[lydakis]

Requested by George · project thread

This bumps agent, agent-client and sonny from 0.5.2 to 0.6.0 (Cargo.toml and Cargo.lock only), and adds a note to /usage in sonny/README.md that it reads retained history, as the outside review asked. Once it merges, v0.6.0 is tagged on the merge commit. The release notes follow and become the GitHub release text.

---

Everything since v0.5.2: Sonny is full screen, with a margin where agents
keep notes beside the conversation, several leads per folder, and goals
that keep an agent working until a check passes; it reaches your MCP
servers through one exec script tool, draws pictures, diagrams, HTML and
folder trees in the side pane, guesses your next message, shows /usage,
and runs as a T3 Code provider and in cmux. The daemon keeps companions
running for as long as it runs, and Sonny's lookout and approver are now
two of them. Bots fetch pages on every provider, type into running
programs, search their own history, run trusted commands outside the
sandbox, and go to a fallback model when a request is declined. Triggers
work on Linux, the lookout hears sources you set up, and long histories
open, scroll and find faster.

Before you update

• Update the daemon and its clients together. The socket protocol is now
  23 (was 21), for declared tools that take free-form text (Declared tools may take free-form text, as apply_patch does (protocol 22) #307) and the
  usage read (Sonny /usage: tokens spent over time, by agent and by model #302).
• The store moves to schema 62 (A request a safety classifier declined goes to the bot's fallback model, that request alone #349 replaces a bot's fallbacks flag
  with its fallback model; What a declined request added goes to later requests as a note naming the refusal #361 adds declined; Daemon companions: commands the daemon keeps running for as long as it runs #370 adds companions).
  The 0.6.0 daemon migrates the store on first start, and 0.5.2 can't open
  it after that, so copy ~/.agent/state.sqlite while the old daemon is
  stopped if you may go back.
• SandFS and Errand are gone from Sonny (Remove SandFS from Sonny #301, Sonny: remove its Errand support #314). Tasks always get a
  git worktree, and outside git they work in the lead's folder; /copies,
  --copies, the copies setting and sonny task changes are removed, as
  are /status's "On Errand" section and the errand skill.
• Agents no longer sign their messages with a From line (Leads carry what one task lacks to whoever has it; the model reads who sent a prompt #295). A prompt
  another bot or a named client sent reaches the model with a part naming
  its sender. --hide-senders, or senders = "hidden" in
  ~/.agent/config.toml, keeps the words alone.
• Sonny's chime and view live in ~/.agent/config.toml (Sonny: /config, chime and view in config.toml, and a side chat that can fork with every tool #315), as
  [sonny] chime and verbose. The same keys in
  ~/.agent/sonny/settings.json are no longer read.
• Sonny is full screen only (Sonny: full screen only, buttons, folds, quiet agent messages, tasks in the pane, pane keys and a little color #326). The inline mode and /fullscreen
  are gone.
• The lookout has no judge (Sonny: the lookout forwards everything and judges nothing #310). A subscription is now name,
  tasks_of, bots and sources; ask, look_for, context, on and
  updates are gone, as are sonny lookout --judge and --replay.
  [judge] stays for the approver.

