# Changelog Newest first. Each entry is what changed, what it makes possible, and what to do about it. `journal upgrade` prints the entries since the version you had; a session started on a newer version than the last one it saw is handed the same. ## 2.269.6 — The start block says how many to-dos you could actually take - The start block counts the to-dos ready to take, not every open row. A blocked row, one waiting on another, one asking the user, one in a helper's hands and one already under way are left out of the number. - The funnel that picks the next row now leaves out a to-do handed to someone else or already under way, so auto mode cannot offer a row somebody is holding. - The hook answers a set of warm-up calls in a folder of its own after a restart, so the first hook of a real session does not pay for the code and caches it meets first. - The spool is found from the stamp of its folder, through the listing the replay already takes. ## 2.269.5 — A migration and a writer no longer wait for each other - A write takes the shared hold of the record before any lock of its own, and a migration takes no record lock, so a migration and a writer never wait for each other. This is what made requests stall for seconds at a time under load. - The look at what an agent's writes changed waits five times as long as the last look took, and the place git keeps a repository's index is asked once, so a large project is no longer scanned all the time. ## 2.269.4 — One live session per environment, and the resume line said once - A session that starts on an environment where another session is already live waits in an environment of its own, and the user is asked whether to take the busy one over. Taking over unbinds the session that was there, so an environment has one live session and a message, a nudge and a resume have one place to go. - The line 'The journal has updated and resumed you now' is typed once. A resume asked for an agent that is not paused is taken and not said, and a pause left on a row for longer than any update takes is treated as the leftover it is, so the agents are no longer asked to resume every second. ## 2.269.3 — A session start no longer scans every attached file - A session start asks for the files still without tags in one line when there are many, instead of writing a row for each, and looks for them again only when the rows of a type changed. The work a hook leaves behind after it has answered is a fifth of what it was. ## 2.269.2 — The hook keeps quiet when it cannot reach the server - A hook that cannot reach the server no longer writes 'The journal did not answer within 2 seconds' into the agent's transcript. It keeps the event as it already did, and the server reads it when it next can. ## 2.269.1 — The mascot blinks without an error - The mascot's blink no longer throws 'blinkOf is not defined' in the viewer; the blink it uses is imported. - A plugin service whose check cannot run, such as a missing program that answers 127, is listed as failing with the command's error and told once, instead of being shown as not needed. - A plan under review says on its card and in its inspector how many reviewers are still reviewing. ## 2.269.0 — One repository for rows, kept in memory - Every row is read and written through one repository per type. The repository keeps parsed rows in a rolling memory sized per type, parses the eager types at start, and answers unread, by type and by relation from indexes it keeps in step. Its funnel methods are final, and a row file written past it is refused. - In a server, the marks of folders and rows come from memory, renewed once a second by the watch loop. A warm read stats no file and parses no row twice, a row another process rewrote is read again within a second, and a board load makes each thing once per request. - Totals per type are saved beside the index and corrected by the housekeeping. A removed or renamed folder lets go of all it held, and a cached row that will not read back is dropped and read again from its file. - Loaded rows survive an update's restart: each type has a version, the rows are kept across the restart, and only a type whose version changed is read again. A release can ask for a full restart in its changelog entry. - The server stops slowing itself under load: - One prober per journal judges its health. - A late heartbeat means busy, not dead. - A hook answers first and replays spooled hooks after. - Work after an answer runs on a pool sized to the machine. - An agent's row is written at most once a second. - Lock files stay open between uses. - The seat file is written only when it changed. - A check can carry an instruction that the agent answers with pass or fail, so it can use an MCP server the agent already reaches. - A rollback to the previous build reads the index this build wrote, and writes give up on a busy repository lock instead of waiting on it. - Uploads through the server work again; two functions of one name had shadowed each other. ## 2.268.30 — Codex sessions no longer crash on the menu funnel - Since 2.268.22 a Codex agent's session could crash: the menu funnel reads the screen through a pattern that Codex's driver lacked, and the screen read raised an error. Every driver now reads its screen. - The upgrade tests hold an upgrade back on a skill file changed by hand, since the journal's own code folder no longer holds one, and the crash-grace test writes the failed hooks that are reported. ## 2.268.29 — An agent is told once that the journal resumed it - After an update, an agent was told 'The journal has updated and resumed you' again and again, about once a second: a paused agent's lines wait until it continues, so the line sent while it was still paused counted as unsent and was sent again on every tick. The pause now ends first and the agent is told once. ## 2.268.28 — Mascots that blink, and steadier chats - The mascots blink as people do: every two to six seconds at an irregular gap, now and then twice, on an eye layer of their own, so a blink also happens during a move. - A mascot enters only once every image of it has loaded, so no piece pops in late. It checks several times a second where the top bar is, so it follows the bars down when something under them goes. - The chat no longer throws 'prepending is not defined' after paging back. - A pull request or an issue named by its number, such as PR 161, is no longer asked to say what it is. - Counting a folder's rows no longer fails when another thread updates the counts at the same moment. ## 2.268.27 — The Squire serves his sovereign - The Squire calls you my liege or sire, as a squire speaks to his sovereign, never knight, king, or your title or name. - When a new bar pushes the mascot up, it is flung a little past the bar and drops back onto it, landing in its own way. - An inspector loads only the view that is open (transcript, terminal, files changed or to-dos), and another loads when you open it. ## 2.268.26 — Scrolling back finds your own messages - Scrolling up the chat loads the messages before the oldest one shown, yours and the agent's alike. The window used to start at a weeks-old open message, so only the agent's notes of that stretch came back and your own messages never did. - The main chat, a subagent's chat and a helper's chat page back through one funnel, a page at a time, with skeleton rows while a page loads. - A sitting voice hops up onto its seat on entry and hops off on exit, its legs showing all the while and each in its own manner, instead of its legs vanishing and popping back. ## 2.268.25 — A calmer mascot and plainer notices - The mascot enters only after the agent has stayed idle for three seconds, and leaves only after it has stayed busy for three seconds, so a quick command no longer makes it pop in and out. - The line that tells an agent of a new message names who wrote it and their words once, with the tag that answers it; the journal's tag is not doubled and the message's row is no longer dumped into the chat. - A report's bar above the message box says when the report has questions for you. ## 2.268.24 — Busy agents say so, and paused ones stay put - An agent whose turn is running reads as working, between commands too, whatever wait it has stated. The header and the message box show the live activity strip, and Waiting shows only for an idle agent. - While an agent is paused, for an update or by you, the lines the journal holds for it are kept until it continues, instead of each starting a turn that the pause then stops. - In the demo, the buttons on the lesson-done card always work; the replay lock no longer holds them behind its hint. ## 2.268.23 — A restart warms up faster, and Linear stays signed in - Once the viewer is warm after a start, the command parsers, the rows and the transcripts warm side by side instead of one after another; a failure in any of them still ends the start, so a broken build rolls back. - An expired Linear sign-in renews itself with its refresh token when the agent connects, so you do not sign in again. ## 2.268.22 — A program's own menu never holds an agent up - A menu that a program puts on an agent's terminal is recognised by its shape, numbered options above a selection footer, and answered by its words, never by a number. A folder the journal launched the agent into is trusted; consent and data sharing are declined; anything else is closed. A note says what was chosen. - A menu you opened yourself, or one listing past messages, sessions or checkpoints such as Claude Code's Rewind list, is never touched, and neither is the agent's own permission question. ## 2.268.21 — Linear works once connected, and restarts are quicker - After the journal's own Linear sign-in, the agent's