Repository navigation
Releases
spark v1.113.2
- S4: the reference and the brief promised that nothing leaves the
machine, and that spark never leaves your LAN, though your words go to
the model server you chose and a program spark starts may use the
network itself. They say what is sent and to where now, as README's
"What leaves this machine" does, and docs_test refuses the blanket claim
in docs/REFERENCE.md and AGENTS.md too. spark serve --print-client tells
you to keep the admin token here, where it promised that it stays (#170)
spark v1.113.1
- S3: the prompt line knew nothing only the desktop knows, its
workspaces, screens, keys and commands, nor what is in use now, so
it picked a near tool or answered in words. A program of the machine
writes those in spark's machine folder now: the lines that hold ride
the line's cached picture, the lines true now ride each request
beside the folder's path, and a stale file is skipped (#167) - S3: a tool call written with a bare value,
spark mcp call apps apps_open mpv, was refused, and one that copied the tool list's
placeholder,name=NAME, landed as a call. A bare value goes to the
tool's required arguments still missing, in order, and a placeholder
is asked again like any wrong command. A call that names no listed
server, or none,spark mcp call spark-shell theme_palette, takes the
one server that lists its tool, and a tool its server does not list
takes the one of its tools that ends the same way;spark notes add milkandspark mcp call reminders snoozeare notes_add and
reminders_snooze where those servers list them. A check the [now]
lines answer is still a command, and what uses the memory prints with
ps, never opens btop (#168)
spark v1.113
- The prompt line knows the machine it runs on (#166). Its prompt said
"Keyboard only, no GUI" on a sway desktop, so a small model answered
that nothing could reach the windows. Its first line says now what
the session is, read from the environment: the terminal, the desktop
and Wayland or X11, or a text console or an ssh login. A picture of
the machine follows spark's own commands. It names the apps and how
each came, which common programs are not installed, the home
folders, the services that start at boot and the network
interfaces by kind. Nothing is run to read it and nothing in it is a
state, so the prompt is the same bytes and its cache holds. - The prompt line carries every MCP tool, one server a line, with its
arguments. From a trusted server it carries the first words of each
tool's description and the values its arguments take, as the
server's own rows name them.spark mcpwrites those to
mcp-valueswhen it lists a server; the line reads files and starts
no server. A window's number, a level or the value in use never
rides. Until now it carried 25 of a desktop's 122 tools, cut at
1,500 characters.spark doand the other shell modes keep that
list. - On a slot of 4,096 cells, a machine that serves a chat model too,
the tool block gives way so the reply keeps its room: its values
first, then its purposes, then whole servers from the end. - A Wi-Fi network around that is not saved never rides: a call that
still asks for a password reaches a place the machine holds none
for, so its name is a neighbor's. - After a request of the launcher's the line waited 17 s: the engine
has one slot, and each prompt pushed the other out. On a machine of
24 GB or more serving one model, the engine keeps a replaced prompt
in RAM now (--cache-ram 2048with--swa-full), and the line
answers in 2.1 s after one. On the box it costs about 2.8 GB of RAM:
1.8 GB for the whole window cache, and the kept prompts, 0.6 GB
after three, 2 GB at most. Below 24 GB, or with a second chat model,
nothing changes. The check's engine row knows the rule. - The tool list no longer goes stale. Each server's tools carry a
stamp of its program. When the program changes, a desktop's update,
the timer's check lists that server again; a check you run says so,
withspark mcpas the remedy. - The line's brief ends on one rule: a check or an action on this
machine is a command, never words, even when the picture holds the
answer. "Download URL" and "remove the packages nothing needs" had
come back in words. - A reply that is a listed tool's bare name, such as
out_muteor
spark out_mute, is itsspark mcp callline now, before the judge
and the danger checks read it. - Fixed on Void: the line answered
timedatectl, which runit does not
have; the picture says it is not installed. It answeredsvlogtail -f, which the script answers with its usage; the prompt says
svlogtail follows the log and takes no option. - The line audition asks a desktop's 227 wants as Void, 75 of them
held out, with a desktop of made-up names: its tools, their values
and its machine picture. They come from a set made for this, its
items with no private data alone. In every run the picture is the
OS snapshot's, never the machine's the audition runs on. The report
keeps the bar to the OS's cases and spark's. - Measured on the Linux machine with gemma4-26b-a4b, against v1.112:
the owner's 43 requests 26 right (20), the new set's held-out part
62 % (42 %) and its tuned part 68 % (34 %). The line audition, a
report: Void 56 of 60 (57), spark's held-out 24 of 24 (21), the
desktop's wants 71 % tuned and 64 % held out. A warm reply takes 2.8
s at the median (2.9 s). #166's bar, 29 requests and 72 %, is not met
yet: the rest needs what the prompt does not carry here, a state
such as the window open or the output in use, and the desktop's own
facts.
