Skip to content

Jason 0.29.0

Choose a tag to compare

@github-actions github-actions released this 05 Nov 11:07
· 60 commits to main since this release
a01ed50

Jason 0.29.0: the runtime, the CLI and the plugin host, one self-contained executable per platform, built
at this version and released only after its tests passed.

What is new since 0.28.0

The runtime's journal is now a bus for plugins: a plugin that a person grants journal kinds is handed each line of those
kinds after the runtime wrote it, and may act back only through the public Runtime API and only as far as the person
granted.

  • The journal's kinds are published. Every kind the runtime reserves for its own lines has one document under
    docs/contracts/journal/, saying what a
    line of the kind records, which things it names and whether a plugin may subscribe to it, and every line the runtime
    writes is checked against its document as it is written: the test suite refuses a line off its document, and an
    installation writes it and its log names the kind. jason journal kinds reads them, a page at a time, in kind order.
    The contracts page
    states the rules a document follows.
  • A plugin subscribes by declaration and grant. A manifest lists the kinds it wants under subscriptions; you admit
    it to some or all of them in Plugins:Grants:<id>:Subscriptions of settings.json, and * admits every kind it
    declares. Where several plugins are admitted to one kind, they run one after another, in plugin-id order unless
    Subscriptions:Order:<kind> lists some of them first. Grants and orders are frozen at each load, so a change takes
    effect at the next jason plugin reload. A plugin that does only this is the new plugin kind subscriber, which
    implements no operation and delivers no notification; a provider or a notification plugin may subscribe as well.
    The plugin guide
    has the manifest, the grant and the order.
  • What a handler is handed, and how it runs. The runtime hands a handler one line of a kind at a time, in the order
    the journal holds them, with the envelope the handler before it in the chain returned; the handler answers an
    envelope, which goes on to the next handler. The runtime never reads what an envelope means. A transient failure is
    tried again, up to Subscriptions:TransientRetries times, after a wait of Subscriptions:RetryBackoffSeconds that
    doubles; any other failure ends the chain at that handler for that line, writes one subscription_failed line naming
    the kind, the line, the plugin, the class and the code, and the kind moves on. What a handler does is never handed to
    another handler as a new line. Delivery is at least once: a line may be handed twice, after a restart in the middle of
    a chain too, and a handler tells a repeat by attempt in its input, which counts how many times that plugin has been
    handed the line in this chain's record, this time included, so it is above 1 after a transient retry or a restart; a
    rewind opens a new chain, so it starts at 1 again, and that repeat is the person's own. The numbers under Subscriptions apply live. See
    the plugin guide.
  • A handler reads and writes Jason through host.call. A handler of a plugin you granted a Runtime API family calls
    Runtime API operations through a door the runtime opens for that one invocation, about that one line. The families
    are the ones the manifest's capabilities.api requested and you listed in the plugin's Api list under
    Plugins:Grants: read, campaign, contact, workitem, suppression, report and journal.append. It acts as
    the plugin, never as a person, with the line as the reason, so the chronicle names the plugin and the line. The
    installation's own doors, a person's decisions and an executor's verbs are in no family, so no grant opens them.
    The guide's capabilities section
    lists them.
  • jason subscription list, runs and rewind. list shows, for each subscribed kind, the plugins of its chain,
    where its position stands, the journal's last line of the kind, how many of its lines are due, and the members the current
    load cannot run. runs --kind <kind> pages the records of that kind's chains, newest first, each handler's outcome
    included. rewind --kind <kind> with --line <id> or --since <moment> moves the position back so the kind's lines
    are handed again; only a person may, so it takes --actor human:<your-name>, it refuses a target ahead of the
    position, and each rewind writes one subscription_cursor_rewound line. jason doctor gains a subscriptions
    section with where each kind stands and how its newest chains ended, and it never carries an envelope.
  • The journal-forwarder plugin pushes each granted line to one URL. It is a reference subscriber: for each line of
    a kind you granted, it POSTs the line, the envelope and the attempt as JSON to the URL named in
    JOURNAL_FORWARDER_URL, and answers the envelope as it was handed. It is in the repository's plugin marketplace under
    plugins/journal-forwarder,
    not a release asset: you copy the package into ~/.jason/plugins, as its README says. What leaves the machine is what
    the journal says, which includes names, fields as they changed and the reasons people typed, so the receiver is a
    place you trust with that. Its manifest names localhost:8787; a receiver elsewhere is your edit of the package, as
    the README says, and the copy is yours to keep current.
  • Two skills teach the new verbs. operating-the-installation and troubleshooting-jason now teach
    jason journal kinds and jason subscription, and both are draft again until a host reads their new text. The
    skills marked draft in this release are those two, researcher-in-session, roles/responder and the nine skills of
    the business pack.