Linear tools work at once: its MCP entry fetches the token when it connects, and the token is never written into the config. Before the sign-in, the card says a browser approval is still needed. - An upgrade builds the new archive with its bytecode already compiled and runs the rest of the install from it, so the new code is not compiled cold. The restart's steps, to the first answer and until warm, are timed in runtime/upgrade-steps.jsonl. - In the demo, Look around after a lesson leaves every control usable, chat included, and the lessons page scrolls to its last cards. - A ticket claims an environment that already carries its name, so a ticket's environment is never listed as a main one. ## 2.268.20 — The mascot perches on the bars, and idle agents rest - The mascot sits on top of the highest bar above the chat (a plan, report or dump bar) instead of behind it. When that bar goes, it falls to the next one, or back onto the chat box, and each voice lands in a way of its own. - The sweep of a project's tickets runs once for the project instead of in every environment's engine, and its git work runs once a minute, so idle agents no longer keep the machine busy. - A merged ticket's agent is stopped only once its turn has ended and nothing waits for it, so follow-up work sent to it is not cut off. ## 2.268.19 — Updates take seconds, and can be put off - When the journal updates by itself, the blurred cover counts down ten seconds with a Not now button; Not now puts the update off, and it is offered again later. - An update waits only for running journal commands, never a test run or a dev server, and for twenty seconds at most; the new build is checked in a process of its own without asking a server; every step's time is written to runtime/upgrade-steps.jsonl. - The changelog loads ten releases at a time, newest first, with a button for older ones. - The agent bar's second share arrow was Import a layout; it has an upload arrow of its own, and Share names the layout it shares. ## 2.268.18 — Questions asked under an environment show with its own - The Questions tab, window and count carry the open questions of this environment and of every environment launched from it, helpers and tickets included, each named for its environment and answered in that environment's inspector. - Under reduced motion the mascot stays visible in its rest pose, with no blinks, moves, entrance or exit, instead of being hidden. ## 2.268.17 — A removed environment stays removed - Once an environment is removed, a late read or a plugin keeping its place in the event log no longer creates its folder again, so a second removal never finds a half-written folder and fails to pack it. ## 2.268.16 — The journal never opens or answers Claude Code's Rewind menu - The journal's Escape presses, to stop a turn or decline a permission, come at least a second apart, so two never land as the double Escape that opens Claude Code's Rewind menu. - The journal no longer presses an option when it reads the rewind dialog's words on an agent's screen; those words in any output, such as a search printing them, made it type 3 into the session. ## 2.268.15 — The Log in button opens its browser on a machine new to it - Log in on a browser-login card downloads Playwright's Chromium first when that build is missing, as it is on a machine whose Playwright only drives Chrome, so the browser opens instead of the card saying no login was saved. - When the browser still does not open, the card says Playwright's own reason. ## 2.268.14 — A hook never fails on a stale pause - A hook that met an agent paused for an update raised an error instead of clearing the pause, because the age a pause counts as stale after was missing; it is ten minutes, and the pause clears there as intended. ## 2.268.13 — An upgrade is not held by the journal's own code folder - An upgrade no longer stops because a file in the journal's own code folder changed: an upgrade replaces that folder whole. A start that finds an upgrade held by changed project files says so instead of counting it as installed. ## 2.268.12 — A start installs a new version once, never in a loop - When `journal claude` installs a newer version before it starts and the restarted start still runs the old one, it starts on the old one and says so, instead of installing the same version again and again. ## 2.268.11 — An update's pause never outlives it - The server clears an update's pause on the agent rows of every environment as soon as no upgrade runs, so an environment with no engine to take the resume is let go too; a sweep clears any update pause older than ten minutes, and a call that meets such a pause goes through at once. ## 2.268.10 — The mascot sits only on the main chat - The mascot no longer shows in a subagent's or worker's inspector, which reuse the chat's composer; only the main chat shows it. ## 2.268.9 — The installer asks before it installs Python - When the machine has no Python 3.11 or newer, the installer asks whether to install Python 3.13 with uv before it does; no means it stops and says what to install, and with no terminal to ask in it never installs on its own. ## 2.268.8 — A plain Mac installs without a Python of its own - When the machine has no Python 3.11 or newer, as a Mac with only its own Python 3.9 does, the installer fetches Python 3.13 with uv (installing uv first when it is missing), with no Homebrew and no admin rights, and installs with it. ## 2.268.7 — The installer finds a Python the journal runs on - The installer picks the newest Python 3.11 or later on the machine (python3.14 down to python3.11, then python3) and stops with a plain message naming what to install when there is none. It used to accept Python 3.10, on which the journal fails at its first import. ## 2.268.6 — A finished sequence never holds writes - A sequence step that is no longer in hand, because the sequence finished, was abandoned or never ran, no longer holds the agent's writes: the write gate asks whether the run still stands before it refuses an Edit or a command. ## 2.268.5 — Agents paused for an update are always resumed - A pause for an update that outlived it clears itself: an agent still marked paused after the update has finished is resumed, a ticket agent is found by its own session, and a resume that did not reach the agent is asked again. ## 2.268.4 — The journal never interrupts a working agent, and messages, waits and tickets behave - The journal no longer presses Ctrl-C to see whether a silent agent is still there, which cancelled a long think or command as if the user had interrupted it. It reads the screen instead: an agent back at its prompt is idle, and anything else is left working. - A line the journal holds back no longer makes every later line repeat after an update, so stale notices about settled rows stop. - A message typed to an agent counts as taken once the screen shows the agent at work, so a note to an idle ticket agent no longer reports that it stayed in the input box. A chat message whose event was numbered under one lock and logged under another is no longer lost. - A long command moves to the background in the second its threshold passes. In orchestrator mode a command goes to the background after 5 seconds. Each work mode has its own switch and number of seconds in Settings. - A wait the agent declared clears at the next command it runs. With auto mode on, an agent that has stopped is told to carry on once the user's own last word is over 15 minutes old. - Plan review and unsticking no longer fire for tickets that are finished, or whose plans were approved and wait on their dependencies. Closing a ticket checks its repositories once. - A server on a build other than the installed one restarts itself. A hold on a sequence step no longer in hand releases itself. Upgrade, heal and the other recovery commands are never held by a step. - An agent's status is never overwritten by an older one, so the mascot shows whenever the agent is idle. A conversation that starts stops the rows of earlier ones whose process is gone. - The sequence marks are hidden in the chat by default. A subagent's cell says only Subagent, with no colour. - The chat shows its last messages without waiting for pictures or the rest of the start, with no loading skeletons. - The chat asks for new events every second, also while its live stream is open, so a message the stream missed appears within a second instead of after a reload. - An upgrade restarts every environment's engine on the installed build, and no supervisor on an older build starts old engines again. - Speed: an agent has one command in flight and never waits on another agent's. Message search is one scan. Hooks answer before replaying kept events. Rows and sessions are read from their change files instead of whole folders. Budget notices count working time only, with a starved machine noticed separately. The shim never reruns a command cold after a timeout. - Claude Code's rewind dialog near the context limit is answered with Summarize up to here. A [journal] line raises one event and never starts a chat turn. journal services list works again. ## 2.268.3 — Sessions written at the same moment are seen, and healing keeps a running build - A session or state file another process wrote at the same moment is read at once, so the write gate no longer says nothing is open while a ticket's agent has work open. - `journal heal` never rolls back a build whose server is running and answering, or one an upgrade is still installing; it rolls back only a build that does not answer. Repeated heals had put a journal's server back on an older build while its viewer was new, so pages such as the voice settings never loaded. - The viewer shares one profile