Adds a configurable pre-execution contract and dynamic reframing for long tasks.
- Adds
intakeMode: off | adaptive | guidedwith an 8-step default threshold.offpreserves fully autonomous behavior; adaptive lets the user choose guided clarification or autonomous execution per task. - Uses Harness's real user-question seam, limits guided intake to high-impact unknowns, requires exact contract approval, and persists a session/step/digest-bound intake receipt.
- Records boundary, horizon, desired outcome, facts, decisions, assumptions, unknowns, invariants, changeable forms, directional forces, and minimal causal variables before decomposition.
- Adds
lattice_reframefor material facts discovered during execution. It renews the human policy, bumps the graph revision, and preserves existing nodes for explicit reconciliation. - Fixes context-tool presentation so the model sees the exact copyable
receiptIdandexpectedRevision. A real-model run exposed the missing receipt as an execution blocker. - Adds a public three-arm real-model evaluation protocol and grader. The first valid smoke scored 100/100 for native sparse, native full, and Plan Lattice guided; this is non-regression evidence, not an uplift claim.
Validation: 20 tests, strict TypeScript, build, publint, diff check, zero-finding staged TruffleHog scan, a 100,000-node scale proof, and real DeepSeek Harness execution. The guided run completed all 7 graph nodes, made 39 lattice calls with zero lattice errors, generated 57 passing fixture tests, and scored 100/100 externally.
Artifact proof: the attached tarball installed through dsh plugin --profile web add into a fresh DSH_HOME. Both default and adaptive Web profile boots returned HTTP 200. SHA-256: 22782f506264b93c177c10bb1eb0ec2c8b0939a1c8e8ebb157cc2a4aa7e4c2fc.
Exact source commit: fc55e593c03f99c0ef62ba5948d3e4f719059cdc.