Cooler brand palette, compact CST-agent home with tagline and framed
composer, session topbar (model + chair chip), sidebar off by default.
Screenshot pipeline in tools/tui_shot.py; docs/screenshots for before/after.
Co-authored-by: Cursor <cursoragent@cursor.com>
Run inbound peer handling on its own thread so outbound dials cannot starve
:9100. Failed peers cool down exponentially. Also land --join (member seat)
and realm/peers.json seed discovery for Tailscale meshes.
Co-authored-by: Cursor <cursoragent@cursor.com>
Peer dial used per-step read timeouts so a stuck handshake could block the
single-threaded daemon for tens of seconds; inbound hellos piled up unread
(CLOSE-WAIT). Use a wall-clock dial budget, drain accepts before each dial,
skip self, and only one side initiates idle polls.
Co-authored-by: Cursor <cursoragent@cursor.com>
Seats publish via parliament_note, sync pushes on write, member join without
stealing the chair, and worker tools are hard-gated when parliament is active.
Co-authored-by: Cursor <cursoragent@cursor.com>
CLI/Docker assert AEAD sync; status shows sealed channel; headless denials no longer pretend the user refused.
Co-authored-by: Cursor <cursoragent@cursor.com>
Pidfile records instance+port; login restarts the service when the seat
changes. TUI /realm status|login|stop; prompts discourage filler chatter.
Co-authored-by: Cursor <cursoragent@cursor.com>
Prompt for Realm password when CST_REALM_PASSWORD is unset; map worker
request-gate exhaustion to a clear budget error with partial progress;
hard-skip identical tool failures after 3 tries. Add docs/parliament.md.
Co-authored-by: Cursor <cursoragent@cursor.com>
Inject shared parliament context into the agent environment, switch
prompts away from in-process workers when parliament is active, clear
login tickets on chair sync/new-project, and surface realm/chair/service
in /status. Tighten empty-reply handling and prefer-tool principles.
Co-authored-by: Cursor <cursoragent@cursor.com>
Ignore/clean host objects, add linux-headers, and include sys/time.h
so discover builds under musl. Docker chair↔member handshake now passes.
Co-authored-by: Cursor <cursoragent@cursor.com>
Login claims the chair, writes a session ticket, and starts a background
service for discovery plus revision-based context sync. Password Realm
opens require --login (or CST_REALM_PASSWORD); --service/--service-stop
manage the resident agent.
Co-authored-by: Cursor <cursoragent@cursor.com>
Derive Realm keys from a user password (Argon2id, no private key on disk).
Login claims the sole chair with handover; peers share parliament context;
--new-project clears it. CLI peer/discover helpers and Docker alpine recipe
included; headless empty-answer retries raised for unattended runs.
Co-authored-by: Cursor <cursoragent@cursor.com>
Exchange enrolled certificates and Ed25519 challenge signatures so
same-Realm instances can authenticate each other after discovery.
Co-authored-by: Cursor <cursoragent@cursor.com>
Advertise instances with a multicast beacon and a config-dir registry so
same-machine and LAN peers that share a Realm fingerprint can find each
other without TOFU pairing.
Co-authored-by: Cursor <cursoragent@cursor.com>
Introduce a user Realm keypair (Monocypher) that signs instance
certificates so same-domain peers can later auto-discover and connect.
Co-authored-by: Cursor <cursoragent@cursor.com>
Pending earlier-revision jobs no longer skip completion repair; late
cancels preserve real exits; begin clears pause; partial undo keeps
snapshots; permissions and worktree failure paths are hardened.
Co-authored-by: Cursor <cursoragent@cursor.com>
Building the Windows TUI by hand -- the first local build since the CI file
went away -- stopped in src/tools/session_tools.c: `getline` was an implicit
declaration there, which this GCC treats as an error. src/session.c had a
private fgetc() shim for exactly that; the session-list scan is newer and used
the POSIX call directly, and only the Windows build notices.
The shim moves to os_compat as os_getline(): getline(3) itself on Unix, the
fgetc() loop on Windows, one prototype in the header so a caller reads the same
everywhere. src/session.c drops its copy and calls it; the scan does too (its
`ssize_t` becomes `long long`, which Windows has).
make check is green on Linux, the Windows cross build now finishes and
packages (dist/cst-agent-0.9.7-windows-x86_64.zip plus the NSIS setup), and the
static Linux tarball was rebuilt with it.
The release files are built on this machine from now on, so the workflow that
did the same steps on a runner -- TinyCC build and unit tests, the static musl
Linux tarballs, the mingw Windows zip, the Gitea release upload -- is gone.
.gitea/ held nothing else.
Everything the workflow called stays and is runnable by hand exactly as it ran
there: scripts/ci_linux_toolchain.sh is what package_linux.sh uses to pick its
musl toolchain, scripts/ci_tcc.sh builds the second compiler the test job used,
scripts/ci_windows_cross.sh and scripts/bootstrap_windows.sh are the mingw
build, and scripts/release.sh is the upload (GITEA_TOKEN=... scripts/release.sh
FILE...).
A phone's back key did not close the settings dialog. A window-level QML
Shortcut never sees the key while a modal popup is up, and a desktop run says
so plainly: with the dialog open, Back left backLayer() at "settings", while
invokeMethod(win, "stepBack") closed it on the spot and Escape -- the popup's
own CloseOnEscape -- closed it too. So the key reaches the popup and never
reaches the shortcut map, which is why the drawer (no popup) closed fine and
hid the problem on a phone.
A Keys handler on the window's root item is no help either: nothing reached
it, the focused TextField keeps the key. So the key is read in C++, on the
application object but only for this window's events -- a filter installed on
the window itself was never called once, while the same key closed the drawer
through the shortcut, hence the application object.
With a layer open, the window's own backLayer() names it and stepBack() closes
it, press and release both taken; with nothing open the key is passed on
untouched, which is what leaves the window on a phone and does nothing on a
desktop. Main.qml keeps only those two functions: the Shortcut and
backHasLayer are gone, --uitest installs the same handler main() does, and the
back checks now drive the code the phone runs instead of a shortcut only the
test had.
The last check asks the phone's own question and can end the run without the
closing line -- the activity leaving takes the process with it, and every line
is flushed as it is written.
Desktop: 106 ok, 0 FAIL. Phone: 104 ok, 0 FAIL (two clipboard checks are
skipped on Android), and "after the last back: window visible 0".
Open and Share failed on the phone for a reason that was in the harness,
not in the app: the card was where isVisible() says yes, but the transcript
clips it away -- the run's own note says it found contentY at -14 with the
card's row below the visible part -- so the click (and the tap after it)
landed on whatever was at that point instead, and the header stayed empty.
The run now stops the view following its tail (ChatView.stickBottom, which
the replay sets), scrolls the card's row into view with the numbers in a
note, and only then clicks. ChatView gets an objectName for that.
The checks also say what the card hands the engine (the name the store
keeps the file under, from the strip's own model) and what the card itself
answers -- which is what turned "it did nothing" into this answer:
note the card is out of the viewport (contentY -14, card at 433 of 615):
scrolling to it
ok Open hands the file to the system (opened report.md (test mode);
cards hold 'img-...png, art-....md', on screen, card says '')
ok Share names the file (shared report.md (test mode); card says '')
The whole device run is green now: 97 ok, 0 FAIL (it was 88/17 when this
session's work started, and every one of those was real).
The two clipboard-image checks failed on the phone for a reason that is not
a defect: Android's clipboard carries text and URIs, not images -- setImage
does nothing there -- so they are skipped on Android with a note saying so,
instead of reporting a phone that is broken because it is a phone.
Open and Share still fail on the device, and now say how far they got:
neither the synthesized mouse click nor a touch tap leaves a line on the
card's buttons, while the same click reaches the header and the sidebar.
The engine's own path is fine (openArtifact() called directly answers as it
should), and the only way for the header to stay empty is openArtifact()'s
early return for a name it cannot resolve -- so what is left to look at is
the name the card hands over (the model's `rel`) and whether the card is
inside the visible part of the scrolled transcript at that moment, since
QQuickItem::isVisible() does not account for clipping.
Device run after this: 95 ok, 2 FAIL (was 17 before this session's fixes).
A Controls ToolBar hands its children to a style-owned contentItem, and on
the NX789J that item sat 35 px below the bar itself: the header was at y 0
h 42 while its own gear was at y 56 -- drawn on top of the sidebar's first
lines, and clicked there too, so the settings dialog could not be opened
from the gear at all (eleven checks that follow it failed with it). A
desktop never showed it: same style (Basic), same numbers, gear at y 21.
A plain Rectangle has no such indirection -- its children are where it is.
Nothing else changed; the background and the border line are the same.
--uitest prints what it sees at that point now (style, window, header,
gear, safe area), which is how this was found:
note style Basic, window 1280x853, header y 0 h 42, gear centre y 21
Before: gear centre y 56, `FAIL the gear opens the settings dialog`, 17
failures on the device. After: 95 ok, 4 FAIL (Open/Share and the two
clipboard-image checks, which Android has no answer for).
The settings dialog check failed only on the phone, with no QML error and
nothing to read: this build's Qt messages never reach logcat (verified: a
plain qInfo from the app is not in the buffer). Two ways in, then:
- --uitest installs a message handler that writes Qt's own warnings into
uitest.log, next to the checks. What the phone did say, it says there.
- --uitest --shots saves window grabs at the interesting points into the
workspace. A device run cannot be handed CST_APP_SHOT (an app started
from adb gets no environment), so what a desktop run asks for with an
environment variable has to be an argument here. main.cpp takes it, so
`--uitest --workspace DIR --shots` works from adb.
