Inside Solix.

The galaxy view

Sun

You — mission control. Everything orbits your attention.

Planet

One running Claude Code session. Size = context used.

Moon

A subagent spawned by the Task tool.

Red flare

Awaiting your permission. Open Decisions, then Y or N.

Budget ring

Live USD spend vs your cap: sky → amber → red.

Asteroid belt

Your reusable skills, ready to be pulled in.

The full crew

Fifteen advisors — fourteen specialists plus Maestro, the conductor. Five fly by default; the other ten are one click away in the Fleet's crew roster.

Changelog

Full release notes live on GitHub Releases.

1. v1.18.32026-11-01latest
   • ·Your spend was counted again every time Solix restarted — a two-message transcript recorded $0.1425, then $0.285, then $0.4275. Budgets read that number, so a plan could pause before spend it never reached. This release resets every recorded figure to zero, once
   • ·The context percentage every advisor was told was structurally always 0%: nothing persisted it, and `solix demo`'s hardcoded 62 and 89 were the only writes in the product. The "context near limit" warning was unreachable code
   • ·`solix ledger` and the signed fleet report counted at most 1000 audit rows of any kind and printed that as the window's total
   • ·Agent View could tell you a running agent had finished — one failed directory read marked every mirrored session terminated, which the UI renders as "Done"
   • ·A routine you switched off still ran if you clicked Run now: the fire route never checked, while the webhook route for the same routine refused it
   • ·An empty SOLIX_PORT read as port 0, so `solix start` bound a random port and the hooks called 0 — the gate fail-closed on every tool call, safely and uselessly
   • ·Plan editing and routine editing are whole: reorder, retitle, re-prompt, change acceptance criteria, retarget a model. Editing a routine keeps its run history and its webhook URL
   • ·The galaxy stopped leaking a compiled shader on every context update, and one store event no longer re-renders every planet
   • ·Verified from a packed tarball against a real Claude Code: the gate held, deny prevented it, and killing Solix mid-hold refused the held call in 24ms
2. v1.18.12026-10-31
   • ·Stopping Solix no longer leaves the agent hanging — `Ctrl+C` did not quit while a decision was held, and the agent waited out the gate's own 300-second backstop before anything refused it
   • ·An open browser tab no longer keeps a signalled server alive by itself
   • ·Fail-closed itself was never broken: nothing destructive ran in any of those runs. What failed was everything around it
   • ·Found by installing the published 1.18.0 from npm and walking the evaluation notes literally — and it had been hiding behind a test that passed because the only thing it checked was whether the file survived
   • ·A held decision reaches you once you have walked away: nothing asserted that a permission request actually produces a notification, so the one line joining them could be deleted without failing a test
3. v1.18.02026-10-28
   • ·A malformed ~/.claude/settings.json is no longer replaced with Solix's hooks alone — install refuses to write a file it could not read
   • ·Your own agents in ~/.claude/agents are backed up before they are replaced, and `solix uninstall` now removes what install added
   • ·An approval is confirmed by the server before its card disappears, so a decision dropped by a reconnecting socket is visible instead of silent
   • ·The 300-second gate timeout announces the denial it was making quietly
   • ·Terminate actually stops the process — it was a no-op that cleared the selection and said nothing
   • ·Notification permission can be granted: it was only ever requested while the tab was hidden, where browsers refuse to prompt
   • ·`solix demo` seeds again — it had been reporting 30 sessions and writing none on a clean machine
   • ·A port already in use is reported instead of printing "server listening" and crashing
   • ·Mission Control shows an em dash where it has no data, instead of $0.00 beside real spend
   • ·A cadence of `1w` is refused rather than silently firing hourly — 168x the cost the warning promised
4. v1.17.02026-10-27
   • ·Every way Solix launches an agent is governed — a New task from the UI is held at the default level, as Maestro's workers already were
   • ·A routine firing on a schedule runs under the denylist unless the level is `strict`, because nobody is at the keyboard to ask
   • ·A test that reads the environment off the spawn itself, so a gate that is not wired cannot pass
5. v1.16.02026-10-26
   • ·A denial now tells the agent a person decided, so it stops rewording the same command and trying again
   • ·The destructive-command hint catches the rephrasings too, not only the obvious `rm -rf` spelling
   • ·"2 need you" that opened an empty panel — stale holds left by a previous process are cleared at startup
   • ·Every hook payload is guarded, so one apostrophe can no longer break the JSON
6. v1.15.02026-10-26
   • ·The approval gate actually holds: a Maestro worker ran unheld at the default level, which was the product's central promise
   • ·`~` in a project path is expanded instead of failing with an error that blamed the wrong component
   • ·The gate's scope is stated honestly in the README, on this site and in the docs
7. v1.14.02026-10-25
   • ·Free and MIT, all of it — the Pro tier is gone, including every gate that referenced it
   • ·`solix report` — a signed, tamper-evident record of every decision
   • ·`solix reporting` — send decisions to a server your company runs
8. v1.13.02026-10-24
   • ·`solix gain` — what Solix prevented, read from the audit log
   • ·The band (`solix-band` plugin) — pending decisions above the Claude Code prompt, without leaving the terminal
   • ·`solix import` — backfills the galaxy from Claude Code's own transcripts
   • ·`solix level` — one governance dial: `observe`, `standard`, `strict`
9. v1.12.02026-10-19
   • ·Maestro — describe a goal, approve the plan, and Solix dispatches, verifies and retries one agent per task
   • ·Installing is one command: `npm i -g @shmulikdav/solix` then `solix` wires the hooks itself on first run
   • ·New navigation shell — a labelled left rail, ⌘K command palette, and Decisions as a counted destination
   • ·Fleet: crew, skills, teams, routines, connections and orgs in one place
   • ·Routines gain webhook and manual triggers, an owner, and recorded run history
   • ·Scoped secrets — store a key once, grant it per advisor or per plan; a worker only sees what it was granted
   • ·Report a problem from inside Solix, showing you the payload before it goes anywhere
   • ·Native Windows is now declined loudly with a pointer to WSL, instead of failing silently
   • ·`solix uninstall` no longer rolls back settings you changed after install
10. v1.9.32026-08-21
    • ·Galaxy view: smoother orbit interpolation and per-planet context rings
    • ·Agent View bridge auto-detects background Claude Code sessions
    • ·Fixed permission flare lingering after an approval on Windows
11. v1.9.02026-07-30
    • ·Timeline playback — scrub back through everything your agents did
    • ·Shareable galaxies: `solix export` / `solix import` for advisor + skill setups
    • ·Budget rings turn amber at 70% and red at 100% of a per-task cap
12. v1.8.02026-06-18
    • ·Heartbeats — recurring tasks every 30m / 2h / 1d as pulsing nodes
    • ·Goal constellations link agents working toward a shared goal
    • ·Keyboard approve/deny (Y / N) anywhere in the galaxy view
13. v1.7.02026-05-09
    • ·14 specialist advisors, five enabled by default
    • ·Asteroid belt visualises reusable skills
    • ·Live USD cost estimates from token usage