Repository navigation
Releases
v1.4.2
v1.4.2 — Chat pings and a neon chat frame (2026-11-03)
- The room, the inbox and a DM thread wear one frame. A title row with clickable buttons
(Room,Inbox ●N,DM ▸,@ pings on|off,← Back), heavy accent rules above and below,
a colour chip per person (six neon tones by a hash of the name, the same everywhere; you amber,
the model violet),@namedrawn in that person's colour, a bar on any line that names you,
and dividers where a day or ten quiet minutes passed. Inbox rows carry the unread mark, the
chip, presence, the time,N newand the last line; an opened thread bars the lines that were
unread. The bottom line counts mentions since you last looked (@ (2)), and the compact
layout now shows the unread-DM count too. Compact terminals keep the plain footer and
one-line inbox rows. - Chat pings. A person's room line that names you with
@nameprints a dim
User replied: <who> (chat)in your transcript, on your terminals only, so someone watching
the agent work knows to look; the rest of the room, and the model's own replies, print
nothing. A DM to you while you are attached printsNew DM (<who>)./chat pings on|offis
your own switch for the mention lines, saved in the host's config aschat.pings; the room's
footer sayspings offwhile it is. - A setting changed from inside a session —
/theme,/chat pings,/update check on|off,
/coworkers add|remove,/online add, and the standingmodel_reloadconsent, in both UIs — is written by
loading the config file afresh and saving that, so the session's own copy, which carries the
akey's "always run shell commands" and "approve every write" for the session under the
same keys, is never written to disk by it. Any of those used to make a session'sa
permanent. - A job's own compaction summary — a model call that streams nothing back — no longer counts as
silence for the stall clock: it is a wait like a tool call, shown aswaiting (compaction),
on the run-time meter, until it has run for twicestall_after(a hung summary call is
still reported). A long chore used to earn one spurious stall line. - A second session on the same workspace no longer closes at startup a job the first session
is still running: each session's work record (work.enabled) now lists its running jobs,
and an open job another live session reports is left as it is with a notice naming that
session. A job nobody reports is closedblocked: interruptedas before — which includes a
job run by a session still on v1.4.1 or earlier, whose record carries no job list. - When the main model's next request waits behind a co-worker's request on the shared server,
the status line says so after two seconds (waiting for the shared server: big's request is in flight; the request goes out when it ends) until the lane is free, instead of pausing in
silence. - The VS Code extension is in the repository. The packaged
be-code-vscode-<version>.vsix
is committed underextensions/(refreshed by thevscodebuild target) and attached to
every release from v1.4.1 on; the releases before it shipped without the file.
v1.4.1
v1.4.1 — Jobs for any co-worker (2026-11-03)
- Jobs. The main model hands a scoped job to any configured co-worker with one tool call,
dispatch(who,job,scope, optionalcontextandwait); nosub_agent: trueis
needed. The job runs the same scratch agent, consent, scoped registry andask_maina
sub-agent gets, and its step is filed under the active task, rendered asjobby/task.
scopeis required: an array, a bare path or a comma list;"."is the whole workspace,
accepted only while nothing else runs. A scope that overlaps a running job or an assigned
step, or falls outside the co-worker'smax_scope, is refused; one co-worker may run several
jobs, all counted againstsub_agents.max_concurrent. The hand-back arrives as a later
message or, withwait: true, as the tool's own result. - The main model steers its jobs.
jobs status|wait|reply|extend|stop: wait for one job
or for any, answer a pending question, clear a stall, end it. A job's question, hand-back and
stall line reach the main model asjob <id> …lines through the same queue a mid-task
message uses, so it reacts on its next turn; a headlessbe-code runkeeps taking turns on
them (up to ten) instead of ending with a job still out. - A stalled job wakes the main model, not the operator. A supervisor watches every job:
aftersub_agents.stall_afterseconds (180) with no tool call or output the job is marked
stalledand the main model is told —job 2.3 (big) has been quiet for 3m: 4 tool calls, last read_file 3m ago. jobs extend 2.3 to give it more time, jobs stop 2.3 to end it, jobs reply 2.3 <text> to steer it, or jobs status 2.3 to look.— and decides. The clock pauses
while the job waits on a question, the shared server lane, a consent prompt, a write approval
or editor review, or a tool call in flight. A third stall stops the job (blocked: stalled)
and the operator is told in the transcript; a headless run stops at the second.