What they said about the phone (NX789J, 1280x853):
ok ... the gear is reachable, centre 1191,56 of 1280x853
ok ... and openSettings() opens it when called directly
so the button is there and the handler works: the synthesized click is what
does not open the dialog. The run's own picture of that moment shows why --
the header is drawn over the sidebar, so the click lands on what is drawn
above the gear. On a desktop the same window has the header above the
sidebar and the click lands. Not fixed yet (nor the Open/Share checks on the
device, which are the other two failures that are not about the dialog).
The settings checks now go on either way, so a phone run reports the dialog
instead of eleven lines of "not there".
A tool that takes minutes -- a build, a wide search, a command nobody can
predict -- read as a window that had hung: the card said "running" and
nothing else, forever. The chip now counts from the point the card starts
running (from five seconds on, before that it is noise), which is where a
person looks when a call seems stuck; the status line and the Stop button
were already there. A card that appears already running (a replayed
transcript) starts its clock too.
--uitest: two checks for it (a card four minutes eighteen in says
"running · 4:18"; a finished one stops counting).
And the run now writes uitest.log in its workspace, every line and the
verdict. That is the only way to read a run *on a phone*: stdout reaches
nobody there, an app started from adb cannot be handed CST_APP_SHOT or any
other environment (so the shots a desktop run can take are out), and logcat
does not carry this build's Qt messages at all -- checked: a plain
`qInfo` from the app is not in the buffer.
First device run (NX789J, 105 checks): 88 ok, 17 FAIL, every failure
pre-existing and unrelated to this work -- the settings dialog does not
open from the gear (so nothing in it is reachable: the five dialog checks,
the two provider-form ones and the four clicks that follow), Open/Share
leave no status line on the device, and the clipboard-image checks have no
clipboard. All the md_to_pdf and card checks pass on the phone, including
"black ink on paper (darkest pixel 0)" from 4c69ef3.
The hook added in da04913 asked for getStringArrayListExtra, but `am
start --esa applicationArguments "--uitest,--workspace,/sdcard/x"`
stores a String[] -- that call answers null for an array, so the window
still started with no arguments at all. Ask for the array first, then
the list (Qt's own loader, putStringArrayListExtra), then the single
comma-separated string.
With this the uitest runs from adb on the NX789J: the fixtures and
selftest.pdf land in --workspace, which is how a device round trip is
possible without touch injection (adb cannot tap this window).
A QTextDocument takes its default text colour from the application
palette, and markdown names no colour at all: in a window running a dark
palette -- which is what a phone in dark mode hands Qt -- every plain
paragraph was laid out #eeeeee on a white page. Printed, that is a blank
document; on screen it reads as "the font is oddly faint". The first
PDFs off the phone were exactly that.
- PdfTools: paperDocument() sets the default stylesheet (body colour
#000000) before setHtml, and writePdf() lays that out. A colour the
document does name still wins, so an .html handed to the tool keeps its
own -- checked with a #cc0000 document under both palettes.
- --uitest: draws that document on a white image under a dark palette and
looks for ink (238 before the fix, 0 after).
Verified on the NX789J: md_to_pdf on the device now writes CID TrueType
NotoSansCJKHK-Regular/-Bold text whose page renders at min=0.
`am start -n <pkg>/org.qtproject.qt.android.bindings.QtActivity --esa
applicationArguments "--workspace,/sdcard/x"` left the window on its default
workspace on the NX789J (the plain-string extra too), so the extra is read
from the intent directly. That is what makes `cst-agent-app --uitest
--workspace DIR` reachable on a device with no keyboard -- and adb cannot
inject touch into this window either (tap, touchscreen tap, mouse tap and
motionevent all do nothing; key events do work), so it is the only way in.
Everything degrades to "no arguments": an invalid object, a missing method or
a context that is not an activity all come back empty.
needs_confirm() (src/agent_loop.c) decides whether a tool asks by reading
cfg->confirm_shell / confirm_write / confirm_net; the approval mode only
relaxes those. The CLI fills them in (main.c: parse_confirm), the TUI and the
WebUI do too -- the Qt window never did: AgentCore::open() memset the whole
config and cst_agent_create() keeps what it is given, so every tool ran
without asking while the sidebar said "ask". That is why nothing was ever
asked on the phone (or on a desktop).
- AgentCore: one mapping, applyApproval(), used by open() and setApproval();
the mode is remembered like the TUI's /approval and read back on start
(else "ask", the only mode where the window's consent dialog is reached).
- --uitest: two checks that the mode reaches the engine as those flags.
Version 0.9.7 (local).
The SAF picker is not 'still to come' any more (v0.9.6 did it): the list keeps
the back gesture and the icon glyphs. In its place, what v0.9.7 changed on the
phone -- a consent question refused in the background and bounded in the
foreground, the bounded GUI hop, the phase in the header, and the window's own
md_to_pdf.
The rewritten cst_agent_store_info() counted a transcript line that ran past a
whole 64 KiB read block a second time, when the newline ending its tail
arrived: a session with one very long line (a tool that read a file back)
reported one message too many -- exactly what the sidebar shows the person.
- src/agent_store.c: the tail's newline ends a line that was already counted.
- tests/test_store.c: test_info_scan() holds the scan down -- a 70 KiB line is
one line, the first user message comes from the head of the file, and a user
line past the window is not called the first one.
- tests/test_tools_ext.c: a byte budget stops the session_search walk, says so
in the result, and never reports a match it did not reach; the same query
without a budget finds it.
make check and make asan are green.
- consent: on_permission / on_ask_user wait at most two / five minutes, and in
the background (a phone off screen, a window that is gone) the answer is
refused at once with a reason -- a dialog nobody can see blocked the worker
for good, which is what "the app hangs while the agent works" looked like.
- finishTurn() no longer joins the thread it is called from; the event loop
reaps the Turn instead.
- AndroidTools: the hop to the GUI thread pings first and gives up after 3 s
instead of waiting forever on a busy GUI thread.
- the session list stopped reading every transcript: cst_agent_store_info()
counts newlines in 64 KiB blocks (first_user only within the first 256 KiB),
and AgentCore::sessions() reuses a row while (mtime, size) is unchanged.
- session_search gets a budget (max_bytes 32 MB, max_seconds 5 s), skips lines
over 64 KiB, and says when the budget stopped the walk.
- the header shows the phase ("shell: make -j8"), so a long turn no longer
looks like a stuck one.
- md_to_pdf: the front end's own Qt tool (owner app:qt, every platform):
markdown or a workspace .md/.html becomes an A4 PDF with 15 mm margins,
paginated by hand, written into the workspace and delivered as a file card.
No PrintSupport, so the phone gets it too.
- --uitest runs md_to_pdf through the registry with Chinese text and checks the
bytes are a PDF; app/CMakeLists.txt no longer drops src/web_md.c from the
standalone engine build, which the window needs to render markdown.
Version 0.9.7.
app/linux/cst-agent-app.{desktop,svg} plus one CMake component ("app") in
app/CMakeLists.txt: `cmake --install --component app` puts the binary in bin/,
the entry in share/applications/ and the icon in hicolor/scalable/apps/. The
entry's file name stays the app_id Qt sets (QGuiApplication::applicationName),
or the launcher draws a generic icon; only the front end installs, the CLI
keeps coming from scripts/bootstrap_linux.sh.
default.nix is that install for nix users: nixpkgs' Qt 6 (qtwayland for a
Wayland session, qtsvg for Breeze's SVG icons), wrapQtAppsHook so the plugins
and QML modules resolve outside a nix-shell and survive nix-collect-garbage,
a filtered src (third_party/ and the build directories are gigabytes this
build never reads), and the version read from include/cst_version.h. doCheck
runs ctest minus test_sandbox, whose Landlock assertion cannot hold inside the
nix build sandbox (the other 40 pass there; make check runs all 41).
Verified: nix-build, nix profile add --file ., then --selftest and --uitest in
a clean environment (no nix-shell, no XDG_* from the session).
Four tools and a document writer, plus the storage half of the phone: until
now the picker's answer (a content:// URI) reached the engine as a path, and
it said "cannot read" -- the app had no way to hand a phone's file over.
- md_to_html (new, src/tools/web.c + cst_web_md_document()): markdown in, one
self-contained HTML document out -- doctype, utf-8, viewport, escaped
title, a small stylesheet; no script, no external request. It is written
into the workspace and delivered as a file card (for_user, so it never
enters a request), the way deliver_file does. The default name comes from
the source file's stem or, for pasted markdown, from the title with what a
path cannot carry turned into '-'. It has a preview, so an approval prompt
says "render x -> y.html" before anything is written.
- download (src/tools/web.c): an http(s) URL into a file in the workspace,
streamed with a 128 MB ceiling; a redirect chain, the same CA handling and
the same refusals as web_fetch, but the bytes stay on disk (a half file is
removed on any failure, including HTTP >= 400). The name comes from the URL
when the caller names only a directory.
- read_pdf (src/tools/fs_tools.c): the text layer and page count of a PDF,
which read_file refuses because the bytes are binary -- and a PDF is
exactly what arrives from a phone. Same extractor an attachment uses, its
own 4 MiB input ceiling, 256 KiB of text by default with a note when cut.
- session_search (new, src/tools/session_tools.c): every stored session --
this one and all the ones before it -- searched for a substring, reporting
the session's id, message count and title, then the matching lines with
their number and role. Case-insensitive by default, capped at 20 matches.
- device_info (same file, perm_name "device"): version, system/arch/CPU
count, workspace, sandbox, model and permissions, config/temp/program
directories, pid -- so the model stops guessing where it runs.
- Tools 3 and 4 come with tests/test_tools_ext.c (also in both build files):
the document is a whole page, a script tag and a javascript: link do not
survive, a hostile title stays text and cannot write outside the workspace,
the preview names the file, read_pdf/download refuse what they must (no
network in a test), session_search finds a fixture session and says so when
nothing matches, device_info names the workspace. README's tool table grew
five rows.
