README.md
August 21, 2026 · View on GitHub
deepseek-dovetail brings eight OpenCnid Dovetail workflows to DeepSeek Harness as one out-of-tree Cordis bundle. DSH remains the sole agent runtime; this package mounts an immutable, package-local skill tree through DSH's published filesystem provider.
Why use it?
- One install, eight workflows: prompting, skill authoring, delegation, evaluation, self-play, agent steering, and session handoff.
- Small runtime seam: one Cordis row and one effect-owned filesystem provider—no replacement agent loop or skill registry.
- Quiet on load: no network access, update check, script execution, watcher, or writes to project/user skill homes.
- Reproducible: upstream content is commit-pinned and hash-locked; package, integration, UI, and bounded behavioral evidence live in the repository.
Quick start
Requirements: DeepSeek Harness v0.1.0-rc.7, Node.js ^22.19.0 || >=24, and pnpm 11.x.
git clone https://github.com/OpenCnid/deepseek-dovetail.git
cd deepseek-dovetail
corepack pnpm install --frozen-lockfile
corepack pnpm run build
corepack pnpm run verify
corepack pnpm run pack:artifact
Install the generated tarball into any DSH profile:
dsh plugin --profile web add {ABSOLUTE_PATH}/artifacts/deepseek-dovetail-0.1.0.tgz
dsh --profile web --dump-config
Replace web with another profile name as needed. Remove the bundle with:
dsh plugin --profile web remove deepseek-dovetail
In a DSH session, type / to browse skills or invoke one directly, for example /prompt-engineering. Six skills are model-discoverable; spark-steering and upsum are intentionally user-only.
Skill gallery
Each original SVG gives the workflow a visual shorthand. Expand Real DSH output for an example generated through the installed bundle in the pinned DSH web profile—not a mockup. Six skills are model-discoverable and slash-invocable; spark-steering and upsum are intentionally slash-only.
|
|
|
|
|
|
|
|
The captures use the assembled deepseek-dovetail@0.1.0 profile with GPT-5.6 Terra. They are illustrative runs, not universal evaluation claims; exact prompts and capture conditions are recorded in the example manifest. A machine-readable UI snapshot covers all eight catalog entries.
Architecture
DSH profile
-> deepseek-dovetail bundle row
-> deepseek-dovetail Cordis plugin
-> @deepseek-ai/dsh-skill-filesystem
-> package-local dist/skills
-> existing DSH catalog, agents, tools, sessions, and UI
The provider is named dovetail, excludes default roots, disables watching, and resolves its bundled directory from installed code. Project and user skill roots keep their normal precedence.
Work on the bundle
Edit DSH adaptations in ports/dsh/; never hand-edit pinned vendor/dovetail files or generated dist/skills output. See AGENTS.md for repository rules and the complete validation workflow.
corepack pnpm run typecheck
corepack pnpm run lint
corepack pnpm run test
corepack pnpm run build
corepack pnpm run verify
corepack pnpm pack --dry-run
To refresh the pinned upstream subset from an exact, clean checkout:
corepack pnpm run sync:upstream -- --source {ABSOLUTE_PINNED_DOVETAIL_CHECKOUT}
The sync fails closed on a wrong commit, dirty source, symlink, path escape, unclassified runtime file, or manifest/hash mismatch.
Evidence and deeper documentation
- COMPATIBILITY.md — pinned host contracts and the skill-by-skill port matrix.
- evidence/REPORT.md — package, profile, UI, removal, and behavioral acceptance ledger.
- evidence/behavioral/LIVE_REPORT.md — bounded real-provider comparisons and known gaps.
- THIRD_PARTY_NOTICES.md — provenance, retained licenses, notices, and code-license status.
Public-source status and limitations
This GitHub repository is public. The npm manifest deliberately remains private: true, no npm package has been published, and no owner-selected public license currently covers the new adapter/build/test code. Public visibility does not imply a license grant; review THIRD_PARTY_NOTICES.md before reuse or redistribution.
Initial assembled acceptance is Windows-only. The changed/unchanged live upsum lifecycle remains UNMEASURED on the pinned Windows restricted-subprocess path, and the recorded behavioral comparisons are bounded evidence—not universal effectiveness claims.