sub_agents.max_runtime(1800) caps running time, waits excluded (blocked: ran out of time); 0 disables either. /agentslists every job with its state (running,waiting (tool: shell),asking,
stalled,done,blocked,stopped). A job's step closesblocked: session endedat exit
andblocked: interruptedat the next start after a crash;/task assignrefuses a job id;
a scheduled event's turn refusesdispatchand the mutatingjobsactions.- The e2e suite gains a
dispatchscenario: a job handed to a scripted co-worker, its
hand-back, and a stall the main model clears withjobs extend. - Anthropic and Gemini work as online main models. Anthropic's current models refuse any
temperature, so none is sent toapi.anthropic.com; Gemini 3 returns a thought signature on every tool call (extra_content) and demands it back on later requests, which the client now carries, stores with the session and echoes; and a request error after a native tool call has already worked no longer downgrades the session to embedded tool calls. The Gemini preset carries its 1,048,576-token window.
v1.4.0 — Packs closed, apps served, keys stored
v1.4.0 — Packs closed, apps served, keys stored (2026-11-02)
-
A pack is checked against the BE Skill Pack Format 1.0 contract when it loads.
pack.json
is a closed document now: every key must be one the format names,descriptionand
core.budget_tokens(100–2000) are required, every chunk needstriggers, a verify check
needs aparser, and a timeout is written30s,5mor2h. A violation names the field
(pack.json: chunk "errors" is missing required field "triggers",unknown key "extend" (use "extends" instead)) andbe-code packs validate <dir>reports it where a pack author looks.
detect.allis gone: any glob indetect.anymatching is enough. The nine packs in the
publicbe-skill-packsrepository conform.If you installed a pack under v1.3.0, it may not conform — it is dropped at session start
withpack in … was not loaded: pack.json: …among the startup notices, its rules leave the
prompt and its checks stop running.be-code packs update <name>fetches the current version,
which does. A pack you wrote yourself: runpacks validateon it and fix what it names. -
The loader accepts the schema's own shapes.
requires.parsers,hooks,packsandtools
are arrays, andmcp_serversmay be the schema's array of{name, command}as well as the
name-keyed object the first packs used; both were refused as "not valid JSON" by the typed
parse before the validator ran. The format'sconformance/minimal-packnow gets as far as its
parser: the universaljsonl-diagparser and{shell}substitution the standard also asks for
are not in this release. -
An unapproved project pack no longer takes your own pack of the same name out of force. A
pack committed to a repository under the name of one in~/.be-code/packsused to replace it
at discovery and then sit unapproved, so the repository had switched off your rules and checks
without a prompt being answered. The global pack stays active until the project pack is
approved (the standard's REQ-SPF-081);/packssays so. -
A pack
packs updatemoved aside and could not remove (<name>.replacing-<stamp>, a file held
open on Windows) was discovered as a pack and, loading last, took the updated pack's name. It
is skipped now. -
The trust hash covers a file a check names anywhere in an argv word —
sh -c "sh {pack}/chunks/run.md"— as it already did for the whole-word form, so rewriting that chunk
re-asks. -
Validator messages: an error inside the fourth check says
verify[3]; the 200-character
limits ondescriptionand chunk summaries count characters, not bytes; which chunk is
reported first no longer depends on map order. -
Two fixes from kmesiab (#1,
#2), which v1.3.0's notes credited
but which had not been merged: the compat (embedded) tool-calling instructions document the
<tool_result>format a result comes back in, and an empty model reply is nudged up to three
times, the count resetting on any productive turn, instead of ending the run after one nudge.
The final error names both things it can mean — a model that believes it is done, or a backend
that returned nothing for a prompt larger than its window. -
The Linux and macOS installer verifies what it downloads.
install.shchecks the release
binary against the release'sSHA256SUMSthe wayinstall.ps1and the in-app updater already
did, withsha256sum,shasumoropenssl, whichever the system has; a download that cannot
be verified — no listing, no line for the asset, a mismatch — is never installed. -
An application API. Other programs on the machine can ask a running session for an
answer over~/.be-code/api/api.sock: a plainquery(no tools, no history, optionally with
an active pack's rules and ajson_schemathe reply is checked against), OpenAI-compatible
chat/completionsthat existing clients reach with a URL change, and a query answered with
a session's conversation as context, addressed by a token the session mints and can rotate.
Callers are registered apps (be-code apps add, a key shown once, a profile of tiers, models,
packs and budgets); browsers are refused; one session serves per machine; headless runs never
serve./api,/appsand/session token|rotatein both UIs;api.enabledin config. -
Add an online provider from inside the harness.