- Android storage: the picker (SAF) answers with a content:// URI, and the
engine opens files by path, so the app now copies what it points at into
its own directory first -- app/android/src/xyz/whcself/cstagent/
AttachFile.java, called through a new app::materializeAttachment()
(ToolProvider.h: one API for every platform; a desktop strips file:// and
refuses content:// with a reason). The copy needs no permission at all: the
grant travels with the URI.
- "All files access" (MANAGE_EXTERNAL_STORAGE) is the other half: it is what
lets the engine read a path in shared storage that a task names rather than
picks. Android has no runtime dialog for it, so main.cpp opens the one
Settings screen once (AttachFile.java's requestAllFilesAccess, with the
all-apps screen as a fallback) and never insists: without it everything
works, one picked file at a time. On the NX789J the screen came up, the
app-op reads back "allow", and a relaunch no longer asks.
- Composer.qml hands a content:// URI through untouched (it is not a path,
and decoding it would corrupt it); AgentCore::attachFile() turns whatever
the picker gave into a path first.
Desktop: build clean, make check 36/36 (the new test included, also under
ASan), --uitest all green, --selftest and cst-app-selftest OK.
The APK builds and installs (86 MB, release-signed) with the new Java class
in its dex.
Version 0.9.6.
A session's transcript lived beside its attachments, both named after an id
that was a timestamp. Two things fell out of that: on a workspace two
machines share, a timestamp is not an id (the same second on both is enough
to hand out the same name), and "delete this conversation" had to chase the
jsonl, the title and a directory in a shared attachments/ tree, guessing what
belonged to whom. A phone's app data has nobody to tidy it by hand.
Now one session is one directory:
<workspace>/.agent/sessions/
<uuid>/session.jsonl the transcript
<uuid>/title the window title, one line
<uuid>/attachments/ images and delivered files (img-…, art-…)
<uuid>/sub-1.jsonl an agent role's own transcript
CURRENT the id the next run opens
- src/agent_store.c: new_id() makes a uuid v4 (16 bytes from /dev/urandom,
clock+pid only as a fallback), dir_for_id/flat_for_id/nested_for_id/
shape_of_id decide which shape an id has, so every reader takes either.
An id is a directory name, so it is checked as one (no "/", no "..",
nothing starting with ".", not CURRENT or attachments). path_for_id() keeps
the shape a session already has and makes the directory for a new id, so
nothing had to change in the callers. store_list() lists the directories
that hold a transcript plus the flat files, newest first. delete() is one
rm_tree of the session's directory, plus the old shared attachments/<id>
and CURRENT when it pointed there.
- cst_agent_store_migrate() (--migrate-store, or plain cst-agent for a
session): moves a flat store in — <id>.jsonl to <id>/session.jsonl,
<id>.title in, the shared attachments/<id>/ to <id>/attachments/. rename
first, copy when the filesystem says no, and counts what it moved, so a
second run is a no-op that says "nothing to move". On the real workspace:
6 sessions moved, the shared directory gone, --list-sessions reads them
back unchanged, a second run moved 0.
- cst_agent_store_id_of_path(): the id of a session path. This is the bug the
whole change had been hiding: cst_agent_session_id() took the id from the
file name, and the new file name is session.jsonl -- every session would
have called itself "session" (the sidebar, CURRENT, undo, an agent role's
directory, and the metrics file, which now also goes through the helper).
- src/attach.c: a session that owns its directory keeps its payloads inside
it (<id>/attachments/); only the flat layout needs the shared directory and
the id stem. Attachments behave exactly as before otherwise.
- tests/test_store.c: uuid shape and uniqueness, a small buffer refused, a
new session's path and directory, write/list/info/title read back, the
payload landing in <id>/attachments/, delete taking the directory with it;
then an old-layout store: readable before the move, moved byte for byte,
its payload readable through the new path, listed once, a second migration
a no-op, deletion covering the migrated payload. Also in the Makefile and
CMakeLists test lists.
- app/src/ui/UiTest.cpp: the image directory check wanted "/attachments/",
which the new layout has no trailing slash for.
Desktop: build clean, make check 35/35, --uitest all green, --selftest and
cst-app-selftest OK, and --selftest opens a session of the migrated workspace.
Version 0.9.5.
A phone freezes an app it cannot see, and the front end had nothing holding
it: a turn that took minutes -- a long shell command, a provider that thinks,
a build -- stopped quietly the moment the user pressed home, and the answer
was only there if the window happened to come back. The engine's own /attach
flow made this visible: an APK turn survived, a pocket did not.
- app/android/src/xyz/whcself/cstagent/CstService.java: a foreground service
started with an explicit component (nothing outside the app can reach it,
no intent filter) for as long as a turn runs. Being a foreground service
makes the process one Android neither freezes nor kills; a partial wake
lock (30 min ceiling) keeps the CPU awake with the screen off. Its ongoing
notification says what is running -- "CST-agent is working" plus the first
line of the prompt -- on the same cst-agent channel the notification_send
tool uses, and taps back into the app. A phone that refuses the foreground
start (Android 12+ does, from the background) loses the protection, not the
turn: the failure is caught and the service stops itself.
- Manifest: FOREGROUND_SERVICE, FOREGROUND_SERVICE_DATA_SYNC (API 34+ wants
the typed one) and WAKE_LOCK, plus the <service> with
foregroundServiceType="dataSync". The first build of this failed with
"Expected '>', but got ' '" -- a manifest comment cannot contain "--".
- Front end: app::holdForeground(text) / app::releaseForeground() in
ToolProvider.h, one call for every platform -- Android starts and stops the
service through JNI, a desktop has nothing to hold (its window is the app).
AgentCore::send() calls it when the turn starts (firstLine() of the prompt
is what the notification says) and finishTurn() releases it, so the
notification exists exactly as long as the work does. CST_APP_NO_SERVICE=1
keeps a test run from touching the phone.
- On the NX789J, with the app sent home and the screen off: the service read
back isForeground=true with its notification, the wake lock cst-agent:turn
was taken at the start of the turn and dropped at its end, run_shell
"sleep 60" finished with the screen dark, and the service was gone after.
Desktop: build clean, --uitest all green, --selftest and cst-app-selftest OK
(the hold is a no-op there, so nothing about the Qt side changed).
Version 0.9.4.
The tool was the last Android one that never worked on the NX789J: it would
fail with "Notification.Builder.build() failed: no exception detail" and the
shade stayed empty. The detail was the whole story -- Qt clears a Java
exception before C++ gets to look at it, so a build() that throws and a
build() that returns null look exactly the same from here, and the only way
to catch the difference is to let Java catch it.
- app/android/src/xyz/whcself/cstagent/Notifier.java: the builder, the
PendingIntent, build() and notify() in one try/catch that returns "" or
the throwable's class and message. androiddeployqt copies app/android/ with
the rest of the package source and Qt's Gradle template puts its src/ on
java.srcDirs, so a class dropped there needs no build-script change (the
first one is this; the foreground service of the next step goes there too).
- AndroidTools.cpp keeps what it can see from C++ (the context, the channel
on API 26+, the small icon looked up by name, the launch intent) and calls
Notifier.post(); the result line now carries the Java failure when there is
one. The javaErrorText() helper of the previous attempt is gone: reading a
pending exception from C++ is exactly what does not work here.
- On the device: the tool answers "notification sent", the system holds the
record (channel cst-agent, importance 3, the framework small icon, the tap
intent back into the app). Earlier attempts in the same session are in the
transcript next to it: four failures from the old path, then this one.
Docs: docs/android.md gets the Java-helper mechanism (and why the tool has to
be wrapped in Java at all) and no longer lists notification_send as missing;
app/README.md's platform table and the Android note say where a Java class
goes.
Version 0.9.3.
A turn could write a report and say "here it is" -- and there was nothing to
open. The engine had no way to hand a finished file to the person, and the
front end had no way to show one. This release adds the whole path: a tool to
deliver, a store that keeps the copy, a card in the transcript, and open /
share through the platform. It also fixes two bugs the phone exposed while it
was being tried there.
deliver_file (engine)
- cst_attachment_t.for_user marks a copy that belongs to the person, not to
the model; cst_attach_import_artifact() stores any bytes up to
CST_ARTIFACT_MAX_BYTES (64 MB) with the mime taken from the name's
extension. Attachments are unchanged (content decides the type, 16 MB).
- It is a read-policy engine tool: path resolves inside the workspace, a
directory is refused, title (optional) is the display name.
- The wire is what keeps it out of the conversation: parts_add_attachments()
skips for_user entries without even a "not sent" note, and a turn that only
delivered a file no longer invents an empty user message. tests/test_artifacts.c
pins that down end to end, including a session-file round trip.
The card (Qt front end)
- filesText() replaced imagesText(): one line per file -- kind, stored name,
name, origin, mime, bytes, pixels -- decoded by ImageStrip.qml. A for_user
entry is a file card (extension badge, name, type and size, Open, Share);
a picture stays a thumbnail and opens on tap. Both shapes are read, so
sessions written by an older version still draw.
- AgentCore::openArtifact() / shareArtifact() accept a name only when one of
the session's own messages carries it, and hand the display name to the
platform layer for the status line.
- Android opens through the app's FileProvider (content:// + ACTION_VIEW,
resolveActivity() checked, read permission granted) and shares through the
share sheet; a desktop uses QDesktopServices / the clipboard.
CST_APP_NO_OPEN=1 makes --uitest run without opening anything.
Two bugs the device found
- The session writer had its own copy of the attachment JSON and forgot
`for_user`: the file survived the file, but came back as an attachment --
no card, and put back in front of the model. It now uses the one writer
(cst_attachments_to_json()), with the round trip covered by a test.
