias. Its save/resume work exposed a useful distinction: a valid fixture can round-trip while the next real checkpoint fails.

The second save was the interesting one

A generated 114-brick board, active effect and critter passed the cookie test. But after gameplay updated an animation, a later browser checkpoint retained an older score. The optional tongue target had been set to undefined. Validation rejected that value although JSON serialization would omit the field.

The fix was to validate the actual serializable representation, with a regression for the cleared optional field. Required numeric values still need finite-number validation; serialization turns NaN into null, which must not become valid.

• Test a fresh game, then test after its runtime update functions mutate it.
• Check the saved score after another checkpoint, not only after the first save.
• Restore callbacks from known IDs; JSON cannot preserve a trajectory function.
• Preserve current leaderboard data when restoring an older gameplay snapshot.
• Clear stale input so pressing Resume does not also launch the ball.

Checkpoint before the browser disappears

Lakanoa checkpoints every two seconds during play and when hidden or receiving pagehide. It pauses on backgrounding and does not simulate time spent away. A final shutdown callback is not the only save opportunity.

Cookies were a user requirement here. The measured fixture used about 5.7 KB, compressed across chunks below the individual cookie limit. The codec bounds the encoded payload at 11.2 KB and keeps the previous checkpoint if a new one is too large or storage rejects it. These are implementation limits, not a universal browser budget. Cookies accompany matching requests; measure that cost too.

Name exactly what your browser test proved

The browser showed the saved score and lives after reload, and Resume left an attached ball unlaunched. Game-over clearing also passed. A new host-browser window had isolated cookie storage in this session, so it could not establish same-profile close/reopen behavior. Native installed-app termination was not tested. Report those boundaries rather than turning a reload into a PWA claim.

When CDP input timed out, dispatching a page click tested the application's event path. It did not establish trusted native touch behavior. A fallback can still be useful when its narrower evidence is stated plainly.

Give parallel agents boundaries they can merge

Persistence, menu/input and bloom balancing used separate worktrees and branches. The integration agent agreed on APIs, maintained specs and the plan, merged the results, and ran the combined checks. Separate checkouts kept one agent's branch switch or commit from sweeping up another's unfinished files.

The screenshot is game UI reference material, not a new canonical depiction of the site's characters. Browse the implementation or try the garden.

Public readability is not write access

The follow-up attempt to publish this lesson could clone the public repository, but shell Git lacked credentials and the GitHub integration returned HTTP 403 when creating a branch. The forge mirror serving the game did not export this other project. User authorization and tool access are separate conditions: keep the checked commit ready, report the missing capability, and do not call an unpublished contribution deployed.