load instead of starting the same requests again. ## 2.268.2 — The mascots' pictures ship, ticket agents can write, and paused agents resume - The voices' pictures and the mascots' rig parts ship with the viewer: the release before left them out, so the mascots and the editor's moves showed empty frames. The old sprite sheets are gone. - The mascot shows on the chat box whenever the agent is not working, and a voice with a rig shows only its moves in the animations editor. - A ticket's environment is at home in its own worktree, so the write gate lets its agent write there. - Every agent an update paused is resumed after it, in every environment, and a resume that was not submitted is asked again. ## 2.268.1 — The Kanban board and the polled routes answer at once - The Kanban board works out each card's waits and lane once and formats its texts through the cached formatter: on transportklok's board of 49 cards it answers in a third of a second after a start and in 8 ms after that, where it took up to 3 s. - Formatting a text that names a command builds only the list of query names, never the parser of every command. - The services, identity and agents routes answer from state kept in memory, and a GET never claims or probes a port; the seat listing reads the session folders only when one is added or removed. - The server raises its open-file limit to 65,536, never lowers the one it started with, and keeps 16 packed archives open instead of 64. ## 2.268.0 — Your voice's mascot lives on the chat box - Each voice has a mascot drawn as a cut-out rig of parts, played from keyframed moves that turn, shift and restack its parts: it enters and leaves in its own way (climbing up from behind the chat box, jumping, fading, marching), blinks, and plays one of its moves at the idle interval while the agent waits. - Settings shows a voice's animations: its moves play from the rig in the editor, beside its sprite sheets. - A new-feature announcement can offer a button that turns its feature off. - The core flows of the journal (messages and replies, to-dos from messages, helpers, tickets, long commands, questions on the agent's screen, an upgrade and restart, nested sequences) run end to end against a real scratch journal before every push. ## 2.267.85 — Session files are read through one cached funnel - Every session file, seat file and per-session state file is read through one funnel that keeps a copy and reads the file again only when its stamp changes: a warm hook, summary or agents list opens none of them, where the summary alone read every session file per environment every few seconds. - A check refuses any other read of those files, and any folder scan, process spawn, git call or path resolve on hook, route and engine paths outside a short allow-list. - A skill on writing lean, reused code (journal-lean, after law L7) is loaded at every session start in every project. ## 2.267.84 — Tickets start at once and the queue starts what others wait on - journal ticket start answers at once: the branches of every nested repository and the agent's launch run behind the answer, and a launch that cannot start says why in a comment on the ticket. - Stopping a ticket and starting it again launches its agent; journal ticket tell answers with the note it delivered. - A stopped ticket frees its running slot at once, a ticket waiting on an open ticket holds none, and the queue starts first the tickets others wait on. - The file search no longer follows a link out of the project into another repository. ## 2.267.83 — Every slow command is caught, and replies to many messages land - The journal shim times every command with curl's clock and reports any over 50 ms behind the answer, so commands the server never timed (such as helper say) file their budget notices too. - A reply naming several messages reads the ones not yet read, answers each it can, and names any it could not; it is refused only when none can be answered. - Hooks answer in a few milliseconds warm, and the work after their answer lists rows once per event instead of reloading every open row and message. - Closing a pin hides it at once; the server is told behind it, and a refused close brings it back. - The journal's law L7: write the least code that solves the whole problem, reuse what already does the job, one funnel per kind of thing, never the same logic twice. ## 2.267.82 — Commands stop waiting on screens, git and stale timers - journal helper say returns at once: the words are typed to the helper's agent behind the answer, so the dispatcher never waits on its screen. - journal ticket board answers from what each card already holds and refreshes the git look behind the answer: on transportklok's board of 37 tickets it went from over 40 s to under a second. - A long command is moved by the call that is actually running, with a press sent at once and dropped when that call has ended; its card names that call and its real time, and one place makes the card. - A conversation started again in a new process is no longer marked stopped before its first report, so its environment no longer shows Idle. - The status line says Working while a command runs, even under a standing wait, and opens with its chevron again. - A Home cell whose plan is approved but waits on an open ticket says so instead of asking for approval. - A start restores each conversation from its kept fold and reads ahead only agents heard from lately. - A command started while an upgrade writes the new build waits for it instead of running a half-written tree. - A standing nudge closed by another thread first is left as it is instead of raising. - Test runs use at most 3 workers. ## 2.267.81 — The engine never waits on its own upkeep - The engine beats every second on its own, and the clock's upkeep (the ticket upkeep's git work among it) runs on a thread of its own: on a loaded machine with many tickets a round could take minutes and hold back state updates, message delivery and moving a long command to the background. A long command now moves within seconds of its limit. - An answered question reaches the agent's own question prompt on screen, for every provider, and Claude's question screen is recognised as a question. - A reply or wait that fails keeps its words in the chat and says why, instead of vanishing. - The update cover shows at once on a page loaded while an update runs. - A request no longer rewrites the trigger counters of stopped sessions, and the start block is rebuilt behind the answer, not in it. ## 2.267.80 — The viewer answers in seconds after a start, and worktrees share the project's packages - A start warms the first dashboard first and opens the viewer on it; command parsers, open work, messages and search texts warm behind it. The first answer after a start came at about 35 s under load, now about 9 s. - A worktree cut for a helper, a ticket or by hand links the project's installed vendor and node_modules, nested repositories' too, instead of installing them again; a job that changes dependencies asks for its own install with --packages (or the ticket's own_packages field). ## 2.267.79 — A kept helper keeps its worktree, and a first start announces nothing - A helper's worktree stays while the helper is kept: new work sent to it rebases the worktree onto the working branch first, and the worktree goes, with the agent's temporary folder, only when the helper is finished or retired. - A journal's first start counts every new-feature announcement as seen, so only later updates show one. ## 2.267.78 — The journal cleans up after itself - Once an update has started, the journal removes the builds it left behind; it keeps only the one running and any a live process still runs from. An older version comes from git. - A closed ticket's worktree is removed by the minute sweep once no agent runs in it and nothing in it is unsaved, together with the agent's temporary folder; its branch stays, and starting the ticket again cuts it from there. - A helper worktree whose work was taken is removed by itself, temporary folder included, and the same name is cut again instead of a second folder. - A start reads the text of every to-do and message in its warm-up, so the first search after a start is as fast as the rest. ## 2.267.77 — A sequence interrupted by another waits at its step - A sequence run cannot be moved on while another run of the same agent is in hand: next names the run in hand and says the interrupted one waits at its step, and it is handed back there when the newer run ends, so a refused action is never skipped. ## 2.267.76 — A reopened ticket stays open and starts again - A reopened ticket whose branch was merged is not closed again by the next sweep: it counts as merged only after new commits. - A reopened ticket's environment is restored as the ticket's own, so it stays out of the environment lists; a plan the environment no longer holds counts as no plan, so the board lists every ticket and the start asks the agent for a new plan. - A note to a busy ticket agent is left as a message in its environment and reaches it through the journal, instead of being reported stuck. - A hook that finds the server busy under load waits up to 3 seconds and spools the event, instead of calling the server down; misses are reported after five. - A long command moved to the background names the call that ran long, with its real duration, for every provider. - A refused chained command names only the commands that actually ran. ## 2.267.75 — Every subagent shows in the agent grid while it runs - The subagent list is read again whenever the files in the session's subagents folder change, not only when the orchestrator's own transcript grows, so a subagent dispatched while its