Skip to content

Horizon AIOS — Development Pipeline

Placeholder backlog for tracking known work items, gaps, and research threads before formal issue management (GitHub Issues, Linear, etc.) is integrated. When that integration happens, migrate items here into the tracker and replace this file with a pointer.

Items are grouped by status, not priority — priority is implicit from the group.

In Progress / Recently Landed (this branch: foundational_misc)

  • AIOS switcher — complete. aios switch <name> rewrites all machine pointers, updates system PATH on switch, aios shortcut in bin/. See documentation/system/aios_switching.md.
  • AIOS uninstall — sbin/uninstall.ps1 / sbin/uninstall.sh landed. Section-by-section mirror of bootstrap; requires elevation; covers ACLs, git hooks, PATH, registry. Invokable via the aios uninstall [--dry-run] [--yes] subcommand, which delegates to the platform script. Both scripts support a no-elevation --dry-run preview and reject unknown arguments.
  • Future harness scope note — philosophy.md addition in progress. Acknowledges Claude-only current scope; names Ollama, Codex, LM Studio as future exploration targets.

Needs Merge

  • foundational_misc → master — contains: skills index, dependencies/footprint doc, Windows <brain-name>_group fix, system PATH on switch, aios shortcut, uninstall script, future harness note, and several other items from this session.

Owner-Handled (not tracked here)

  • ACL verification (C-1) — run bootstrap.ps1 as Administrator on the Windows reference machine with at least one provisioned brain, then horizon_aios_doctor.py to confirm Deny ACEs are applied and non-inherited. This is a verification step, not a code change.
  • QC Round 6 — post-QC5 consistency pass was interrupted. Owner will run this after the current branch stabilises. Scope: validate monitor brain-dir changes, log identity prefix, and any new surface from this session.

Known Gaps — Code

  • Skills for core sbin utilities — landed. Wrappers now exist: /doctor and /monitor as user-callable (skills_bin/); /harden, /create-brain, /remove-brain as admin-only (skills_sbin/). Uninstall is intentionally not exposed as a skill — it is an owner/system-level operation, invoked directly (aios uninstall / uninstall.ps1), not something an AI or an uninformed user should run via a slash command.

  • horizon_aios_remove_brain.py uninstall integration — verify that horizon_aios_remove_brain.py correctly deletes <brain-name>_group (not <brain-name>) on Windows following the group naming fix landed this session.

  • horizon_aios_maintain_logs.py scheduler — horizon_aios_setup_sync_schedule.py handles sync scheduling; no equivalent for horizon_aios_maintain_logs.py. A scheduled-task / cron setup step for log maintenance is undocumented. Consider a setup_log_maintenance.py or fold into horizon_aios_setup_sync_schedule.py.

  • Brain automation — Linux linger path unverified — the scheduled automation tier (horizon_aios_create_brain.py --automation scheduled) applies loginctl enable-linger on Linux, and horizon_aios_remove_brain.py runs loginctl disable-linger on teardown. Both Windows tiers (scheduled/daemon, LSA logon rights) are verified end-to-end, but the Linux linger path has only been code-reviewed, not run on a live Linux host. To close: on a real Linux machine, horizon_aios_create_brain.py <name> --automation scheduled then loginctl show-user <name> --property=Linger must report Linger=yes; horizon_aios_remove_brain.py <name> --yes then re-check must report Linger=no (or the user gone). Until then the Linux row in tested_configurations.md stays Partial. See deployment/brain_automation.md.

  • macOS — effectively undeveloped/untested — macOS support is the weakest of the three platforms: it exists only as POSIX-compatible code paths that have never been run on a real Mac. Specifics: bootstrap.sh is POSIX but unvalidated on macOS; horizon_aios_create_brain.py's macOS branch (dscl / dseditgroup / createhomedir) and horizon_aios_remove_brain.py's dscl -delete are unrun; the ACL model needs macOS-native equivalents (no NTFS ACLs); and brain automation has no applied path on macOS — horizon_aios_brain_logon_rights.py is Windows-only and there is no launchd equivalent of the Linux linger step, so both tiers are guidance only (LaunchDaemon/LaunchAgent). Closing this needs a dedicated pass on real Apple hardware: validate provisioning/removal, define the ACL approach, and implement

  • verify the launchd automation path. The macOS row in tested_configurations.md stays Untested until then.