- engineToolImages() looked the card up with findTool(), which only finds a
card that is still running -- and the engine reports the result *before*
the files, so the card was already complete and the files were dropped.
A live turn therefore showed no card while a replay of the same session
did. It now matches on the call id.
- A tap that could not work said nothing: ImageStrip.qml's errorText was
never drawn. The card now carries the reason ("no app on the phone opens
report.md").
Also
- The saved provider key reaches the engine on every start: AgentCore::open()
was handed an empty key and the engine never resolves credentials itself
(only the CLI did), so every start asked for it again. It resolves the same
order main.c uses (saved key, provider environment, CST_API_KEY) and wipes
the copy it made; --selftest checks it against a scratch credentials file.
- Android: notification_send reads the Java exception text into its result,
and keeps failing on the NX789J -- the build() returns null and Qt clears
the exception before C++ can read it, so the call has to be wrapped in Java
next (docs/android.md says so).
- --uitest grew the delivered-file case (card, Open, Share, a forged name
refused), the live order of result-then-files, and the startup row count.
Version 0.9.2.
A workspace is never empty on the second run: cst_agent_open() resumes the
last session of the store, so the engine holds a transcript before the window
exists. The window only filled its view on a session *switch*, and a start is
not a switch -- the pane came up empty next to a sidebar pointing at the very
session it belonged to (on the phone: "hello, 6 messages" highlighted, no
transcript under it).
- Main.qml: Component.onCompleted reloads the session list and replays the
open session, the same call the sessionChanged handler makes.
- AgentCore::switchSession(): a click on the row of the session that is
already open emits sessionChanged instead of returning quietly, so the row
redraws its transcript rather than being a dead spot.
- --uitest: the window is now created over a seeded session (a fixture is
written and switched to before the QML loads, which is what a restart looks
like) and the check "the open session is on screen at startup" pins it --
it failed with 0 rows before the fix. Two qWaits in the settings-dialog
section, whose clicks were racing the dialog's layout: the extra startup
work made that visible (5 failures that had nothing to do with sessions).
Verified: --uitest 79/79 (the new check among them), cst-app-selftest OK, and
on the phone -- the APK rebuilt and installed, a restart of the app draws the
whole stored session (messages, tool cards, thinking blocks, copy icons).
Version 0.9.1.
mode = coding
Two strands land together: the window draws pictures now, and the whole front
end builds, installs and runs on the phone with the engine linked in-process.
Images (the "make a picture actually show up" part)
- Engine, additively: cst_agent_hooks_t.on_tool_images(call, list, n, ud) fires
when a call left attachments behind, and cst_agent_pending_rel() /
cst_agent_pending_is_image() read that list without the app touching the
store. tests/test_attach.c::test_agent_pending pins the three cases (an
image, a stored name that passes cst_attach_rel_ok, out-of-range NULL).
- App: image://cstattach/<stored name> (app/src/ui/SessionImages.cpp) serves
only the current session's attachment directory, checked with
cst_attach_rel_ok, decodes down to the requested sourceSize and never
upscales, and holds a mutex because requestImage() runs on the render
thread while the directory can change underneath it. AgentCore gained
attachmentDir(), pendingImages(), the toolImages()/promptImages() signals
and hookToolImages(); history() gives replayed rows their images too.
- QML: ImageStrip.qml (a thumb strip, <=150 px tall, shaped by the picture,
tooltip with name/size/origin, the file's name when it went missing),
wired into MessageCard and ToolCard. Qt 6's ListModel cannot hold a
list-valued role (append() reads back undefined, setProperty drops it), so
a row carries one image per line as <stored>|<name>|<origin>|<w>x<h> and
ImageStrip decodes that.
Android (the engine is already in the APK; this is the front end)
- app/src/main.cpp: on Android the workspace, config, tmp and HOME go under
the app's own data directory before the engine opens anything, because
QDir::currentPath() is "/" there and /tmp is not writable.
- app/android/: the manifest (QtApplication/QtActivity, INTERNET and
POST_NOTIFICATIONS, adjustResize, no storage permission -- the picker will
be the system one), strings, a vector icon; minSdk 24 and targetSdk 35 come
from app/CMakeLists.txt through QT_ANDROID_MIN/TARGET_SDK_VERSION.
- app/src/platform/android/AndroidTools.cpp: the Android file behind
PlatformInit(), owner "app:android" (notification, clipboard, share sheet,
file picker are stubs for now, so the tool list does not lie about them).
- app/tools/selftest.c decides its scratch workspace with os_temp_dir(), so
the C smoke test runs on the device too.
- scripts/build_apk_android.sh: sources to installed APK in one command
(environment checks, qt-cmake configure, --target apk, adb install), and
the host-side plumbing the NixOS + qt.io combination needs -- the NDK
sysroot include dirs for Qt's FindGLESv2/FindEGL (they only look under
<sysroot>/include), -L$PREFIX/lib on the shared/module link lines as well
as the executables', the Android debug keystore created on first use with
QT_ANDROID_SIGN_APK + QT_ANDROID_KEYSTORE_* (androiddeployqt copies the
unsigned gradle output and signs it itself), and the #!/bin/bash shebang of
the build-tools wrappers, which NixOS does not have.
- scripts/gradle-android-init.gradle (installed into ~/.gradle/init.d/): the
release lint task is switched off -- lintVitalAnalyzeRelease resolves its
own tool jars straight from dl.google.com, bypassing every repository the
build configures, and fails the build after all the real work is done.
- Icons: app/icons/*.svg (Breeze's 24@3x files, the 19 names the window asks
for) are compiled in as :/cst-icons and ThemeIcons.cpp falls back to them
when the platform has no icon theme. On the phone the Qt theme lookup found
nothing and the QML glyph fallback was an empty box in the phone's fonts.
- docs/android.md: "The APK (Qt front end)" -- the SDK assembled without
sdkmanager, the host Qt's missing libraries on NixOS, the Gradle mirrors,
lint, signing -- and a Status that says what runs where. app/README.md
gains the images and APK lines, and stops claiming Android platform tools
that are still stubs.
Verified: engine `make check` green; desktop `cst-app-selftest: OK` and
--uitest all checks pass; phone (NX789J, Android 16, arm64-v8a) -- the APK
installs, starts on the workspace under /data/user/0/xyz.whcself.cstagent/
files/, a turn reaches a configured provider over HTTPS from inside the ART
process and renders the reply, rotation keeps the session and docks the
sidebar, and the toolbar/card icons draw.
Version 0.9.0.
mode = coding
The settings dialog could pick a preset, a model and a key and nothing else.
Two things were missing: a provider the presets do not know, and what the
model is asked to do while it answers. Both live under "Advanced options"
now, next to the entry they belong to.
Custom providers (engine)
- cst_models_user_set_provider() / cst_models_user_drop_provider() write the
engine's own user layer, <config>/cst-agent/models.json:
{"providers": {"myllm": {"api_base": ..., "key_env": ...,
"default_model": ..., "models": [...]}}}
and reload the registry, so the provider is at once in the provider list,
in the wire models[] and in the CLI's /provider -- the file a hand-written
override would go into is the same one, so the GUI and an editor agree.
The name is checked (1-31 of letters, digits, '.', '_', '-'), a preset's
name overrides that preset (a self-hosted endpoint, a proxy), everything
else in the file (model rules) is kept, and a file that is not a JSON
object is never overwritten. Dropping only removes the user's own entry.
- AgentCore::providers() gained "custom" (source == "user"), and
addProvider() / removeProvider() drive the two calls above.
Model parameters (engine + wire)
- cst_agent_config_t and cst_llm_request_t carry temperature, top_p and
max_tokens. 0 means the field is not sent at all -- an absent field is the
provider's own default, which "temperature": 0 is not (deterministic
decoding is a real choice), so cst_llm_build_body() only adds what was set.
tests/test_llm_body.c:test_gen_params pins both halves.
- settings.json keeps them per provider ("params": {...}) via
cst_settings_remember_params(); cst_settings_remember() carries them along
with the provider choice, and cst_agent_create() reads them, so a CLI or
TUI run sends what the window was set to. An all-zero config (the CLI)
leaves what the GUI saved alone.
- AgentCore::genParams() / setGenParams(), and useProvider() applies both the
remembered parameters of the provider it lands on and a base URL that was
edited in the dialog.
The dialog
- Rewritten around one scrolling body with a fixed footer: a long model list
used to push the key field and the buttons out of the dialog. The model
chips are filtered (a filter field appears past twelve) and capped at 40
with "... +N more", which is what a router with hundreds of models needs.
- "Advanced options" folds out the provider entry (name, base URL, key
variable, default model, models; Save provider / Remove from models.json)
and the model parameters (temperature, top-p, max tokens; empty = not
sent). The fields follow the picked provider and are what gets saved, so a
hand-typed edit is never lost to a stale property.
Verified: --uitest 67/67 (15 new checks: parameters round-trip through the
engine, a provider added through the form and removed again, a bad name
refused, a preset not removable), --selftest, ctest cst-app-selftest, and
make check with the new body test.
Version 0.8.15.
mode = coding
The window drew its icons with unicode glyphs -- a clip emoji for attach, a
gear, a hamburger, a chevron, a multiplication sign for delete. Those render
in whichever font happens to cover them (colour emoji in a grey toolbar) and
are a box on a machine without that glyph. They now come from the icon theme,
the way every other desktop app gets them.
Icons
- app/src/ui/ThemeIcons.{h,cpp}: a QQuickImageProvider for
image://csttheme/<name>?color=%23rrggbb -- QIcon::fromTheme() in the
desktop's theme (whatever the platform theme set, else breeze-dark, breeze,
hicolor), tinted with QPainter's SourceIn and rasterised at the requested
size and the screen's device pixel ratio. Qt only builds its theme search
paths from XDG once a theme name is set, so the provider fills them itself
(GenericDataLocation/icons, ~/.icons and, on NixOS,
/run/current-system/sw/share/icons, for a shell started outside the desktop
session). An unknown name yields a null image, which QML reports as
Image.Error.
