* fix: never auto-select embed-only models for chat
Ollama lists nomic-embed-text alongside chat models. Auto-picking
models[0] / recommending the only available id selected the embedder
and every generation failed with HTTP 400 "does not support chat".
- Filter embed-only ids out of GET /v1/models (chat picker)
- Exclude them from /v1/recommended-model; return empty when none left
- Frontend setModels prefers chat models and clears a bad embed selection
- Regression tests for mixed, embed-only, and classifier cases
* fix: harden chat model capability filtering
---------
Co-authored-by: Elliot Slusky <elliot@slusky.com>
* fix: use a signal-free liveness probe for the daemon on Windows
`_read_pid()` probed the recorded pid with `os.kill(pid, 0)`. That is a
POSIX idiom: on Windows signal 0 is `CTRL_C_EVENT`, so the call routes to
`GenerateConsoleCtrlEvent` rather than testing for existence, and raises
`OSError` (WinError 87, "The parameter is incorrect") for any pid that is
not a live console process-group leader — which includes both dead pids
and the detached server `jarvis start` creates.
That single call produced three symptoms. `jarvis status` propagated the
error and crashed. `_read_pid`'s `except OSError` swallowed it for a
running server, so `status` and `stop` reported "not running" and deleted
a live pid file. And because the probe *sends* a console control event
rather than merely asking, running `status` against the daemon could
terminate it.
Add `_pid_alive()`, which opens a process handle and checks it on Windows
and keeps the signal-0 probe on POSIX, and use it for both liveness
checks. `SIGKILL` in the stop path is now reached on Windows for the
first time, so guard it — it is POSIX-only, and `SIGTERM` already maps to
`TerminateProcess` there.
The existing round-trip test mocked `os.kill` to succeed, which is why
this passed CI on Linux while failing on every Windows run. Point it at
the new seam and add `TestPidLiveness`, which exercises real pids so the
platform behaviour is actually covered.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* style: format daemon tests with CI Ruff
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: Elliot Slusky <elliot@slusky.com>
Previously the ws_bridge send loop blocked forever on queue.get(),
never learning the client left. When the server service stopped,
uvicorn waited for open WebSocket tasks until systemd SIGKILLed
after TimeoutStopSec=90s. Now each iteration races recv+send; a
completed receive means the client disconnected => break the loop.
Every other built-in tool is imported here specifically to fire its
@ToolRegistry.register() decorator at package-load time; these two
were missing, so their test_registered tests only passed when some
unrelated test (via agent_manager_routes.py, channels_cmd.py, or
deep_research_setup_cmd.py) happened to import the module first in
the same process. Under pytest-xdist that's worker-distribution
dependent, so adding an unrelated test file could flip either test
from pass to fail.
register_cron() ran unconditionally on every 'jarvis serve' startup and
create_task() persists to scheduler.db, so each restart added another
copy of the daily proactive cron. On a real install 68 duplicates
accumulated; when due they fired back-to-back and monopolized the
single-slot local inference queue, stalling interactive chat.
register_cron() is now idempotent: an existing active task is reused and
surplus duplicates are cancelled. Also stop the notification channel
from calling connect() — a second getUpdates poll loop on the same bot
token makes Telegram return Conflict and kills the main listener.
The Vite dev proxy forwards `/v1` as plain HTTP with no `ws: true`, so the
WebSocket upgrade for `/v1/agents/events` is never proxied. The socket does not
open, does not error, and does not close — it just sits silent — so every live
agent view is empty under `npm run dev` while working in a production build,
where the frontend is served from the same origin as the API.
That covers the per-agent live trace on the Agents page, which subscribes via
`useAgentEvents` (`frontend/src/lib/useAgentEvents.ts`).
The silence is what makes it costly: with no error to see, it reads as "no
events are being emitted" rather than "the transport never connected", so the
search starts on the server side.
Verified on Windows 11 with `jarvis start` running: before, a
`new WebSocket('ws://localhost:5173/v1/agents/events')` from the dev page never
fired open, error or close within 6s. After, it opens, and a real agent tick
delivers 8 events (agent_tick_start, inference_start/end, tool_call_start/end,
agent_tick_end).
`changeOrigin` is set alongside so the upgrade request carries the target's
host, which some setups require when the API is not on localhost.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
``SessionStore.__init__`` passed its ``db_path`` straight to
``secure_create()``, which touches the path and chmods it. ``:memory:`` is
a SQLite sentinel, not a filename, so this tried to create a file literally
named ``:memory:``.
On Windows ``:`` is illegal in a filename, so construction raised
``OSError: [Errno 22] Invalid argument: ':memory:'`` and the two
``tests/server/test_channel_bridge_deep_research.py`` tests failed there.