/online add(andbe-code online add, andbe-code setup) takes a pasted key or, for OpenRouter, a browser sign-in, validates it against the provider's listing, and stores it in~/.be-code/credentials.json(0600). The list opens with a quick sign-in: OpenRouter signs in through the browser, and OpenAI, Anthropic and Gemini open their own API-keys page in the browser before the masked field (no provider issues a key to a third-party app through a login, so the page is the honest shortcut). An environment variable still wins./online revoke,doctorlines, and a/online addhint on a rejected key. Anthropic keys that are not scoped to a workspace are asked for their workspace ID once (providers.<name>.headers), and a rejected key shows the provider's own message. -
Co-workers from inside a session.
/coworkers add(and Add a co-worker in/menu) picks a configured provider, a model from its live listing and a name, saves the entry and registersconsultat once;/coworkers remove <name>is the inverse. The sub-agent role still takes effect at the next launch. -
scripts/fix-provider-config.pyrepairs~/.be-code/config.json, the file be-code loads; it
no longer honours aBE_CODE_CONFIGvariable be-code itself never read. -
uninstall.ps1points the-Purgehint at.\uninstall.cmd -Purge, the launcher a default
Windows policy can run.
v1.3.0 — Packs, online models, a browser, schedules
v1.3.0 — Packs, online models, a browser, schedules
Highlights:
- Skill packs can be installed into a project (
be-code packs add <name>,/menu,/packs add). - Cross-session workspace awareness with conflict detection on writes and recoverable pre-overwrite backups.
- Sub-agent hand-back names what changed under it and where its starting tree lives.
- Structured verification output — name a parser (
generic,gcc,msvc,ubt) so the repair turn gets the root cause, not the log tail. - Online models, a browser tool, schedules, and sub-agents are all first-class.
Full changelog: https://github.com/BE-AI-Research/be-code-redux/blob/main/CHANGELOG.md (v1.3.0)
Install
curl -fsSL https://raw.githubusercontent.com/BE-AI-Research/be-code-redux/main/CHANGELOG.md # see README for install
Verify the checksum of your download against SHA256SUMS.txt.
v1.2.5 — The Windows interface, restored
v1.2.5 — The Windows interface, restored
A hotfix for Windows. If you are on v1.2.4, update.
What v1.2.4 broke
v1.2.4 reported a Windows terminal as UTF-8 capable by asking the console — which answers yes,
because BE-Code had just put it into UTF-8. That switched every Windows terminal from the ASCII
glyph set it had always used to real box drawing, arrows and the context wheel: the characters a
Windows console is least likely to draw at the width the layout measured. A row drawn wider than
it was measured wraps, which makes the frame taller than the terminal, which scrolls the alt
screen — so the top bar disappeared and the input and status rows kept stale copies of
themselves.
Capability now comes from the locale again, exactly as it did before v1.2.4, so a Windows
terminal is drawn with ASCII glyphs. The console's code page is still set to UTF-8, so UTF-8 in
a model's reply or a tool's output decodes correctly; what is no longer claimed is that the
terminal's glyph widths can be trusted.
Also fixed
- The ghost rows, the original bug v1.2.4 was meant to cure and did not. Every row of a frame
is padded to exactly the terminal width, and Bubble Tea adds no erase to a row that already
fills the width — it expects the write itself to cover the old cells. With the console's
wrap-at-end-of-line enabled, writing that last cell moved the cursor to the next row, which on
the bottom row scrolled the alt screen. A client now clearsENABLE_WRAP_AT_EOL_OUTPUTfor the
length of an attach. (v1.2.4 setDISABLE_NEWLINE_AUTO_RETURN, which governs line feeds rather
than the wrap, and is now left alone: it staircases the plain-text closing lines the host writes
through the same console, the resume code among them.) - The one-line install from cmd.exe, which installed nothing and said nothing.
install.cmd
chooses between theinstall.ps1beside it and the published installer on GitHub, but that line
ended… & exit /b— and cmd.exe runs what follows&whether theifmatched or not, so
downloaded on its own it skipped the local branch and then exited anyway. The PowerShell
one-liner was never affected.
Nothing changes on Linux or macOS.
The full list is in CHANGELOG.md.
Install
curl -fsSL https://raw.githubusercontent.com/BE-AI-Research/be-code-redux/main/install.sh | shOn Windows, in PowerShell:
irm https://raw.githubusercontent.com/BE-AI-Research/be-code-redux/main/install.ps1 | iexYou can also update with /update inside BE-Code. Binaries are provided for linux/amd64,
linux/arm64, darwin/arm64 and windows/amd64, along with the VS Code extension
(be-code-vscode-1.1.1.vsix). Checksums are in SHA256SUMS.
v1.2.4 — A readable terminal on Windows
v1.2.4 — A readable terminal on Windows
A hotfix for Windows shared sessions.