orchestrator waits shows at once. - What was announced as dispatched or returned is kept on the agent row, so a restart no longer announces every subagent again. ## 2.267.74 — A working agent is no longer marked stopped after an update - The silence of an update's pause no longer counts against the agent: the engine measures silence from its own start and from the resume, so an agent working right after an upgrade is not marked stopped and its Home status no longer says Idle. - A resumed subagent is found by the id its launch answered with, so it stays in the agent grid; a flag sent as the word "false" is read as false for every provider. - A message another Claude Code session sends is read from the queue where it arrives, with the sender's name, and shows as an agent's card. - No feature names a provider: the default provider is one fact in the provider seam, and the .claude/settings.json check holds back only a Claude helper. - A search counts the conversations its time limit left to load behind its answer and says so. ## 2.267.73 — The family view reads transcripts again - The family tree and the conversation query hand a transcript to its provider as a path, so opening an environment's family no longer fails after 2.267.72. - The journal's law L4 says a helper is reused by its whole session: one whose agent ended is resumed in the session it ran, never started again under the same name. ## 2.267.72 — A ticket's final report always reaches its orchestrator - A ticket agent that ends its turn with a report, with no message written, still reaches the orchestrator, and an orchestrator waiting on its board is woken by it. - A helper whose agent ended takes new work by resuming its own session, with its context, instead of starting fresh. - A provider's question on the terminal screen, such as Claude Code's "Do you want to proceed?", is noticed and routed for every provider, in bypass mode too. - While messages are typed because the Claude channel failed, the journal checks every 30 seconds whether the channel is back and switches to it. - Every search mark gets its own result and opens it; a message from another Claude Code session shows as an agent's card; an identifier with underscores no longer turns chat text into italics. - While the update cover shows, the viewer asks for nothing but the update's status; a dialog over a row stays open while results arrive. - Search keeps each conversation's turns on disk and loads them newest first within a time limit; a chat opened again serves its kept turns; saving settings no longer resolves every path on disk. ## 2.267.71 — A ticket's edits show, nested repositories included - A ticket's inspector finds its environment's agent the way a helper's does, so Files changed and Terminal no longer say "No agent yet" while the ticket's agent works. - A project that is a repository and holds nested repositories feeds the edits of each, not only the root's. - Pictures an agent makes show in its Files changed, even where the project's .gitignore ignores them. - A plugin that declares its own command ends every notice it sends with the exact command to run inside the repository that holds the file, so an agent never installs it to check a sin. - A Codex agent's cell says what it last said. - A file of up to 100 MB can be sent; the viewer names the file that is over the limit. - The chat mark for a compaction reads Conversation compacted; the file feed's loading rows fill the panel; the update cover keeps its spinner above the title in every step. ## 2.267.70 — The row index lives beside the rows, not among them - A type's row index is saved in its own folder, so saving it no longer makes the rows' folder look changed and the next read no longer re-scans every row. ## 2.267.69 — Faster search - A search reads each type's rows once instead of twice, keeps each conversation turn's searchable text, and finds conversation files with one check each. ## 2.267.68 — Dashboard reads never wait on a lock - A dashboard read never writes: a due row-index flush is left to the server's background thread, so a read no longer waits on the write lock while another write is busy. - A reply writes the message it answers once instead of twice. ## 2.267.67 — No cold dashboard after a formatting change - A formatting setting change warms the dashboard in the background, so the first request after it no longer formats every row view on the spot. ## 2.267.66 — Plans as rows in a collection, and one folder scan per request - A collection's Plans tab lists each plan as a full-width row: its title, the ticket it comes from, the phase it is in, a progress bar with done out of total, and its state; a click opens it, on the collection page and in a share. - A dashboard request scans each type's folder once and hands the rows to the listing and the counts. ## 2.267.65 — The dashboard reads its open rows from what it holds - A listing's open rows come from the held list of open rows instead of a walk through the whole history, so a type with few open rows no longer slows the dashboard. ## 2.267.64 — Agent cells say which ticket - A ticket's or plan's agent cell is named after it, such as Ticket 15 or Plan 3, and its kind reads Ticket agent or Plan agent. ## 2.267.63 — Counts that follow the change - A type's counts on the dashboard follow the rows a change touches instead of walking every row again, so a busy type such as messages no longer slows the dashboard. ## 2.267.62 — The viewer keeps its address - A project can set its viewer's port (viewer.port), and a restart waits for that port while the journal's own previous server still holds it, instead of moving to the next free one. ## 2.267.61 — A ticket closes on merge only with commits of its own - A ticket counts as merged only once a commit made on its own branch is in the target; a branch fast-forwarded onto the target's history stays open, and a ticket whose agent is working is never closed by the minute check. - journal ticket reopen brings a closed ticket back with its environment, its approved plan and its conversation, and journal ticket start resumes its agent there. ## 2.267.60 — A quicker dashboard on a long history - A listing of the newest rows and the open ones is found from the newest end of the history, so a long history no longer slows every dashboard request. - Reopening a closed board for a started ticket no longer changes how a board is reopened by hand. ## 2.267.59 — A closed board's orchestrator still decides for its open tickets - A board's orchestrator approves plans and decides for every open ticket on it, also after the board has closed; starting a ticket on a closed board reopens it. - An upload's body is cut at its boundaries instead of read line by line, so dropping files is quick. - A command that waits on its agent is known by the method behind its word, so journal helper finish is not reported as slow. ## 2.267.58 — No more hook errors on edits, helper marks that open, and a Copy button on every code block - The line for what an agent does counts added and removed lines as text, so a hook after an edit no longer logs an error. - Every chat mark about a helper or a subagent uses the agents' yellow, and a click opens that agent's dialog. - Every code block in the chat has its own Copy button. - The dashboard keeps each row as the JSON it is sent as and encodes only the rows whose view changed. - A starting server never waits for another process's migrations. ## 2.267.57 — The orchestrator answers its agents' permission requests - With auto mode and orchestrator mode both on, a helper's or ticket agent's permission request goes to the orchestrating agent, which allows or denies it with journal agent permit; the agent's cell says it waits on the orchestrator, not on you. - The viewer waits a minute before it says the journal is taking too long to respond. ## 2.267.56 — Updates that never take the viewer offline, quicker replies, and agent cells that say what is happening - An update's backup is copied before the lock and only its changes under it, a starting server never waits for another process's migrations, and the dump-files step runs in one quick pass, so the server keeps answering through an update. - A reply no longer scans every message and comment: who links a row comes from an index kept with the row summaries. - A tool-use hook rewrites its session file at most every few seconds, and the ticket nudges skip their work when no board is orchestrated here. - A cell on Home shows what the agent last said and a plain line for what it is doing, never a bare tool name. - A running command's card in the chat has its icon beside its label at the top. ## 2.267.55 — The Board chip - The board menu beside the mode switch is one chip: Board, a thin divider, then the board's name; a click anywhere on it opens the menu. ## 2.267.54 — Files in collections, a lighter dashboard, and agents that carry on after an update - A collection holds files and shows them on a Files tab, on its page and in its share: images previewed, other files with their name, size and a download. Files dropped on a dump land in its collection, and earlier dumps' files are added on upgrade. - A row another process changed is patched into the held lists without walking every row twice, and a repeated search no longer rereads every row. - An agent paused for an update hears nothing while paused, and is told on resuming that its open work stays open, so it carries on instead of parking it. ## 2.267.53 — Two tickets' worktrees of one repository stay apart - A worktree is recognised by git's own list of its repository, whatever its folder is called, so two tickets' worktrees of one nested repository no longer collapse into one. - A retry removes a half-made