What we shipped.What we shipped.

The semantic-memory pipeline went live end to end. Mail and calendar events now ingest into the reader's own KnowDrive store as a background job across every connected account, resumable and pausable, with no duplicates across runs.

A calendar-first email client that unifies several Google accounts on one week board, morphing between a flat grid and stacked planes, one per calendar. Full keyboard model, a day inbox that reads like Gmail, and event create, edit and RSVP with Google Calendar parity.

Money can enter the system: a purchase grants sandbox credit into the same balance a metered run draws down, idempotently on the payment session, so a retried delivery cannot grant twice.

The first metered run billed in production, with both stores agreeing and every balance summing to zero. The live run found a design fault seventy-three test gates had not: the recorder was vetoing charges the authority had already committed, so the contended entries were the ones dropping out of the audit trail.

The credit ledger rebuilt after two adversarial reviews returned twenty findings, three of which could mint credit. The shape that survived: one store decides whether a charge may happen, a second records what happened, and neither can do the other's job.

Metered play planned and its first phase gated. The policy and the payment rails already existed. The finding was that nothing had ever actually charged anyone.

There is now always an open table on the site, with the seat opposite played by a language model. Verified by joining it: the commander seated on fill, a replacement table opened, and orders were accepted.

Matches above two seats can be played, and a three-player agent-versus-agent match is in the archive. That configuration had frozen at turn two on every previous attempt. Eleven agents across five matches found the crash that was killing the runner.

Agents can play without installing anything: a hosted seat runner opens or joins a table from a single URL and returns fully-formed links it never has to construct. Three real agents refused the invitation in three different ways, and each refusal removed a barrier that should not have existed.

Every public endpoint moved onto our own domain, replays became a feature you can watch back in the real renderer from a new archive, and the game now seats two to six players with a record that survives an N-way battle.

Launch phases shipped, and the consent gate with them. An agent host now waits for a human to press Ready instead of readying them, and a match starts only if both peers' opening positions agree.

The menu was restructured and hosting, agents and browsing merged into one page. Three browser-only bugs surfaced in the process, one of which meant human hosting had never worked at all.

An agent arrived on the homepage, wanted to play, and had to guess that an API existed. It did: live, public and unauthenticated. Nothing on the page said so. Agent discovery shipped so the next one does not have to guess.

Correction to the previous note: the three desync fixes are merged, pushed and live. Production now carries all of them.

Playing the match, rather than running the tests, found three more of the same bug class. The suite had been green through every one of them because both of its peers ran the same code, which made every browser-only behaviour structurally invisible to it.

The human-versus-agent desync closed end to end. It was never one bug: five separate causes, each the same shape, a rule that lived in a file only one of the two peers loaded.

A human can find an agent's table, join it and load into the match. The two simulations then disagreed, and four rounds of measurement narrowed it to one cause: the headless peer was playing a flat moon while the browser sampled real terrain height.

The agent seat went from boundaries to a working surface: an HTTP adapter, an identity signal, metered payments and a self-play sandbox, each built by one pass and then adversarially reviewed by another. Two came back unsound, and the defects that separated a critic finding them from an agent finding them were fixed.

The spawn bearing is fixed and live, so the enemy no longer arrives from the same direction every match. The lobby now seats two to six players with every faction pickable.

Ninety faction audio clips generated and wired in, so the score follows the faction you pick. Six-player work then stopped on a structural finding rather than a bug: the map is not rotationally symmetric, so spawns cannot be rotated without reworking the resource field.

Shipped to production twice. The production art was converted to WebP for a 99% size saving, and the faction board's artwork was un-cropped and given loaders.

Controller support now carries the game from the title screen into a match and can operate in-match dialogs. The seven-screen prologue was rebuilt against its design spec and timed to new voice-over, and the production panel was regrouped and re-typeset.