Upgrading from 0.28.0

  • The executable: jason update apply, or run the install line again. Installed with Homebrew, it is the three
    lines jason update apply prints instead: jason runtime stop, brew upgrade jason, jason runtime start. The
    install line and Homebrew place the executable alone, so after either, jason skills update, jason ui install
    and jason plugin install reply --replace follow.
  • Skills: the update redeploys them wherever a deployment was recorded; after the install line or
    brew upgrade jason, run jason skills update. Two skills changed their text and their status line is draft
    again.
  • The data directory: this release adds one database migration, SubscriptionRunner, which adds the tables that
    hold each subscribed kind's position and the records of its chains. The first start on 0.29.0 runs it after taking a
    backup, as every migration does. Your projects, campaigns, work items and chronicle stay as they were.
  • Nothing subscribes until you say so. No plugin is handed a journal line until you install a subscriber and grant it
    kinds, so an upgrade with no subscriber changes no behaviour.
  • The Reply plugin is unchanged.
  • From 0.27.x or earlier, the notes of each release in between apply as well. From 0.18.x or earlier, Jason's own
    installation updates through 0.19.0: jason update apply --version 0.19.0, then jason update apply; from a build
    older than 0.18.0, the first of the two can stop at started, as the known issue about that update below says.

Known issues

reply skills install --project does not see an old reply-adapter@reply-skills plugin installed at user scope on
Claude Code, so that plugin keeps loading beside the project copy of the skills.

reply skills install --dry-run and reply skills update --dry-run describe what they would do in the past tense, as
if it had been done; nothing is changed.

The click on a notification and its Copy prompt button are Windows only. Where another program already holds jason: links,
Jason leaves that registration alone and the toast shows without a click or a button.

On Windows, jason session open starts the console host, which hands the window to Windows Terminal where that is
your default terminal; where your default terminal is the console host itself, you get a plain console window; neither
is yet measured on a machine.

The macOS form of jason session open, a script opened in Terminal, has not been watched on a Mac.

Antigravity CLI's hook route has not been run on an account. The host also reads your own global AGENTS.md and
GEMINI.md under ~/.gemini for every role.

Two installations on one account, a build of a checkout and the installed one, take the jason: link from each other at
each toast, since the registration rewrites a key that is Jason's and names another executable. Uninstall leaves a key
that names another installation alone, so nothing is lost.

On Linux, kitty, foot, gnome-terminal, xfce4-terminal, mate-terminal and wezterm are started in the form each takes;
any other terminal program is given the host after -e as separate arguments.

Under Windows PowerShell 5.1 and cmd.exe the inner quotes of --value '{"host":"codex"}' are lost, so give the value in a file:
jason preference set agent.interactive_host --value-file host.json --actor human:<your-name>.

If Claude Code's npm postinstall never replaced its placeholder executable, jason session open reports that it started the host
and the window closes at once.

