Repository navigation
Releases
v0.15.0
Minor Changes
d071a83Thanks @TheBinaryGuy! - Choose the model the daemon's runs use: for the daemon as a whole (model), per agent
(agents.<name>.model) and per channel or signal (connections.<name>.model), each with an
optional thinking level (anthropic/claude-opus-5-5:medium). Send/modelon a channel to pick
one with buttons, or/model opus medium; in the TUI, ask and the agent uses the newmodelsand
set_modeltools. Changes apply from the next run, without a restart.
Patch Changes
d071a83Thanks @TheBinaryGuy! - The daemon's log no longer shows routine news as errors: a paired account, MCP servers that
connected and the pi extensions it loaded are info, and only failures are errors. A task you hand
the daemon from the terminal (tell_agent,/sedge tell,/sedge bg) can message you on your
channel when you ask it to, instead of refusing because you started it in the terminal.
What's Changed
- chore: release by @github-actions[bot] in #22
Full Changelog: v0.14.0...v0.15.0
v0.14.0
Minor Changes
-
f98029eThanks @TheBinaryGuy! - Security fixes from adversarial testing:- The sandbox now hides all of sedge's own files (database, conversations, logs, packages, not just
the config and secrets) and blocks connecting to its socket, through which a sandboxed command
could ask the daemon to run something outside the sandbox. Commands in it also no longer get
environment variables named like credentials. - sedge's own files are off limits to file tools by resolved path (
../config.jsonfrom the
workspace, links), not only by their full path, andsend_messagewon't attach them. - A run woken by your answer to an approval, or by a pick of a button, keeps the stricter rules
when the run that asked was woken by outside content. Buttons the agent makes can no longer pass
for sedge's approve and deny ones. - Approval messages show the whole call: a command in full, a write's content, an edit's changes,
whether the run was woken by outside content, and say so when it's cut. - Runs woken by outside content may write in their own folder and the workspace again under your
usualpermissions.write; only writes elsewhere ask, so a signal's agent can do its work without
asking at every file. ~paths (~/.sedge/config.json) are resolved to the home folder in every check, as the tools
read them;findandgrepskip sedge's own files in every folder they search; credential
variables are taken out when a sandboxed command starts; runs woken by outside content can't set
named values or write into a project's.pifolder without asking.
- The sandbox now hides all of sedge's own files (database, conversations, logs, packages, not just
What's Changed
- chore: release by @github-actions[bot] in #21
Full Changelog: v0.13.0...v0.14.0
v0.13.0
Minor Changes
a490824Thanks @TheBinaryGuy! - pi extensions in the daemon: list them inpiExtensionsin config.json, and their tools, their
tool_callchecks and their model providers work in runs you're not watching. Nothing loads
unless listed. What needs a person doesn't carry over: dialogs answer no, so a check that asks
before something risky blocks it.
What's Changed
- chore: release by @github-actions[bot] in #20
Full Changelog: v0.12.0...v0.13.0
v0.12.0
Minor Changes
8ee26a1Thanks @TheBinaryGuy! - Named values: small facts likemode = awaythat signals' conditions read asstate.<name>
(when: state.mode = away). A signal withrememberkeeps its latest value under a name instead
of waking the agent, and the daemon's agent sets them withset_state("I'm away till Sunday").
They survive restarts; the TUI'sstatetool lists them.
What's Changed
- chore: release by @github-actions[bot] in #19
Full Changelog: v0.11.0...v0.12.0
v0.11.0
Minor Changes
-
202e7f0Thanks @TheBinaryGuy! - Runs woken by outside content get stricter permissions: a signal's events, and what a run woken by
one passes on (a message to another agent, a subagent, a timer), use the stricter ofpermissions
and the newpermissions.signals, which asks for everything by default. This changes one default:
such runs now ask before writing files. Loosen it with"permissions": { "signals": { "write": "allow" } }. Your own messages and terminal tasks keep the usual rules.Signals can also have
labels: the cheap triage model tags each event with one ("person",
"nothing"), the agent andsedge dashsee it, andwakeOnwakes the agent only for the labels you
choose.
What's Changed
Full Changelog: v0.10.1...v0.11.0
v0.10.1
Patch Changes
-
c3c53a2Thanks @TheBinaryGuy! -sedge install <name>scopes only the official names (telegram, slack, github, web, local, mqtt)
to@sedge-hq; any other name installs as written. Filteringsedge dashby agent includes its
subagents' events and runs. The sandbox also hides MCP sign-ins,~/.npmrcand git credentials,
and the pi logins inSEDGE_AGENT_DIRwhen it's set. The TUI's agent knows about limits, the
sandbox and subagents. -
c3c53a2Thanks @TheBinaryGuy! -sedge dashshows today's tokens and cost in its header, against your daily limits when set. The
platform packages on npm get a README that points to@sedge-hq/sedge.
What's Changed
- chore: release by @github-actions[bot] in #16
Full Changelog: v0.10.0...v0.10.1
v0.10.0
Minor Changes
0159e37Thanks @TheBinaryGuy! - An optional sandbox for the daemon's agent, off unlesssandbox.enabledis set in config.json.
Commands run in the operating system's sandbox (Seatbelt on macOS, bubblewrap on Linux) and file
tools keep to the same rules: writes stay in the run's folder, the workspace, temp, package caches
and the folders you list; credentials (~/.ssh, cloud logins, pi logins, sedge's secrets) and the
paths you list can't be read; the network can be turned off. Where no sandbox is available,
commands don't run while it's on.
Patch Changes
0159e37Thanks @TheBinaryGuy! -sedge dash runsshows the tokens each run used, and their cost at the model's listed price.
New docs pages: Limits and costs, and Sandbox; Extensions, MCP and skills now says which pi
extensions load where.
What's Changed
- chore: release by @github-actions[bot] in #15
Full Changelog: v0.9.0...v0.10.0
v0.9.0
Minor Changes
38174d1Thanks @TheBinaryGuy! - Optional limits on what the agents use, all off unless you set them inlimitsin config.json:
runs per hour for each agent, tokens and money per day, and subagents at once. When one is
reached, events from signals, timers and agents wait in the inbox (nothing is dropped), your own
messages still get answers, and sedge tells you once. A new limit applies without a restart.
What's Changed
- chore: release by @github-actions[bot] in #14
Full Changelog: v0.8.0...v0.9.0
v0.8.0
Minor Changes
41c898aThanks @TheBinaryGuy! - Signals get three more ways to decide what reaches the agent.whenlets through only events
whose details meet a condition, liketemperature > 30 and room = kitchen.throttlestays quiet
for a while after an event gets through.triagehas a cheap model read each event with your
instruction first, and the agent wakes only when it should; skipped events show in the inbox with
the reason.triageModelin the config picks the model; the cheapest of your provider's is used
otherwise.
What's Changed
Full Changelog: v0.7.1...v0.8.0
v0.7.1
Patch Changes
6578312Thanks @TheBinaryGuy! -mqttis now one of the official packages sedge offers to install, for sensors, home automation
and devices that publish to an MQTT broker.
What's Changed
- chore: release by @github-actions[bot] in #11
Full Changelog: v0.7.0...v0.7.1