The symptom
In a shared session on Windows, a terminal's text slowly duplicated and became hard to read as
the session ran, while the layout itself held together.
The cause
Nothing had taken ownership of the client's console. A shared session renders in the detached
session host, which has no terminal of its own, so the Windows console setup Bubble Tea performs
when it owns a terminal never ran — and raw mode covers only the input handle. The console
therefore kept its OEM code page, which draws every non-ASCII byte of a frame as two or three
cells. Lines outgrew the width they had been padded to, the console wrapped them, the alt screen
scrolled, and every repaint after that landed one row out of step.
The fix
- A client now takes its console's output for the length of an attach — VT processing on, no wrap
at the last cell, output in UTF-8 — and puts back exactly what it changed on the way out. - A Windows terminal is no longer reported as non-UTF-8. It was being asked for the POSIX locale
variables Windows does not set, so every Windows client was drawn with ASCII glyphs.
Nothing changes on Linux or macOS: a POSIX terminal needs no such setup, and its character set
still comes from the locale.
The full list is in CHANGELOG.md.
Install
curl -fsSL https://raw.githubusercontent.com/BE-AI-Research/be-code-redux/main/install.sh | shOn Windows, in PowerShell:
irm https://raw.githubusercontent.com/BE-AI-Research/be-code-redux/main/install.ps1 | iexYou can also update with /update inside BE-Code. Binaries are provided for linux/amd64,
linux/arm64, darwin/arm64 and windows/amd64, along with the VS Code extension
(be-code-vscode-1.1.1.vsix). Checksums are in SHA256SUMS.
v1.2.3 — A cut-off tool call is never the answer
v1.2.3 — A cut-off tool call is never the answer
The fix
On a long run, a local model writing a whole source file into one write_file call could hit
the output limit mid-call. The unfinished call parsed as no call at all, so the raw fragment —
pages of half-escaped JSON — was shown as the model's final answer and the run ended there.
Now:
- A reply the output limit cuts off inside an embedded tool call is discarded, never shown as
the answer. - The model is told once to re-issue the call in smaller steps (write the file in parts:
write_filethe first part, then extend it withedit_file). - If the retry is cut off too, the run ends with a clear error naming the output limit —
raisemax_tokensor the backend window, or ask for the change in smaller steps.
Both the <tool_call> JSON form and Qwen's <function= XML layout are recognized.
The full list is in CHANGELOG.md.
Install
curl -fsSL https://raw.githubusercontent.com/BE-AI-Research/be-code-redux/main/install.sh | shYou can also update with /update inside BE-Code. Binaries are provided for linux/amd64,
linux/arm64, darwin/arm64 and windows/amd64, along with the VS Code extension
(be-code-vscode-1.1.1.vsix). Checksums are in SHA256SUMS.
v1.2.2 — One session on Windows, and a one-line install
v1.2.2 — One session on Windows, and a one-line install
One session on Windows, whatever terminal you're in
- Joining. A second
be-codeon the same project — say Visual Studio's terminal and an SSH
login — now joins the live session instead of starting its own on a copy. The two shells can
spell the folder differently (case, a trailing separator, an 8.3 short name); workspaces now
match the way Windows does. - Surviving the SSH login. A shared session started from an SSH login used to die when the
login ended, because Windows' SSH server ends every process its session started. The session
host now breaks away from that job; where the job forbids it, the launch says the session will
end when you disconnect. - Elevated and non-elevated terminals. A live session started from an administrator's login
looked dead to a non-elevated terminal, so its record was deleted and a second host started on
a copy. The aliveness check now uses the limited query Windows allows across integrity levels
and treats "access denied" as running — as Unix already did.
Windows one-line install
- PowerShell:
irm https://raw.githubusercontent.com/BE-AI-Research/be-code-redux/main/install.ps1 | iex
- cmd.exe:
Both download the current release's binary and check it against the release's
curl -fsSLo "%TEMP%\be-code-install.cmd" https://raw.githubusercontent.com/BE-AI-Research/be-code-redux/main/install.cmd && "%TEMP%\be-code-install.cmd"
SHA256SUMS
before installing (BE_CODE_VERSIONpins a release,BE_CODE_NO_SETUP=1skips the wizard).
When/updatecannot write the install folder on Windows, it now names the PowerShell
installer instead ofcurl … | sh.
The full list is in CHANGELOG.md.