Codex CLI on Windows needs its elevated sandbox set up once by an administrator before a role can run a command; until
then every command a role runs is refused, so the role cannot report, and the attempt ends as executor_exited and is retried against the same refusal; Jason does not check for the
setup when the attempt starts.

Whether jason reaches the runtime from inside Codex's sandbox has not been checked on Linux, macOS or Windows, and on
Windows it is not yet checked whether the elevated sandbox reads only the list Jason names. Codex also reads its own global
AGENTS.md and offers every role the skills in ~/.agents/skills, whatever the profile says, so those two reach every
Codex role. Holding Codex's network to loopback rests on its network proxy, an experimental feature in 0.166.0.

Whether Cursor CLI runs the allow-listed jason command without asking, in print mode, has not been seen on a signed-in
run. If it does not, a role cannot call the runtime back, so it cannot report, and the attempt ends as executor_exited and is retried against the same refusal. Cursor CLI also reads more than the
memory files Jason names, since it takes the union of every configuration layer it finds, yours included: your own allow
rules of any kind, a shell command or a write among them, stand beside Jason's. Each attempt leaves one trust marker under
~/.cursor that Jason does not remove.

Gemini CLI reads the whole folder of each level a role's memory files are in, and the client level's folder is the harness
root, so a role reads every memory file under it. A role on Gemini CLI is also refused when the host's system policy folder
holds a .toml file, since the host would then ignore the policy Jason gives it.
Gemini CLI's --session-id and --admin-policy, which Jason passes, are in the 0.66.0 source and not in its options table.
Gemini CLI also reads your own ~/.gemini/GEMINI.md, includeDirectories and skills for every role.

GitHub Copilot CLI's reads outside the work directory, which a role needs for its memory files, are documented but have not
been seen to reach there.
Whether Copilot CLI 1.0.105 accepts every flag Jason composes is not said by its changelog, so its launch is composed from the
reference and not from a run.
A role on Copilot CLI runs under your own user-level hooks, and a COPILOT_ALLOW_ALL in your environment reaches it as the
host's own allow-all, minus what the deny list names.

A proxy flag or a proxy variable is yours to set: Jason sets none of them and reserves none of them. A flag that names a
host's own server, its credentials or its certificate check is reserved where the host has one, as the Cursor CLI section
says, and a host whose arguments are an allowlist takes no flag outside it.

On a build older than 0.18.0, the update through 0.19.0 can stop at started, saying that a library failed to load.
That update is still performed by the older build, and it happens when the two builds carry different .NET patches.
Type jason update apply again: it runs 0.19.0, which finishes the update, and one more jason update apply then
takes the newest release. From 0.18.x this does not happen: 0.19.0 is built on the same .NET patch as 0.18.0.

Reply's routes behind 0.22.0's send features - a send's moment and its sending schedule, reading and cancelling a send
by its reference, and the inbox change feed - reach Reply's accounts on Reply's own schedule. Until an account's Reply
has a route, the Reply package answers capability_not_offered for that operation or argument, and the account's
replies are read thread by thread in place of the feed, as its README says; nothing in Jason needs updating when the
route arrives.

For a sequence email, Reply keeps at most one reply to each email it sent and drops a finished contact's later replies
on that thread before it stores them, so they reach none of conversation.list, conversation.get and the change
feed, and Jason's inbound sync says nothing of them. After this side has answered on the thread through
conversation.reply, the contact's next replies on it are stored and read as usual; a second reply that arrives
before any answer may never be seen. A LinkedIn thread keeps later messages. All of this is read from how Reply is built
and has not yet been observed on an account.

A send or an invitation through Reply carries no sending time in its own result: sent_at is null, because Reply's
answer to a LinkedIn send gives a local time marked as UTC and its answer to an email gives none. message.status
reports when a send went, and a message's own moment is read from its thread, where it is true UTC.

Through Reply, a sending schedule times a LinkedIn message only once the account's Reply times deferred LinkedIn sends
itself. Until then the Reply package answers a LinkedIn message that names one capability_not_offered, and nothing is
sent, whether the account's Reply has no direct sends yet or has them without a schedule on LinkedIn.