- app/qml/ThemeIcon.qml: the one component that asks for an icon. Every use
still names the old character as `glyph:`, so a host with no icon theme
degrades to exactly what the window looked like before instead of going
blank. installThemeIcons() is called by main() and by --uitest, so the test
loads the same icons the app does.
- Replaced: nav (application-menu), settings (settings-configure), attach
(mail-attachment), copy (edit-copy), session delete (edit-delete), new
session (list-add), expand/collapse (arrow-right/arrow-down), jump to the end
(go-down), the plan checklist (task-complete/box). A tool card also shows
what kind of tool ran -- Theme.toolIcon(): pencil for patch/write, text page
for read, folder for list/glob, magnifier for grep, braces for a shell,
globe for web, list for a plan. Toolbar icons are 22 px (Breeze's toolbar
size), card and row icons 14-16 px.
Colours (the "some text is unreadable" part)
- textFaint #6d757b -> #8d959c: ~3.4:1 -> ~5:1 on the card fill. It carries
hints, provider URLs, session metadata and placeholders, so it is text, not
decoration.
- A selected provider row and the open session row write on the accent fill;
they now use accentText / accentTextDim / okOnAccent instead of faint grey
and green on blue (named "accentText", not "onAccent": QML reads any `onFoo`
property as a signal handler and refuses the assignment).
- The filled buttons (Send, Use this) take their ink from Theme.accentInk.
Composer
- The frame is one Theme.pad around its column on every side (anchors.fill,
implicitHeight + 2*pad). With the top margin alone the last row -- the clip
and Send -- hung over the lower border, which is what put the clip on the
frame instead of inside it.
app/README.md gains an "Icons, colors" section (the provider, why the theme
search paths are filled by hand, the qtsvg requirement for rasterising Breeze,
the tool map, and the contrast rules).
Verified: --uitest 52/52, --selftest OK, and the 0.8.14 window relaunched in
the Plasma session (no QML warnings).
Version 0.8.14.
mode = coding
Two things the window got wrong once a real turn ran through it, plus the
switch that was invisible on a dark theme.
Tool cards (app/src/core/AgentCore.*, app/qml/Main.qml, ToolCard.qml):
- A turn that called the same tool twice left the first card on "running"
forever and gave its result to the second card: both were paired by name.
cst_agent_hooks_t hands the call (ctl_tool_call_t) to every callback, so the
id travels with the signals now (toolEvent(phase, id, name, detail),
toolResult(id, name, exitCode, output)) and findTool() matches on it, in the
same way AgentCore::history() fills it from tool_call_id / tool_calls_json
for a replayed session. Only a call without an id -- an old text-protocol
session -- still falls back to the name.
- Nothing may keep spinning after a turn: closeAllRows() runs when the turn
ends and settles every open row; a call that never came back shows "no
result" (state "unfinished") and no longer claims an exit code.
Streaming text (app/qml/Main.qml, ChatView.qml):
- The answer was plain text while it streamed and turned into markdown only
when the turn ended. A 120 ms timer now re-renders the open answer with the
engine's own renderer (AgentCore::renderMarkdown -> cst_web_md_html), so a
list, a table or a code fence fills in formatted as it arrives. A row keeps
its own "html" field; a replayed session renders once, where its rows are
built.
QML details:
- the Plan mode Switch got a drawn indicator: the Basic style's pill is a dark
pill with a dark knob on this theme, which made the control look disabled.
- Every engine signal handler is a plain window function now
(engineToolEvent, engineToolResult, engineTurnFinished, engineAssistant,
...), which is also what lets --uitest drive the exact path; rowJson(i)
returns a transcript row as JSON for it (an object would not survive
QMetaObject::invokeMethod).
Tests: 6 more checks -- a call opens a card, two calls of one tool are two
cards, the result lands on the card of its own call, the other card is
untouched, the turn end settles an open call, a delta opens a streaming answer
and markdown is rendered while it still streams. 52/52 pass, --selftest,
ctest cst-app-selftest.
Version 0.8.13.
A prompt can carry files now, and they come in the three ways a desktop user
expects: the 📎 button in the composer (QtQuick.Dialogs' FileDialog), files
dropped anywhere on the window (a DropArea over it, with an accent border and
"drop to attach" while something is dragged), and an image on the clipboard
(Ctrl+V). Text keeps the editor's own paste: the composer asks
AgentCore::clipboardHasImage() first and only takes the keystroke when there is
an image.
AgentCore (app/src/core/AgentCore.*):
- pending is a property now (QStringList) fed by cst_agent_pending_describe();
attachFile() and attachClipboard() wrap cst_agent_attach_file() /
cst_agent_attach_bytes(), clearPending() wraps cst_agent_pending_clear(), and
seesImages() reports whether the model is shown images at all.
- the engine's description says what the model gets ("png 12.0 KB"), not where
it came from, so the window remembers the names it attached and puts them in
front: "note.txt — text 21 B".
- pendingChanged() fires on attach, on clear and when a turn ends (the prompt
consumed the list), which is what makes the chips move.
Engine (src/attach.c):
- cst_attach_describe() printed every sub-megabyte size as "%.0f KB": a 21-byte
text file came out as "text 0 KB". It is "B", "KB" with one decimal or "MB"
now, and test_attach still passes.
QML: the composer shows one chip per pending file with "clear all"; the hint
line mentions 📎/Ctrl+V and says when the model is text-only (images then
travel as notes); Main.qml owns the DropArea and the drag feedback.
Verified: 44/44 UI checks (8 new: attach ok, one pending, the name is shown, a
missing file is refused, one chip in the composer, clear all empties it, the
clipboard image is seen and attached), --selftest, make check rc=0 (test_attach
included), and screenshots of the chip row.
Version 0.8.12.
The Basic style paints its own controls: the two text fields came out as white
boxes on the dark panel and the buttons looked nothing like the rest of the
window. Both fields now carry the surface, border and focus colors the chat
cards use (same shape as AskDialog's field), and the buttons follow the
composer: Close is a quiet bordered button, Use this is the accent one.
No behavior change; 36/36 UI checks still pass and the dialog was looked at
again through CST_APP_SHOT.
Version 0.8.11.
The window could be started, but it could not be pointed anywhere: a provider,
a model and a key only came in on the command line. The gear in the header now
opens a dialog that changes all three on the engine's live config.
AgentCore (app/src/core/AgentCore.*):
- providers(): the engine's preset table (cst_provider_count/at), with base
URL, the remembered or preset model, whether a key exists and where it comes
from (cst_creds_resolve: "saved", or the environment variable), the preset's
key variable and the source ("built-in" / "user"). A key is read into a local
buffer only to answer "is there one"; it never leaves the C++ side.
- modelChoices(provider): the provider's own list (cst_provider_models), the
fetched cache (cst_ps_cached -- which writes through its out-parameters, so
it takes a real int, not NULL: passing one cost a segfault) and the
remembered pick. No network. modelInfo() is cst_model_info + its "200K ·
text only" label; usage() is cst_agent_usage for the session and the turn.
- useProvider(provider, model, key) -> "" or a reason: a different provider
goes through cst_ps_switch_provider() (the key is verified against /models
before it is saved; offline it is kept unverified and the header says so),
a key for the current one through cst_ps_set_key(), and the model is
remembered in settings.json. syncConfig() takes the engine's config back
over the app's copy, so provider/model/approval read correctly afterwards.
The engine's session is never touched by a switch.
- open() no longer starts with an empty provider: it takes the one the engine
used last (settings.json), else the built-in default, plus that provider's
base URL and model. Without this the first turn could not be sent at all and
no entry in the dialog was "in use".
QML (app/qml/SettingsDialog.qml, Main.qml):
- a modal dialog with the themed surface the other dialogs use (the Basic
style's light panel washes out the dark theme), the provider list with its
key state, a model field with the info label and clickable candidates, a
password field for a new key, the session's token counters, and a note that
reports what the last "Use this" answered.
Tests: 8 more checks -- the presets are readable, an unknown provider is
refused, the counters are readable, the gear opens the panel, the rows exist, a
model typed into the field reaches the engine, the note says so, Close puts the
dialog away. 36 checks pass. --uitest also points XDG_CONFIG_HOME at its
scratch directory now, so a run cannot rewrite the user's settings.json or
credentials, and CST_APP_SHOT writes shot.png at the end plus
shot-settings.png while the dialog is open (that is how the dialog above was
reviewed on a machine without a screen).
Verified: 36/36 UI checks, --selftest, ctest cst-app-selftest, make check rc=0,
and a screenshot of the dialog.
Version 0.8.10.
A conversation is now something the window can pick. The sidebar lists the
workspace's stored sessions and switching replays one into the view.
Sessions (app/src/core/AgentCore.*, app/qml/SessionRow.qml, Sidebar.qml):
- AgentCore::open() resolves the workspace's current session through
cst_agent_store_resolve_default() before the engine starts. Without it the
engine opened session.jsonl in the process's working directory, next to a
session list that stayed empty.
- sessionId property; sessions() (id, title, firstUser, messages, mtime in ms,
current); newSession(), switchSession(), deleteSession() (moves the engine
off a session it has open first), renameSession(); sessionChanged and
sessionListChanged signals. No mutation runs during a turn.
- history() turns the open session back into the rows the live signals build
(kind/role/text/name/state/detail/output/exitCode/complete): system prompt
skipped, notes and attachments shown, assistant reasoning as a think row,
tool_calls_json as one card per call, and each result paired with its call
the way findTool() pairs the live stream. A call with no result keeps state
"unfinished" ("no result"), so a replay never shows a spinner that cannot
stop. Main.qml replays on sessionChanged.
