v0.22.0

Highlights

• Chats: pick the model and effort. Two lists under the chat box choose the model and effort from the next turn on; Default leaves the agent's own setting. The choice holds for the chat until you change it.
  • The companion reports what each agent offers on that pc: Claude Code's aliases (fable, opus, sonnet, haiku), Codex's listed models (codex debug models) with their effort levels, and agy's models (the list agy's ACP offers once a chat ran there, agy models before that).
  • The server refuses any model or level the pc did not report.
  • The turn gets the choice the agent's own way: flags (--model/--effort, Codex -m and -c model_reasoning_effort) or over JSON-RPC (Codex turn/start, agy session/set_model).
  • agy has no effort list: Gemini's model names carry the level and agy's other models refuse --effort.
  • Codex keeps a turn/start model for the rest of the thread, so Default is named explicitly (config/read).

Needs this plugin version on the pc that runs the chat; with an older companion the lists stay hidden. No new settings, no breaking changes.

Image: ghcr.io/to0wnn/knowledge:0.22.0

v0.21.1

Highlights

• Chats: messages wait in line. A message sent while the agent answers is queued (up to 5 per chat) and goes in when the turn ends. With messages waiting and an empty box, the button turns into Force: it stops the running turn and the first waiting message goes in at once. Stop ends the turn and drops the line.
• Chats: Ctrl/Cmd+Enter sends, Enter is a new line.
• Help page checked against the code: folder linking from the companion page and the web UI, Codex hook trust by the companion, the import offer at first setup, more troubleshooting rows.
• Plugin: the offline context and the new-memo hint name ## Status / ## Decisions.
• UI: calmer working spinner, scroll position kept in chats, layout fixes.
• Session-start context is built once for all its parts.

No new settings, no breaking changes.

Image: ghcr.io/to0wnn/knowledge:0.21.1

v0.21.0

One design system for the web UI and the companion page.

• Buttons: one look in two sizes (36 px, 28 px; 44 px on phones), variants primary / secondary / quiet / danger / icon, 8 px radius. Inputs and selects match button height.
• Clickable rows and cards: hover tint, underlined title, a › when they open something. Disclosures always show ▸/▾. One link style, one focus ring.
• Status chips are the only pill shape and never look clickable.
• Measured: 14 → 6 button looks, 0 radii off the scale, 201 → 0 clickable elements that did not look clickable, 0 browser-blue links.
• Fixes: the usage "resets" chip only on Inbox and Chats; the companion page no longer shows "undefined live".

Image: ghcr.io/to0wnn/knowledge:0.21.0

v0.20.2

(0.20.2: the hooks run again on macOS's Python 3.9.) Follow, steer and take over a worker; and a secret-guard fix.

• Follow: a running or ended worker opens as a read-only conversation (Tasks, Inbox, Flow; #/chats/worker/<project>/<ID>). Output is sent only while someone follows it, masked, capped; the last events are kept when it ends.
• Steer: send a message to a running worker; it arrives after its next tool call and the worker carries on in its own session, worktree and permissions.
• Take over: stop the worker and continue its own session as an ordinary chat.
• Tasks: start, follow, stop and assign a task (to another agent or person, optionally pinned to one pc) from the Tasks list; the Machines list no longer starts workers.
• Secret guard: a known secret in a URL query, after & or before a literal \n is refused.
• Usage limits of Codex/agy stay visible between sessions until their window resets; a worker whose exit is unknown after a companion restart is judged by its outcome.

Image: ghcr.io/to0wnn/knowledge:0.20.2

v0.19.3

Usage limits stay visible between sessions.

• Codex and agy only report their usage while a session runs; the companion now keeps the last reading until that window resets (usage cannot go down before then), so the Machines list no longer says "unknown" between sessions.

Image: ghcr.io/to0wnn/knowledge:0.19.3

v0.19.2

Remote actions reach the pc in about a second.

• Long poll for machine actions: the companion holds its ask for actions open (/machine/actions?wait=, at most 25 s), so a click in the web UI (YOLO on/off, workers, Stop, Merge, …) is carried out in about a second instead of up to 10 s; an older server keeps the 10 s rhythm.
• Busy buttons: a machine action button shows a spinner and "Turning on…" / "Turning off…" until the pc reports back; the companion reports its new state right after an action.
• The companion finds agents installed in user folders (npm, mise, bun, volta) and reports each agent's installed plugin version.

Image: ghcr.io/to0wnn/knowledge:0.19.2

v0.19.1

Review finished work from the Inbox, and a tighter permission set for agents without YOLO.

New

• Review card: when a worker hands a task back with its branch, the Inbox shows a Review item: branch and target, commits ahead, changed files, the last commit message, what the worker said is done/left and about tests, and a "View diff". Merge runs on that pc only (git merge --no-ff into the branch the worker started from, never a push); it refuses when the main checkout is not clean or not on that branch, when a merge/rebase of yours is in progress, when the branch adds a file that already exists on disk (e.g. an ignored .env), when the repository's git config changed since the worker started, or when the branch touches .gitattributes/.gitmodules/.lfsconfig; a conflict is aborted ("merge by hand"). Reject keeps the branch and can send the task back to the agent on that same pc with your reason ([Agent@pc, ID-fix]).