The approval preview shows a send's send_after and not the provider object beside the arguments, so a send that also
names a sending schedule goes at the schedule's first opening at or after that moment, which may be later than the
moment the approver read.

A template held for a later moment, or queued by Reply behind the account's other sends, is filled in when it goes: a
name Reply cannot fill then fails the send after a person approved it, which message.status reads as failed with
Reply's reason.

Reply converts a LinkedIn text from HTML to text after filling it in, so markup in a LinkedIn message or InMail, a
final text or a template, does not reach the person as written; what Reply makes of a < that starts a tag there is
not known yet, and the Reply package does not refuse one on LinkedIn as it does on email.

Reply does not consult its blacklist when it takes a LinkedIn direct send, and does when it makes it: a send to a
person its blacklist holds is accepted, and fails when Reply makes it, which message.status reads as failed with
Reply's reason. Jason's own do-not-contact register is checked at the claim as before; record such a stop there with
jason suppression add as well, so the send is refused before it is made.

Through Reply, schedule.list does not say which campaign runs on which schedule: campaign.get names no schedule,
so which one a campaign sends on is read at Reply.

Through Reply, a reply on a thread, conversation.reply, is not held to the rule on a < that starts a tag yet, and
its text goes as written; whether Reply cuts a reply as it cuts a direct email has not been seen, so write an address
in a reply without its angle brackets.

The responder's skill ships marked draft: no agent host has read its text in the state it ships in. Its commands parse
against this build, as every skill's do, but how an agent follows its loop on a real reply is not yet seen.

A fault of the runtime's own that stops a subscribed kind's turn is logged, and the kind is offered again at the next
scan with no pause, from the same line, so a fault that lasts repeats at every scan until it is gone.

Assets

File For
jason-win-x64.zip Windows on x64
jason-linux-x64.tar.gz Linux on x64
jason-osx-arm64.tar.gz macOS on Apple silicon
manifest.json what a running Jason reads to learn that this release exists
checksums.txt the SHA-256 of each archive and of the two packages, in sha256sum format
install.sh macOS and Linux: the install script this release was published with
install.ps1 Windows: the install script this release was published with
jason-plugin-reply.zip the Reply plugin's package: the files of plugins/reply at v0.29.0, which jason plugin install reply places in ~/.jason/plugins/reply
jason-ui.zip the browser UI's package: ui/ built at v0.29.0, which jason ui install places as the default UI

Each archive holds the executable, LICENSE (the MIT licence, which covers the whole repository, the runtime
included) and THIRD-PARTY-NOTICES, nothing else. The Reply plugin's package is installed by
jason plugin install reply, which places it in ~/.jason/plugins/reply once it matches its line in
checksums.txt and the manifest's content digest; step 3 of docs/golden-path.md runs it. The UI's package is
placed by jason ui install, which the install guide runs and jason update apply runs at its end.
On Linux the executable needs the ICU library (libicu), which desktop distributions carry and minimal
images may not; the install script stops, with the runtime's own message, rather than install a program that does
not start.

Installing

One line, on macOS and Linux:

curl -fsSL https://get.jasonharness.ai/releases/latest/download/install.sh | sh

and on Windows, in PowerShell:

irm https://get.jasonharness.ai/releases/latest/download/install.ps1 | iex

Jason is released under the MIT licence: https://github.com/reply-team/jason-ai/blob/v0.29.0/LICENSE.