- SessionRow.qml is a file of its own: as an inline component it would not see
the Repeater's modelData. Session controls are disabled while a turn runs.
Tests (app/src/ui/UiTest.cpp, app/src/main.cpp):
- the fixture session is written through the store's own API (a tool call with
its result), then the checks click its row, assert the five replayed rows and
the paired result, start two new sessions (the view empties), delete one by
its button (the open one stays), and switch back. 28 checks in total.
- item() falls back to the visual tree: delegates a model owns are not always
in the window's QObject tree, which is why the rows were invisible to
findChild even though they were on screen.
- --uitest drives a scratch workspace unless --workspace is given; it creates
and deletes sessions. CST_APP_SHOT=file saves the window as a PNG when the
run ends and CST_APP_DUMP=1 prints the session rows seen; both are how the
layout gets looked at on a machine without a screen.
QML: ThinkBlock's click target is a TapHandler on its row (a MouseArea anchored
inside a layout is undefined behavior and Qt said so).
Verified: 28/28 UI checks, --selftest, ctest cst-app-selftest, make check rc=0,
and a screenshot of a replayed session (cards, tool card with its result, the
think block, the sidebar list).
Version 0.8.9.
The window is no longer one file pretending to be a chat box. The transcript,
the run panel, the composer and the two consent dialogs are their own QML
components over one Theme singleton, and the shell around them adapts to the
screen: docked and scrollable on a desktop, a slide-out drawer on a phone.
QML (app/qml/, new except Main.qml):
- Theme.qml: single QML singleton, dark Breeze-flavoured tokens (surface,
border, accent, text/textDim/textFaint, radius, pad, gap, sizes, monospace);
diffLineColor() for unified-diff rows, compactBreakpoint 880, touchTarget
34, segmentHeight 30, headerHeight 42.
- ChatView.qml: a row-type dispatcher over the transcript model (message,
thought, tool, note), follow-the-tail scrolling with a "jump to latest"
button once the user scrolls up, and an empty state that says where the
agent runs and what it can do.
- MessageCard.qml, ThinkBlock.qml, NoteLine.qml, ToolCard.qml (status pill,
arguments, collapsible output, 500-line cut, diff colouring, spinner while
the call runs), TodoPanel.qml.
- Sidebar.qml: workspace/provider/status, mode and approval segmented rows,
plan switch, the engine's todo list, version footer; scrolls, so a short
window still reaches every control.
- Composer.qml: auto-growing field, Enter sends / Shift+Enter newlines,
red Stop while a turn runs, autocorrect hints off, finger-sized targets.
- ConsentDialog.qml (Deny / Once / Always -> CST_PERM_*) and AskDialog.qml
(the model's options become buttons, free text stays possible).
- Main.qml: shell with header, draggable sidebar splitter (hover feedback,
200-560 px, double-click resets), transcript, composer, the two dialogs,
Esc to cancel and Ctrl+L to clear. Under 880 px the sidebar becomes a
drawer with a scrim instead of squeezing the chat away.
AgentCore / main.cpp:
- provider, model, approval and planMode are properties now; approval and
plan mode write back into the engine's config, and renderMarkdown() calls
cst_web_md_html(), so a table or a code fence looks like the WebUI's.
- the todo hook writes m_todo on the GUI thread (it was written from the
worker thread before: a data race on a QString member).
- --size WxH for the window (the layout tests drive it), and the Basic style
is chosen by default: the Breeze Quick style registers its types with
relative URLs, which floods the log and breaks the composer.
UI tests (app/src/ui/):
- --uitest drives the real window with synthesized input and exits with the
number of failures: 15 checks over docking/drawer transitions, mouse clicks
and touch taps reaching the engine's mode, the plan switch, the composer
arming the send button, and a 460x900 -> 1200x820 resize round trip.
- every control the tests touch carries an objectName; the sidebar's
implicitHeight binding loop is gone.
Tests: cst-app-selftest and --uitest pass in the standalone app build,
ctest cst-app-selftest passes in the repo build (38/38 on the host).
Version 0.8.8.
The engine can be embedded now instead of only driven from a terminal: the
same objects link into a Qt front end (app/) that owns a cst_agent_t inside
its own process -- no port, no token, no second process -- and that front end
now runs on the desktop and, engine-side, on the phone.
Capability metadata (include/cst_tools.h, src/tools/registry.c):
- cst_tool_def_t gains perm_name, platforms and available (a probe that
returns a reason), so a permission is addressable by name and a tool can
say why it cannot run here.
- cst_platform_name(), cst_tool_platform_ok(), cst_tool_available(),
cst_tool_meta_reload(), cst_tool_perm_name() and CST_PLAT_DESKTOP[].
- cst_tool_register() may take over a built-in name that is not runnable on
this platform, so a front end replaces it with its own implementation.
- the registry entries carry their class (delete_file -> "delete"), the
desktop clipboard tools carry a clipboard probe; find/active/run gate on it.
- /api/settings.tools[] exposes perm_name, platforms, available, why and the
declared parameters.
Engine as a library (Makefile, CMakeLists.txt):
- libcstcore.a from the CORE_OBJ set, and a CMake cstcore target shared by
the CLI and by app/ (-DCST_BUILD_APP=ON): the engine is compiled once.
- Android branch for the vendored sashidown: a per-ABI archive built with the
NDK's aarch64-linux-android24-clang wrapper, because CMake's compiler is
plain clang and keeps --target/--sysroot in its own variables.
Qt front end (app/, new):
- AgentCore wraps one cst_agent_t: turns run on a worker thread, every
cst_agent_hooks_t callback becomes a Qt signal, and the two hooks that need
an answer (on_permission, on_ask_user) block the worker until the GUI
replies; mode, approval and plan mode go through the engine's config.
- ToolProvider hands the front end's tools to the engine (owner "app:*"); the
Linux one is notification_send with a probe for notify-send/D-Bus.
- Main.qml: transcript, tool blocks, consent dialog (Deny/Once/Always ->
CST_PERM_*), ask_user, mode switch, Esc to cancel; --selftest stays
headless and runs the embedding checks without QML.
Android (scripts/, docs/android.md, new):
- scripts/build_third_party_android.sh builds static mbedTLS + libcurl for an
ABI (CA path /system/etc/security/cacerts); build_android.sh builds
cst-agent, libcstcore.a and the test binaries for the phone, can push them
and run the whole suite there.
- sandbox_exec.c keeps the plain exec path on Android: the seccomp supervisor
compiled in and the probe reported "seccomp", but writes outside the
workspace were not denied on the device -- claiming a sandbox that does not
hold is worse than reporting none.
- tests/test_diff.c compares layouts only against real GNU diffutils: toybox's
diff prints the same hunk header with its own hunk/context rules.
Tests: test_tool_meta (metadata, platform gate, probe reasons, takeover of a
built-in name) and test_embed (in-process embedding contract) are new; the
existing registry/llm_body assertions now follow availability. 38 test
binaries pass on the host, 36/36 on arm64-v8a Android 16, cst-app-selftest
passes in both the standalone app/ build and the repo build.
Version 0.8.7.
Working mode: --mode coding|general (CST_MODE, settings.json, /mode,
WebUI run settings). general swaps the default framing for a general
assistant one -- workflow, safety notes, tool wording (run_shell, the
subagents) -- and /status shows it; minimal (the paper profile) stays
byte-frozen. The prompt state carries the mode so a switch leaves a
system note in history.
Clipboard: clipboard_read / clipboard_write tools (16 MiB cap read,
default 64 KiB window; writing is a WRITE permission with a preview) on
pbcopy / wl-copy / xclip / pbpaste / wl-paste / xclip and the Win32
clipboard. In the TUI, /attach paste falls back to clipboard text when
there is no image on it.
PDF: a dependency-free text extractor (src/pdf_text.c): inflate,
Flate/ASCIIHex/ASCII85, PNG/TIFF predictors, layout-independent object
scan, object streams, page tree walk with inherited resources, content
lexer (Tj/TJ/'/" and the Td/TD/T*/Tm spacing heuristics), per-font
ToUnicode CMaps and inline images skipped. Encrypted files and pages
with no text layer are refused with the reason.
Attachments: /attach (a path or files dropped on the terminal) now take
images, PDFs and text. A PDF is stored as its text layer with the mime
still saying application/pdf and its page count; anything else has to be
UTF-8 text. Documents travel inline as text parts, so CST_NO_IMAGES and
text-only models no longer drop them, while images keep their old path.
Tests: test_pdf_text (plain, bare and zlib-framed Flate, inline image,
CID + ToUnicode, object stream, error paths), documents in test_attach,
inlined documents in test_llm_body, mode in test_config.
Process timeline: a run of consecutive thinking and tool calls is one
block that folds to a single line ("思考 4 次 · 工具 6 次"; while running
it shows the current action) and opens into a compact timeline, one row
per thought or call. Visible text, a note or the next prompt closes the
block; history and the live stream follow the same rule. show_thinking
now also applies while streaming instead of forcing thoughts open.
Layout: the avatar gutter and the flex-wrap that dropped it under long
answers are gone, so answer, process blocks, notes and composer share one
left edge. The text column and composer are both 720px (--content); the
rail and run panel grow with clamp() on wide screens. A rebuild after
each turn no longer replays every entrance animation, and the turn time
survives it. Message actions are icon buttons and only appear under text.
Images: a prompt's images sit above the bubble, right-aligned (one keeps
its shape, several are 152px rounded tiles) and open in a lightbox (Esc,
arrow keys). Images uploaded from the page show as thumbnails over the
composer and appear on the prompt as soon as it is sent.
