argus skill
scriptsProcesses equirectangular panorama images to produce depth maps, GLB point clouds, camera poses, and intrinsics.
- GitHub stars
- 12
- Files
- 51
- Last commit
- Nov 10, 2026
- License
- —
- Allowed tools
- —
What does the argus skill do?
argus processes equirectangular panorama images to produce depth maps, GLB point clouds, camera poses, and intrinsics.
Summary generated from the skill's SKILL.md.
About the argus skill
Converting equirectangular panorama images into 3D spatial data is what this skill enables for developers working with spatial reconstruction. By processing between one and 99 local images or a single ZIP file, the tool generates depth maps, camera poses, intrinsics, and merged GLB point clouds. It is designed for users who need to transform 2:1 aspect ratio JPEG, PNG, or WebP files into structured geometric outputs.
An agent should load this skill when a user provides panorama images for reconstruction. The workflow involves checking local prerequisites with check-local.mjs, submitting a task via run-argus.mjs start, monitoring the status, and finally using collect to retrieve the results. The process relies on a remote service and requires specific credentials including REALSEE_APP_KEY and REALSEE_APP_SECRET. Once the data is downloaded, the inspect-result.mjs tool can be used to validate the integrity of the artifacts and generate a verification report.
What an agent can do with it
- 1Reconstruct 3D scenes from panoramasSubmit a collection of local equirectangular images to generate a merged GLB point cloud and depth maps.
- 2Verify reconstruction integrityRun an offline inspection on a completed workspace to validate artifact references and archive integrity.
- 3Batch process image archivesProvide a single ZIP file containing multiple images to be processed as a single reconstruction task.
- 4Check environment readinessValidate that the local system has the necessary Node.js and npm dependencies before starting a task.
Written from the skill's SKILL.md. Every use case is checked against a quote from it.
How to install the argus skill
Each command copies the skill folder into the client's skills directory. Restart the client or start a new session to load it.Install the argus skill in Claude Code
~/.claude/skills/argus/SKILL.md
$ npx degit realsee-developer/skills/arkclaw/argus ~/.claude/skills/argusInstall the argus skill in Claude Code (project)
.claude/skills/argus/SKILL.md
$ npx degit realsee-developer/skills/arkclaw/argus .claude/skills/argusInstall the argus skill in Codex
~/.agents/skills/argus/SKILL.md
$ npx degit realsee-developer/skills/arkclaw/argus ~/.agents/skills/argusInstall the argus skill in Gemini CLI
~/.gemini/skills/argus/SKILL.md
$ npx degit realsee-developer/skills/arkclaw/argus ~/.gemini/skills/argusInstall the argus skill in Claude.ai
npx degit realsee-developer/skills/arkclaw/argus argus && zip -r argus.zip argus # claude.ai → Customize → Skills → + → Create skill → Upload a skill → argus.zip
Automated checks
Ran Nov 16, 2026Full report →Not verified yet. A skill needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachablerealsee-developer/skills/arkclaw/argus is public and reachablegithub.com/realsee-developer/skills/tree/HEAD/arkclaw/argus ↗Pass
- Open source license (OSI)No licenseFinding
- Commit or release in the last yearLast commit 2026-11-10github.com/realsee-developer/skills/commits ↗Pass
- No download-and-run or credential theft in scripts5 of 23 scripts read; none of them downloads and runs code, the rest were not readFinding
- No hidden instructions in SKILL.mdNo override phrases, safeguard switches, credential requests or hidden charactersgithub.com/realsee-developer/skills/blob/HEAD/arkclaw/argus/ ↗Pass
- Pre-approved tools are scopedNo pre-approved tools; the agent asks before using onegithub.com/realsee-developer/skills/blob/HEAD/arkclaw/argus/ ↗Pass
- SKILL.md follows the Agent Skills formatValid name and description; says when to use it; 152 linesgithub.com/realsee-developer/skills/blob/HEAD/arkclaw/argus/ ↗Pass
SKILL.md
View on GitHubargus
Run Argus through scripts/run-argus.mjs. <skillDir> below means the directory containing this file.