spark v1.112
- The engine serves one slot:
-np 1on every server command and
parallel = 1in the router's presets. Left to itself llama-server
opened four and shared one cache ofSPARK_CTXamong them, so two 4k
prompts filled it and a third purged a slot's warm prefix, eight times
in a night on the box. The whole context is one request's now, and
callers take their turn, as the page already made them. The room
staysSPARK_CTX, by hand (#157). - Fixed:
spark readread first the part of a long text a rare word
ranked highest, and never the part that answers. Asked how to delete
at the destination the files gone from the source, rsync's manual
read part 2 and answered--updateand--remove-source-files,
never--delete. Words are counted over passages now, and a part
scores its best three (#160). - Fixed: the prompt line checked options only up to a sub-command, so
an inventeddnf repoquery --owns,systemctl restart --frobnicate,
zypper install --frobnicateandbrew install --frobnicate
passed. An option there is held to the program's manual now, and to
the command's own page where there is one: dnf reads dnf5's pages,
and-ystays right afterdnf5 install. git keeps its rule. A
--no-FOOis right where the manual names--FOO, sosystemctl list-units --no-legendis no longer doubted (#161). - Fixed:
read,do,recallandedit ? --sourcesent a credential
line as it was when its word followed an underscore, a dot or a
quote: the shapes an env file, a shell export and JSON give it
(DB_PASSWORD=,export AUTH_TOKEN=,"password":). They hold it
now; a word inside another word (passwordless) still passes. - Fixed: a slow disk could make the page refuse everyone with 503
busy. Each request wrote its line toforge.log, under one lock,
before its slot came back, so a stalled write filled all 64 slots.
The page's log lines go through a queue of 1,024 to one writer
thread now, in order, and no request waits for the disk. A line that
finds the queue full is dropped, and the next line says how many
were. A stop writes every queued line (#165). spark vernames the license and the repo again: under the logo,
the version line ends with MIT and the repo follows, then one blank
line before the prompt.spark ver --creditsand the end ofspark helpadd the engine and its repo, then the notice with CREDITS.md
before its last line. The install notice is two lines: the own-risk
line said again what "no warranty" and "read a command before you
run it" already say (#164).- Tests: the CI flakes that still hit
try/runs wait on real
conditions. policy_test sends a busy 503 once more after
/api/healthanswers; widget_pty waits for the prompt row under the
failure line; forge_smoke waits for the forge's root line and takes
a new port pair on its warning; the 1 s caps of the danger reading
and the secret hold time process time; presence prints its exit
codes; the tests that readforge.logwait for their lines (#165,
#144).
spark v1.111
- Fixed:
spark docalled a step that starts with a shell keyword
"not on this machine"; a loop, a test or a group (for,while,
if,[[) is offered now (#111). - Fixed: under zsh an unmatched glob stopped a whole step, so
mv *.jpg *.jpeg images/moved nothing; a zsh step runs with
+o nomatchnow, as in bash (#109). - Fixed:
spark edit ?marked a fix written as"was" -> change to "now"as[not in the text]; a span after an arrow and up to four
words of instruction is a proposal now, marked[proposed](#112). - Fixed: the
spark edit ?review ended without a newline, so the next
prompt joined its last line; it ends with one newline now, asspark readandspark askdo, never a second (#113). - Fixed: with
COLUMNSexported, ps cut the engine's command line, so
servewarned of old settings and serve off missed its own engine.
Every ps that reads a command line now asks with-ww(#119). - Fixed: setup at a terminal asked for sudo with no step needing root,
so a standard Mac account could not set up a client. Now it asks only
for the dry run's(sudo)rows, and names them (#121). - Fixed: after clear or Ctrl-L spark's row was drawn over a first-row
command line, hiding a!command; both widgets now open the row
above the prompt first (#116). - Fixed: the page's chat showed a reply's
**bold**marks raw; it now
draws them as bold by the terminal's rule, built as text nodes, so
*.txtand**/stay as typed and a stored thread reads the same
(#120). - Fixed: a
??with "top 3" in its words carried top's manual and
du's "-s display only a total", and dropped the*it followed. A
??now names its command and carries only a new tool's manual.
