v0.100.0

Install or update: npm i -g cotal-ai@0.100.0

Changes in this release

• cotal channel remove no longer says a channel was removed when it had no saved entry, such as a mistyped name or one already removed. That run is now a no-op: the receipt reads Nothing removed with Metadata: no saved entry, and result.removed is false. With --purge, the channel's retained messages are still purged and the receipt says so. deleteChannels now returns the channels it removed, and clearChannel says whether it removed the channel's entry.
• --creds reads and checks its file before dialing. A missing, empty or malformed file (CRLF line
  endings, a truncated JWT, no seed block) and a directory are refused naming the file and the cause,
  instead of reporting a serving broker as unreachable or dumping the command's help. A file other
  accounts can read is still used, with a warning that names chmod 600.
• A broker that accepts the connection and never answers, as a stopped or overloaded one does, is no longer reported as not running, and status no longer hangs on it. ps, who, send and history refused such a broker with "no mesh running at ; ... recorded at but not running". They now say it "accepted the connection but is not answering", keep the registry entry, and name checking that the broker is not stopped or overloaded. cotal meshes add and a raw --server connect use the same wording. status printed its report and then never exited, because its manager service probe dialed the broker without first reading its greeting and the client library left that socket open. status and status --components now read the greeting first, report the manager service as unchecked or refused, and exit.
• A channel subscription that the broker refuses after the join was confirmed, for example after a
  permission reload, now reaches the endpoint's error event once, as left channel "<channel>", with
  the broker's permission violation as its cause. It was also reported a second time as a generic
  NATS permission denied error, so an embedder could not count or alert on it once.
• With the space's manager stopped, cotal run and its hosted verbs and cotal models refuse at once, as cotal ps does: TARGET_UNAVAILABLE (exit 2) with the Mesh header, the registry's own answer as the native detail, and bare cotal up as the next step. They used to wait about 10 seconds for a manager answer that could not come and then print a refusal about the ep.v1 rail with no code or next step. An unpinned operator call reads the service registry first and refuses only when it lists no live manager; an unreadable registry, a pinned call and a user's manager view go to the manager as before. Core marks the registry's empty answer with a new ai.cotal.ep.no-live-instances detail, so a caller can tell it from an unreadable registry.
• Create a missing persona catalog .cotal/agents with mode 0700 from every writer. Under a
  group-writable umask, space create, up and setup made it 0775 while seeding the default
  persona, so until the first persona create tightened it, a group member could replace
  default.md. The setup log's .cotal is created 0700 the same way. An existing directory keeps
  its mode.
• A persona file with CRLF line endings or a UTF-8 BOM now reads like the same file with LF endings, so persona create --file, persona list, persona show, persona edit and launches accept it. persona create --file now names the file it read when it refuses a definition, not the catalog path it would have written.
• Observers and chat sessions no longer hang silently after a broker outage the client could not ride out. watch --plain and console --plain now resume when the broker comes back, as history --follow already did: the endpoint re-subscribes their feed on every reconnect, where before it subscribed once and stayed silent while reporting itself connected. A chat --stdio or --plain session whose inbox is gone from the broker after the outage, as when cotal up restarted the space and retired it, no longer reports reconnecting forever. It ends by itself with that reason, its --stdio end record and close carry it, and it exits 5. An endpoint stopped from its own warning listener also exits promptly instead of waiting out the reconnect backoff.
• cotal run show and cotal run journal now say why a step failed. A failed or refused step's journal row carries its recorded errorMessage beside errorCode in --json, and the text view lists each such step with its outcome and message under Step errors, escaping control characters. Before, both views printed only failed (L4000), and the reason, such as a hosted run that is not admitted to read a channel, was only in the manager's log. The manager's run-status reply schema admits the new field.
• cotal send --creds with a credential that may not read the space's roster now refuses with PERMISSION_DENIED as soon as the broker denies the read. It used to wait about five seconds and report an unknown timeout. A batched Direct Get in core now ends on the broker's permission denial for its request.
• cotal who on a mesh where nobody is present now answers at once that nobody is there, and exits 0.
  It used to wait 3 seconds and report the roster as partial. A presence observer that does not
  register now completes its snapshot on an empty bucket, so its view is current.
• ADMITTED_NODES, STRUCTURAL_NODES, RESERVED_NAMES, FORBIDDEN_GLOBALS and MUTATING_METHODS are now exported as frozen arrays instead of Sets, and CATALOG is frozen. Code that imports the package could call add or delete on one of those sets, or assign a CATALOG title, and change what validate accepted or how its errors were titled for the rest of the process. The tables can no longer be changed: a write throws in strict-mode code, such as an ES module, and changes nothing elsewhere. This changes the exported type, so a caller that used .has or .size on one of these tables uses .includes or .length instead.
• A declared adapter path that other local writers could change is refused with every unsafe entry
  at once, the reason group write is not accepted, and the one chmod that fixes them. Before, the
  refusal named only the shallowest directory, with no reason and a Next that pointed at
  cotal service. An install made by npm i -g under umask 0002 left the bundled nats-server and
  each of its directories group-writable, so a controller up refused one directory per attempt.
• cotal service add now says plainly when a declared adapter path crosses a symlink, such as --handler /bin/sh or --handler /usr/bin/python3 for the exec-endpoint preset. It names the link and what it points to, and its Next names the physical path to use instead. The refusal used to report the symlink's own mode, 0777, as if group or others could write it, and a symlinked file was refused as DEPENDENCY_UNAVAILABLE. A symlink on any part of the path now refuses as PERMISSION_DENIED, because a link could be repointed after the check.
• Output is no longer colored on a terminal whose TERM is dumb, unless FORCE_COLOR asks for it,
  and FORCE_COLOR set to false in any letter case, such as FALSE, now turns color off as false
  does.
• A command given --creds <file> and no --server, such as cotal send or cotal history, no longer presents that credential to the default address nats://127.0.0.1:4222. It goes to the broker the --space is registered at (main when no space is named), with that registration's TLS requirement. A space not registered here is refused with TARGET_NOT_FOUND before anything is dialed. Before, the registration was ignored and the credential was offered to whatever listened on the default port.
• COTAL_DEBUG=1 now prints a stack trace for a command that fails with a reported error, such as a failed ep describe, as docs/config.md says. It prints the failure's stack and then the stack of each error that caused it, so an endpoint call's underlying error, for example no describe reply ... within 10000ms, is visible. It used to print nothing for these failures. The JSON envelope is unchanged.