Sonny

• Full screen only, with buttons and folds (Sonny: full screen only, buttons, folds, quiet agent messages, tasks in the pane, pane keys and a little color #326). The inline mode and
  /fullscreen are gone; /status, /background, /side and /mcp open
  in the pane. Calls of one kind in a row fold into one row, a question's
  options are buttons, messages between agents and task updates are one dim
  row a click opens, and reactions sit under the message they answer. Tasks
  the lead starts show as a roster with live states, and a click opens a
  task's own Sonny in the pane. Ctrl+[ and Ctrl+] go back and forward
  through what the pane showed, and Ctrl+\ moves the keys between the
  conversation and the pane. Colors use the terminal's 16, each with one
  meaning, and the context count turns yellow at 100k tokens and red at
  200k.
• An agent call is one row (Sonny: an agent call is one row saying what it did, opened on a click #369). Starting, messaging, steering,
  queueing for or waiting on an agent reads as one short row saying what
  happened, such as • Steered calc.neg ▸; another agent's message reads
  ↳ From calc.neg ▸, and a click opens the details.
• The margin (Sonny: the margin, where agents keep notes beside the conversation #308). Agents keep notes beside the conversation with a
  margin tool: a diagram, a results table, the file being discussed, drawn
  with the same Markdown, Mermaid, HTML and pictures as the pane. The plan
  sits at the top, a glance line shows tasks, what needs you and context,
  and × takes a note away. It shows by itself from about 141 columns; in a
  narrower window the footer counts the notes, Tab shows the margin, and
  each note is drawn inline where it was posted. A lead can show its tasks'
  notes in its own margin (the lookout tells it when a task puts one up),
  and a counter under the composer shows agents at work and what needs you;
  a click opens /status (Sonny: a lead shows its tasks' margin notes, and the footer counts agents at work #367).
• Your MCP servers, through one exec tool (Sonny: agents reach the person's MCP servers through one exec script tool #321, Sonny: sonny mcp tool, and scripts' saved files kept a week #327, Sonny: four exec scripts run at once in the lookout, the rest wait #334).
  [sonny.mcp_servers.NAME] in ~/.agent/config.toml takes Codex's keys;
  sonny mcp add|remove|list keeps them and /mcp shows them. Agents get
  one gated exec tool with Codex's code mode: JavaScript in QuickJS, each
  MCP tool as tools.mcp__SERVER__TOOL. The lookout hosts one connection
  per server for every agent, stops idle servers after ten minutes, and runs
  four scripts at once.
• The side pane draws more (Sonny: draw pictures, HTML and diagrams in the side pane #303, Sonny: draw \[ \] and \( \) math, in the chat and the side pane #313, Sonny: a folder as a tree in the side pane, and back and forward #319). Pictures, SVG, Mermaid
  and HTML files are drawn pane-tall, and a click on one in the chat opens
  it there; \[ \] and \( \) math reads like $$; a folder opens as a
  tree with git's state per file; ‹ › go back and forward. One Markdown
  renderer now draws every surface, decision cards and /status included
  (Sonny: one Markdown renderer for every surface #325).
• A guess at your next message (Sonny: a guess at your next message, for Tab to take #329). After a turn you started ends
  well, the empty composer shows a dim guess; Tab takes it in, unsent.
  [sonny] suggest = false turns it off.
• /config (Sonny: /config, chime and view in config.toml, and a side chat that can fork with every tool #315) lists every key config.toml takes, with its value
  or default, and writes a change right away, keeping your comments.
  [sonny] side_chat = "fork" makes /side a copy that keeps every tool.
• /usage (Sonny /usage: tokens spent over time, by agent and by model #302) draws token usage from retained history over the last
  day, week or month, by agent and by model; agent usage prints the same.
• Profiles (Let leads and tasks run as other profiles, and profiles build on others #296). A project's profile / threads_profile, or
  [sonny] lead_profile / task_profile, pick the profile leads and tasks
  run as. Profiles can extends another and carry skills. An opt-in
  superlead profile ships, used only when a project sets it.
• T3 Code and cmux (Lay Sonny's work out in cmux #300, Sonny: sonny acp for T3 Code, and cmux's Automation mode in the cmux skill #323, sonny acp: read the person's words out of each block T3 Code sends #337, sonny acp: T3 Code's model picker lists your models and switches the lead #339, sonny acp: a task opens as the lead thread's child in T3 Code #352, T3 Code opens Sonny threads again: the model option carries id #358, T3 Code: models as /model lists them, an effort picker, and a new thread leaves the lead's model alone #363).
  sonny acp makes Sonny a T3 Code provider: a thread is its folder's lead,
  each task opens as its own thread, approvals are asked in T3, and T3's
  model picker lists your models as /model lists them, with an effort
  picker, and a new thread leaves the lead's model alone (T3 Code: models as /model lists them, an effort picker, and a new thread leaves the lead's model alone #363); a task's
  thread opens as the lead thread's child (sonny acp: a task opens as the lead thread's child in T3 Code #352). Run in a cmux tab, Sonny
  lays its tasks out as cmux tabs and workspaces, and agents get a