Also: ?theme= is no longer overridden by app.js, the user bubble is
visible in the light theme, and the narrow top bar no longer wraps.
Sessions: rename, delete and search; /api/retry regenerates the last
answer, editing a prompt resends from there; /api/export downloads the
transcript as Markdown (cst_session_export_md). Session messages carry
note/tool_result markers, so a reload rebuilds the tool steps with
their arguments, output and exit status; subagent progress is shown.
Diff and settings: /api/diff shows the real diff; /api/settings covers
protocol, max_steps, tool switches and confirm; /api/permissions edits
the rules; sandbox status is reported. Session rename no longer reads
past a short title.
Commands: slash commands with completion in the composer (/model,
/provider, /new, /retry, /undo, /compact, /export, /diff, /init,
/skill and more), a Ctrl+P palette, Shift+Tab cycles the approval
mode, F1 or ? shows the keys. "Thinking expanded" is saved in the
settings (show_thinking), "expand all tool output" in the browser.
Extensions: GET /api/extensions lists MCP servers, plugins, skills,
agent roles and subagent runs; POST /api/extensions/reload and
/api/extensions/trust; a workspace switch reloads them. /api/init and
/api/skill run on the server (cst_init_prompt shared with the TUI).
Errors, statuses and [cst] notes are shown in Chinese with a readable
reason instead of rc=; a failed turn's error survives the reload.
Tests: scripts/e2e_webui_sessions.py, e2e_webui_workspace.py and
e2e_webui_ext.py, run from e2e_webui.sh.
Streaming: Chinese reasoning showed up garbled in the WebUI. The agent
loop now feeds visible/thinking text in runs instead of byte by byte,
and the WebUI holds an incomplete UTF-8 tail until the next chunk
(cst_utf8_hold_feed), so every think_delta/assistant_delta is whole.
src/provider_switch.c: model list fetch/cache, key verify-and-save and
provider switching moved out of tui.c and shared by TUI and WebUI.
WebUI API:
- GET /api/providers, GET /api/models[?refresh=1] (401 on a rejected
key, still listing built-ins), POST /api/provider, POST /api/key
(key checked against /v1/models, saved to credentials 0600, never
echoed in replies, events or logs)
- /api/config validates everything before applying, persists model,
effort and approval (not yolo), warns when effort does not apply;
plan_mode no longer forces read-only
- state: turn_usage, api_base, key_set/key_from, model_info,
effort_applies, sees_images, max_steps, tool_profile, perm_timeout;
permission events carry their timeout
Frontend: searchable model picker with refresh and manual entry,
provider picker and key dialog, effort/vision hints, toasts instead of
silent failures, Chinese error text, real protocol in the settings and
usage panels, approval countdown from the server's timeout.
Tests: utf8 hold unit test, scripts/e2e_webui_settings.py (fake
/v1/models, wrong key, switching, persistence, Chinese thinking) run
from e2e_webui.sh, scripts/tui_keys/provider.py.
Build: <fcntl.h> in tui.c, format-truncation in http_srv.c and /web.
WebUI:
- refuse cross-site requests (Origin must be the page itself) in the
default no-token mode: any page could POST approval=yolo and a prompt
- session/workspace/undo/compact/attach/model changes are 409 while a
turn runs (use-after-free, messages landing in the wrong session)
- ask_user returned "dismissed" at once instead of waiting
- cancel and shutdown wake a turn parked on a permission or question;
SIGTERM mid-turn no longer hangs or crashes
- a page opened while a question is pending still gets it;
Last-Event-ID resumes the stream; event seq matches the SSE id
- token mode: page assets load without the token, images via cookie
- --web-open works, --web-port/--web-token are validated
- frontend: gate keys ignored while typing, clear-images button,
reasoning no longer wiped while streaming
HTTP server: auth and size cap before reading the body, strict
Content-Length, NUL in url_decode, 413 on keep-alive, idle timeout,
absolute-form only without a leading '/', case-insensitive Connection,
header deadline, 400 for long request lines, unlink before close,
poll() instead of select(), SO_EXCLUSIVEADDRUSE on Windows.
Markdown sanitizer escapes quotes in attribute values.
/web: passes this session's provider/model/base/effort/approval (key
via the environment), keeps output off the TUI, shows the real URL and
keeps CURRENT on the TUI's session.
Tests: scripts/e2e_webui_ask.py, scripts/e2e_webui_http.py (run from
e2e_webui.sh), sanitizer quote test.
The front end is rebuilt on the design tokens the desktop app ships (surfaces,
color-mix borders, alpha hovers, accent): a 30-icon inline SVG sprite, dark and
light themes, layout switches as deep links (?side=0, ?panel=0, ?theme=light)
resolved before the first paint. The empty page is one line and leaves with the
first turn; the right rail keeps three cards - todo, file changes (+N -M) and
context (tokens, prompt cache hit rate). Both rails animate in and out, and a
collapsed rail widens its grid track to 0 instead of being removed, so the
transcript never slides into a zero-width column.
- src/webui.c: POST /api/workspace - re-root the agent at another directory
(permission rules, metrics, AGENTS.md follow) and start a session there, as
sessions live per directory; state also reports cached/cache_miss tokens
- src/agent_loop.c: cst_agent_set_workspace(), rebuilt prompt and re-measured
budget; the auto-compact threshold is 75% of the window (was 60%)
- src/tui.c: /web forks a WebUI next to the TUI - separate process and a new
session, because cst_webui_run() blocks and two writers corrupt one session
- src/os_compat.c, src/os_win.c: run_shell runs without a time limit unless
CST_SHELL_TIMEOUT asks for one, so builds and long clones finish
- src/main.c, include/cst_webui.h: one CST_WEBUI_DEFAULT_PORT for --help, the
docs and /web; Termux opens the browser through termux-open-url
- README.md, docs/webui.md, docs/env_openai.md: /web, /api/workspace, the new
state fields, the 75% threshold and the layout deep links
tests/e2e_webui.sh green (permission round-trip, token paths, --web-dev);
make check green.
A second implementation of cst_agent_hooks_t next to the TUI: the turn runs on
a worker thread, the hooks publish events into a ring buffer that SSE clients
read, so the engine itself needs no changes.
- src/net.c, include/cst_net.h: POSIX/Winsock sockets, non-blocking connect
- src/http_srv.c, include/cst_http.h: HTTP/1.1, keep-alive, chunked streaming,
body caps (413), Expect: 100-continue, connection cap, exact routing
- src/web_md.c, include/cst_web_md.h: streaming markdown -> sanitized HTML
(tag/attribute allowlist, javascript: URLs and off-site images dropped)
- src/webui.c, include/cst_webui.h: event ring + SSE broadcast, all routes,
token (Bearer/Cookie/query) and loopback Host checks, 3 s auto-deny when no
client is attached
- src/webui_stub.c, CST_WEBUI=0 / CST_BUILD_WEBUI=OFF: build without the WebUI
- web/: no-build front end (transcript, approvals, todo, sessions, image
upload, status bar)
- docs/webui.md: engine API v1 - events, requests, security model
- tests/test_webui.c, scripts/e2e_webui.sh: unit tests plus an end-to-end run
driven by a recorded SSE stream (no API key needed)
- src/main.c: --web --web-port --web-host --web-token --web-dev --web-open
and CST_WEB_PORT/HOST/TOKEN/ROOT
An empty POST body is now accepted without Content-Length (RFC 9112 6.3) so
curl-driven clients can reach cancel/undo/compact/quit; CMakeList's core tests
learned about sashidown and test_webui, and CST_WEBUI=0 defines CST_NO_WEBUI.
TUI keys: register both CSI and SS3 forms of the cursor/Home/End keys and
the xterm modifier sequences when terminfo lacks them, so an arrow sent in
the "other" form no longer turns into Esc + a typed letter. Up/Down move by
visual (soft-wrapped) row, and the draft survives a trip through history.
scripts/tui_keys drives the TUI in a pty to check all of this.
Images: text inside a multimodal content array is now a {"type":"text"}
part; DeepSeek rejected the bare string with HTTP 422. deepseek-flash
(and its deepseek-chat / deepseek-reasoner aliases) read images, so the
registry marks them vision; deepseek-v4-pro stays text only.
CI: models.c lost _DEFAULT_SOURCE in 0.7.0 and gmtime_r broke the musl
build; tests sleep with poll() instead of usleep for the same reason.
Landlock no longer handles TCP, the Seatbelt profile drops (deny network*),
and the seccomp filter stops refusing IPv4/IPv6 sockets. Sandboxed commands
run without asking in auto mode on every backend (Low-IL no longer asks for
network commands), and connection errors are not treated as sandbox blocks.
Images
- vision gate: text-only models get a note instead of image parts and no
view_image tool; unknown models retry once without images on HTTP 400
- paste: Ctrl+V and /attach paste read the clipboard (wl-paste, xclip,
pngpaste/osascript, Win32 PNG or CF_DIB converted to PNG); pasting image
file paths attaches them
- MCP image content is stored in the session and sent with the tool result
- attach: rel path validation, base64 decode, /retry keeps images
Model registry
- built-in src/models_builtin.json embedded at build time, merged with
<config>/models.json and cached GET /models metadata (context, vision,
pricing); most specific rule wins per field
- /model labels, /status source, --list-models [provider]
Cache and cost
- stable request prefix (frozen tools, notes as messages), DeepSeek prefix
completion, OpenAI/Anthropic cache hints, /cost and hit rate in the status
After a successful windows-cross build on main or v* tags, upload the zip
and .sha256 via scripts/release_windows.sh (creates/updates v$VERSION).
