DeepSeek Harness Configured Assets

Use this reference for the bounded DeepSeek Harness (DSH) configured-assets provider. It is qualified against DSH 0.1.1-rc.2 at source b150a551b8d465e31e418e1b2eaf5e79bbb7d28e; later versions are not implicitly qualified.

Inventory

Run the read-only CLI from an authorized workspace:

better-harness agent-customize inventory --provider dsh --workspace <path> [--cwd <path>] [--dsh-home <path>] [--include-user-home[=true]]

workspace is the analysis-selection and cwd-containment boundary. cwd defaults to the workspace and must be that directory or lexically beneath it. DSH walks upward from cwd to the nearest .git marker; that directory is the project root, or cwd is used when no marker exists.

User-home collection is off by default. Without --include-user-home, the provider does not inspect DSH or Agents user Skill roots, DSH's global AGENTS.md, or an ambient bundled Skill root. --dsh-home selects a home but does not authorize reading those sources by itself.

Filesystem Skills

Only DSH's native one-level forms are discovered: <root>/<entry>/SKILL.md and <root>/<entry>.md. Valid YAML frontmatter supplies the Skill identity and description. Duplicate declared names use this filesystem precedence:

1. project .dsh/skills;
2. project .agents/skills;
3. programmatic custom Skill roots, in declaration order;
4. authorized DSH user Skills;
5. authorized Agents user Skills;
6. an explicitly authorized or opted-in bundled Skill root.

Only the winning filesystem definition enters manage.skills; malformed and shadowed candidates remain bounded diagnostics. The runtime/in-process rank between project and custom Skills is not enumerable, so the result is never a claim about the complete runtime winner.

Instructions

Authorized Instructions are considered in native order: DSH's global AGENTS.md, then every directory from project root through cwd, with base AGENTS.md/CLAUDE.md candidates before local AGENTS.local.md/CLAUDE.local.md candidates in each directory. Regular-file, source-byte, same-directory content-deduplication, and aggregate rendered-byte limits decide which sources appear in manage.rules. Better Harness returns paths and non-content descriptions only; it never serializes Instruction text, digests, rendered framing, or symlink targets.

Evidence Boundary

Every result is configured-not-observed. Filesystem configuration and applicability do not prove runtime use. Runtime/in-process or scoped Skill providers and active Cordis, Profile, and Preset composition remain unresolved. DSH advertises sessionAnalysis, agentCustomize, assetPractices, harnessReport, reportRendering, and evidenceBundle. The shared Asset Baseline and Evidence Bundle freeze a canonical cwd, retain compact configured provenance, and keep current configuration distinct from historical Session observation. This does not change configured-assets semantics or claim historical runtime use. Durable /better-harness output reuses the portable html mode and publishes findings.json, report.md, and report.html at the generic Better Harness-owned root <target>/.dsh/better-harness; this is not a native DSH storage root. Explicit inline/no-files operation remains write-free. This does not add Checkup, Canvas, Studio, lifecycle, managed-shell, public Quickstart, or full DSH support.

Repository contributors can compare the Better Harness collector with pinned native DSH behavior without credentials or a model request:

npm run test:dsh-configured-assets-native

The #101 configured-assets specification remains canonical for native DSH configured-assets semantics and Skills / Instructions collection qualification. The #104 shared-analysis specification is canonical for shared Harness analysis admission, Asset Practices qualification, Asset Baseline v2, Evidence Bundle v3 with frozen cwd, and Harness Report admission. The #112 portable-rendering specification is canonical for DSH portable durable report routing and explicit no-files preservation.