  cmux-tabs skill.
• Leads (Leads carry what one task lacks to whoever has it; the model reads who sent a prompt #295, Sonny: a lead reacts 👀 to the person's words in a task it routed elsewhere #330, Sonny lead: the lead picks the emoji for a pickup in a task #340). A lead carries what one task lacks to
  whoever has it, splits a message of several asks into a task per goal, and
  reacts in a task when it routes the person's words elsewhere. A task knows
  its lead.
• Several leads per folder (A folder can have several Sonny leads: the last one written to opens, /lead switches #365). sonny opens the lead you last wrote
  to in a folder; /lead lists the folder's leads, /lead NAME switches,
  and /lead new [NAME] makes one (sonny --lead NAME,
  sonny --new-lead). Leads are peers, and each window and /status shows
  its own lead's tasks. In T3 Code a new thread is a new lead named after
  its first words.
• Triggers on Linux (Sonny: triggers on Linux, kept by systemd's user manager #344, Sonny: fix main's build; restamp edits a systemd trigger too #351). systemd's user manager keeps them,
  with the same CLI, tool and fires as on a Mac.
• The lookout forwards everything and judges nothing (Sonny: the lookout forwards everything and judges nothing #310). It passes
  on every message someone else writes in the agents a lead watches, and
  each finished turn with the task's own reply, marked as the task's words.
  The person's words in a task cut into a busy lead at once; the rest waits
  until the lead rests.
• Lookout sources (Let the lookout hear sources the person sets up outside the daemon #305). [sonny.sources.NAME] runs a command you
  write (for example a Slack poller); each line it prints is an event, and
  agents subscribe by name. Every event is told; one marked urgent steers a
  busy subscriber.
• Unsandboxed commands (Run trusted commands outside the sandbox, as Codex's allow rules do #309). [sonny] unsandboxed = ["mcpx"] lets
  those command prefixes run outside a lead's or task's sandbox without
  asking, when a script runs nothing else. In full mode tasks now run
  unconfined, as their lead does.
• Long histories (Sonny: older history drawn a page at a time above the rest #331, Sonny: code coloured with Oniguruma, about three times faster #336, Sonny: a picture sent to the terminal the first time it is shown #338, Sonny: Ctrl+F reads each line once, and places only the lines it is in #343, Sonny: lay out only near the view (pages let go, cheap repaints) #353, Sonny: Home goes to the first turn, and Ctrl+F finds what pages let go hold #364). Older history
  loads a page at a time above the view, at a steady cost, and every turn
  can be reached; pages far from the view let go and are drawn again as it
  nears them, so a 2,000-turn history holds 62 MB instead of 232 MB, and
  /verbose or a theme switch repaints in 0.45 s instead of 12 s; code
  colours about three times faster with Oniguruma; pictures go to the
  terminal only when first shown; Ctrl+F costs about 8 ms a key on 12,000
  rows instead of 130. With nothing typed, Home goes to the first turn in
  about a second on 2,000 turns and End comes back, and Ctrl+F also finds
  matches in pages the screen had let go (Sonny: Home goes to the first turn, and Ctrl+F finds what pages let go hold #364).
• Memory (Memory records which facts agents read, and the cleanup checks ones unread for 30 days #354, Memory marks facts saved after outside content, and the nightly cleanup saves what the person said #356). The memory tool and script take read NAME,
  counted in each scope's .reads; the nightly cleanup is told of facts
  unread for 30 days, and removes one only when something newer shows it no
  longer holds. A fact saved after a web read, a fetched page or an MCP call
  is marked external and read as a lead to check, never an instruction;
  the cleanup also reads what you wrote since its last run and saves what a
  later agent would otherwise ask again (Memory marks facts saved after outside content, and the nightly cleanup saves what the person said #356).
• Goals (Sonny goals: keep an agent working until a check command passes #341). The goal tool sets a check command for an agent or one
  of its tasks; each time a turn ends, the lookout runs the check and, while
  it fails, gives the agent another turn with its output. The goal ends when
  the check exits 0, the token budget runs out, two turns in a row call no
  tool, or the agent says it is blocked. Its lead hears only how it ended.
• Your edits to Sonny's roles and skills survive an upgrade (Sonny: offer a merge when an upgrade meets an edited skill or role #357). When
  Sonny has new text for a lead.md, task.md or skill you edited, it
  writes the new text beside yours and the next agent you open offers to
  bring its changes in, keeping yours.
• Smaller: Shift-Enter starts a line, and the composer takes the mouse
  (Sonny composer: Shift-Enter starts a line; mouse selection to type over or delete #297); an idle window starts no programs (Sonny: stop starting programs while a window sits idle #311); a first run sets itself
  up, and Linux gets an install script (Sonny: the welcome heads every lead's history, as Claude Code's does #320); the welcome heads every
  lead's history (Sonny: the welcome heads every lead's history, as Claude Code's does #320); a lead made for a subfolder reads its own project's
  memory (Sonny: a lead made for a subfolder reads its own project's memory #298).