worktree of its own failed start with git worktree remove and deletes its branch; a folder no worktree owns is moved aside, never deleted. ## 2.267.52 — A ticket start that fails cleans up after itself - A ticket start no longer fails with an invalid path on a board with a branch. - A start that fails partway removes the worktrees, branches and folders it made, and a retry of the same ticket clears what a failed start left and starts over. - A nested repository's worktree starts at that repository's checked-out head, never at the board branch's merge-base. - A command's exit message is shown as a message, never read as a number. ## 2.267.51 — An automatic update shows itself from start to finish - Every open viewer shows the update cover by itself when an update starts on its own: the countdown, Starting the update, each step, then a reload once the new version runs, with no gap in between. - While an update is counting down or running, the viewer asks for its state every second, so short steps are seen. ## 2.267.50 — The first reply after a restart is quick - The server reads the messages and comments while it starts, so the first reply after a restart no longer pays to load them. ## 2.267.49 — A countdown before an update, agents paused while it runs, and voices that introduce themselves - An update that starts by itself first counts down ten seconds with a Cancel; Cancel skips that release, and the next one asks again. An update you start yourself begins at once. - While the journal updates, every agent is paused and told why, and resumed and told so once the new version runs; the agent whose own command runs the update is left alone. - Each voice introduces itself on its card, in Settings and in the first-start dialog, which now shows all five voices in one row on a wide screen. ## 2.267.48 — The Board menu - The menu beside the mode switch reads Board and lists the project's boards by name, with No board for to-dos. - The notice after a pick tells the agent to file new requests as tickets on the board picked, never as to-dos. ## 2.267.47 — Pick a board for new work - With orchestrator mode on, a New work menu beside the mode switch picks where new requests go: None, or one of the project's open boards. With a board picked, the orchestrating agent files each new request as a ticket on it and starts it, then reviews, approves and merges its plan; with None it files to-dos as before. ## 2.267.46 — A shared plan shows its rows loading - A plan opened in a share shows skeleton rows in its phases and timeline until its to-dos have arrived, never an empty list in between. ## 2.267.45 — Faster updates - An update re-packs only the files that changed, reusing the rest of the previous archive. - The record is copied before an update only when a migration will change it, without compression, and not again within the hour; Preparing the update no longer waits on it. ## 2.267.44 — Refusals that say who decides, and a calmer update check - A refused ticket or board decision names who may take it and how, and never sends an agent its board lets decide back to the user. - The update check runs every half hour by default. ## 2.267.43 — A ticket's plan always reaches its board's orchestrator - Who orchestrates a board is its orchestrator field alone: a ticket plan waiting for approval starts the review for that environment, and it may approve the plan, also after the board's run has ended. ## 2.267.42 — Documents searched on the server, a lighter dashboard - Searching the library for words inside documents runs on the server, so the dashboard sends documents as summaries too and the preview reads a document whole when it opens. ## 2.267.41 — Voice pictures you can see - Each voice's picture sits large on top of its card, with its name and sample below, in Settings, in the first-start dialog and in the phone's voice list. ## 2.267.40 — A member from another environment names its ticket - A collection member that lives in another environment carries a badge naming the ticket it belongs to, such as Ticket 3 · Fix the login page, or the environment's name when no ticket has it, on its card and in the page or panel it opens, in the collection and in a share. - The plan buttons of the Tickets and Boards tabs read Open plan of ticket N, with the ticket's title in the tooltip. ## 2.267.39 — A lighter dashboard, a quicker work await, and option lists in one group - The dashboard holds each type's rows apart, so a type that changes every few seconds rebuilds only itself, and sends reports, dumps and plans as summaries that the viewer reads whole when one is opened. - Releasing a write hold no longer reads every agent that never held one, and the start block is rewritten only when its text changed. - Every option list in Settings is one group that wraps onto a second line when it does not fit, never a column of cards. - Codex models are named family first without the GPT- prefix: Sol 6.1, Astra 6, Sol 6, Luna 6, Sol 5.6, Terra 5.6, Luna 5.6. ## 2.267.38 — A picture for every voice, and board columns that scroll - Every voice has its own picture, on its card in Settings and in the first-start dialog where you choose a voice; a copy of a voice keeps its picture. - A board's columns keep a maximum height and scroll their cards, with each column's heading in view, on the collection page and in a share. ## 2.267.37 — A seen announcement stays seen - An announcement leaves the queue the moment you answer it, whether with its button, Not now, Escape or by closing it, and never shows again after a reload. - A seen mark the server did not take is reported, kept in the browser and sent again, never dropped in silence. ## 2.267.36 — Helpers are reused, and a retired helper keeps its conversation - Sending work to a stopped helper starts it again with its context, and a new dispatch is refused while an idle helper that already knows the files exists. - Retiring a helper takes a reason, read back first; the chat marks a helper Reused, Continued or Retired. - The helpers list sorts helpers into Working, Waiting for work, Closed and Retired, and shows how often each was reused and when it was first dispatched. - A closed or retired helper's inspector shows its conversation. ## 2.267.35 — The Squire calls you Sir Knight, and a new voice says hello - A voice can address you in words of its own instead of your title and name: the Squire calls you Sir Knight, and now and then my liege. - When you change the voice, the agent is told at once, even while idle, and answers with one opening line in the new voice. ## 2.267.34 — Announcements wait until seen, and an agent's finished to-dos - New-feature announcements wait until you have seen them: one missed while you were away shows the next time you open the viewer, several show one after another, oldest first, and a first install announces none. - An agent's To-dos tab lists the to-dos it finished too, under Finished, the latest first. ## 2.267.33 — The Squire, new-feature announcements, update steps and blue read ticks - A new chat voice, the Squire: every task is a quest, the code is the realm and every bug a foe; he never drops the role in the chat and names his helpers as knights of the realm. Code, commits and docs stay plain. - A release can announce a new feature in a dialog shown once after the update, with its own artwork and a button that turns the feature on. - An update shows the step it is on, refuses a second update while one runs, and waits at most a minute for running commands before it swaps the build. - A message the agent has read shows blue read ticks, in the chat and on the phone. - A comment shows on its row and in the chat the moment it is written, and never twice. - Replying to a message is faster: linking to-dos to their messages is skipped when no to-do was made lately. ## 2.267.32 — A board in a shared collection, and a faster dashboard - A collection can hold a board: its page and its share show the board's goal, its done-when clauses and its lanes with each ticket's card and stage, read-only in a share. - Each ticket's plan opens from its card, in the Boards and Tickets tabs and in a share. - The dashboard sends its lists and counts from one held body, made again only when a list of rows or the settings changed. ## 2.267.31 — Shared views, chained commands part by part, and no lost hook events - The agents view and the chat can be shared read-only on a private link that follows them live. - A chained Bash command runs part by part: each part reports its start and end and marks the chat once it has run. - A call the transcript shows as answered is closed and never moved to the background, whatever happened to its end hook. - A hook event that decides nothing is kept in a spool when the server is slow to answer, and the server replays it in order. - The chat stays where you scrolled it: nothing that arrives pulls it to the bottom once you scroll up. - The instruction-files check runs on a fresh session start only, never after a compaction. ## 2.267.30 — Updates no longer re-read every transcript - The transcript cache is keyed by each transcript's path and the code that shapes a turn, never by the providers' code, so an update re-reads no transcript unless turn shaping changed; a cache that cannot be read or written is simply no cache. - After a restart the server answers before it warms up, and warms up in the background. ## 2.267.29 — Quieter hidden tabs, and no error for a file an update removes - A viewer tab that is out of view or out of focus polls ten times less often, and returns to normal when shown again. - The managed-files check counts a file removed between listing and reading as removed instead of raising, and waits while an update holds its mark. ## 