Prefer secrets.RELEASE_TOKEN; fall back to GITHUB_TOKEN. Skip on PRs.
os_win.c needs cst_buf_* / cst_utf8_valid from util.c after os_compat was
added to the small replay CMake targets. Release packaging only needs
cst-agent.exe, so bootstrap_windows.sh builds that target alone.
windows-cross died cloning busybox-w32 through a global gitclone.com
insteadOf (HTTP 502). Try ghfast/ghproxy/gitclone/github per repo instead
of forcing one proxy for every fetch.
gcc -O0 keeps unused cst_test_util.h helpers that need os_*; clang GC'd
them so only the gcc CMake step failed. Prefer ghfast/ghproxy.net for the
llvm-mingw tarball (gitclone 500 / github unreachable on the runner).
clang under strict C11 left strcasestr undeclared (returns int → pointer
truncation → SIGSEGV). Parse Content-Length portably; keep -Werror on
make check. Drop Nix for windows-cross (root /nix fails) and use pinned
llvm-mingw UCRT instead; Alpine musl gets linux-headers for audit.h.
test_subagent segfaulted in the Docker runner: the fake SSE handler took
user + strlen(user) - 1 with an empty user string. Guard JSON access and
format the child report with an empty-safe last character. Remove the
sanitizers job entirely (if: false still queued it and blocked the runner).
- Replace actions/checkout (pulls from github.com) with scripts/ci_checkout.sh
fetched from this Gitea over raw/commit.
- Remove the macOS job (no runner).
- Disable sanitizers until the Docker runner has SYS_PTRACE; it was wedging
the only slot with AddressSanitizer:DEADLYSIGNAL.
- Add windows-cross: Nix UCRT mingw via TUNA, bootstrap_windows.sh, verify
dist/cst-agent-*-windows-x86_64.zip. Third-party git clones go through
gitclone.com.
The sanitizers job wedged the only runner by printing AddressSanitizer
DEADLYSIGNAL forever (Docker without SYS_PTRACE). Concurrency now cancels
in-progress jobs on a new push, and make asan is wrapped in timeout -s KILL
480 with shallow ASAN_OPTIONS.
Also copy the path with memcpy after the length check so -O1 asan builds
do not warn about snprintf truncation into OS_PATH_MAX.
CI passes CFLAGS=-Werror on the command line; that wiped the library's
CFLAGS ?= ... -Iinclude -Isrc, so fsm.h could not find sashidown.h.
Always append the include paths via override CPPFLAGS.
Install was pulling bookworm packages from deb.debian.org; the domestic
Gitea runner crawls that for minutes. Rewrite sources to TUNA before
apt-get/apk, with CST_CI_MIRROR to override.
With Yama ptrace_scope=2 and no CAP_SYS_PTRACE, process_vm_readv and
/proc/<pid>/mem both fail, so sc_decide cannot read path strings and every
workspace write gets EACCES. The old probe only checked the USER_NOTIF
round trip, so CST_SANDBOX=seccomp and the forced-seccomp test pass looked
available and then failed on the first write.
sc_can_peek() forks a child, places a marker on its stack, and retries the
same sc_peek path. sc_probe() now needs that too, so those hosts fall back
to none (or keep landlock) instead of a silent deny-all sandbox.
os_temp_dir() used $TMPDIR or "/tmp" without checking either. On Termux
/tmp usually does not exist (or SELinux denies it) unless TMPDIR points at
$PREFIX/tmp, so plugin and MCP temp files failed and shell children that
need a temp file (heredocs) broke.
* os_temp_dir() picks, once, the first of $TMPDIR, $PREFIX/tmp, /tmp,
/var/tmp and ~/.cache/cst-agent/tmp where a file can really be
created; if none works it keeps the old answer
* os_temp_dir_export() at startup writes the choice back to TMPDIR, so
shell children and the sandbox's writable temp dir agree
* tests/test_tmpdir.c covers the fallbacks in fresh forks
Build
* cross build switches to pkgsCross.ucrt64 (shell-win.nix); curl (Schannel),
winpthreads and ncurses 6.6 are linked statically, so the release is
cst-agent.exe + busybox.exe (busybox-w32 Unicode) with no extra DLLs
* windres embeds a manifest with activeCodePage=UTF-8 and a version resource
* ncurses 6.6 Windows console driver patches (mouse button state, A_DIM,
wide-char declarations): scripts/patches/
* scripts/bootstrap_msys2.sh builds the same release natively in MSYS2
UCRT64; PR #1's MINGW64 notes now point there
* vendor/tre-regex (musl's TRE) provides <regex.h> on MinGW
UTF-8
* paths cross the system call boundary as UTF-16 (*W APIs); console code
pages set to 65001 and restored on exit; files are read and written in
binary mode
* child output that is not valid UTF-8 is decoded from the console, OEM or
ANSI system code page
* read_file hides CR; edit_file, apply_patch and write_file keep CRLF line
endings and a UTF-8 BOM
OS layer and processes
* os_dir_* and cst_fnmatch replace the POSIX-only directory walks in
subagent, plugins, skills, undo and fs_tools; realpath / path_under and
symlink rejection work on Windows
* every child goes through CreateProcessW with its own environment block
inside a Job Object, so timeouts and Esc end the whole process tree
Sandbox (auto mode, "auto · lowil")
* run_shell runs with a restricted Low integrity token; the workspace gets
an inheritable Low label once, the sandbox TEMP is labeled Low, and the
credentials file gets a no-read-up label
* drive roots, the profile and system folders are never labeled
* network commands still ask, since Low IL does not block the network;
Access denied and MSYS2's 0xc0000142 offer a rerun outside the sandbox
* /sandbox lists labeled folders, /sandbox reset-labels restores them
TUI
* raster console fonts are switched to Consolas / NSimSun while running
* the home logo falls back to text when the console font lacks U+2580
* mouse drag copy goes to the Windows clipboard
Distribution
* scripts/install_windows.ps1 installs a Gitea release into
%LOCALAPPDATA%\Programs\cst-agent and adds it to the user PATH;
cst-agent-update.cmd reruns it; scripts/release_windows.sh uploads the
zip with the token taken from the environment
* scripts/run_windows.bat is gone; docs/README-windows.md is rewritten
Tests
* new test_lowil, test_win_utf8 and test_os_dir; tests made portable
(cst_test_util.h)
* scripts/win_test.sh and scripts/winvm/ run the suite on English and
Chinese Windows Server VMs as Administrator, a standard user and a
Chinese user name: 29/29 on all six
- Drag, double-click (word) and triple-click (line) selection in the log,
auto-scroll at the edges, soft-wrapped lines joined on copy.
- sashidown ANSI renderer: delimiter runs with flanking rules, so
intraword underscores (test_apply_patch.c) stay literal; unmatched
* _ and backticks are shown as-is; * and _ do not close each other.
- tests: test_textsel, test_markdown.
Query CSI ? u at startup and push the disambiguate flag on the alternate
screen when the terminal answers; CSI u reports are mapped back to legacy
input (Shift/Ctrl+Enter insert a newline, Ctrl+Backspace deletes a word).
Ctrl/Alt+Left/Right move by word, Ctrl/Alt+Delete delete the next word.
Our define_key codes move above the range ncurses uses for terminfo
extended keys.
- Child installs a seccomp filter with a listener; a supervisor thread
resolves the path of write-type syscalls (cwd, dirfd, .., symlinks,
/proc/self, /dev/stdout, both ends of rename/link) and answers
CONTINUE or EACCES; IPv4/IPv6 sockets are refused
- Probe order Landlock -> seccomp -> none, checked in a throwaway child;
CST_SANDBOX=landlock|seccomp|off
- curl connect failures count as sandbox refusals
- test_sandbox runs the detected sandbox and forced seccomp, with escape,
symlink, rename, link, chmod, stdout, ENOENT, background and network cases
- Docs: platform matrix, limits (not a security boundary, Yama and
background jobs)
- bootstrap_termux.sh installs pkg-config/binutils, builds with make,
runs the tests and installs into Termux's $PREFIX/bin.
- Makefile drops -lpthread on Android (pthreads are in bionic libc).
- TUI assumes UTF-8 on Android: nl_langinfo needs API 26, Termux
targets 24. All sources compile for aarch64-linux-android24 (NDK 27).
Tools
- Table-driven tool registry with permission classes and default/minimal
profiles; new tools: edit_file, list_dir, glob, grep, delete_file,
move_file, todo_write, web_fetch, ask_user (plus read_file paging).
- Built-in Myers diff; write tools return unified diffs.
- Persistent permission rules (.agent/permissions.json), per-turn undo
snapshots (.agent/undo), AGENTS.md project notes, auto-compact at 80%
of the model context, cancellable run_shell with process-group kill.
TUI
- Table-driven slash commands: /provider /key /effort /status /tools
/permissions /yolo /undo /retry /diff /export /copy /rename /delete
/init /max-steps /todo, with / and @ completion.
- Permission modal with diff and prefix rules, ask_user modal, todo
panel, header with model/effort/context, bracketed paste, line editing,
wide-char cursor, NO_COLOR and ASCII fallback; Esc never quits.
Startup and configuration
- API keys are entered in the TUI (hidden) and saved per provider in
<config>/credentials (0600); env variables are only a fallback.
- Last provider, per-provider model and reasoning effort are remembered
in <config>/settings.json; --effort/-e and /effort send
reasoning_effort (low|medium|high|max) to models that accept it.
- DeepSeek preset uses deepseek-flash / deepseek-v4-pro.
- Session flags --new/--continue/--resume/--list-sessions, --yolo,
--confirm.
Compatibility and tests
- ncurses detection, macOS/Homebrew support, make install, XDG config
dir, POSIX cleanups; Gitea CI (gcc/clang/asan/musl/macOS).
- Unit tests for diff, fs tools, registry, undo/perms and config;
e2e fixtures multi_file_refactor and ask_or_todo; vendored cJSON.