• backup cleanup-fallback now retires the journal of a restore made with --accept-missing-source.
  That restore replaced a missing source and keeps no old store, so cleanup refused with an unknown
  outcome, every other recovery route refused too, and the root could never start another backup
  cycle. Its plan, prompt and receipt now say that nothing is removed, and clean restore-fallback
  does the same.
• A controller stack no longer stays blocked after a cotal down exits partway through a member's stop, for example because it was killed or its terminal closed. A manager stop that the dead down had not yet handed to the manager is settled by the next bare up or down, and the stop reservation that down left is released once its pid has exited. A reservation held by a live process is refused with that pid named. A broker or delivery daemon that the dead down left stopping is named in the refusal with cotal down --abandon-stop <request-id> --yes, which now takes that stop once the process is gone. Before, every up, down, --abandon-stop and service stop refused.
• cotal down of a controller stack keeps a selection that names its space. The stopped space stays prepared at its root, so the next command run without --space refuses with TARGET_UNAVAILABLE and names the root to start it in, as the docs describe. It used to clear the selection, so that command quietly ran against another running space.
• A paired endpoint caller file from cotal auth mint <name> --endpoint <class> --out <file> now works with ep describe, ep invoke and invoke under --creds <file>. The mint recorded the endpoint's wire token in <file>.issuance.json, and reading that intent back refused the token, so every such call failed with "the endpoint contract read did not complete" before any request was sent. The intent now keeps the dotted endpoint name. An intent written by an earlier release is refused with SCHEMA_INVALID and a remedy to issue a new caller.
• config, config get and status now warn with PREFERENCE_INVALID when a variable such as
  COTAL_DETACH_KEY, COTAL_DEFAULT_AGENT or COTAL_DEFAULT_PERSONA holds a value config set
  refuses, naming the variable and the reason. The variable still wins.
• A command such as cotal channel run in the folder of a forgotten prepared space, with no broker at the default address, now says that no space is registered or prepared for that folder and names cotal space create <name> --root <root>. It used to say that mesh main was recorded there on 127.0.0.1:4222 and to run cotal up to restart it, though nothing recorded main there.
• After space forget on a stopped controller stack, status and every command in that root now say
  to prepare the space again with cotal space create <space> --root <root> before cotal up. Before,
  they said to run cotal up, which refuses there until the space is prepared again.
• A backup create stopped by Ctrl-C now says what it left and how to clear it, and the claim it leaves no longer locks the cut for 30 minutes. Interrupted after it claims the cut, it exits 130 (143 for SIGTERM, 129 for SIGHUP) and prints its attempt id, the partial artifact and cotal backup recover <attempt> --yes. backup recover now takes a create's backup-<uuid> as well as a restore's id. Once the backup process and the broker it started are proven gone, which takes seconds, it removes the partial artifact, the clone and the broker files, each matched by inode, and returns the cut to ready. A create to another directory does the same first. While the claim is live, up, another create and backup recover refuse LIFECYCLE_BLOCKED (exit 2) and name the backup process and the deadline, where up used to say "follow the recorded recovery" without naming one and a new create exited 3. cotal status shows the create's own attempt and its next step. A create naming the partial artifact refuses CONFLICT and names the recovery, and backup verify or backup inspect on a directory with no manifest.json calls it an incomplete artifact.
• A backup restore (or up --restore) stopped by Ctrl-C no longer locks its root for 30 minutes without saying so. Interrupted inside its pre-commit window, it now exits 130 (143 for SIGTERM, 129 for SIGHUP) and prints its attempt id and cotal backup recover <attempt> --yes. The claim it leaves is stale as soon as its restore process and every broker it started are proven gone on this host, which takes seconds because each broker's watchdog stops it when the restore process exits. So that recovery, clean restore-attempt, or a rerun of the restore rolls it back at once. The 30-minute deadline still applies when a broker cannot be proven gone. While the claim is live, up, a repeated restore and backup recover refuse LIFECYCLE_BLOCKED (exit 2) and name the restore process and the deadline, rather than exiting 3 with an unknown outcome. A restore or recovery that finds the root's maintenance lock held names the pid holding it. The status and receipt hints now say --yes, which the apply needs, and status no longer offers clean store --force while a maintenance journal holds the store.
• Under a locale whose character set is not UTF-8, such as LANG=C, POSIX or ISO-8859-1, reports,
  refusals and usage errors now write their marks and punctuation in ASCII, as --plain does, and
  keep their color. With no locale variable set, the marks stay UTF-8.
• A managed verb such as cotal ps, stop, attach or spawn that cannot choose between several of your live managers now shows the choice. It used to refuse with only the instance ids inside one sentence. It now lists each manager on its own row with its instance id, host, activity (which names its runtime) and the time its process started, and its Next gives your own command again with --on <instanceId> for each one. The code and the exit code stay as they were, and --json carries the rows under error.cause.detail.managers.
• cotal auth mint --expires-at refuses a time that is not in the future, before anything is written. It used to write a credential the broker already refused, under a label that was then taken, and doctor auth reported it as expired.
• cotal persona --server <address> with an address that no registered space uses no longer turns the current folder into a catalog. Outside a project it listed an empty catalog for a space named main, and persona create wrote .cotal/agents/<name>.md into whatever folder you were in. It now refuses with TARGET_NOT_FOUND and writes nothing. In a project it still reads that folder's catalog, as launch does. @cotal-ai/workspace exports isGenuineSpace for this check.
• config set and config unset in a preferences folder you cannot write are now refused with
  PERMISSION_DENIED, exit 2, naming the folder and how to make it writable. Before, they reported
  a raw EACCES as an unknown outcome, exit 3, although nothing had changed.
• A preferences.json that is a symbolic link is now refused, naming its target, as a private
  record's link is. config set used to replace the link with a regular file while the file it named
  kept the old value.
• A selected prepared space whose project folder was deleted now says the folder no longer exists and
  names cotal space forget <space>. Before, use, status, ps and up --space said to start it
  with cotal up in the missing folder, persona refused with no path or remedy, and space forget
  said the files under that folder were kept.
• A command that targets the selected space from inside a project whose root has a different
  registered space now says so before it acts. It warns on stderr and in its JSON warnings with
  SELECTION_NOT_FOLDER_SPACE, naming both spaces and the --space that targets the folder's space.
  send, status, who, ps, whoami, endpoints and a named down <service> used to act on the
  selection there with only a Mesh: header to show it.
