Camera Control

Agent Vision

Use Agent Vision when the user explicitly asks for local Mac camera context.

Execution Discipline

Agent Vision camera requests are not repository tasks. Do not orient on the workspace, inspect files, check git state, read README or AGENTS files, or summarize the project before acting.

For /agent-vision snapshot, /agent-vision roast, and /agent-vision mood, the first shell command must be the camera capture command:

mkdir -p "$HOME/.codex/agent-vision/frames" && OUTPUT="$HOME/.codex/agent-vision/frames/agent-vision-$(date +%Y%m%d-%H%M%S).jpg" && "$HOME/.codex/plugins/cache/local/agent-vision/1.0.3/dist/agent-vision-capture-file" --output "$OUTPUT" --json

If this rule is violated, report that as command-dispatch behavior. Do not reinterpret repository inspection as part of the camera workflow.

Workflow

This skill controls the local camera. Do not inspect or roast the repository, source files, git state, README, or workspace unless the capture command fails and the exact failure requires local debugging.

For /agent-vision snapshot, /agent-vision roast, and /agent-vision mood, the first shell command must create the frame directory and run agent-vision-capture-file. Do not run git status, rg, find, ls, sed, cat, or any repository/workspace inspection command before the capture command.

For one-shot camera context, materialize the camera image to a JPEG file. Create $HOME/.codex/agent-vision/frames, choose an absolute output path inside it, then run:

"$HOME/.codex/plugins/cache/local/agent-vision/1.0.3/dist/agent-vision-capture-file" --output "$OUTPUT" --json

This is the normal image-access path. It calls the installed Agent Vision capture stack, decodes the returned image content, writes exactly one JPEG file, and prints JSON with the saved path and metadata.

For roast mode, materialize a JPEG file with the same command, then run a separate image-input pass:

codex exec --ephemeral --skip-git-repo-check -i "$OUTPUT" -- "Write exactly one playful roast of 400 characters or fewer based only on visible non-sensitive details in the attached image. Do not infer or attack protected traits, body size, age, disability, or other sensitive attributes."

Return the saved image link and the roast text from that image-input pass. The final response must include a Markdown image link using the captured absolute JPEG path, followed by the roast text. Do not write the roast in the current agent from Markdown, metadata, or memory. If the separate image-input pass fails, report that exact failure instead of roasting from metadata.

For mood mode, materialize a JPEG file with the same command, then run a separate image-input pass:

codex exec --ephemeral --skip-git-repo-check -i "$OUTPUT" -- "Analyze the attached Agent Vision camera image for current interaction-state calibration only. Return strict JSON and no prose. Use exactly these keys: presence, interaction_state, confidence, observable_basis, assistant_adjustments. presence must be one of present, absent, uncertain. interaction_state must be one of focused_neutral, frustrated_or_blocked, tired_or_overloaded, curious_or_exploratory, skeptical_or_evaluating, high_stakes_or_cautious, absent, uncertain. confidence must be a number from 0 to 1. observable_basis and assistant_adjustments must be arrays of strings. Apply these gates: if the user is absent, occluded, multiple people are visible, image quality is unusable, or confidence is below 0.40, return interaction_state uncertain or absent and use no mood-conditioned behavior; if confidence is from 0.40 through 0.69, include only low-risk clarity adjustments; if confidence is 0.70 or higher, include state-specific response delivery adjustments. Do not infer medical, psychological, intoxication, crisis, protected-trait, identity, or safety-state categories. Mood changes only pacing, verbosity, clarification threshold, evidence density, tone, and repair behavior; it must not change facts, permissions, approval behavior, user intent, or task scope."

Do not display the saved image, do not display the strict JSON, and do not explain the visual analysis or confidence band in the final response. Use the JSON only as ephemeral response delivery calibration for the current response or current task phase. User correction overrides the visual estimate. If the separate image-input pass fails or returns invalid JSON, report that exact failure instead of estimating mood from metadata, Markdown, or memory.

For streaming mode:

1. Do not call a tool.
2. Do not launch AgentVision.app, agent-vision-mcp, or agent-vision-capture-file.
3. Reply exactly: Agent Vision streaming is temporarily disabled in 1.0.3 while the runtime is being moved to an explicit start/stop design. Snapshot, roast, and mood still use one-shot capture and exit after the requested frame.
4. For stop-streaming requests, reply exactly: Agent Vision streaming is disabled in 1.0.3, so there is no Agent Vision streaming session to stop.

Slash Commands

• /agent-vision snapshot: materialize one JPEG file, then display it with a Markdown image link using the absolute path.
• /agent-vision streaming: report the temporary disabled message and launch no process.
• /agent-vision roast: materialize one JPEG file, run codex exec --ephemeral --skip-git-repo-check -i "$OUTPUT" -- "...roast prompt...", then return the saved image link and the roast text. The final response must include a Markdown image link using the captured absolute JPEG path.
• /agent-vision mood: materialize one JPEG file, run codex exec --ephemeral --skip-git-repo-check -i "$OUTPUT" -- "...mood JSON prompt...", then silently apply only permitted response-shape adjustments to the current response or task phase. Do not display the saved image or raw JSON unless the user explicitly asks to debug the mood analysis.

Agent Vision 1.0.3 has no streaming session. When the user asks to stop camera use or stop streaming, report that there is no Agent Vision streaming session to stop.

Treat requests such as "streaming off", "stop streaming", or "turn off the camera" as stop-streaming requests that launch no Agent Vision process.

Guardrails

• Agent Vision is macOS-only.
• The plugin uses the built-in Mac camera only in version 1.0.3.
• Snapshot, roast, and mood mode intentionally wait for a usable frame. If the camera returns black warm-up frames, the tool keeps the camera on, waits 5 seconds between attempts, and tries up to 3 total attempts before returning an error.
• Agent Vision 1.0.3 must not start agent-vision-mcp, AgentVision.app, or any Agent Vision camera-capable helper process at install, plugin load, idle Codex startup, unrelated prompts, streaming requests, or stop-streaming requests.
• Streaming is temporarily disabled until it has an explicit start/stop runtime independent of Codex plugin-load MCP lifecycle.
• Snapshot mode captures one usable frame, saves it as a JPEG file, and turns the camera off.
• Mood mode captures one usable frame, saves it as a JPEG file, analyzes that file through codex exec -i, and changes response delivery only. It must not change factual standards, permissions, approval behavior, user intent, or task scope.
• Mood mode must not create mood history, a training dataset, background recording, separate image archive, or any raw-image persistence beyond the existing saved JPEG frame path.
• Do not mention internal readiness metadata such as brightness values unless reporting an error.