Both find this machine's archive, check it against checksums.txt before unpacking, and put jason on the
PATH — in ~/.local/bin, or %LOCALAPPDATA%\Programs\jason — with its licence and the third-party notices
beside it, as jason-LICENSE and jason-THIRD-PARTY-NOTICES. Running the line again over an installed Jason of
the same version changes nothing and says so, unless a licence file is missing beside the executable, which it
then writes. Two options, for a script that must be told:

curl -fsSL https://get.jasonharness.ai/releases/download/v0.29.0/install.sh | sh -s -- --version 0.29.0 --no-modify-path
& ([scriptblock]::Create((irm https://get.jasonharness.ai/releases/download/v0.29.0/install.ps1))) -Version 0.29.0 -NoModifyPath

Upgrading from 0.1.0

If this machine ran 0.1.0 with autostart enabled, run jason runtime autostart enable or disable once from an
ordinary prompt — if that answers Access is denied, follow its message first — then jason runtime stop once, then
jason runtime start (or sign in again, with autostart enabled). 0.1.0's logon task started a runtime that outlives a
sign-out and holds the data directory; until it stops, the runtime this version starts at sign-in finds the directory
taken and exits.

Checking a download by hand

sha256sum --check --ignore-missing checksums.txt

On Windows, compare Get-FileHash of the archive with its line in checksums.txt.

If you then put the executable in place by hand, keep LICENSE and THIRD-PARTY-NOTICES beside it, named
jason-LICENSE and jason-THIRD-PARTY-NOTICES.

What is signed

The macOS executable is signed with Reply's Developer ID and notarized by Apple. Before anything is published, the
release's macOS job has spctl assess the executable's signature and stops unless that signature is accepted from the
source Notarized Developer ID; the same job starts a copy carrying the quarantine flag a browser puts on a download,
and its log says whether that copy started and opened its database. The ticket is not stapled to the file, which a
bare executable has no place for; for a file without a stapled ticket, Apple's documentation says Gatekeeper finds the
ticket online. The quarantine flag is put there by the browser that downloaded the file, not by curl, so the
one-liner's download has none and install.sh clears nothing; an archive fetched with a browser and unpacked by hand
keeps the flag.

Who signed it is yours to read: codesign -dv --verbose=2 on the unpacked executable prints an
Authority=Developer ID Application: line naming Reply and its team id in parentheses, the same id the
TeamIdentifier= line gives. spctl accepts the signature of any notarized Developer ID from the source
Notarized Developer ID, so it is that name and that team id, not the acceptance, that say the executable is Reply's.

The one native library the executable carries, libe_sqlite3.dylib, is signed with the same Developer ID and
timestamped before it is bundled; the executable unpacks it under ~/.net/jason/ the first time it runs, and
codesign -dv --verbose=2 on it prints the same Authority= and TeamIdentifier= lines. So the executable keeps the
hardened runtime's library validation, which loads only a library signed by Apple or by the executable's own team, and
it ignores DYLD_ environment variables: the only entitlements it carries, com.apple.security.cs.allow-jit and
com.apple.security.cs.allow-unsigned-executable-memory, let the .NET runtime compile code as it runs.

The Windows executable is signed with Reply's certificate through Azure Artifact Signing, and timestamped under
RFC 3161, so the signature stays valid after the certificate expires. Who signed it is yours to read, in Windows
PowerShell or PowerShell alike: Get-AuthenticodeSignature on the unpacked executable says Valid, and, run where it
was unpacked, (Get-AuthenticodeSignature .\jason.exe).SignerCertificate.Subject prints the subject the certificate
is issued to, which begins CN=ReplyApp Inc.; signtool verify /pa /v, where the Windows SDK is installed, accepts it
and prints the timestamp. An archive a browser downloaded carries the mark of the web, and
SmartScreen looks at the executable unpacked from it the first time it runs. SmartScreen's reputation for a new
publisher builds over downloads, so in the first weeks of signed releases a download may still show its dialog;
install.ps1 clears that mark from what it unpacks.

The Linux executable is not signed, and nothing else a release carries is either.

Both scripts check the archive against checksums.txt before unpacking. On Linux that check is the only provenance
there is; on macOS and Windows the signature says who built the executable as well.