Elsewhere it succeeds and is merely wrong: it leaves a stray ``:memory:``
file in the working directory, and because ``Path(":memory:").parent`` is
``.``, ``secure_mkdir`` chmods the working directory itself to 0o700.
``KnowledgeStore``, ``TelemetryStore`` and ``TraceStore`` already guard this
exact case; ``SessionStore`` was the one store missing the check. Apply the
same guard, with the same comment.
Adds two regression tests: one that an in-memory store is usable, one that
constructing it creates no file. The second fails on every platform without
the fix, so the bug cannot silently return on Linux or macOS.
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
``TestMemoryRoutes.test_search`` and ``test_stats`` assert the status code
is in ``(200, 500)``. That list dates from the initial commit; #527 later
made the memory routes raise 503 when the native ``openjarvis_rust``
extension is missing, so both tests now fail on any checkout where the
extension has not been built — which is every contributor who has not run
``maturin develop``.
The failure is spurious: these two tests only check that the routes are
wired up, and their own comment ("May fail if SQLite not set up, that's
ok") says an unavailable backend is tolerated. 503 is exactly that case,
and it is already asserted deliberately in ``TestMemoryRustMissing``
directly below.
Add 503 to the tolerated set via a named constant, so the reason is stated
once rather than repeated as a bare literal.
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Someone asked in Discord how to give Jarvis access to their whole
machine and hit a confusing failure. With no ~/.openjarvis/config.toml,
tools.enabled and agent.tools both come back empty, and SystemBuilder
builds the agent with no tools at all. It reads like a permissions
problem but it's just missing config, and nothing in the docs points
you anywhere useful.
Adds docs/user-guide/system-access.md, covering the empty tool list as
the usual cause, what shell_exec and the file tools actually reach,
which entry points prompt for confirmation and which quietly
auto-approve, Full Disk Access on macOS and which process needs it, and
the fact that there's no computer use at all, so Accessibility and
Screen Recording grants buy you nothing on their own.
Also adds a full-system-access.toml example to copy from.
Fixes two config tables that describe security.enforce_tool_confirmation
as requiring confirmation before tools run. The loader accepts the key
but nothing on the execution path reads it, so anyone setting it gets
assurance they don't actually have.
``jarvis start`` spawned the server with ``start_new_session=True``. That is
POSIX-only — CPython's Windows ``_execute_child`` names the parameter
``unused_start_new_session`` and ignores it — so on Windows the server
inherited the launching console instead of detaching from it.
Closing that console, or logging off, therefore delivered CTRL_CLOSE_EVENT
to the server. Observed in the wild as the daemon dying overnight, with
forrtl: error (200): program aborting due to window-CLOSE event
in server.log (the Fortran runtime under NumPy handles the event and
aborts). ``jarvis start`` looked like it worked: it printed a PID, wrote the
pid file and exited 0, and the server ran for as long as the console stayed
open. Registered as a log-on scheduled task, this means the machine comes
back up with no backend.
Pass DETACHED_PROCESS on Windows so the child gets no console at all, plus
CREATE_NEW_PROCESS_GROUP so a Ctrl-C in the parent console cannot reach it.
POSIX keeps start_new_session.
Verified by attaching to each spawned process with AttachConsole():
start_new_session=True attaches successfully (the child shares a console);
DETACHED_PROCESS fails with ERROR_INVALID_HANDLE (no console exists).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
TelemetryStore opened SQLite without WAL, so concurrent readers (server, aggregator, dashboard) hitting the database under inference load raised SQLITE_BUSY, and every insert committed immediately, paying fsync on each record.
- Enable PRAGMA journal_mode=WAL with synchronous=NORMAL and busy_timeout=5000, matching TraceStore.
- Batch inserts in memory under a lock and flush via executemany() when a batch reaches batch_size (default 50), when a batch goes stale, on any read through the store, and on close().
- Run a background flusher thread (default 5s interval) so a partial batch written just before traffic stops still becomes visible to other connections; close() stops the thread with an ordering that prevents touching a closed connection.
- Tests cover batching deferral, read-triggered flushes, stale-batch flushes, and close() behavior.
Fixes#560
Co-authored-by: Elliot Slusky <elliot@slusky.com>
The chat area re-armed autoscroll whenever the user was within 100px of the bottom, so scrolling up during a streaming response fought the incoming content ticks and produced jitter.
Autoscroll now disengages on any upward scroll (direction-based, no distance threshold), re-engages when scrolled back within 2px of the bottom (tolerating sub-pixel rounding at fractional zoom levels, where the at-bottom residual can reach 1px), and ignores sub-1px upward movement so macOS elastic-bounce settling does not disengage it. Sending a message pins the view to the bottom even if the user had scrolled up to read earlier messages.
goatoush