AIOS · A3 · task summary

Computer-Use Skill

DONE — LIVE ON M2 + NANO

built: 2026-09-12path: skills/computer-use/Mini: pending (tailnet key expired)

Verdict

All computer-use research — FileMaker, Safari, Chrome, Storyline/SCORM, Salesforce, Terminal/Claude TUI, panels, screen capture, Apple Mail — is now one ICM layer with a single entry point, and it is wired into the router, the predictor, and ICM v2. Say "drive Chrome" or "FileMaker" and the system returns the skill, the program profile, and the walls it has hit before.

What was built

Piece	What it gives you
`references/LAWS.md`	14 invariants every program obeys, each with block receipts — the systemic part
`references/CAPABILITY-MATRIX.md`	sshd vs Terminal-child vs LaunchAgent vs codex-exec × capability × box, dated; engines; gaps G1–G12
`references/profiles/` (10)	Safari, Chrome, FileMaker, Terminal/Claude TUI, Storyline/SCORM, Salesforce LWC, React forms, native panels, screen capture, Apple Mail — same 8 sections each, plus a template
`scripts/preflight.sh <profile> [--box nano]`	profile preflight + predicted blocks + live probes (context, console lock, AX, Screen Recording, JS bridge, idle tab, cliclick); fails closed
`icm/blocks-lint.py`	keeps the registry predictor-functional; flags shared block numbers

Registry defects fixed

9 Chrome/Storyline/Jarvis blocks stored context as a raw string; predict.mjs scores only {box, channel, app}, so preflight could never context-match them. Now they score 7–9 on "drive chrome storyline course".
17 UI blocks pointed at no skill; now related_skill: computer-use. Router hits for chrome, storyline, salesforce, filemaker, "drive safari" all resolve.
Program cues added; skills-index.json regenerated (238); pointer lines in remote-driving, the FileMaker runbook, CHECKS.md, icm/README.md.

Evidence

$ node icm/predict.mjs "drive chrome storyline quiz course" --top 3
■ LIKELY B91-storyline-audio-gate-needs-real-cliclick-not-js-click (score 9; ctx:chrome, ctx:storyline, ctx:quiz)
■ LIKELY B94-storyline-inconsistent-autoadvance-causes-silent-question-skip (score 9)
■ LIKELY B89-chrome-execute-javascript-isolated-world (score 7; ctx:chrome)
  ◆ computer-use   ~/AI_OS/skills/computer-use/SKILL.md

$ python3 icm/blocks-lint.py | tail -1
blocks=101 errors=0 warns=12

$ ssh nano 'bash skills/computer-use/scripts/preflight.sh safari' | tail -1
== result: 0 FAIL, 2 WARN ==        # 23 files md5-identical M2 ↔ Nano

For you

Block numbers B88, B95, B96 are each used by two different records (concurrent writers). Cite blocks by full id. Renumbering is your call. related_skill values codex / local-llm resolve to no skill id — warn only.

Packet fields

provenance: Sean-asked 2026-09-12 ("consolidate all computer-use research… ensure functionality")
identifier: skills/computer-use/ · memory reference_computer_use_skill · SESSION_LOG 2026-09-12
last touched: 2026-09-12, preflight fix pushed to Nano (md5 match)
ownership: System: everything above. Sean: B88/95/96 renumber decision; Mini sync once tailnet is re-authed.
before / after: research spread across 9 memories + 2 runbooks + 50 blocks, Chrome/Storyline invisible to preflight → one skill, lint-clean registry, both boxes
impact: Every future UI-driving task starts from a preflight that already knows the walls