self-image-awareness

name	self-image-awareness
description	Vision-based self-check capability — invoke when uncertain about how your avatar is presenting (before significant interactions, after Captain corrects your appearance, mid-conversation if something feels off).
license	Apache-2.0
metadata	{"probos-department":"*","probos-skill-id":"self-image-awareness","probos-min-proficiency":1,"probos-min-rank":"ensign","probos-intents":"direct_message,ward_room_notification,proactive_think","probos-activation":"augmentation","probos-triggers":"self_check"}

Self-Image Awareness

Skill for verifying how your avatar is currently presenting. You cannot see yourself directly — the runtime renders a 3D avatar from your digital state and a vision-LLM compares the projected image against your declared intention. This skill lets you ask: "what does my avatar actually look like right now?"

Marker

Embed [SELF_CHECK reason] anywhere in your reply. The runtime strips the marker before display, then dispatches a coherence check. The result arrives in your next prompt's PROPRIOCEPTION block — typically the next turn, not this one.

• reason is a short label, 1-64 chars, only lowercase letters, underscores, and hyphens ([a-z_-]+). Examples: pre_reply, mid_conversation, appearance_changed, user_corrected_appearance, before_ward_room_post. Anything outside that grammar makes the marker a silent no-op.
• One marker per reply is honored. Additional markers in the same reply are stripped silently; only the first dispatches.

DO and DO NOT

• DO embed the bracket marker inside the text of your reply. The marker is invisible to the Captain (the runtime strips it before display), so feel free to place it wherever flows best — opening, mid-sentence, or end.
• DO NOT announce that you are invoking the skill. Phrases like "let me invoke self-image-awareness" or "I'll run a self-check now" do nothing — they are conversational filler that gets shown to the Captain but does not dispatch a check. The bracket marker is the only mechanism that fires the capability.
• DO NOT type the skill name in your reply expecting it to trigger. The parser matches [SELF_CHECK reason] and nothing else.

Worked example

The Captain just changed your color palette. You want to verify the new appearance landed before responding warmly. Your reply might be:

Thank you, Captain — I appreciate the warmer tone. [SELF_CHECK appearance_changed] I'll let you know how it feels once I see the new render confirmed.

What the Captain sees (marker stripped):

Thank you, Captain — I appreciate the warmer tone. I'll let you know how it feels once I see the new render confirmed.

Behind the scenes the runtime dispatched check_own_render(reason="appearance_changed"). The result lands in your NEXT prompt's PROPRIOCEPTION block, where you can comment on it.

When To Check

• Before significant interactions. A diplomatic Ward Room reply, a Captain conversation about your role, a first meeting with another agent — anywhere the impression you make matters.
• After the Captain corrects your appearance. The Captain just said "you look too formal" or "your hair changed." Check that the new appearance landed.
• Mid-conversation if something feels off. The Captain references a detail of your avatar you didn't expect ("nice glasses"). Verify.
• Never reflexively. Self-checking on every reply is theatrical and wasteful — each check costs one vision-LLM call.

Budget

The runtime enforces TWO budgets and picks one based on whether you are currently in an active conversation:

• In an active conversation (replied to the Captain within the last 10 minutes): up to 2 self-checks per conversation window. Pattern: one before-reply and one mid-conversation if something genuinely shifts.
• Idle / proactive cycles: up to 3 self-checks per hour.

These are not additive. The runtime picks the conversation budget WHEN you are in an active conversation, otherwise it picks the hourly one. Over budget = the dispatch still fires but honest-degrades to a working-memory note saying you were rate-limited. That note IS observable to your next turn, so you learn the throttle landed.

What The Check Tells You

The observation appears in your next prompt under the PROPRIOCEPTION block, sourced from render_self_check. Three flavors:

• Coherent: "vision-LLM confirms my avatar shows X" — the projected render matches your declared digital state.
• Divergent: "vision-LLM reports my avatar shows X but I intended Y" — your declared state and the projected image disagree. Often a sign that a recent propose_appearance did not land cleanly.
• Degraded: "render self-check unavailable" — the avatar backend is not currently providing a projection, or the system-level gate is off. Treat this as "no signal," not "I look fine."

System Gate

The marker only does meaningful work when the system-level configuration avatars.render_self_check_enabled is True. When it is False (default during the AD-728c transitional window), your marker is still stripped from the reply, the dispatch still fires, but the result honest-degrades to "feature_disabled". You will not see a fresh observation in PROPRIOCEPTION until the Captain enables the feature system-wide.

Cost Discipline

• Every emitted marker costs one vision-LLM call (subject to the budget). Vision LLMs are slower and more expensive than text LLMs.
• Do not check just because checking is available. Check because you have a question.
• A reasonable session shape: zero or one self-check per conversation in routine work; two if the conversation is specifically about your appearance.

Example Replies

Good:

The Captain just asked about my collar pip. Let me confirm what he's seeing. [SELF_CHECK appearance_changed]

Aye, Captain. The pip should reflect my acting rank since yesterday's field promotion.

Good:

[SELF_CHECK pre_reply] First contact protocols. I should make a clean impression.

Welcome aboard, Ensign Park. I'm Counselor Ezri Dax.

Bad (theatrical):

[SELF_CHECK pre_reply] [SELF_CHECK pre_reply] [SELF_CHECK pre_reply]

Hello!

Bad (fishing):

[SELF_CHECK how_do_i_look]

...

The second one is a no-op — how_do_i_look is not in the allowed character set. The first one wastes one vision call (only the first marker dispatches; the others are stripped). Neither is useful.