2.267.28 — Background cards that close, collection tabs with ticket context, and a faster start - A command that never reported back is closed by the next one, so a 'moved to the background' card closes when its command ends, names only the part of a chain still running, and a working agent is never called silent or stuck. - A ticket agent waiting by design (on another ticket, on its plan's approval, or done and awaiting its merge) is no longer flagged as stuck. - A collection remembers its open tab; its Tickets tab shows each ticket's plan progress, what its agent does now, its stage and its waits; its plan cards show progress, on the page and in a share. - Each opened panel or dialog stacks above the ones already open. - The server starts about twice as fast: it no longer builds every command's parser while warming up. - The changelog check no longer reads compiled Python files that another process is replacing. ## 2.267.27 — Journal commands answered by the warm server, and slow ones reported again - The server answers every journal command that does not start or stop a process, so version, status, help and the rest no longer start a fresh Python each time; on a busy machine that start took 9 to 19 seconds. - A slow request, command or hook is reported whatever the machine's load, with the load beside it, instead of being only logged while the machine is busy. ## 2.267.26 — Tickets plan ahead, closing test cards, and a locked viewer while updating - A ticket that waits on another starts its agent and writes its plan; the wait holds only approving the plan, so implementation begins once the tickets it waits on are merged. - A helper's 'Running tests' card closes when its run ends, also after it was moved to the background or killed. - The viewer is blurred and locked whenever the journal is upgrading, from any upgrade path and also after a refresh, and reloads when it is done. ## 2.267.25 — Collection tabs for plans, to-dos and tickets - A collection's page and its share show Resources, Plans, To-dos and Tickets as tabs of their own; a tab shows only when the collection holds that kind. ## 2.267.24 — A Tickets tab in collections and their shares - A collection that holds tickets lists them under a Tickets tab with each ticket's stage and an Open plan button for the plan in its work environment, on the collection's page and in a share of it; a shared collection carries each ticket's plan with its to-dos and timeline. ## 2.267.23 — Shared plans with their to-dos and timeline, ticket agents on the home screen, plain secrets, faster upgrades - A shared collection carries the to-dos and timeline of a plan it holds from another environment, such as a ticket's plan, and opening one of its to-dos opens it from that environment. - The home screen shows a cell for each ticket's agent. A ticket whose board branch a nested repository lacks starts from that repository's default branch and says so. Listing rows and counting edited lines no longer fail on a row or a blob that vanished. - Saying something to a helper finds it running as soon as it is launched. A slow journal command a helper runs is reported to whoever dispatched it. - A helper's pill stays inside its plan phase card, and a helper's Brief line shows chips and code. - Secrets use plain labels (Allowed programs), a secret made with instructions arrives with its programs allowed, and a request for a key points at an MCP server that already reaches the service. - An upgrade to the version already installed does nothing, and every journal command starts faster (the scrubber reads its word list only when used). ## 2.267.22 — Collections with members from other environments open and count them - A collection's card counts a member of another environment by its own type, so a collection holding `ticket-1/plan:1` no longer breaks the page it is shown on, in the viewer and in a share. - On a shared collection's page, clicking a member of another environment opens that member instead of the collection again. ## 2.267.23 — Tickets in a collection - A collection's page and its share list the tickets it holds under a Tickets tab beside the cards and to-dos, each with its stage, and the page links a ticket to its plan in the ticket's own environment. A ticket can now be shared on its own. ## 2.267.21 — A shared collection shows a row of another environment - A share of a collection that holds a row from another environment, such as a ticket's own plan, now lists it on the shared page and opens it without an error; the page reads every member by its full environment/type:number. ## 2.267.20 — A collection holds rows of other environments - A collection can hold a row from another environment, written environment/type:number, such as `journal collection add 6 ticket-1/plan:1` for a ticket's own plan. The collection's page, and a share of the collection, read that row live from its own environment; a member whose environment is gone is left out. ## 2.267.19 — Check time limits that follow the load - A check's time limit stretches by how far the machine's load exceeds its cores, the same way the boot guard's does, so a check on a busy machine is no longer stopped early. The helper is shared from `engine/load.py`. ## 2.267.18 — Helper stops that wait for the agent, and a steadier test suite - Saying something to a helper, stopping it, finishing it and the stopped notice all ask one method whether the helper's agent still runs, and a stop returns only once that answer is no, so a helper is never reported stopped while its agent is still running. - Closing a login and dropping a phone in a hosted journal go through one vault method. The imports check finds the test files that import from tests/ again. - A plain `pytest` in the repository runs one worker per core: `-n auto` is part of the default options. - Every browser scenario connects to one headless Chromium per test worker instead of launching its own; the browser tests take about a quarter less CPU. Ten fixed pauses in the scenarios now wait for what they were waiting for, and the inspector padding check no longer fails now and then. - Test installs start from the run's first packed build, so each install takes about half the CPU. A test that waits for processes left after a session stops at its deadline instead of hanging on a busy machine, and a scratch viewer server that never answers is stopped instead of left running. The installer's `AGENT_JOURNAL_UNVERIFIED` switch, which only the tests set, is gone. ## 2.267.17 — Open tabs reload, plugins update from the viewer, and whispers stop repeating - An open viewer tab sees a new build and reloads itself: the manifest named the bundle with a pattern that only matched the old bundle name, so no tab ever noticed an update. A tab already open on the old bundle needs one manual reload. - A plugin whose repository holds a newer version shows an Update bar at the top. A daily check of each plugin's repository finds it; Update runs the plugin's upgrade and leaves a chat mark naming the plugin and its version, on your side when you pressed Update. - A keyword reminder (a fact, rule or law) reaches the agent through the channel only: it is not retried, not typed into the terminal, and dropped when the channel cannot take it. The check that a handed line arrived ignores line breaks. - The Closed bar in the helpers list runs edge to edge and folds the closed helpers until pressed. The agent cell's rows are a two-column grid, so every value starts at the same place. - A shared collection no longer carries the dump it was filled from. ## 2.267.16 — Quieter budget notices, and a round of viewer fixes - A request or hook that runs over its budget while the machine's load is above its cores is logged with the load and not filed as a to-do; on a quiet machine it is filed as before. The time a request reports after its answer is the work it left behind, not its wait in a queue. - A plugin's manifest declares each hook as sync or async; an async hook only runs beside the agent, and the install dialog says whether the agent waits. - A plan under review cannot be approved until the review's report is linked and the plan is ready again; its card shows no Approve meanwhile. - Reports stay user-creatable; only a suggestion is the agent's alone, and a type the agent alone makes has no New button. - The Brief and Report bars run edge to edge in the agent dialog. The Other plans popover is as wide as its content and stays inside the window. The agent cell's work line wraps to three lines. - The Files changed view shows a file-shaped skeleton where the cards appear, starts with Show removals, Flush and 5 lines on, and shows view files such as SecretForm.vue; only data files named for a credential stay hidden. - A hook that gets no answer while the machine's load is above its cores is logged with the load and not reported. A fact or rule is whispered once to a session and again only when its words change. A channel queue is no longer cut while its channel reads it, which had replayed every old notice after an upgrade. - A turn the transcript posted first is not posted again by the display hook. A test's journal server ends with the test run, and the restart test cannot miss the restart marker. ## 2.267.15 — A secret's programs are yours to name - Every secret on the Secrets page has Programs it may go to, a list you add programs to and remove them from; the server refuses an agent that tries to set it. An agent proposes a program with journal secret propose_programs , and the proposal waits on that page for you to allow it. The refusal for a secret with no programs points to that control. Shells and interpreters are still never given a secret. ## 2.267.14 — One agent row per session, and the