• cotal space forget now removes the per-space preferences saved under the name it forgets, so a later space with the same name no longer inherits them. The receipt lists each removed key and value with the config set that restores one. The removal takes the preferences file's lock and revision check; when it fails after the record is gone, the run is partial and names the config unset that finishes it.
• When a root's controller stack runs under another COTAL_HOME, status and up now say so, name the
  Cotal home in use and tell the operator to run with the one the stack was started under. status
  said to start the stack with cotal up, and up refused its own members as legacy pidfile-owned
  processes with no next step.
• cotal status in the folder of a forgotten prepared space now names cotal space create <name> --root <root> as the next step. It used to name a bare cotal up, which started the default space main there and then blocked preparing the forgotten space in that folder again.
• A refusal of a mesh target now names the space and root it concerns under target in --json, where they are known. A stopped space's ps, who, whoami, endpoints and history refusals carried "target": null although the message named the space, so a script had to parse the message to tell which space refused.
• cotal up now refuses a project root that other local accounts can write, with the same check space create applies, before it writes anything. In a group-writable folder on a host without a private-group proof, such as one whose /etc/nsswitch.conf names sss, space create refused but up started a legacy main stack, and bare down then refused "the additive attempt inventory could not be read" with no cause and no next step. The refusal from every command that checks a path now names every unsafe entry in its chain at once, with one remedy such as chmod g-w <project> <parent>. Bare down and status also name that cause for a stack an earlier release started there.
• On macOS, cotal setup and cotal update no longer write Cotal's entry into Claude Desktop's
  claude_desktop_config.json while Claude Desktop is running. Desktop reads that file when it starts
  and writes the copy it holds back on ordinary use, such as switching a Code session or changing the
  model, so an entry added while it ran was dropped. Setup now finishes its other steps and reports
  that the Desktop entry was not added, with the fix: quit Claude Desktop, then run cotal setup
  again from a terminal outside it. Windows is unchanged.
• The cotal history docs now say that the text and plain reports show a line break or other control character in a message as its \uXXXX escape, and that --json carries the text as it was sent.
• cotal run validate, plan and simulate now say what each validator warning is: its code, title, position, cause and fix, on stderr and in the --json envelope's warnings, and validate --json lists each one under result.warnings with the same fields. Before, a warning printed only "The language validator reported a warning at this source position.", so a fanOut over a literal list with no key validated with no visible reason and then failed its run with L3021.
• The auth ledger, obligation and manager goal-index scans now run one shared consumer lifecycle: the pre-clean, forced create, bind check, drain and final delete. A manager goal-index scan whose consumer create fails now fails closed as unavailable and names its stream and filter, as the other two scans do. It used to surface the raw JetStream error.
• A backup create that recovers an interrupted create's claim first now backs up the cut's recorded
  store. On a root whose store is a custom --store-dir, it looked for .cotal/nats instead, removed
  the partial artifact and then failed with an unknown outcome, "store path is missing".
• backup create without --store-dir now backs up the preserved cut's recorded source store. It used
  .cotal/nats, so after a restore into another store it refused path-missing although the cut was
  ready.
• A backup create refusal or failure now shows its cause as native detail, such as a broker still
  listening at the preserved cut's endpoint. Since the command line redesign it printed only a generic
  sentence, while backup restore and down --preserve-state kept naming the same cause.
• cotal channel members on an open mesh now refuses with UNSUPPORTED, exit 2, and says why: the delivery service writes the membership feed, and an open mesh runs none. It used to report an unestablished census, exit 4, and tell you to check a delivery service the mesh does not run. The remedy names cotal who for who is present.
• cotal channel and its subcommands now report a target they cannot use, such as a space that is prepared but not started, the way ps, who and history do: a refusal under its own code, such as TARGET_UNAVAILABLE, in one JSON envelope with --json, exit 2, and the remedy under Next. They printed one line and exited 1, the usage exit, with no JSON.
• /help in chat lists the slash commands. It used to be refused as an unknown command.
• /leave of a channel the chat session never joined now answers not on #<channel>. It used to say
  still on #<channel>, as if the session were reading that channel.
• /history in chat works on an open mesh. It reads the channel the way cotal history does,
  because an open mesh runs no delivery daemon. It used to fail with "no responders" and a raw
  delivery subject.
• /quit in chat --plain's numbered recipient list leaves the session. It used to print the list
  again, so only the end of input left.
• In cotal chat --plain, :to <target> now sets the recipient. :to eve or :to @eve became a draft to the current channel, and answering y would have posted that text. :to @eve, :to #<channel>, :to @<owner>.<actor> and :to role:<role> now choose that recipient, and a target in another form, such as a bare name, names the forms it accepts and opens the numbered list. The docs now describe :to, :multi and the other line commands.
• /reconnect in chat reports the outage it causes once. It used to print the gap line twice.
• chat now shows when a static or signed-in session ends: the banner gives the time, and a notice a
  minute before says the session is ending. In --plain, /quit leaves a session that has already
  ended. It used to be answered with "end input (Ctrl-D) to leave".
• A static chat session can /join a channel again after /leave. Leaving used to empty the
  session's read list, so /join refused with a message saying the session was opened on no channel.
• cotal chat --stdio now answers every line in the order it arrived, as the docs say. A request refused at validation, such as an unknown type, a line that is not JSON or one too long to read, was answered at once, ahead of the requests queued before it. A request reusing the id of one still queued was refused under that id before the first request's result, so a client matching by id read the first request as refused. Each refusal now waits its turn, so the first request's result comes before the duplicate's refusal, and the refusal says the id was still in flight when the request arrived.
• cotal completion install with a symlinked startup file, such as a ~/.bashrc that stow links into a dotfiles folder, now names the file, says it is a symlink and gives the line to add to the file it points to. It used to refuse with "could not be opened safely" and no path or next step. Bare cotal completion also warns that the file is a symlink.
• cotal completion install and uninstall now name the directory they refuse and why. The refusal said only that completion paths "cross a symlink, untrusted owner or writable ancestor", with no path and no next step. It now names the directory and whether it is a symlink, not a directory, owned by another user, or writable by group or others. For a writable one, Next gives chmod go-w <path>. Otherwise Next says to wire completion yourself from cotal completion script <shell>.