Runtime

• web_fetch for every provider (Add web_fetch for every provider, and AI_AGENT for every command #306), Bedrock Claude included, gated
  and judged like other tools. Shell commands get AI_AGENT=agent.
• Type into a running program (Let a bot type into a running program: shell tty and write_stdin #322). shell with tty=true runs in a
  pseudo-terminal and returns a handle; write_stdin types into it, polls
  it or kills it, as Codex's unified exec does.
• Declared tools may take free-form text (Declared tools may take free-form text, as apply_patch does (protocol 22) #307), as apply_patch does:
  a custom tool on Responses, a one-string function elsewhere.
• A fallback model for a declined request (A request a safety classifier declined goes to the bot's fallback model, that request alone #349). A request Claude's
  safety classifier declines can go to a fallback model for that request
  alone, then the bot goes back to its own model (--fallback, or
  [fallbacks] in ~/.agent/config.toml). A request that is still declined
  goes to later requests as a short note naming the category, so the
  conversation can go on; the original stays readable through the history
  APIs, and turn_finished counts what was left_out (What a declined request added goes to later requests as a note naming the refusal #361).
• Companions (Daemon companions: commands the daemon keeps running for as long as it runs #370). agent companion add NAME -- COMMAND keeps a
  command running for as long as the daemon runs, started again if it exits
  and stopped when the daemon stops; rm and ls manage them. Sonny's
  lookout and approver now run as companions (sonny-lookout,
  approver-auto), added on Sonny's first open, so they live as long as the
  daemon rather than the window that started them.
• history searches earlier turns for text (Let the history tool search earlier turns for text #299).
• unsandboxed bot setting (Run trusted commands outside the sandbox, as Codex's allow rules do #309): command prefixes that run outside
  the sandbox when a plain script runs nothing else;
  agent run --unsandboxed CMD.
• Six rough edges (Fix six rough edges from a day of running bots #318): every Claude bot Sonny runs keeps its prompt
  cache for an hour; triggers tell their agent when a fire is missed; an
  empty reply ends a turn; a cut write or edit still names its file for the
  approver; a task that resolves itself still reports.
• Fixes: a stopping daemon ends background commands at once (Fix two test_wait flakes: a stop that waited on background commands, and a test lock that held the service loop #342); a
  companion that exits has what it left running stopped before it starts
  again (Stop exited companions’ descendants before restarting #371).

Docs and CI

• Pull requests run the Rust tests on macOS too (CI: run the Rust tests on macOS for every pull request #346), where the release
  builds; advisory for now.
• The README's status describes the sandbox, and EVIDENCE lists what
  concurrent tool calls have yet to show (Docs: the README's status names the sandbox, and EVIDENCE says concurrent tool calls are built #360); Sonny's README describes
  T3's model and effort pickers (sonny README: T3 Code's pickers as #363 left them #366).

🤖 Generated with Claude Code

https://claude.ai/code/session_013jkdGEK7HMibaxgHFDtGXA

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-29T09:54:31.444531Z	ffcea95	New commits
🔒 Security Review	✅ Completed 2026-10-29T09:55:09.026910Z	ffcea95	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.