journal block once - An answer to a message from another session is never kept out of the chat, even when a journal line started the turn. - An upgrade folds an agent's duplicate rows for one session into the row that reported last, keeping every compaction and mark; the context popup's last compaction is right again. - The journal's block is written once, into AGENTS.md. CLAUDE.md keeps your own text and imports AGENTS.md, and a copy of the block left in it is taken out at the next upgrade. - The agent cell's running row is labelled Running. ## 2.267.13 — Lists, chats and cells follow what you do - Each tab of a list page fetches its own first page when opened and keeps its own Load more count; the Closed tab lists the closed rows newest closed first. The Files page shows a skeleton and asks for each tab's files a page at a time, filtered on the server by kind, by what they are attached to and by the words searched. - The inspector's chat stays where you scrolled when lines arrive and shows a Jump to latest pill; only your own send moves it down. A chat message's header stays on one line. - A turn that answers only journal lines stays out of the chat, even when a journal line hands over a message the agent already answered, and a turn is sent once. A helper's working notes stay in its inspector; only what it writes to you reaches the chat. - A message processed into a row is handled at once and its check marks colour. Delete is offered only on messages you sent, and the server refuses any other. A row's panel shows its comments once. - Connecting, disconnecting, logging in to and logging out of an integration, and a failing check or the run that clears it, show in the chat on your side. - The agent cell's last row is Running for or Last active, and a helper's cell shows its latest instruction. The context popup shows the real last compaction. The waiting dots light one at a time. The Brief and Report bars are flush. The Gmail browser case no longer loses its address under load. ## 2.267.12 — journal secret run works again - journal secret run works again: a word that takes a list of words, such as the command after the secret, no longer overwrites the command line's own command. ## 2.267.11 — An upgrade's restart is not a hook error - A hook that gets no answer while an upgrade restarts the server, and during the new server's first seconds, is no longer told to the agent as an error. - A server that crashes and comes back still shows its hook failures: the restart marker is taken away once the new server has started. ## 2.267.10 — Journal commands stay inside their budget - Journal commands no longer start a git process each time: the server reads the git user name once at start and the commands read it from the runtime folder. - The first seconds after a start are not held against the budget for commands, as they already were not for requests. ## 2.267.9 — The engine waits out an upgrade quietly - While an upgrade is migrating the record, the engine's writes wait for it and are tried again on the next pass; nothing is reported as an engine error or told to the agent any more. ## 2.267.8 — Test cards, file changes and chat marks follow what an agent really does - A test run in the chat is a card for the test runner's own part of a command, named for what it tests, and a run started in the background gets its card when it ends, with its tally read from its output; a finished run with no readable result says Tests ran. A python -m runner counts as a test run. - An agent's file changes show in its dialog when it works in a worktree or a checkout of its own, and any shell command that is not only a read or a search is checked for changes. - The helper dialog shows its brief and report as one line each with an Open button, and its Chat and Transcript tabs lose their description bars. The chat box loses its waiting glow and shows three moving dots in its waiting badge, on the phone too. - A search mark in the chat names the term it searched for, and a click opens what the search found in a dialog. - The Linear and Gmail cards show the switch, the login with its state, direct use by agents and the state line; a logged in card offers Log out; reading into tickets is off by default; a card reads its state as soon as it is switched on; a switch with its label beside it has no tooltip. - A reply tag no longer runs twice and reports that it did not run: a message a handler consumed stays marked as sent. - Every create ends its output with one line naming what it made, such as todo 12. - A finished plan reads done, and an upgrade fixes the finished plans that kept parked or active. - A request that passes its budget keeps a stack sample of every thread beside its profile. ## 2.267.7 — Tooltips stay while the chat scrolls, and the first command after a start is quick - A tooltip now closes on a scroll only when the scrolled area holds the button it points at, so a chat that scrolls itself no longer takes it away. - The server prepares the command parsers and the open work and to-dos before it prints its address, so the first command after an install or a restart answers in under 10 ms instead of 200. - The tests that start a scratch journal end every process they started and fail when one is left running. ## 2.267.6 — The status bar fits a phone's width - At phone width the status bar's agent controls wrap onto a second line and the connection label shortens with an ellipsis, so the page no longer scrolls sideways when a connection label is long. The browser checks of the Integrations page and its Linear and Gmail cases no longer depend on the order they run in. ## 2.267.5 — The Gmail browser check waits for each answer - The browser check of the Gmail card waits for the answer to each settings write and to the settings read after a reload, instead of a timer, so it no longer fails when the machine is busy. Nothing changes for you. ## 2.267.4 — Filing a breach to-do frees the agent's writes - Filing or closing the to-do for a breach of the time budget now releases every agent session whose writes that breach holds, in every environment, so an agent held by it can write again at once instead of at the next breach. ## 2.267.3 — A settings save keeps the formatted texts - Saving a setting that no formatting reads no longer clears every formatted text, so a save in Settings is fast and the chat does not redraw; a switch of a formatting feature still refreshes the texts it changes. - Finishing a helper right after stopping it waits up to five seconds for its agent to leave instead of answering that the agent is still running. ## 2.267.2 — The faults hold finds the to-do in every environment - A to-do filed for a breach of the time budget now counts wherever it sits: the hold on the agent's writes lifts once an open to-do of that title exists in any environment, and a closed one starts its count again, so installing the journal and the agent's writes are no longer refused while the to-do is open. ## 2.267.1 — A finished helper never keeps a to-do - Finishing a helper gives its to-dos back before it marks the helper finished, so a finish that breaks later, as when releasing its worktree fails, no longer leaves a finished helper holding rows nobody can start. An upgrade gives back the rows that finished helpers still hold. - A reply tag followed by tool calls in the same turn is checked to be picked up as a reply; nothing changes for you. ## 2.267.0 — Integrations, starting with Linear - A new Integrations page and sidebar item list the outside services the journal can reach for you, each off until you switch it on and pick its key. The key is a secret only you pick; it is read by the journal's own process, sent only to the service's own address and never given to a command, a subprocess or an agent. - Linear: issues assigned to you in the teams you choose become tickets on a board you pick, each comment once, checked every five minutes or by Check now, and by Linear's webhook through your tunnel when you give it a signing secret. Their words are wrapped as untrusted for agents and read plainly by people; hidden characters are removed and long text is cut. An issue that leaves what you chose or is deleted keeps its ticket and says so. - Only you start or confirm a Linear ticket, even in auto mode. You can map each board stage to a Linear state and send status changes to Linear, and an agent can propose a comment that is posted only when you press Send. - Log in from the Linear card opens Linear's own sign-in in your browser and keeps the token as a secret. A second switch adds Linear's own MCP server to Claude's and Codex's project config; what an agent reads there is not marked untrusted, so it stays off until you turn it on. - Gmail: the mail you choose, by a Gmail label or search, becomes tickets on a board you pick, wrapped as untrusted text. It signs in to Google's mail servers with an app password you pick from your secrets, and only you start such a ticket. An agent can propose an answer to the sender; it is sent only when you press Send, with exactly the text shown. Google offers no MCP server for Gmail, so the card has no such switch. ## 2.266.1 — A hosted journal's login page refuses the pages that never run from outside - A journal on a server, whose login page runs apart under its own user, now refuses /api/run, /api/stop, /api/upgrade, agent hooks and service changes from outside at the login page itself; before, a logged-in browser's request for them was passed on to the journal. Members' own requests are answered again in that setup. Upgrade the server; nothing else to do. ## 2.266.0 — An Agent tab in Settings, a login button in the chat, search that finds everything, and a viewer that asks the journal's