• cotal completion install with a group-writable ~/.bashrc or ~/.zshrc, as a umask of 0002 leaves
  one, now names the file, its mode and the chmod go-w that fixes it. It used to refuse with "shell
  integration needs an owned, bounded regular file" and no path or next step. One refusal names every
  unsafe startup file, stub and folder on the way to them, and nothing is created before it.
• cotal config set persona now refuses a value that can name neither a persona nor a persona file. It saved anything that was not empty, such as " spaced " or a name holding a terminal escape. A name must use letters, digits, _ and -, as persona create requires, and a path must hold no control characters or surrounding spaces. Whether the persona exists is still checked when a launch uses it. The no harness refusals of config set harness and setup --harness now escape the value as the other keys do, so a control character in it no longer reaches the terminal.
• cotal console --plain and cotal chat --plain now write their lines in ASCII, as every plain report does: no status marks, and ·, … and → as -, ... and ->. They drew the marks and punctuation under any locale. The console line stream that a pipe gets does the same under a locale that is not UTF-8, as LC_ALL, then LC_CTYPE, then LANG decide.
• cotal console now shows a control character in a peer's message body, name or other text as its \uXXXX escape, in both --plain and the TUI, as history and watch --plain already did. console --plain wrote a body to the terminal as it arrived, so a message could clear the screen, set the terminal title, plant an OSC 8 hyperlink, ring the bell or overwrite its own line with a carriage return. The TUI passed the title and hyperlink sequences through too. The TUI still shows a body's line breaks.
• A controller-stack up now says when another space stays selected, with the same line a legacy up
  prints: "<space>" up; current is still "<selected>" - cotal use <space> to switch.
• A copied or moved controller project folder no longer gets an "unsupported shape" refusal. The stack journal records the absolute folder it was written in, and in a cp -a copy or a renamed folder it named the old path, so up (and every other command that reads the journal) refused SCHEMA_INVALID with a literal <root> in its Next. A well-formed journal that names another folder now refuses CONFLICT (exit 2) and names that folder: a copy is told to use the stack from the original, and a moved folder to move back there first. A malformed journal still refuses SCHEMA_INVALID, and its Next now names the journal's real path. space create in a copy whose records name another root as their space's root says that the records were copied or moved from there, rather than calling the copy that space's root while status names the original.
• status no longer suggests the deprecated cotal deliver when the delivery responder is not bound;
  it names cotal up. The delivery daemon's missing-credential error named cotal go, which does not
  exist, and now says that cotal up mints and writes the credential.
• cotal doctor auth names a credential file that other accounts can open by its mode, with chmod 600 on the file as the repair, instead of reporting a parse failure. It also parses each file the way --creds does, so a file the NATS client cannot parse, such as one with Windows (CRLF) line endings, is a finding rather than healthy.
• cotal down now runs to its end when the reader of its output goes away, as cotal down --yes | head -1 does. It used to exit 0 at its next write, partway through a member's stop, and the next down or up refused LIFECYCLE_BLOCKED because that stop was never confirmed.
• cotal down run in a folder that holds no stack, while --space or the selected space names a controller stack rooted in another folder, now refuses with TARGET_NOT_FOUND and names that space and its root, with a next step that runs the same down from there. It used to blame --yes or --space, or say there was no stack, and point at a down that stopped nothing.
• A bare cotal down --space <name> or cotal down --dry-run --space <name> run outside the folder where that space's stack is recorded now refuses the same way for both, TARGET_NOT_FOUND (exit 2), naming that folder and the stop or preview to run there. The stop used to exit 1 with a usage error and the whole help, and neither named the folder.
• A repeated cotal down --with-agents --yes now reports that the stack is not running and exits 0 when the stack's last completed down already took --with-agents, and so does --with-agents on a root that never started a stack. It used to refuse with LIFECYCLE_BLOCKED and ask to start the stack first. After a down without --with-agents it still refuses, because agents that run outside the manager may remain.
• cotal down --yes now works on every root. On a root without a controller stack, such as a prepared space that never started or a stack up -f started, it gives the same answer as cotal down: it stops the running stack, or reports that the stack is not running. It used to refuse with UNSUPPORTED because --yes applied only to a controller stack.
• auth mint --endpoint, ep describe and ep invoke now say when no instance of the endpoint
  answered its contract read: they refuse with TARGET_UNAVAILABLE, exit 2, and their Next says to
  start the service that serves it. They used to fail with "the endpoint contract read did not
  complete", exit 5, and no Next.
• cotal endpoints now shows each participant's name and role in separate columns, as cotal who does. The Participant column held <name>/<role>, such as manager/manager, and following the Next line with that value, cotal endpoint describe manager/manager, was refused. The Participant column now holds the name endpoint describe takes. The --json rows are unchanged.
• cotal ext remove creates a missing process-record folder under a mesh root's .cotal with mode
  0700. Under a group-writable umask, the reservation it takes for a provider whose records sit in a
  folder of their own left that folder 0775.
• A first cotal up that launched nothing no longer wedges its folder. When another NATS broker already answered at the address, the refused up left a stack journal that every later command treated as a stack that once ran: up, up --server <other>, reset space and down --yes all refused LIFECYCLE_BLOCKED. A journal whose members never launched now pins nothing. The next up plans a first start, at the same address once it is free or at another --server, and down reports that nothing runs there and exits 0. reset space and reset store refuse TARGET_NOT_FOUND and name cotal up, because no broker wrote a store.
• A foreground cotal team boot that another command's cotal down releases now exits 0, as a foreground cotal up does. It used to keep the exit code 4 of its partial start receipt, so a script waiting on it read a normal stop as a failure.
• Ctrl-C or SIGTERM on a foreground cotal up while the controller stack is still starting now stops
  what was started. The start settles first, then the same guarded whole-stack stop as bare
  cotal down runs, and the command exits 130 or 143. Before, the interrupt ended the command at
  once with no output and left the broker, delivery and manager running.
• space forget now removes the stopped stack definitions cotal up saved for the space at its root, so another folder can start a stack under the same name. When a member was not stopped cleanly, the definitions are kept and the result warns with STACK_DEFINITIONS_KEPT, naming the route that releases them. A bare up that meets another folder's stack definitions names that folder and every definition saved for it.