  • Brain isolation test — Unix live probe unvalidated — horizon_aios_verify_isolation.py (criterion #5) is verified end-to-end on Windows; its safe ACL check works on Unix, but the --live run-as-the-brain probe (root → runuser/su/sudo -u) is a framework that has not been run on a live Linux/macOS host. To close: on a real Linux/macOS box, horizon_aios_verify_isolation.py --live must report BIN=READABLE SBIN=DENIED. Folds into the Linux/macOS verification passes above. See security/brain_isolation_test.md.

  • Uninstall Section 8 (Unix) — implemented, not validated on real hardware — uninstall.sh now has the Section 8 parity of uninstall.ps1: it strips the active_env.sh line from the owner's ~/.bashrc/~/.zshrc and unsets the two global git include.path entries aios setup writes (framework gitconfig + git_identity.local.gitconfig), acting as $OWNER (via sudo -u, falling back to HOME=). Syntax-checked (bash -n) but not yet run on a live Linux/macOS host. To close: on a real Unix box, run aios setup then uninstall.sh and confirm the profile line is gone and git config --global --get-all include.path no longer lists either entry (and unrelated include.path values are untouched). Folds into the Linux/macOS verification passes above.

Known Gaps — Documentation

  • Sound system — sounds/ directory, sounds.map, resolve_sound.py, and platform audio playback are undocumented. No entry in utilities.md; no authoring guide for adding new sounds or cues.

  • Sync system — RESOLVED (2026-07-04): sync_setup.md is now the dedicated operational reference for the two-lane sync workflow (what is synced, official vs. personal lane, overwrite vs. local-wins handling, the DCO --no-verify exception); system/distribution_and_updates.md and getting_started/updating.md cover the model. Prior gap: the workflow had no dedicated doc and sync_setup.md was stale relative to the two-lane implementation.

  • Brain authoring guide — no doc on what goes in a brain's CLAUDE.md, how to scope its tools/permissions, or how to design a brain persona. The template has defaults but no narrative guide.

  • tested_configurations.md — exists but content may be stale relative to current bootstrap and ACL changes. Should be validated after ACL verification (C-1) is done.

Research / Future Work

  • Ollama harness support — local model runtime. Open questions: harness config location, hook model, context-loading equivalent. See documentation/philosophy.md for the scope note.

  • Codex / OpenAI CLI harness support — Open questions: same as Ollama plus cloud vs. local auth model.

  • LM Studio harness support — local GUI with OpenAI-compatible API. Open questions: headless / CLI mode availability, hook equivalents.

  • GitHub / project management integration — when this repo moves to formal issue tracking, migrate items from this file into the tracker. Replace this file with a pointer to the project board.

  • Brain marketplace / registry — a mechanism for discovering, installing, and versioning shared brain configurations. Currently file_structure_invariants.md §7 notes that marketplace plugins coexist with AIOS skills; no brain equivalent exists.

  • Multi-operator brain sharing — server.md documents the multi-operator pattern (each human operator gets their own OS account); no mechanism for two operators to share access to a single brain. May not be desirable by design, but worth an explicit decision.

Completed This Session (reference)

For full detail see the session handoff. Short list:

  • Worktree cleanup (9 stale branches removed)
  • context_cost.py + /context-cost skill
  • Context loading documentation (documentation/context_loading.md)
  • @-import clarification (documentation/authoring/claude_md_authoring.md)
  • Utilities index (documentation/utilities.md)
  • Skills index with onboarding/offboarding reference (documentation/skills.md)
  • Dependencies and system footprint reference (documentation/getting_started/dependencies_and_footprint.md)
  • Windows per-brain group as <brain-name>_group (code + docs)
  • Bootstrap system PATH (Machine-scope Windows, /etc/profile.d/ Linux/macOS)
  • AIOS switcher system PATH update on switch
  • aios / aios.ps1 shortcut wrappers in bin/