On the box the*stayed in 16 of 16, from 2 of 16 (#110).
spark v1.110
-
A Linux step of
spark do --sandboxruns under a seccomp filter:
sandbox.SECCOMP, 37 named syscalls with their reasons (ptrace, bpf,
the keyring, perf, userfaultfd, the modules, mount, the clock, raw
ports, file handles, setns and unshare, device nodes), built as bytes
per architecture and handed tobwrap --seccomp. The spike measured
git, gcc, make, pip, npm and cargo unchanged under it on Debian 12,
Ubuntu 24.04 and arm64; a step that sandboxes itself now fails with a
readableEPERM(#156). -
The
serverow says the engine's slots and the cells a slot, read
from/props(wire.props_slots):4 slots of 8192 cells. llama.cpp
opens four slots of its own accord and shares one unified cache of
SPARK_CTXamong them, so two 4k prompts fill it and a third purges
a slot's warm prefix; the next such log is read from the row (#157,
the engine's flags stay the owner's call). -
CI on main no longer runs the gate twice. lint's first step asks
whether the commit already passed on itstry/branch
(tests/land.sh --passed, the one question); when it did, lint
stops, every other job skips andgreenaccepts it, so main keeps a
run of its own for the landing's time and spends one minute, not six.
A commit that reached main any other way still gets the full run
(#145). -
The workflows move github/codeql-action to 4.39.1 and actions/cache
to 6.1.0, Dependabot's #114 and #115. Each sha is the commit the
upstream tag points at (#117, #118).
spark v1.109
spark do --sandboxruns on bubblewrap 0.8 to 0.10, Debian 12's and
Ubuntu 24.04's, where--overlaydoes not exist yet. There the
project is copied into the run and bound over itself (the copy mode,
sandbox.linux_mode), and the review and the apply read the copy
against the project as macOS's clone is read. The probe's refusal
under Ubuntu's AppArmor userns switch names the switch and the way
round it. CI proves it in debian:12 and ubuntu:24.04 containers,
tests/gate.sh full sandboxand thesandboxrow (#155).- A trust by hand is pinned.
spark mcp trust NAME onwrites
MCP_SERVER_<NAME>_PIN, the program's sha256 and each tool's flag as
listed then. A program that changed, or an older trust with no pin,
reads as not trusted and themcprow says so with the remedy. A
tool whose flags ask less than the pin's did, or one the pin never
saw, takes the typedyesagain until the person trusts the server
again (#154). spark setupfinds a llama-server of yours on llama-server's own
default port, 9931 since b11960, as well as onSPARK_PORT
(setup.ENGINE_PORTS), and offers the one that answers.SPARK_PORT
stays 8080 (#153).- The engine pin moves from llama.cpp b10689 to b11960 (2026-10-28,
0.8.0): the six tarballs fetched, their sha256 taken and their build
attestations verified against ggml-org/llama.cpp. Every flag spark
passes is in the new server's own table, and the glibc floor is the
old one, 2.34 on x86_64 and 2.38 on aarch64. llama-server's own
default port is 9931 there; spark's stays 8080 (#152). - The engine starts with a clean environment:
engine.ENV_KEEPand
engine.ENV_PREFIXESname what it gets (its path, its locale, the
GPU variables), and everyLLAMA_ARG_*of a shell is dropped, so
none can point it at another host. Its command carries--offline.
Theserverow warns on a connection the engine holds to another
host (engine.outbound) (#151). voice.envsays why the runtime is held at sherpa-onnx 1.13.8: 2.0.0
builds espeak-ng out and ignores the data_dir Kokoro reads
espeak-ng-data through, so the pt-BR voice would go silent. It moves
only with phonemes made on this side, for English and pt-BR both.
tests/voice_test.pyfails a bump before that (#150).