• A controller up whose broker address another broker already answers now says where to go next. The refusal said only "something already answers at the stack's broker address", without the address, and its next step was to inspect the stack's members. It now names the address and any space this machine registered there. For a first start, the next step is cotal up --server at another port, or stopping that space and running cotal up again. A stack that ran keeps its recorded address, so the next step is stopping that space, then a bare cotal up.
• Bare cotal now offers cotal up first for a prepared space that has not started. Before, it offered
  cotal spawn for a selected prepared space, which refuses until the space runs. With nothing
  selected, it also names the space prepared in the current folder, including right after down
  there, rather than saying no space is selected and offering cotal setup. With no agent application,
  status Next no longer says setup prepares a space when one is running or prepared, and on a
  running space it still lists ps and use.
• join --creds or --token and feedback-intake --creds with no --server now dial the broker the
  space is registered at, keeping its TLS requirement, and refuse an unregistered space before any
  dial. Before, they sent the credential to the default address.
• On a running legacy root, the reset store and reset space verdicts now name cotal down and then
  cotal clean store --force or cotal clean all --force from that root. Before, they named the reset
  again after down, which refuses because a legacy down drops the registration a reset selects.
• A bare cotal up on a stopped legacy root now starts it at the address its last start used, and in
  open mode when that start was open, and says so. Before, a legacy down dropped the registration
  that kept the address, so the next bare up started on 127.0.0.1:4222 in static auth. After
  clean all on a root deployed with up -f, a bare up that would start another address, mode or
  space refuses and names cotal up -f <manifest>.
• A legacy project folder that was renamed or moved after up -f no longer leaves a stack that no down stops. The launch summary that up -f writes records its absolute folder, and after a move it named the old path, so it was counted as an unreadable record. status then exited 4, team show <id> refused SCHEMA_INVALID, and bare down refused LIFECYCLE_BLOCKED while the broker kept running. A summary that reads cleanly but names another folder is now listed by status with the folder it was recorded at, blocks no up or down in the new folder, and team show <id> refuses it with CONFLICT, naming that folder. A summary that does not parse is still reported unreadable and still blocks. A copy of a running folder still refuses down, so the original keeps running.
• A bare cotal up in a prepared project folder that was moved no longer starts the legacy launcher under the default "main" space. A prepared space records its folder by path, so after the move nothing in the folder names it, and up fell back to the default. When the selected space is prepared in a folder that no longer exists and the current folder names no space, up now refuses TARGET_UNAVAILABLE (exit 2) with nothing written, and says to forget the space and prepare it again in its new folder, or to pass --space to start another space there. space create <name> for a space whose recorded folder no longer exists now says so and says to forget it first, rather than "start it with cotal up there".
• A project folder moved after down --preserve-state no longer blocks resume and restore with an unknown outcome and no path. The maintenance journal records the preserved store by absolute path. When that path is gone and the folder's own .cotal/nats is the same store by device, inode and generation, up, backup create and backup restore (with or without --accept-missing-source) now refuse CONFLICT (exit 2), naming the folder it was moved from, and say to move it back there first. status gives the same next step instead of "resume with bare cotal up". A store that is missing from its recorded path is refused TARGET_NOT_FOUND with that path, rather than exiting 3, and backup restore --accept-missing-source still restores over a lost store. Previously up exited 3 with "store path is missing", restore exited 3 with "preserved source is missing", the flag gave "store path must be absent", and backup create blamed a --store-dir that was never passed.
• cotal persona edit keeps your edited draft when it refuses with CONFLICT because the persona changed or was removed while the editor was open. It deleted the draft, so the edit was lost. The refusal now names the draft and how to use it, and a persona file that is gone is reported as removed.
• cotal persona edit with an editor the shell cannot find or run now refuses with DEPENDENCY_UNAVAILABLE, exit 2, as it does when no editor is set. It failed with OPERATION_FAILED, exit 5, and "editor exited with status 127", with no next step. The refusal names the editor and says to set VISUAL or EDITOR to an installed editor or run cotal config set editor vim. It also notes that the value runs as a shell command, so a path that holds spaces needs quotes. An editor that runs and exits non-zero still fails as before.
• persona create <name> --file <path> with a file that names another persona now says which name the
  file has and points to cotal persona create <file's name> --file <path>. Before, it said only that the
  name must match the catalog name.
• cotal persona create --harness <name> refuses a harness this machine does not know, with USAGE_INVALID and the list of known harnesses, the same check cotal config set harness makes. It used to save the pin as given, so a typo surfaced only when the persona was spawned.
• A refused persona name now says which rule it broke. A name over 128 characters said only invalid persona name; it now gives the limit and the name's length, and the character rule names the limit too.
• cotal persona with no space selected, outside a project, now says that no space is selected and this folder is not a space's project. It said no mesh running, though the catalog is read offline and needs no running mesh. Commands that read the same catalog, such as cotal team, give the same message.
• cotal persona create, edit and remove, and the legacy personas new, edit and rm, refuse a persona catalog folder you made read-only, such as .cotal/agents set to mode 0500, with PERMISSION_DENIED, and leave the folder and its files as they were. They used to set the folder back to 0700 and write into it without a word. edit refuses before the editor opens.
• cotal personas list prints the YAML diagnostic of an unparseable persona as lines: the reason, then the excerpt with a caret under the fault. It printed them on one line, with each line break shown as \u000a.
• personas list and personas list -v now escape control characters in the persona names, roles,
  models, descriptions and bodies they print, as persona show does. persona create and
  personas new refuse a --role or --model that holds control characters.
• personas show on a terminal now escapes control characters in the persona file, as persona show
  does. Piped or redirected, it still writes the file byte for byte.
• cotal help personas <subcommand> and cotal personas <subcommand> --help print that subcommand's usage for list, show, edit, new and rm. help personas new and help personas list said the subcommand did not exist, although personas runs them.
• down --preserve-state now cuts the store the mesh record names, such as a
  backup restore --store-dir target. It used .cotal/nats unless --store-dir was passed, so after
  a restore into another store it stopped the delivery daemon, left a mismatched prepare intent and the
  manager fenced, and exited with an unknown outcome. A --store-dir that differs from the recorded
  store now refuses CONFLICT before anything stops.
• A re-entry up that adopts the surviving listener of an interrupted restore or resume now
  recreates the space's memory-only buckets before the daemons start. When the first up stopped
  after binding its listener and before it made them, the delivery daemon exited with "stream not
  found" and the re-entry reported a degraded control plane.
