v0.8.9 — Every session starts knowing what you decided, four new agents, videos you can direct, and no JSON at setup

What you can now do. Open any session and have it already know your trust contract and what was learned lately. Report something about AIOS upstream without your vault leaving the machine. Ask for an explainer of a release, a quarter's objectives drafted from your vault, a dry run of publishing a book on Amazon KDP, or a short 3D video from a one-line ask. Give a spawned worker an effort level or an advisor. Paste text into another app and have the accents arrive intact. Set AIOS up without approving a wall of JSON.

Contributions from operators, each weighed by who it serves (#204, #205, #206, #207, #208): setup no longer stalls after /add-dir, and the App's setup path skips the GitHub step it cannot use; every uv run a command prescribes works inside a sandbox; an interrupted /close-session resumes instead of starting over, and uncommitted work counts as work; /today re-reads the note before writing it back, so it no longer overwrites what another session added; /today and /aios:housekeeping commit only their own files.

Every new session starts with the floor already loaded. Measured across a week of sessions, the instruction to run the context floor did not deliver it: some sessions never ran it, and most of the rest worked from its short map. A new hook (floor-at-start.py, SETUP Hook E) hands each new session a compact floor instead: first the path to the full floor and the instruction to read it, then the rules in INTENT.md that govern what the session may do, then the titles of the newest learnings. It stays under 9,000 characters, because Claude Code shows a longer hook message only as a 2 KB preview. Unattended routines get those INTENT.md rules, because they commit and send with nobody watching. It fires only for sessions started inside your vault, so a session in any other repository never receives your private context. A resumed session gets nothing new. AIOS_FLOOR_AT_START=0 keeps a session out.

Four new agents (40 bundled):

• aios-feedback turns something you noticed into one issue on the right repo. It checks for duplicates, removes names, paths and private content from the draft (it also catches names hidden in ways that only show once GitHub renders them), shows you the target repo and the full text, and files only after you say yes. Name a test repository to try it safely.
• explainer builds the cheapest thing that makes you understand a mechanism — usually one interactive HTML walk, every claim tied to its source. /close-session offers one when a lot shipped; /aios:ingest offers it beside the infographic.
• okr-coach drafts a quarter's objectives (at most three) from your vault as a draft you edit, in one file next to the quarter's first week, and scores them once at quarter end. It also builds an interactive page of the quarter (scores, weekly confidence, what each objective says no to), regenerated from that file. /7plan, /drift and /weekly-learnings read that file when it exists.
• kdp-publisher takes a finished book to Amazon KDP: checks the files, fills an answer sheet once per author, and names the decisions you cannot undo (DRM, KDP Select, categories, price). A dry run uploads nothing.

Make videos (motion-lab). Ask for "a 10-second 3D animation of our product", or a launch loop, a logo sting, a title card, an all-hands intro, and get back a finished MP4 plus a square cut for social. By default the subject floats on a dark ground, the camera moves, the moment that matters fills the frame, and every move lands on a spring. The session first proposes the whole brief for your yes (goal, story, look, camera, what to avoid), with a recommended answer for each. Then it renders frame by frame and checks its own work by measurement: framing, abrupt starts, seams, glitches, the brief's own numbers. It needs Python Playwright with Chromium, Pillow and ffmpeg, and says which is missing.

Workers you can tune. spawn --effort low|medium|high|xhigh|max sets a worker's effort. spawn --advisor <rung|model> gives a cheaper worker a stronger model to consult at hard decisions (Windows: -Effort, -Advisor). MODEL-ROUTING.md says when that pays (on the scale and fast rungs) and when raising effort is the better buy. No rung has a default advisor. The scale and fast rungs now run Sonnet 5.5 and Haiku 5.5; --tier mechanical is unchanged.

You are told when a session finished on a model nobody chose. A flagged message can move a session to an older model for the rest of its run. A new hook (Hook F) logs each switch, and /close-session and /close-day name any session that ended on a model other than the one it asked for.

Text for another app lands intact. On macOS, copying from a routine or a script could turn →, — and accents into ‚Üí, —, √°. Sessions now copy through hooks/clip, which forces UTF-8 and reads the clipboard back. If there is no clipboard, they give you the file path instead.

Setup without JSON. hooks/install-settings.py lists in plain words what each hook does for you and changes nothing; with --apply it merges them into your Claude settings, keeps everything else, backs up the old file and keeps its permissions. New installs also leave the framework's own test suite out of the working tree (git sparse-checkout disable brings it back).

Smaller fixes you may notice. /aios:housekeeping adds four checks. They report when the model ladder falls behind Anthropic's catalog, which skills cost context without being used, when a claude.ai-synced skill collides with a bundled one, and which old explainers can be pruned. Hooks read and write text as UTF-8 on Windows, where some read accents wrongly. Hook scripts check out with Unix line endings on Windows, and your notes are left untouched. Design skills point to Refero Styles as a second catalog of DESIGN.md files. A prompt the spawn inbox delivers again after two weeks is labelled as the inbox's, not read as your typing. A message sent through the inbox to a session that is busy now waits for its next pause (up to 24 hours) instead of being lost after 30 minutes, once your App or Glass supports it: the inbox contract and its tests in this release say how. A vault running a branch other than the framework's main one (a release being tested) now says testing that branch instead of offering an update that would undo it, and goes back to normal once that branch is merged. spawn given a flag with no value now stops with a message instead of hanging. install-settings.py keeps a symlinked settings file as a link, and refuses settings it cannot read safely, in plain words, before changing anything.

Action required — check each first; do only what applies:

1. Wire the two new hooks (E and F). Run python3 ~/aios/hooks/install-settings.py (Windows: python or py -3) and read the operator what it prints. It changes nothing; if it lists anything to add, run it again with --apply. Expect one window asking to allow the command; it writes only ~/.claude/settings.json and a backup beside it. If it says nothing to change, skip this.
2. One installer changes: hooks/claude-identity/install-wrappers.sh (Windows: install-wrappers.ps1). It is modified, not new, and writes the same file as before (~/.zshrc, or your PowerShell profile). The new spawn flags it installs may, when used, write one small settings file per advised worker in your temp folder. Show the diff and recommend yes: without it, spawn lacks the new flags and the corrected worker instructions.
3. One new skill (motion-lab) is registered by skills/setup.sh, which /aios:update runs when a bundled skill changes. Nothing to approve beyond that script's usual run.
4. Optional, for a vault cloned before today: leave the test suite out of the working tree with git -C ~/aios sparse-checkout set --no-cone '/*' '!/tests/'. Offer it once; skip it if they contribute to AIOS. Undo: git -C ~/aios sparse-checkout disable.
5. Start a new Claude session after updating, so the new hooks, commands and agents load.