Talus Nexus Agent Skills

Reusable engineering skills for building and diagnosing Talus Nexus applications. Stable examples, companion scripts, and test fixtures live in this repository, while version-sensitive source is limited to anonymous public repositories and read-only Sui testnet evidence. A skill-only installation does not include the repository-level companion scripts.

Install from the public repository

Install all five skill directories or select individual skills with the public skills CLI:

npx skills add Talus-Network/skills

Choose your coding assistant and the skills needed for your project in the installer. Once installed, ask your assistant to use the relevant skill by name.

Skills

Before following any helper or fixture command from an installed skill, use its references/consumer-setup.md to fetch the pinned public companion checkout and set SKILLS_BUNDLE_ROOT. The installer does not set that variable or supply the repository-level scripts. Repository-maintainer commands below run from a full checkout. Live v2.0.0 execution is covered in the TAP workflow; Tool verification, payment reconciliation, and Task diagnosis include their own exact-FQN and Testnet evidence guidance.

Skill	Use it for
nexus-onchain-tool-development	Build and validate an on-chain Tool in Sui Move.
nexus-offchain-tool-development	Build, test, and validate an off-chain HTTP Tool service.
nexus-tap-development	Build a TAP and its artifacts, and a Nexus API dashboard/dApp/backend for an existing TAP.
nexus-onchain-task-debugging	Trace Task, Occurrence, Execution, Tool, and payment evidence.
nexus-cli-payment-tracking	Reconcile funding, reserves, charges, refunds, and Tool revenue.

TAP applications and Tool scaffolding

For an application over an existing TAP, use Nexus API application development: hosted REST projections, paginated execution history, replayable SSE, a server-only provider key, and deterministic frontend/relay tests. App-only work does not require a Move package or CLI/source setup. The public Nexus API guide and provider reference define the selected deployment's consumer contract.

The Tool skills include the full scaffold-to-implementation workflow and completion checks for Rust HTTP Tools and Sui Move Tools. Start with the released public CLI scaffold, then implement real behavior and verify it using the matching skill. No separate plugin runtime is required.

Safety model

The skills default to repository-owned examples, isolated build/test work, dry-run validation, and read-only inspection. Publishing, registration, deposits, refills, settlement, scheduling, upgrades, and asset movement require explicit authorization and a fresh network/signer/configuration preflight. Examples use placeholders rather than credentials or deployment-specific capability IDs.

Portable source evidence

The source helper downloads only three reviewed anonymous public archives when bundled material is insufficient: nexus-sdk, nexus-move-packages, and the pinned Sui framework source matching sui 1.78.0-d8459684b41e. It rejects local paths, SSH URLs, unknown repository selectors, and unapproved source roots. The Move package archive supplies the six public interface packages and their supporting closure; the Sui archive supplies only the framework packages needed for local Move compilation, and neither declaration set is an executable mock.

Public Move Registry dependencies

For a network-facing Move consumer, declare the six reviewed Nexus packages through the public Move Registry and let Sui resolve the deployment-specific package address. The package pages are nexus-interface, nexus-primitives, nexus-registry, nexus-tool, nexus-scheduler, and nexus-workflow.

[dependencies]
nexus_interface  = { r.mvr = "@talus/nexus-interface" }
nexus_primitives = { r.mvr = "@talus/nexus-primitives" }
nexus_registry   = { r.mvr = "@talus/nexus-registry" }
nexus_tool       = { r.mvr = "@talus/nexus-tool" }
nexus_scheduler  = { r.mvr = "@talus/nexus-scheduler" }
nexus_workflow   = { r.mvr = "@talus/nexus-workflow" }

Commit the generated Move.lock and keep the package revision, network, and published dependency addresses matched. The public Move Packages repository remains useful for source inspection, provenance, contribution, and explicitly offline copied closure; it is not the preferred installation source. Do not invent a registry name for nexus_policy or nexus_kernel: the reviewed registry map records both as unregistered, and nexus_kernel remains supporting source closure rather than a direct application dependency.

Published Setup authority

Version-sensitive instructions use the published Developer Setup, not a source-repository checkout. Check the page anonymously before relying on a release-specific command; the expected release for this bundle is v2.0.0.

python3 scripts/docs_website.py --url https://docs.talus.network/guides/getting-started/setup --expected-version v2.0.0

The check is read-only and fails closed on an unavailable page, an unsafe redirect, malformed content, or a version mismatch. It never installs a dependency or changes a Sui environment; show any resulting setup command to the user as a separate, explicitly authorized workflow.

Read-only Sui testnet evidence

When a question needs deployed package, module, object, or network evidence, use the bundled boundary with the explicit official Sui testnet GraphQL endpoint. It performs only allowlisted read-only GraphQL queries and never uses a wallet, active CLI environment, or state-changing command:

python3 "$SKILLS_BUNDLE_ROOT/scripts/testnet_evidence.py" \
  --graphql-url https://graphql.testnet.sui.io/graphql \
  --package-id 0x... --module module_name

The endpoint, network label, query methods, response digest, and timestamp are retained in the JSON report. An unavailable endpoint, malformed response, wrong network, or missing identifier is a concrete evidence failure; structural fixtures and validators must not be presented as runtime execution or registration proof.

Validation

Run the bundle validator before proposing changes:

python3 scripts/validate_skills.py
python3 -B -m unittest discover -s scripts -p 'test_*.py'
python3 -B -m unittest discover -s nexus-tap-development/scripts -p 'test_*.py'

The standard-library scripts/run_skill_evals.py validates the five catalogs and can plan, run an explicitly supplied command, or replay saved traces. Each run records a selected skill snapshot, source digests, effective argv, raw JSONL stdout, stderr, and declared artifacts in a fresh case directory. A temporary cwd/snapshot organizes files but does not change the explicit command's OS permissions. Offline traces cannot prove Codex automatic selection, live provider or chain state, published-bytecode execution, registration, or settlement; missing activation is unknown for positive and negative cases, and ungraded behavior rubrics keep overall status unknown with manual_status: pending. Use --plan to inspect cases without execution and --replay <saved-run> --output-dir <new-dir> to rescore saved evidence without rerunning a command. The documented Codex command includes --skip-git-repo-check for its temporary non-Git cwd.

See the source helper reference and the testnet evidence reference for the complete evidence and cleanup contract.