• space add and meshes add against a broker that requires auth, with a --root that holds another
  space's trust, now say that the folder holds the trust for that space and refuse with
  DEPENDENCY_UNAVAILABLE. Before, they said to copy the mesh's credentials there and re-run with
  --mode auth. The note on what such a copy carries is now part of the same line, so space add no
  longer prints a literal \u000a in its message.
• A command run from a folder that no longer exists is refused with USAGE_INVALID, exit 1, before
  it does anything, and says to cd to an existing folder. status and other commands used to
  report an unknown outcome, exit 3, with Node's raw uv_cwd error. help, version and
  completion still work there.
• After a crash on a controller stack, cotal down --yes now leaves a root that cotal reset store, cotal reset space and cotal clean store|all --force accept. That down finds every member gone and records it as already stopped, and these commands, which remove the store, now count such a member as stopped. They used to refuse it for having no stop receipt and name a down that changed nothing, so the only way out was a full up and another down. A reset refused before that down now names cotal down --yes. Backup still requires each member's own stop receipt.
• backup restore and up --restore with an artifact of another space now refuse CONFLICT (exit 2)
  and name both spaces. They reported an unknown outcome (exit 3) and sent the operator to check a
  restore attempt that never existed, although the check runs before the preserved cut changes.
• backup restore and up --restore with the preserved source missing and no
  --accept-missing-source now refuse TARGET_NOT_FOUND (exit 2), naming the missing path and the
  --accept-missing-source route. They reported an unknown outcome (exit 3) with a Next about a
  restore attempt that did not exist, although nothing was changed.
• cotal backup restore <dir> --store-dir <path> and cotal up --restore <dir> --store-dir <path> refuse a store path that already exists, or whose parent directory is missing, as USAGE_INVALID (exit 1), naming the path, before anything is staged. Both used to exit 3 with OPERATION_UNKNOWN, "store path must be absent" or a raw ENOENT, and a Next that said the journal names no restore attempt.
• cotal send refuses a lone - and names --stdin and --file -, which read stdin. It used to
  publish the text "-". An empty quoted body is refused as empty. It used to be told to choose one
  body. These refusals now suggest the destination you gave.
• cotal send --file - now reads the message from stdin, as --stdin does and as persona create --from - already did. It was refused with TARGET_NOT_FOUND and "the explicit message input could not be read". A message file that cannot be read is now refused with its path, in the message and in the --json target.
• cotal send now says what to do when its target or a mention cannot be resolved. A name that several participants share is refused with the full identities that carry it, and one that is not present points at cotal who --all. Both reach error.remedy and the human "Next" line.
• cotal space add --mode open now checks that the broker serves a space by that name, as its streams show. A name it does not serve, such as a typo, is refused with TARGET_NOT_FOUND naming the spaces it does serve. Before, an open broker's answer was enough, so the mistyped space was recorded as healthy and every later operation on it failed without naming the name.
• cotal space add outside a project with no --root now refuses with one line and a remedy that says what to pass, instead of printing the whole space help. The line no longer describes the folder only in terms of credentials: an open space keeps none there, so any folder will do.
• space add of a space this machine started, or one a catalog lists, now says so in its refusal
  and points to cotal space show <name>. Before, the refusal offered --force, which is then
  refused for that space.
• space create refuses a name that cotal up cannot start, such as one with a space, a dot or a
  non-ASCII letter, with USAGE_INVALID before anything is written. Such a name used to be prepared,
  and the first up in its folder then refused it and left the folder held by that name.
• cotal space create <name> --root <dir> on a folder this account cannot write, such as /, now refuses PERMISSION_DENIED (exit 2) with " is not writable by this account, so its .cotal could not be made; nothing was written" and a next step, rather than exiting 3 with an unknown outcome and a raw EACCES error. Nothing is written in either case. cotal setup prepares a space through the same step and refuses the same way.
• The space delete receipt now has a Next that names the route back: on a space this machine started, cotal down --yes (or cotal down on a legacy root) and then cotal up start it again with empty streams, or space forget after that stop drops it. A bare cotal up under a broker whose streams were deleted now refuses TARGET_NOT_FOUND (exit 2), saying the broker no longer holds the space's streams and naming that restart, instead of exiting 3 with a raw JetStream code 10059.
• space delete of a stopped stack now names cotal reset store <space> --yes, which removes its stored
  streams, and space forget. Before, it said to start the space and delete it then, which meets the
  running-stack refusal again. A prepared space that never ran names only space forget. The
  --dry-run warning on a running stack no longer says that no broker was contacted.
• When a controller stack's manager or delivery daemon has died, status now leads its Next with
  cotal up --root <root>, which restarts it. ps against a stack whose manager died refuses with
  "no manager is running for this space" and names bare cotal up, as it does for a stack started
  with --no-manager. It used to end with "no manager reachable on the ep rails" and no Next.
• A controller stack's first up records its broker address, and an up --server naming another address refuses CONFLICT. That refusal said only "this root's stack journal names another space or broker address; it was not rebound", with no address and no next step, after a down, after a successful run, and after space forget and space create. It now names the recorded address and the address this up asked for, and its next step is the bare cotal up that starts the stack at the recorded address. A different --store-dir, --max-file-store or --host names each input it changes with its recorded value, and gives the same next step. up --help and docs/cli.md no longer call --server an override: a stack's first up records it.
• status --components no longer fails a healthy stack because the optional web dashboard is not
  installed. With no dashboard extension and no dashboard process record, the web row reads
  not-installed · optional and exits 0. An installed dashboard that is not running still reads
  absent and exits 1.
• cotal status no longer reports the delivery responder as bound right after the delivery daemon is killed. The dead daemon's ready row stays in the lease for up to 30 seconds and names this folder's own credential, so status read it as healthy while the same report showed the delivery process gone. When this folder's delivery pidfile names a process that has exited, status now reports that row as stale: not bound, a dead daemon's record that expires on its own.
• cotal status on an open mesh no longer reports the delivery responder as NOT BOUND with a cotal up remedy. Open mode runs no delivery daemon, so when no delivery lease is recorded the row says the responder is not used, and status --json gives deliveryResponder: null. A lease that is recorded is still reported as before.
• A command whose output cannot be written no longer crashes with a Node stack trace. cotal version >/dev/full printed a 21-line trace and exited 1, and every command did the same on a failed stdout write. It now prints one line, could not write to stdout, with the reason, and exits 3, or with the non-zero status it had already set. A --json refusal whose envelope cannot be written keeps its own exit code, and a reader that closes the output early, as | head does, stays quiet there too.
