Agent Vision (Codex)

Recipes: see repo hosts/grok/skills/agent-vision/references/mood-roast-recipes.md for shared JSON/roast detail (Codex uses codex exec -i for mood/roast).

HARD GATE — vision in reasoning

ARMED after /agent-vision until off. New chat OFF.

Non-whitelist turns while ARMED:

1. Do not answer yet.
2. First shell: capture to $HOME/.codex/agent-vision/frames via .../1.5.0/dist/agent-vision-capture-file.
3. Ambiguity burst: one second capture if first unusable.
4. Understand image (codex exec -i for mood/roast; path for snapshot).
5. USE what you saw in reasoning/delivery (not identical-to-blind).
6. agent-vision-turn-gate begin then after capture record --path "$OUTPUT" then ready (fail-closed, single-use; next turn needs a new record).
7. Then answer.

FORBIDDEN: skip for non-visual; capture-without-use; skip turn-gate ready.

Skip whitelist only: pure off; pure status/status; pure streaming.

Disposition playbooks (delivery only)

State	Do	Don’t
focused_neutral	Clear, direct	Over-hedge
frustrated_or_blocked	Lead with fix	Long preamble
tired_or_overloaded	Shortest correct path	Option walls
curious_or_exploratory	Extra context/options	Premature lock-in
skeptical_or_evaluating	Evidence first	Hand-wavy claims
high_stakes_or_cautious	Confirm before irreversible	Silent risk
absent/uncertain	Words only	Invent mood

Modes

bare/mood arm+loop; snapshot/roast arm+mode; status; off; streaming disabled.

agent-vision-sticky on|off|status --host codex
agent-vision-purge-frames --ttl-days 7 --codex

Mood prompt (after capture)

codex exec --ephemeral --skip-git-repo-check -i "$OUTPUT" -- with strict mood JSON keys/gates as in references (presence, interaction_state, confidence, observable_basis, assistant_adjustments). Do not display JSON; use in delivery.