own commands - Settings has an Agent tab at the top that lists the agent's sessions, rules, skills, work tracking, plans, questions, messages and voice, with a Models group; every model picker offers each provider's own models. - The agent asks for a browser login with a button in the chat, saved logins reach every Playwright tool the agent has, and a site the user lets the agent log in to on its own needs no button. Secrets are a project-wide item in the Project section of the sidebar. - `journal search` finds what the viewer's search finds, every kind of row, conversations and attached files, grouped by type, and `--resources todo,message` narrows it; a search mark in the chat keeps its term and opens the search page with it filled in. - The viewer reads the manifest, summary, events, agent controls, transcripts, links, family tree, file feed, skills and status bar through journal commands the CLI runs too, and a check fails on a viewer endpoint no command uses. - Helpers: kept agents are limited for each type, a checkout a helper is launched into is followed like a worktree, a helper waiting on a question reads Waits for answer or Needs you, a helper's inspector shows its terminal and the brief and messages the main agent sent it, and a helper's report is named while it stands. - Sending a message never blocks sending the next one, and everything that loads shows one kit skeleton until its data is there, on the phone too. - The message box shows nothing once a wait is over, a finished wait no longer stays At work, and a transcript nobody has read answers at once with its newest turns and fills in behind. - Updating the journal locks the viewer under a cover with the version and the step until the new build answers; an update report strikes the items the user has since answered. - Voice profiles name how the agent calls its helpers and subagents and whom they report to; journal law L6 keeps a journal line out of the chat. A first budget breach of a request, command or hook files its own to-do. - A check tells the agent of each pull request waiting; a plugin service starts through one small keeper entry, waits for a command not found yet, and a plugin's update check follows its source's newest version. ## 2.265.0 — A journal on a server that members share, a copy that connects to it, and questions that reach the agent - A journal on a server has members: the owner invites a person by name, the person joins from a one-time link, is a writer or a reader, sees only the environments the owner shares, and loses every login at once when removed. Every row names who made it, a member's words are read by agents marked as untrusted, and a member's phone reaches only what their rights grant. - A copy of a journal on your computer connects to one on a server from the viewer or `journal environment connect`: row and event numbers come from leased blocks, an environment is held by one machine at a time, what a copy writes while the server is away waits and goes when it is back, and the project's code travels through git with its nested repositories. What never travels (phones, keys, secrets, passwords) is written in one list and tested. The bar says whether this journal is the server or your copy, and the viewer shows each failed connection in plain words. - A question a helper or a lent subagent asks reaches the orchestrator at once and again while it waits, and the helpers list and the plan page show a helper that waits on one. - The orchestrator is told when a helper stands idle and is refused a wait that names a helper; a helper's report names the commits its branch holds that yours lacks, lists the tests it wrote, and reads a to-do handed to it by the number you know; a helper or subagent is refused every test run unless the dispatcher allows it; a launched agent's kickoff waits in a file instead of its command line. - A to-do one helper holds moves to another with `journal helper say --todos` in one command, and the plan page groups the to-dos one helper holds under one tag and shows its helpers at once, with placeholders while they load. - Hooks answer the agent before anything that only tells it something, and no longer wait on the record lock: a long transcript is folded behind on a thread of its own instead of holding every hook and leaking sockets, and writers held back by a migration wait together. This ends the hook timeouts and the "too many open files" of a long session. - A screenshot or file the agent names shows as a card in the chat, files dropped onto the chat box are attached to the message, opening an item over a page keeps the page's own address, and a plugin opens on the Plugins page. - Tags and status lines in the viewer read like a label and a value (Helper: Hedy, Waiting: the agent, Helpers: 4), a project rule flags one written as a sentence, and the press-everything run is a browser scenario like the others. - Faster to run: the suite, the browser scenarios and the viewer's unit tests set up less per test; the lesson tests play every lesson on the computer and the phone. - A shared link's password is refused after five wrong tries from one place, a stopped plugin service keeps the reason it stopped, history search is run by the server, and the images of a server install are pinned by digest. ## 2.264.2 — A transcript is never read whole by accident, and the skills check never judges from a stale reading - Every read of a conversation goes through one place that starts where the last read stopped; a read from the very start needs a named reason (only a search has one) and is logged. - A conversation read for the first time after an upgrade is read from its last 64 MB, not its whole length. - The check that skills are loaded never refuses a tool call while its reading of the conversation is behind. ## 2.264.1 — Hooks answer at once and never hold the agent - A hook answers before anything that only tells the agent something; nudges, marks, whispers and the agent's report run after the answer, on worker threads. - A long transcript is folded in the background, so a hook never waits minutes behind it after an upgrade, and the viewer's connections are no longer left open until the server runs out of files. - Writers held back by a migration wait together instead of one after another. - A plugin service's when-check runs in the project's folder. ## 2.264.0 — Logins for the agent's browser, secrets on a server, and to-dos that close when their work lands - `journal secret login ` opens a browser for you to log in once; the session is kept beside the secrets file, owner-only, and the agent's own browser tool starts logged in from its next start. - A journal on a server keeps its secrets in a folder of its own that no backup holds and a restore leaves in place. The owner can take a hosted journal down or upgrade it from the viewer, only signed release images are deployed (the first deploy included), and a nearly full disk ends in a clear refusal with a health address that says so. - A to-do a helper works closes by itself when the commit carrying its trailer reaches main, whatever route it took; `journal todo sweep` closes open rows whose work is already there, and `journal helper say --todos` hands more rows to a running helper. - Helpers and subagents are told never to run the whole suite, and a guard refuses it unless the main agent allows it for that helper. - An update is no longer refused over your own text in AGENTS.md or CLAUDE.md, and the updates page offers Update anyway for anything else it warns about. - An open question is asked with its choices whenever they are known; history searches name the tool and the term in their chat mark. - The plan page shows which helper holds each phase and to-do; the orchestrator grid keeps each agent in a fixed cell with a timer; the phone has one picker at the top; an agent's inspector shows only that agent; the helpers menu shows a helper working again after a follow-up. - Viewer polish: the waiting glow eases again, the waiting badge is smaller, the status bar shows a running command on one line without seconds, project paths are shown relative, pasted images get names of their own, search can include archived items, and buttons show a spinner until answered. - Chat marks for dispatching, stopping and finishing helpers and for worktree changes; a helper's own list holds the to-dos it was handed; a mark when an agent changes a setting; a plugin's settle reaches its card on any agent row. - Requests stay within their budgets after an environment change or a worker reload; the demos open what they teach and pace themselves. ## 2.263.0 — A bar for each active plan, lists that show their real totals, and demos that play to the end - Each active plan has its own bar at the top, the plan you work first. A plan handed to helpers or a ticket's agent is marked: its bar reads 'With helper Hedy' (or 'Delegated' when no one is on it yet) and opens a card of who works on it. Such a plan keeps running beside yours; journal plan delegate sets it by hand. From four plans on, the rest fold into one line below the bars. The phone shows one strip per plan. - Every list shows its real total, such as 'Showing 25 of 1,072', loads more as you scroll to its end, and has a Load more button; the phone's lists too. - The demos play to their end again on the computer, and the phone version starts instead of saying it cannot reach your computer. - The phone tour's highlight ring starts below the status bar, so no edge of it is hidden. - A server whose requests get stuck is restarted by itself: every ten seconds the journal checks the server can still take its record's locks, and after three misses in a row it keeps what each thread was doing (.journal/runtime/threads-