• up, up --restore and backup restore refuse a --store-dir that is a control directory or lies
  in one, such as a root's .cotal or the machine home's .cotal/nats, unless it is the root's own
  .cotal/nats. Before, the broker wrote its store among the root's credentials and records, or into
  the home stack's store, and reset store then could not inspect it or offered to remove the home
  store.
• The existing part of an explicit --store-dir path on cotal up, cotal up --restore or cotal backup restore is now held to the rule space create applies to a project root. A group- or world-writable directory in it refuses PERMISSION_DENIED before anything is written, naming the path and its chmod remedy. up used to start the stack on such a store and leave the directory writable by other local accounts. A non-empty directory is still accepted.
• An explicit --store-dir on cotal up, cotal up --restore or cotal backup restore that names a regular file, or a path under one, now refuses USAGE_INVALID (exit 1) before anything is written, naming the file as not a directory. up used to exit 3 with OPERATION_UNKNOWN "the controller stack did not start every planned member" without naming the path, after writing its stack journal, and the restores exited 3 with "store path must be absent".
• team boot of a manifest that requests no agents now returns ok with a complete receipt. Before, it
  returned partial with exit 4 because agent joins were not checked, though there were none. The docs
  now give the reason every team remove --dry-run is partial.
• team check, team boot and team deploy --dry-run now name each manifest problem by file, line,
  key and reason, and --json lists them under target.issues. An alias bomb is reported as a
  manifest problem. Before, every mistake gave one identical sentence.
• team deploy now names each manifest problem by file, line, key and reason, as team check and
  team boot do. Before, it refused every mistake with one sentence.
• cotal telemetry <level> no longer creates or rewrites the settings file while the environment forces telemetry off. Under DO_NOT_TRACK, COTAL_TELEMETRY_DISABLED or COTAL_TELEMETRY=off it created ~/.cotal/telemetry.json with an install id, though the docs say the file is neither read nor created then, and it reported ok on every repeat while the receipt showed no install id. It now saves nothing and reports a no-op with an ENVIRONMENT_OPT_OUT warning that names the variable to unset.
• cotal up --channels <path> with a file that is missing, unreadable or not JSON now refuses USAGE_INVALID (exit 1) and names the file, before anything is written. It used to exit 3 with OPERATION_UNKNOWN and no path, after writing the stack journal and creating the broker store directory, and every later up in that folder then refused.
• cotal up -f <file> run again in the folder whose mesh for that manifest's space is already running now says the mesh is already running and names cotal team deploy <file> and cotal down before another up -f. It used to report the folder's own broker as one that a fresh boot will not adopt and advise selecting another address.
• up -f on a root that holds a stack preserved by down --preserve-state now refuses
  LIFECYCLE_BLOCKED (exit 2) and names bare cotal up, which resumes it. It reported an unknown
  outcome (exit 3) with no next step, although nothing had started.
• cotal up refuses a URL or port-bearing --host, or one naming another address than --server,
  before anything is written. On a prepared root the controller stack used to take such a value, exit
  3, and leave a stack journal and trust behind, so every later up refused with LIFECYCLE_BLOCKED.
• A cotal up that loses the root to another whole-stack or maintenance command now refuses with LIFECYCLE_BLOCKED, names that command's pid, and gives a next step, as cotal down already did. It used to exit 3 with an unconfirmed outcome and a bare "maintenance lock is held by a live owner".
• A start that names no store no longer opens an empty second store over data down stopped
  tracking. After backup restore --store-dir and a plain down, up and up -f opened a new
  .cotal/nats and left the restored messages unused in the other store, with no warning. They now
  refuse CONFLICT and name the store the last broker served, and up -f accepts --store-dir to
  reopen it.
• cotal up --space <name> --server <url> from another folder, where <url> is the address that space already runs on, now refuses CONFLICT (exit 2), saying the space runs there and naming the folder it was started in, instead of blaming "a broker not started here" and suggesting a free port that would start a second broker.
• The receipt of a controller cotal up now offers cotal spawn --space <space>. up keeps an existing selection, so after starting a second space the bare cotal spawn it used to offer launched into the other, still-selected space.
• up now refuses a tls:// broker address, from --server or a manifest's broker.servers, when no
  TLS certificate and key are configured. It used to start a plaintext broker and record the tls://
  address, which every later dial reads as required TLS. The refusal is USAGE_INVALID before anything
  starts, and it names --tls-cert and --tls-key, or a nats:// address.
• cotal watch --json now names the resolved space and the refusal's own reason in its error record, as the plain form does. watch --json --view dms on an authenticated space used to report only requestedSpace: null and a generic sentence. The plain refusal also begins with its Mesh: line.
• A manager now reports the machine it runs on in its presence, as an agent does, so cotal who --all fills the Host column for manager/manager rows. An owner with several managers can tell them apart by it.
• cotal manager reconcile on a healthy stack now says the gate is open and there is nothing to repair,
  and its Next says nothing needs doing unless you meant another gate. It used to point at the delivery
  liveness oracle. A missing gate reads the same way. The code and exit code are unchanged.
• The manager's teardown target now declares the admitted credential renewal that the static terminal drains before it begins. The free, stop and shutdown paths build that target from the agent's fields, so the agent's handle, seed and restart state no longer ride into the detached teardown. Retirement still drains an accepted renewal before it revokes and cleans up the credential family.
• cotal run check no longer tells you to pass --discard-approvals when an edit drops an answered checkpoint, an override that run check refuses because it decides what a migration commit does. The L5004 reason now says to keep the checkpoint in the edited program, or to discard the decision in a migration commit under the --discard-approvals override.
• A hosted cotal run verb the manager refuses now reports the outcome and code as other commands do, with --json as one refusal envelope. run show of a run the manager holds no record of is TARGET_NOT_FOUND, exit 2, and names cotal run ps. It used to print run status: … on stderr, write nothing on stdout and exit 1.
• A workflow program file that cannot be read is now refused with TARGET_NOT_FOUND, exit 2, by cotal run start, run start --local, run resume --local --file and run check, as run validate, plan and simulate already refused it. run start and resume reported the raw ENOENT as an unknown outcome with exit 3, and run check printed it as a bare error with exit 1. The --json envelope now names the file as its target.
