Skip to content

dsh-tool-creator v0.1.0

Latest

Choose a tag to compare

@VincentJiang06 VincentJiang06 released this 18 Aug 22:30
· 5 commits to main since this release

dsh-tool-creator is a dsh (DeepSeek Harness) artifact-creation pipeline for deepseek-v4-pro: a five-role conductor (composer → guidance → engineer → zipper → battery) that mechanically dispatches confined subagents and ships machine-adjudicated acceptance evidence inside every artifact it creates (skill / plugin / preset). Distributed as a dsh preset plus the dsh-pipeline-executor plugin.

What makes it different

  • Evidence-in-artifact standard. Every created artifact carries an acceptance-manifest.json at its root — per-file sha256 + rootHash, pinned model/dsh version, verdict min-fold, and re-runnable reverify commands. Anyone holding the directory re-proves the byte + in-tree battery slice with node tools/reverify.mjs <dir>.
  • Mechanical guarantees over instructions. Capability-level stamping, per-line manifestSha256 provenance, confined subagent tool-filters, target-aware intake — the properties that matter are enforced by the executor, not requested of the model.
  • Adversarial acceptance battery (coherence / gaming / reality; evidence + foundation opt-in) with SEED discipline and PROVE-OR-FLAG adjudication.

This release was hardened against its own attack battery

An independent five-lens attack battery (see docs/ATTACK-R1.md) was run against the whole pipeline, followed by an independent fix-audit rotation on the fixes themselves. What it found and this release closes:

  • A verdict floor that let a clean battery verdict ship an industrial grade over real P1/P2 findings — now refused mechanically at assembly (the converse floor). Reproduced-before, refused-after.
  • The shipped manifest now discloses its adjudication mode (machine-adjudicated, human veto reserved-not-exercised) and battery mode, so top-grade evidence never travels silent about how it was judged.
  • reverify now skips-with-disclosure the kit/workspace commands on an installed tree instead of a false-RED indistinguishable from tampering; __pycache__ runtime cache is ignored symmetrically so python-harness artifacts reach honest green.
  • Artifact-tree ambiguity is refused rather than silently mislocated.
  • Several evidence-doc overstatements corrected: the delegation-audit synthesis-stage blind spot, default-lens coverage honesty, and O-L3 reframed as a structural floor, not an earned climb.

Disclosed residual (not hidden): a genuinely un-diligent ("hollow") battery reporting nothing isn't catchable by a finding-count; the mechanical guard covers the synthesis's own verdict-relabeling, not lens-artifact integrity against the host-open subagent door. This is now stated in every manifest, with mechanical SEED-gate enforcement tracked for a later release. Fixing what's mechanically closable and disclosing what isn't is the release posture.

Tests

assemble_manifest 17/17 traps + green · validate_decision 10/10 · validate_report 15/15 · validate_structure 12/12 · validate_compression 7/7 · acceptance/assembly (node) 15/15 · dsh-pipeline-executor 111/111.

Companion npm package

dsh-pipeline-executor@0.1.7 (the executor plugin) is live on npm: npm install dsh-pipeline-executor.