Security

• Agents without YOLO (workers and chats) could run git log/git diff with --output=<file> and write anywhere. Claude now gets only fixed forms of git diff/git log without asking; agy no longer gets them without asking; Codex was already confined by its sandbox.

Image: ghcr.io/to0wnn/knowledge:0.19.1

v0.19.0

A finished app shell, a desktop board for your chats, an easier first run and an import of what your agents already know.

New

• The whole web UI in one shell: Inbox, Chats, Tasks (List | Flow), Machines, Notes, Map, Changes, Health, Admin and Help share one sidebar (a tab bar on phones), one header and one style. Inbox groups identical items and supports links to one item (#/inbox/<key>); Tasks shows where each task runs, with Start, Approve and Cancel (with a reason); Machines shows workers with their progress; Changes lists deleted notes with Restore.
• Chats board (desktop): #/chats?view=board shows every chat as a card on a canvas, one frame per project; drag cards (they stay), zoom, F to fit, a minimap. A glance at the board does not mark a chat read; opening it does.
• First run: on first start with no token file the server creates one admin token and prints it once in its log; the README quickstart uses it. The public /setup page has the steps per OS and agent.
• Import what your agents already know: knowledge-agent.py import [--dry-run] (also offered once during setup, default no) imports your CLAUDE.md / AGENTS.md / GEMINI.md and Claude Code's local memory into the right project through the server's own write path. It lists the files first, follows no links out of the expected folders, skips files that look like they hold secrets, and never overwrites.
• Proxy sign-in header is configurable (KNOWLEDGE_PROXY_USER_HEADER, default X-Authentik-Username), always together with the proxy secret. The root compose.yaml passes Matrix and rerank settings through.

Fixes

• The phone header keeps its search button on screen; contrast of amber text ≥ 4.5:1; the admin Projects table loads again inside the shell.

Image: ghcr.io/to0wnn/knowledge:0.19.0

v0.18.1

One app, approvals in chats, safer defaults and honest status. Covers 0.18.0 and 0.18.1.

Breaking / action needed

• Own device tokens need the self flag. A token without an owner no longer counts as the admin. Mark your own devices ("My own device" on the admin page, or "self": true in tokens.json); other owner-less tokens get their own machine record and their task lines are untrusted until approved.
• YOLO on the companion page needs the word YOLO typed, like the web UI. YOLO sets the agent's global setting: it applies to all sessions of that agent on that pc.

New

• One app shell: sidebar (Inbox, Chats, Tasks with List | Flow, Machines; Notes, Map, Changes, Health), a bottom tab bar on phones, a route per view so back/forward and refresh work, Ctrl/Cmd+K command palette, toasts instead of alerts, installable as an app (PWA).
• Approvals in chats: when an agent wants a tool that is not pre-allowed, the chat pauses and shows Allow once / Allow for this chat / Deny (Claude via --permission-prompt-tool, Codex via app-server, agy via ACP). Unanswered after 30 minutes means deny. Git's own files are never auto-allowed and git runs without fsmonitor or hooks.
• Public setup page /setup, and the admin token screen shows the server URL with a ready-to-copy install line per OS.
• "Needs you" messages on Matrix with a link to the item (admin setting needs_per_hour).
• Task flow: a failed worker says why and can be restarted; a worker on a pc that went offline shows under Needs you; results handed back to a person show under Needs you; cancel a task (dependent tasks keep waiting); task lines with an unknown start after ID or a date without a year are refused when written; worker progress (last commit, changed files, last log lines, masked); deleted notes can be restored (API).
• Health groups token-usage notes apart; hints follow the server's language.

Fixes

• Plugin updates report a failure instead of "up to date"; network errors no longer assume a home network.
• The companion runs far fewer git commands (94 → 22 per two minutes); lint is cached; static files answer 304.

Image: ghcr.io/to0wnn/knowledge:0.18.1

v0.17.0

The Agents cockpit gets an attention board, git pills and a mission map.

New

• Attention board as the cockpit's start screen, grouped per project with counts: Needs you (an agent answered and you have not read it, a refused tool call or error, a failed worker, a task waiting for your approval, a woken task nobody took), Running, Open (with where each task will run and a "Start on " button) and Done (24 h). Approve, Start and Stop act in place; Dismiss hides a Needs-you item until it changes (stored per person on the server). A phone shows one list in that order; a chat opens full screen with a back button.
• Git pill on chat cards and workers: task/KN-12 · ●3 · ↑2↓0 (branch, changed files, commits ahead/behind). Tap it for the diff, fetched from that pc through the companion (only the linked folder, capped at 200 kB, withheld when it contains a known secret).
• Mission map (Tools → Mission map): the task lines as an auto-laid-out graph, left to right, one lane per agent or person, edges for "start after", hand-overs, -cont and -done, colours per status. Open, running and the last 24 h of done tasks by default; "Show all" and a project filter. Tap a task for its line, approve, memo and worker. Phones get the chains as a list.

Behaviour

• Chats from before 0.17 count as read; an unread reply needs you for at most 7 days; a turn you stopped yourself does not.
• The companion counts the git pill once a round per linked folder, read-only (GIT_OPTIONAL_LOCKS=0, 10 s timeout).

Image: ghcr.io/to0wnn/knowledge:0.17.0