• The cotal run answer and cotal run amend receipts now name the run in their Next line and their JSON target. They printed cotal r show <run-id> although the command had named the run.
• cotal run show and run journal list a pause under Needs an answer only while its last journal row is pending, and offer answer as a next step only when one is open. A checkpoint that was answered or expired kept its earlier pending rows in the journal, so completed and failed runs still listed it as needing an answer. A pause settled with an accepted answer now prints that answer's value, answerer, artifact when present, time and answer id, then one amended line per later amendment, as the docs describe; before, only --json carried them.
• cotal run simulate refuses a results script whose scripted checkpoint result is not one checkpoint() can return, with SCHEMA_INVALID (exit 1) naming the checkpoint and the problem. A status other than resolved or expired, a missing status, a non-object entry, an unknown field or a scripted at used to simulate as a resolved answer and report ok, and a scripted at was dropped without a word.
• cotal run validate, run plan and run simulate list every problem in a program that does not validate: its code, title, position, cause and fix, as a hosted run start refusal lists them. They used to print only the workflow is not valid cotal-lang with no problem at all, and --json carried only each problem's code and span. The JSON envelope now carries the full problems under result.problems, beside the unchanged SCHEMA_INVALID refusal.
• A workflow wait given a value that is not an event, such as wait("never-sent"), now fails its step with a message that names the four event constructors and the value it was given. Before, the hosted and local runtime failed it with the host's own Cannot read properties of undefined (reading 'split').

What's Changed

• fix(cli): stop an interrupted foreground start and restart legacy roots where they ran by @davidfarah2003 in #4447
• fix(workspace): let a paired endpoint caller call, and refuse symlinked adapter paths plainly by @davidfarah2003 in #4449
• fix(runtime): list validation problems, settle pauses in run show, and refuse at once with no manager by @davidfarah2003 in #4450
• fix(cli): refuse cross-space restores and up -f over a preserved stack, and name another COTAL_HOME by @davidfarah2003 in #4451
• fix(cli): answer chat --stdio in arrival order, set :to targets, and refuse unreadable run programs by @davidfarah2003 in #4452
• fix(workspace): never present --creds to the default broker address, and refuse read-only persona catalogs by @davidfarah2003 in #4453
• fix(cli): escape peer text in the console, and give send refusals a next step by @davidfarah2003 in #4454
• fix(cli): point down, up and status at the right space and root by @davidfarah2003 in #4455
• fix(core): read CRLF and BOM persona files, and name a symlinked shell startup file by @davidfarah2003 in #4456
• fix(cli): point status and the home screen the right way for prepared and dashboard-less spaces by @davidfarah2003 in #4457
• fix(cli): refuse unusable persona preferences and name the input behind five refusals by @davidfarah2003 in #4458
• fix(cli): give space add refusals a next step that works by @davidfarah2003 in #4459
• fix(cli): keep credentials off the default address and name manifest and preference problems by @davidfarah2003 in #4460
• fix(cli): recover a folder whose first up launched nothing and refuse a bad channels file early by @davidfarah2003 in #4461
• fix(cli): refuse unsafe roots and store dirs before up writes, and name recorded and busy addresses by @davidfarah2003 in #4462
• fix(cli): refuse tls:// served in plaintext and unstick restore, re-entry and custom-store starts by @davidfarah2003 in #4465
• fix(cli): escape personas show on a terminal, name team deploy problems, warn on invalid preference variables and respect TERM=dumb by @davidfarah2003 in #4466
• fix(cli): unblock stacks after a crash, a killed down or a lost output pipe, and settle repeated down and up -f by @davidfarah2003 in #4473
• fix(cli): resume observers after an outage, report a silent broker as not answering, and report a failed stdout write in one line by @davidfarah2003 in #4474
• fix(cli): say what an interrupted restore or backup create left, and release it once its processes are gone by @davidfarah2003 in #4475
• fix(lang)!: export the validator's name and node tables as frozen arrays and freeze CATALOG by @davidfarah2003 in #4476
• fix(cli): refuse a past mint expiry, classify hosted run refusals, and name the run and space in receipts and watch errors by @davidfarah2003 in #4477
• fix(cli): list the live managers when a managed verb cannot choose one by @davidfarah2003 in #4478
• fix(cli): preserve and back up the store the mesh record names by @davidfarah2003 in #4479
• fix(cli): check --creds files before dialing and name the target in refusals by @davidfarah2003 in #4481
• fix(manager): declare the admitted renewal on the teardown target by @davidfarah2003 in #4480
• fix(core): report a late channel subscription refusal once by @davidfarah2003 in #4482
• fix(auth): run the three sealed scans through one consumer lifecycle by @davidfarah2003 in #4483
• fix(cli): name the kept selection, the stopped-space delete route and a persona file's own name by @davidfarah2003 in #4484
• fix(cli): name every unsafe adapter path in one refusal and refuse a control-directory store by @davidfarah2003 in #4485
• fix(cli): refuse unstartable space names, removed folders and a bad up --host, and name a selection that is not the folder's by @davidfarah2003 in #4486
• fix(cli): remove a forgotten space's preferences and give channel refusals their own code by @davidfarah2003 in #4487
• fix(cli): let a static chat session rejoin a channel it left and report one outage per reconnect by @davidfarah2003 in #4488
• fix(runtime): show each failed step's recorded error and say what validator warnings mean by @davidfarah2003 in #4489
• fix(cli): make persona, plain chat output and telemetry opt-out say what happened by @davidfarah2003 in #4494
• fix(cli): name the original folder of a moved or copied project, and refuse an unwritable space root by @davidfarah2003 in #4495
• test(workspace): classify the presence-render census drift on main by @davidfarah2003 in #4496
• fix(cli): back up the recorded store on a recovered claim and name unsafe completion files by @davidfarah2003 in #4497
• fix(workspace): refuse a linked preferences file, naming its target by @davidfarah2003 in #4498
• fix(cli): say what reconcile, who, contract reads, chat sessions, send bodies and dead members found by @davidfarah2003 in #4499
• fix(cli): name the way out of a forgotten, deleted, moved or elsewhere-running space by @davidfarah2003 in #4501
• fix(connector-claude-code): skip the Claude Desktop entry while Desktop runs on macOS by @davidfarah2003 in #4500
• chore(release): version packages by @davidfarah2003 in #4448

Full Changelog: v0.99.3...v0.100.0