Install
curl -fsSL https://raw.githubusercontent.com/BE-AI-Research/be-code-redux/main/install.sh | shYou can also update with /update inside BE-Code. Binaries are provided for linux/amd64,
linux/arm64, darwin/arm64 and windows/amd64, along with the VS Code extension
(be-code-vscode-1.1.1.vsix). Checksums are in SHA256SUMS.
v1.2.1 — Updates from inside the app, and long sessions that stay alive
v1.2.1 — Updates from inside the app, and long sessions that stay alive
Update from inside BE-Code
-
The notice. Each session start checks GitHub once, quietly in the background. If a newer release exists,
⬆ vX.Y.Z availableappears on the bottom line. Nothing installs on its own. -
Updating. Use
/update, or Check for updates in/menu. BE-Code checks again and asks first; there's no "always" option and-ynever answers it. Then it:- downloads the release's binary for your platform;
- verifies it against the release's
SHA256SUMS; - replaces the installed
be-code.
Only the binary changes. Your
config.jsonand sessions are never touched. Restart BE-Code to use the new version. -
Fully offline. Update check: off in
/menu(or/update check off) stops the start-of-session check. BE-Code then makes no network call at start and is fully offline with a local backend./updatestill works whenever you choose to use it.
Long sessions that stay alive
- Large tool results. One very large tool result could end a run. For example, a
task showof a big task tree came back at around 75k tokens in a 32k window. Every tool result is now held to the per-call output cap. - Resumed sessions. A session saved before this fix opened at about 300% of its window. Its oversized results are now cut to the same cap when it opens.
- Clearer error. When a request still doesn't fit, the error now tells the two cases apart:
- a model loaded smaller than you configured: reload it;
- otherwise the conversation is too large:
/compactor/clear.
Also
- The TUI header now shows the version taken from the build, so it's always current.
The full list is in CHANGELOG.md.
Install
curl -fsSL https://raw.githubusercontent.com/BE-AI-Research/be-code-redux/main/install.sh | shFrom v1.2.1 onwards you can also update with /update inside BE-Code. Binaries are provided for linux/amd64, linux/arm64, darwin/arm64 and windows/amd64, along with the VS Code extension (be-code-vscode-1.1.1.vsix). Checksums are in SHA256SUMS.
v1.2 Update: The beaver has landed.
v1.2 Update: The beaver has landed. 🦫
BE-Code stays offline-first, and 1.2 opens it up to people without a local GPU. Your main model can now be an online one, while a small local model handles the housekeeping. Nothing leaves your machine until you have said yes to it.
Online main models
- Presets.
be-code setupnow offers "use an online provider", with presets for OpenRouter (the headline), OpenAI, Groq, DeepSeek, Mistral, Gemini and Anthropic. It also works with no local server at all. - Keys. API keys are only read from environment variables and are never written to config.
- Model details. Each model's context window and prices come from the provider's model list.
- Rejected keys and rate limits. A rejected key is reported clearly by name and never retried. Rate limits honour
Retry-After. - Local helper.
local_helperruns compaction summaries, the resume briefing,/init, commit messages and review on your local model. These chores stay local and free. - Outages. If the online provider goes down, BE-Code offers to switch the session to your local model.
Consent before anything leaves
- Per project. Each project is asked once before its code goes to an online model. The answer is remembered per project and provider, and
/online forgetwithdraws it.-yapproves for that run only. - Per site. Page text from the browser,
web_fetchandweb_searchreaches an online model only after you answer for that site.- Your own Chrome asks on every page.
- Pages already in the conversation are asked about site by site when a session goes online.
- No "always".
online_project,share_page,spend_capandswitch_to_localhave no "always" option, and-ynever answers them, exceptonline_projectfor a single run.
Spend
- Estimated spend shows in the status line and
/stats. max_spend_usdasks before the request that would go past it.
Also in 1.2
- Scheduled events. A running session can wake the model later: one-off follow-ups, or recurring upkeep under an allowance you approve up front. A fired event never takes the session's shortcuts. Tools with no gate of their own ask first.
- Your own Chrome. The browser tool can attach to the Chrome you already use (Chrome 144+, remote debugging on). It works only in its own tabs or ones you hand over, and every action outside your allow list asks.
- Budget fix. A
max_tokensas large as the context window no longer swallows the whole budget, which showed up as the 999% context wheel and the apparent freeze.
The full list is in CHANGELOG.md.
Install
curl -fsSL https://raw.githubusercontent.com/BE-AI-Research/be-code-redux/main/install.sh | shBinaries are provided for linux/amd64, linux/arm64, darwin/arm64 and windows/amd64, along with the VS Code extension (be-code-vscode-1.1.1.vsix). Checksums are in SHA256SUMS.