Argus uploads inputs to a remote service. Never print or log credentials, upload tokens, presigned URLs, or raw provider errors. Keep tokens and signed URLs out of persistent state.
Match the requested outcome
| Request | Finish condition |
|---|---|
| Check an installed Argus setup | Run check-local.mjs; report missing prerequisites. |
| Verify a collected Argus result | Run inspect-result.mjs for that workspace. |
| Run the reconstruction, official examples, or a complete verification | Submit once, query until terminal, collect, and inspect the result. Continue without requiring a new user request at each step. |
| Start or submit and return immediately | Submit once; return workspace_dir and current status. |
| Check status | Perform one status request. |
| Collect or continue an existing task | Reuse its workspace_dir; query if needed, collect at terminal, and inspect. |
1. Ensure runtime dependencies
For setup checks or before the first live command, run from any working directory:
node <skillDir>/scripts/check-local.mjs --json
The check works before runtime dependencies are installed. It checks the package, Node.js, npm, dependency resolution, and environment configuration without network access or reading credential-file contents. Read status (ready, needs_configuration, or blocked); exit 0 alone does not mean ready or prove remote service access.
If runtime dependencies are missing, install the lockfile once:
(cd "<skillDir>" && npm ci --omit=dev --ignore-scripts --no-audit --no-fund)
Rerun the check after resolving dependencies and, for live commands, credentials below. If the locked install fails, report the error before uploading; do not substitute an unlocked install.
2. Resolve credentials
Live commands require REALSEE_APP_KEY, REALSEE_APP_SECRET, and REALSEE_REGION (global or cn). The CN-only Arkclaw distribution uses cn.
Use the existing shell environment first. If check-local.mjs reports missing configuration and ~/.realsee/credentials exists, load it in the same shell and rerun the check without displaying the file:
set -a
. "$HOME/.realsee/credentials"
set +a
node <skillDir>/scripts/check-local.mjs --json
If configuration is still missing, request only the missing items through a local shell or secure credential interface. Persist app credentials only when explicitly requested, in ~/.realsee/credentials with mode 0600 outside the repository. Never put credential values in recorded CLI arguments or environment prefixes. Resolve credentials again whenever a live command starts in a new shell.
3. Select the input
Two mutually exclusive modes are supported:
- repeat
--image <absolute-path>for 1–99 local images; or - pass one
--zip <absolute-path>containing root-level images.
The CLI performs authoritative validation and deterministic packaging. Inputs must be JPEG, PNG, or WebP, RGB, 8-bit, and exactly width == 2 * height. A resolution below 2048×1024 emits a warning. Square 1:1 images are rejected; users who require the old square/single-GLB workflow must pin v1.0.2.
ZIP input is safely extracted, validated, Unicode-normalized, sorted, and repacked before upload. Use the returned name_mapping for output IDs rather than deriving or renaming them.
examples/manifest.json lists two first-party example sets with CDN URLs, byte lengths, and SHA-256 digests; the JPEGs are not bundled. For official examples, use only the CN region, and choose a new absolute output directory in the task workspace outside <skillDir>. State the choices; ask only if they cannot be resolved from the task. Create the parent before invoking the downloader:
node <skillDir>/scripts/download-examples.mjs \
--region cn \
--output "/absolute/example-output"
The downloader publishes the directory after all files pass byte-length and SHA-256 checks. Do not create, rename, or replace the output path or its parent while it runs. This CN-only Arkclaw distribution only allows cn; do not offer or attempt a Global download. See the example panorama guide / 中文 for selection details.
4. Start once
Before uploading, require selected inputs and consent to send them to Realsee. An explicit request to upload those inputs is sufficient; file selection or example downloads alone are not consent. If either is missing, ask one question explaining that the files will leave the machine. Reuse existing consent for the same input and scope without reconfirming.
