Repository navigation
v0.1.19
Pre-releaseWhat this build changes
The window says what a bot is doing. A dot beside a bot's avatar now reads working, watching or asleep, taken from the two live tables the kernel already keeps and never stored as a field of its own — a field would be a second answer that could disagree with the loop it describes. A fourth state, needy, is not drawn, because nothing in the kernel produces it: there is no path that asks you and waits for your answer. Three readings that are true are worth more than four, one of which is decoration.
A bot's message says whether it came to find you or answered you. The two sentences are the ones docs/design/agent-lifecycle.md says a reader must not confuse. Nothing here decides which a message is: delivery already computed, once, whether an entry addressed this bot, and the turn now records which entries woke it, so the mark is a read. true is "it came to find you", false is "it answered you", and null is "the record does not say" — a message no turn produced draws nothing at all rather than printing an answer to a question nobody asked. Only a bot is marked; a person's message is answered by the room, not by a wake.
A bot that said nothing now says which kind of nothing it was. A message's silence list carries two readings: not-addressed, where the judgement settled the entry as ignore and no turn ran, and nothing-sent, where a turn ran, called a model and produced no message. Before this, a bot that weighed the room and chose to stay quiet looked exactly like a bot that never woke. Two rows are deliberately not silences: a turn still deciding (ended_at is written in the turn's finally, so a turn without one is in flight) and a turn that ran no model at all — the window says nothing rather than inventing a decision.
bash runs now, and the kernel puts a fence around it. The roster has always listed it and the runtime never executed it, because a shell names no path and "inside the bot's own directory" is not something a name check can say. The bound is enforced instead of refused: one command inside sandbox-exec, with no writing outside the bot's own directory plus one private scratch directory, no network, and no reading of the person's home except where the bot's directory sits inside it. The child gets the system PATH, a private TMPDIR, and nothing of the app's own environment — an exported key would otherwise be readable by env. A machine that will not apply the fence has the call refused rather than run unconfined, and that refusal is kept apart from an exit status: one says the command failed, the other says there is no shell here and nothing typed will run.
A roster left behind by an older version no longer mutes every bot. This is the defect behind a window that said the model is not ready while the model was ready. agents.yml is the person's file and the turn artifact ships with a release, so an install whose roster still said tools: [say] met an artifact offering six and the bind demanded the two lists be equal — every bot refused, and the window had nothing to report it as except a model that was not ready. The row is now a ceiling and the artifact is the offer: a turn is given their intersection, a narrow row costs a bot those capabilities and never its voice, and one row is still refused — the one that lost say, which is the edit the README invites and the only one that leaves a bot with a full box and no way to answer. Measured over a copy of a real world with that unmodified roster: before, stage2 quiet: [Echo, 小李, 小荷, 小胖]; after, stage2 quiet: [], refusal: none, both halves bound.
What this build has not been shown to do
- What the
bashfence blocks is not measured. The machine this work was done on refuses a nestedsandbox-execwithsandbox_apply: Operation not permitted, so the tests pin both branches — the refusal, where the command is never started, and the spawn-and-read path — but whether the profile really denies on a host that can apply it is a reading this build does not have. What it would have to show is a command writing outside the bot's directory and being stopped. - The window still cannot say why a half refused.
stage2.refusaldoes not reach the screen, so any mute that is not the model reads asthe model is not ready— the same false sentence this build fixes the cause of, but not its reporting. That is its own change. - The acceptance case was not re-run for this build. The readings above are the ones taken while each change landed, not one fresh run of the whole room.
What is in the download
Seek-0.1.19-arm64-mac.pkg, 133,419,547 bytes, andSeek-0.1.19-arm64-mac.zip, 134,453,340 bytes.- Signed
Developer ID Installer: anqing shao (Y55RQ6LU24)and notarized:pkgutil --check-signaturereportsNotarization: trusted by the Apple notary service, andspctlaccepts the package and the app inside it, each asaccepted / Notarized Developer ID. CFBundleShortVersionStringandCFBundleVersionboth read0.1.19.- Eighteen prompt artifacts ship inside the app.
work-v13.jsonis the one a turn loads, and it is the first that declaresbash— seven tools, checked against the package.compact-v3.jsonwrites the notes andwatch-v4.jsondecides whether a bot speaks. - The bundled server reports
version: 0.7.0-dev (build 11946, commit 8ca4060b7), built for Darwin arm64. - Launched once from the extracted package in a fresh directory: the window drew in 1.8 s, made its world —
meandEcho (bot)inthe first room— and the bundled server loaded the decision model and named itlfm3-d1.
Checks
Kernel suite: 479 tests, all passing. Interface suite: 326 tests, all passing — on the merged commit as well as on the pull request's head, which is the same tree (6381fc46).
Full Changelog: v0.1.18...v0.1.19