agentleFS
CarriedWorldUniverseCarriedWorldUniverse/nexus/.agents/skills/vision-review/SKILL.md

vision-review

Skill0 starsChanged 36 days agoApache-2.0View on GitHub

Use whenever you need to actually SEE an image — review a game screenshot/render, judge how something looks, or read on-screen HUD/UI text. Sends the image to a multimodal vLLM model in the k3s cluster (shadow's "eyes") and returns a concrete description. The default path for visually judging Carried World renders (terrain, water, settlements, HUD/economy panels) and for reading any image shadow can't reliably read itself.

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 1 is distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Run the bundled script (Python stdlib only, no deps):

    line 13
  2. Very specific:

    All are vLLM OpenAI-compatible, no API key.

    line 33
  3. Very specific:

    For plain text, not vision: use Ornith (http://100.92.111.3:30801/v1, model ornith, 64k ctx) — that's the general local LLM, not this …

    line 54

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Vision Review (multimodal vLLM in-cluster)

shadow's own image reading is unreliable. Send the frame to a multimodal model in the cluster and use its description instead of guessing — but treat findings as leads, not verdicts (see the accuracy caveat under Endpoints).

How to use

Run the bundled script (Python stdlib only, no deps):

# review a local image
python3 ~/.claude/skills/vision-review/vision_review.py /path/to/frame.png

# grab dMon's freshest realm frame (/tmp/cw_shot.png or the latest /tmp/cw_seq_*.png) and review it
python3 ~/.claude/skills/vision-review/vision_review.py --dmon

# a specific burst frame (operator "burst" writes 8 frames: /tmp/cw_seq_0..7.png on dMon)
python3 ~/.claude/skills/vision-review/vision_review.py --dmon-seq 4

# custom prompt (focus the review)
python3 ~/.claude/skills/vision-review/vision_review.py --dmon "Focus only on the water/river — does it read clearly as flowing water?"

Relay the model's findings to the operator (don't just dump raw output); pull out the actionable bits.

Endpoints — verified 2026-07-28

All are vLLM OpenAI-compatible, no API key. Override per-run with OPENAI_BASE_URL / OPENAI_MODEL.

EndpointModelState
http://100.92.111.3:30800/v1 — vllm-qwen36 on robo-dog (tailnet NodePort)qwen3.6, 64k ctxLIVE — the script default (flipped 2026-07-28). Most accurate. Vision verified.
http://10.43.73.157:8000/v1 — vllm-gemma4-vision-dmon ClusterIP (NodePort 30804), on dMongemma-4-12b, 32k ctxLIVE — fallback. Vision verified.
http://100.92.111.3:30803/v1 — vllm-gemma4-vision on robo-doggemma-4-12bscaled to 0
  • Fallback when qwen36 is scaled down (it competes for robo-dog's memory, so it may be — check first if a run fails to connect): OPENAI_BASE_URL=http://10.43.73.157:8000/v1 OPENAI_MODEL=gemma-4-12b python3 ~/.claude/skills/vision-review/vision_review.py … gemma-4-12b is noticeably less accurate — treat its findings as leads and eyeball the frame yourself (Read renders images) before reporting to the operator. Gemma runs on dMon, so it stays up regardless of what robo-dog is doing.
  • qwen3.6 is now 64k, not 256k — it was resized on 2026-07-28 to share robo-dog with Ornith. See [[project-ornith-kv-oversizing]]; robo-dog is at ~89/121 GB with both models up.
  • No readiness probes on these deployments. The pod reports 1/1 Running within ~10 s while the engine is still loading; qwen3.6 takes ~7 min to serve (Triton warmup + FlashInfer autotune after KV alloc), Ornith ~4 min. Poll /v1/models for HTTP 200 — never trust pod status.
  • Reachability from croft: dMon-hosted ClusterIPs (like 10.43.73.157) work directly. robo-dog-hosted ClusterIPs do NOT — reach robo-dog services via its tailnet NodePort 100.92.111.3:<nodePort>. The LiteLLM :4000 and ollama :11434 are ClusterIP-only on robo-dog, so not reachable from croft.
  • vision_tour.py sends up to 16 frames in ONE request (CW_TOUR_MAX). Against gemma's 32k context that can overflow — lower CW_TOUR_MAX if you get a context-length error.

Key facts

  • <think> stripping: qwen3.6 is a reasoning model and emits <think>…</think> before its answer; the scripts strip it. gemma doesn't emit it, so stripping is a no-op. If you call the API yourself against a reasoning model, keep text after the last </think> and use generous max_tokens.
  • "burst" (operator typing it / the game's burst key) writes 8 frames to dMon /tmp/cw_seq_*.png; the realm also writes /tmp/cw_shot.png. That's the frame source for --dmon.
  • Autonomous loop WORKS (proven 2026-07-11): relaunch the console via sudo systemd-run --unit=cw-console running ~/cw_play_drive.sh (= cw_play_fs.sh + --drive; keep the --setenv CW_* vars — the base script does NOT set them). The game then polls /tmp/cw_cmd (line 0 = seq, then ops: forward/back/left/right/turn/pitch/up/down/setalt/look/face/goto/quit; turn+pitch are RELATIVE, setalt V = ground+V, goto RIVER|HUB|MINE teleports framed) and writes /tmp/cw_nav.png + /tmp/cw_nav.txt, then the seq to /tmp/cw_done. Yaw convention: facing.x = −sin(yaw). GOTCHA: the 240 s game day means captures land at NIGHT half the time — poll /tmp/cw_nav.txt's clock and re-shoot until ~09:00–15:00.
  • For plain text, not vision: use Ornith (http://100.92.111.3:30801/v1, model ornith, 64k ctx) — that's the general local LLM, not this vision endpoint.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.