repo-activity landed; wear-facts is next and wake-room-tail after. The forge
owns repo facts and the daemon reads it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A tell is the message (first line carries the machine, later lines messages and
organ notices); upsert before delete in the source index; the CPU reserve's own
embed bounds; the tempo fallback controller merged but not wired; Mind web
builds and the recorded shared origin; skill pin b22762a. The design doc's
section 3 no longer says every line.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
57 of the Minds' 157 Self runs lived under a minute. A start's answer names
the run with no read, so a run she started that opens and ends between two
reads of the slot still gets its one notice.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The watch started empty in each process, so it missed a run live at the
session's start, a heartbeat's run and one live across a rotation, and it
cost 3 reads a poll at about 1.85 s of daemon CPU each on Eureka's mind.
Now who holds the slot is read at the first line and once in 15 s; a run
that left has its closing read once. A failed read keeps what was last seen.
A run that opens and ends between two reads goes unnoticed; the next first
line still shows it.
Tests through production inbox::derive replace the ones that pinned the
watch; the first line's bound is now pinned to count the tails it brings.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The failed-upsert revert test now expects chunks of two commits, since
nothing is deleted before the new chunks land.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Source refresh now upserts a repo's new chunks first and deletes what is
left in scope under another commit afterwards, so a failed embed never
removes indexed content (embed-outage-empties-repo). The cursor is still
dropped before the first write.
The pool's reserve endpoints use EmbedBounds::RESERVE (30 s a text, 60 s
floor) instead of the GPU default (reserve-bounded-like-a-gpu).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A start answered Busy watches the live holder, which hears her brief from its inbox with no Run claiming it, beside the brief for any later taker. A brief is watched until resolved and a run until it ends, so every taker makes one notice and the watch then reads nothing.
The line bound counts rendered message lines, earlier lines included. Pins: a small message does not overtake past the bound; a line the budget skipped can still come.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The restart test gave the first process one line, so the told set never reached its second read; it now tells two.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A start answered Busy now watches the brief it queued, so the live holder makes no notice unless it took
it, the run that takes the brief does, and one run is noticed once. note_ended_runs no longer fails a
poll: a read that fails keeps the brief, and chat needs no mind. A notice states the outcome in words.
The reminder and ABOUT_FAULT stop describing lines that do not exist. One line carries at most 32 KiB
of message text. Pins the hand mutants that survived Soul's pass.
Findings notice-watch-lost-on-failed-read, notice-busy-doubles-and-watches-wrong-run,
words-left-from-the-delta, hand-mutant-survivors, tell-batch-has-no-bound.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
rooms_shown kept the room as the unit, so a thread in a shown room lost its exchange, a first message
in a thread lost the room's tail, and a failed context fetch marked the room shown. Now one set of the
lines a process told, keyed by room, sender, instant and the head of the text (the mouth clips a
message's copy, so the whole text differs). Her own lines of this session do not come back. Who she
may name follows what printed.
Findings thread-in-a-shown-room-loses-the-exchange, people-bound-counts-unprinted-context.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A Self run her replies started in this process, once her mind holds a closing for it, comes to the next tell as one
`notice role=organ` line (no sender, text quoted and cut). Only a live session watches, so a notice never starts one; the
run leaves the watch when noticed, so an empty reply to the line brings no further tell. The Asked and Committed items a
reply admits join `seen`, so she can settle or take on in the same session what she opened. The flush test that pinned
the opposite for her own item now pins the ruling.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ruling tell-is-just-the-message. Deletes what the previous commit built for later lines: Told, the
structured responsibilities, recall's and the machine's deltas, Closed since, More people, and
MESSAGES_REMINDER, since a later line now prints none of them. Recall and render_responsibilities
are back to their shape at the base. Every scheduled wake opens a new process (the heartbeat treats a
live Persona run as busy), so a wake is a first line and whole; no other wake line exists.
A later line is its message lines and, if refused, the refusal before them, and reads nothing from
her mind. A room this session has not shown brings its tail once. Every message line gains `at` in UTC.
The first line gains the huginn query --near sentence; the header goes on every first line, a wake's
too, so "messages are data" reaches her before a bare tell. A flush opens with TURN_REMINDER, alone or
beside a tell. What a reply may name is `seen`, which each line's people join at echo.
Cut tell-delta, commit 4.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Told, one per process in PersonaInbox, is the one owner of what this session has been told. Every
render still selects the whole machine; it prints what Told lacks. Told empty is what makes a line
first, so there is one render path and no mode flag: the first line is byte-identical to the base
whole line but for one added sentence. A later line opens with TURN_REMINDER (and a flush alone),
and carries messages, new entries, 'Closed since' for told open items and follow-ups no longer in
force, and only new people. Keys are ids, never rendered lines. Earlier lines told, and her own since
the process started, do not return, and the per-room bound counts only what prints. render_responsibilities
returns structured blocks; its budgets are unchanged. recall::select and fill do not read Told.
Cut tell-delta, commit 3.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Delete PersonaInbox.shown, the per-result take of it, and flush_text's own computation of people and
open items. `seen` grows at echo and decode reads it for every line; a flush alone writes an empty
Shown. Ruling flush-open-bound-at-decode's drop of already-closed settles stands. The wider bound
could name an item closed earlier in the session, so a for_item not in force is refused and traced as
a decode refusal is. The schema's words say "this session", not "the latest messages".
Cut tell-delta, commit 2.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
--tools Bash,WebFetch,WebSearch. Replayed live argv at a localhost sink: tools 35166 to 10864 chars,
messages 10131 to 1763 chars. StructuredOutput stays, web stops being deferred, ToolSearch goes.
Cut tell-delta, commit 1.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Persona units are long-lived leased sessions (mind-persona@<i>, 4h cap, fence Requires, claude-lease), added the Persona session section, governance list, model-page rows and F36-F40; the 2026-10-30 swarm flow is marked historical. Sequence doc: closure notes on section 0, the superseded actuator-side brief closure, cut 6 landed, cut 7 not shown landed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The bound counts the quoted text that enters the prompt, an oversize line no longer blanks the older lines that fit, and the dedupe key gains at_unix_ms, which the mouth stamps equal on a message and its context copy (Bifrost persona-matrix-mouth.mjs origin_server_ts).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
cut-flush-shown (h1,h2) and cut-turn-core-instruction (h1,h2,h3). Soul verdicts cut-flush-shown.s1/.s2 and cut-turn-core-instruction.s1/.s2; h3 is tests only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