spark v1.108
spark serve https ownwrites the key identifiers into its leaf and
its root. OpenSSL 3 added them by itself and OpenSSL 4 does not, and a
python 3.13 client refused the chain without the authority key
identifier. The test fixture's pair carries them too (#158).- Three project skills under
.claude/skills/:spark-landruns
tests/land.shin a forked context, so the gate's minutes of output
stay out of the session;spark-releaseis the Releasing steps, run
by a person alone;spark-add-check-rowis the four obligations for
a row, with the fixture and the selftest proof (#148). - The full reference moves to
docs/REFERENCE.md.CLAUDE.md, what a
Claude session loads, is a short file now: it importsAGENTS.mdand
names the reference's sections by topic. A session starts with about
7k tokens of context instead of 50k, and reads the rest when a task
needs it.tests/docs_test.pyholds the file under 200 lines and
every section it names to the reference (#147). - CI: the Void job pins the Fastly mirror before xbps runs.
repo-default.voidlinux.org maps to any Tier 1 mirror, and the one it
mapped to on 2026-10-29 failed, then crawled for hours. spark serve offwaits for a unit only when one was running. A
server started by hand does not leave with a unit, so the verb took
the unit's 20 seconds for nothing, and so did its test (#133).- Every
sparkcall starts about 40 ms sooner on Python 3.12 and
newer. The TLS trust handler built a default context at import, and
loaded the system's CAs with it, for a connection it never made
(#134). - The full gate runs its suites side by side:
tests/gate_run.py, the
CPU count at a time, the longest first by the last run's times, each
log printed whole when its suite ends.smoke.pyruns in four parts
(--part line|read|machine|tree) andwidget_pty.pyin three
(session,living,rows), suites of their own. 18 minutes in a
row on 4 cores, about 4 on a 10-core Mac. CI runs them side by side
too, the selftest and chaos first, and stops at the first suite that
fails (#135, #136, #137, #138, #139, #140). - The streamed line's test waits on a latch, not a clock: the stub
held its hint for 1.5 s, and a driver that stalled longer reordered
them (#137). A chat step that waits for a prompt that never comes
fails now; two such steps cost 45 silent seconds (#138). Two tests
silently neededghandcurlon PATH, or a git identity, and
failed without them. - A Claude Code cloud session turns the repository's hooks on and
installs the gate's tools at its start, and names the command that
runs the full gate as a user of its own: the session is root, and
the knowledge index never builds as root (#128, #129, #130).
spark v1.107
spark stepsserves a desktop's whole tree: 128 tools and 32 kB of
tool lines a request (64 and 16 kB before). The tool list stands
first in the request and the words last, so the served prompt's
cached prefix holds the tools and a request pays its own words: on
the box a repeat with other words reused 194 tokens before, the
system message alone, and the tool list of about 1,300 tokens was
read again every time.spark do --tools, the desktop task adapter of v1.102, retires. Its
one client, the desktop's Home, asksspark stepsone bounded
question instead: these words, this tool list, which steps. The
twelve-turn continuation, the inspect step and the finish rule go
with it, and so do its help row, its completion word, its test and
its pages.spark mcp,spark stepsand the MCP client's gates stay
as they are.
spark v1.106
- Pi leaves spark. v1.104 moved Home tasks,
spark doandspark chat
onto a pinned Node runtime and the Pi SDK; v1.106 reverts v1.104,
v1.105 and v1.105.1 whole.spark do,spark chat, the page's chat
andspark do --toolsanswer through spark's own model path again,
as v1.102 did: a grammar over the typed decision, the 60 second wait
for a task decision. No runtime to install, no--agentwords, no
agent check row (42 rows again), no Node or npm in the SBOM, no
runtime job in CI.spark chatworks right after the install again. - The read of the code behind the decision: the host loop (inspect,
ask, act, observe, finish) was spark's and the desktop's since
v1.102; the runtime formatted one model call a request. The same 27
desktop task scenarios scored 14 of 19 and 4 of 8 on the native path
before v1.104 and 19 of 19 and 7 of 8 through the runtime, with the
task brief changed in the same commit, so nothing isolated the
runtime; a model call took 3.9 s median before and 6.2 s with it. spark updatehands the space back: a machine that ranspark do --agent onhas the runtime removed from spark's data directory by
the newagentrow of bootstrap, once, with one line said.- Kept from v1.104:
do.fitkeeps the head and the tail of an
oversized newest step and refuses a goal that cannot fit, and the
task brief's lines on literal note content and reminder titles. - The desktop's inspect control says what it does: it reads a
provider's state and makes its tools available on the next turn.
A task could claim a capability was unavailable when its provider
had not been inspected yet (#125)