Submit the selected images once:
node <skillDir>/scripts/run-argus.mjs start \
--image "/absolute/path/a.jpg" \
--image "/absolute/path/b.webp" \
--workspace "/absolute/workspace-root" \
--yes --json
For a ZIP, replace all --image arguments with --zip "/absolute/path/input.zip". --yes records existing authorization; it does not create consent. Capture the returned workspace_dir for all later commands.
start validates, packages, uploads, submits, and returns. Never automatically repeat it after submission_unknown: submission is not idempotent and a retry may create a duplicate task.
5. Query status explicitly
Run one status query:
node <skillDir>/scripts/run-argus.mjs status \
--workspace "<workspace_dir>" --json
task_status is queued, processing, succeeded, or failed. For a full run, query until terminal with bounded foreground waits of about 30 seconds, within the host's tool timeout. There is no detached poller or --resume mode. After an interruption or transient query/collection failure, continue with status or collect using the saved workspace; never recover by repeating start. If recovery cannot proceed, report the workspace and failed stage. Do not claim ongoing monitoring after the agent stops.
6. Collect a terminal result
When the task is terminal, run:
node <skillDir>/scripts/run-argus.mjs collect \
--workspace "<workspace_dir>" --json
collect retains output.zip, safely extracts and validates its manifest and references, and writes result.json. A failed remote task produces a local error result instead. Repeating a completed collect does not resubmit or download again.
7. Verify and report the collected result
Use the installed offline inspector rather than rewriting Schema or artifact checks:
node <skillDir>/scripts/inspect-result.mjs \
--workspace "<workspace_dir>" --json --write-report
The inspector reuses the canonical validator offline and needs no credentials. It checks artifact references and available archive integrity metadata; --write-report saves verification.json in the workspace. Without that option it is read-only. It preserves the original state, result index, archive, and extracted artifacts.
Report these fields separately:
- Recorded local
phase, remotetask_status, and algorithmresult_status; copy them from the report rather than inferring one from another. - Every
missing_idsentry and all input/result warnings. - Usable local artifact/report links, counts, and file sizes; do not replace file links with shortened
...paths. - Declared depth resolution and whether optional intrinsics were returned.
For partial, the CLI exits 0. Still show a prominent warning and the complete non-empty missing_ids list. For error, surface the sanitized error and treat the command as failed. Do not present temporary result URLs as durable output.
Keep the response concise. Validation covers the artifact contract: binary checks inspect headers without fully decoding GLB/EXR payloads, and point counts/resolutions come from the manifest. Depth resolution is service-defined and may be lower than input resolution. Visual quality and geometric accuracy remain unassessed. When requested, inspect the original GLB with an available local point-cloud viewer; unsupported host previews do not indicate reconstruction failure. Do not claim visual inspection from metadata alone.
References
- For lifecycle implementation details: Gateway workflow
- When checking request/response fields: Gateway OpenAPI
- When interpreting or validating artifacts: Algorithm I/O contract / 中文
- Official example panoramas / 中文
output.jsonJSON Schema- For older inputs or integrations: 2.0 migration guide / 中文
- When a command fails: Troubleshooting
Files in this skill
51argus skill questions
How do I install the argus skill in Claude Code?
Run npx degit realsee-developer/skills/arkclaw/argus ~/.claude/skills/argus. Claude Code loads skills from ~/.claude/skills (all projects) or .claude/skills (one project) and uses a skill when its description matches the task.
Which clients can use the argus skill?
Any client that reads the Agent Skills format: Claude Code, Codex, Gemini CLI and claude.ai (upload as a ZIP). Scripts in the skill run only where the client allows code execution.
What outputs does the reconstruction process generate?
The process produces depth maps, a merged GLB point cloud, camera poses, and optional intrinsics.
Can I use a ZIP file as input instead of individual images?
Yes, you can pass a single ZIP file containing root-level images as an input.