v5.7.0 — harness-design reference

Added

• skills/graph-engineering/references/harness-design.md: the harness-design
  reference the execution-graph grammar lacked — the minimal-harness ladder
  (prompt/retrieval -> fixed workflow -> agent loop) with the walking-skeleton
  build order; the pattern catalogue (augmented LLM, prompt chaining, routing,
  parallelization, orchestrator-workers, evaluator-optimizer, agent) mapped onto
  the grammar's shapes and edges; voting vs. sectioning with the decorrelation
  caveat; tool-surface (ACI) design — document for the model, poka-yoke the
  arguments, iterate on observed misuse; scoping the tool surface; server-side
  context injection; layer attribution for failures; and the anti-framework
  stance. Distilled from Anthropic's Building effective agents and Hugo
  Bowne-Anderson's Stop overengineering your agent harness, adopted as method
  with numbers re-derived locally.

Changed

• skills/graph-engineering: the description, three new short sections
  (enter at the least-agency rung; fan-out's two rationales; design the tool
  surface), three new common-mistake rows, and the reference linked. The agency
  ladder is named as the shape ladder viewed from the agency axis.
• AGENTS.md: §4 routes to the minimal-harness ladder — enter at the least
  agency that closes on evidence, a prompt or fixed workflow before an agent
  loop, the smallest shape before a graph; §11's graph-engineering row lists the
  new reference.
• skills/teamwork: the spawn-brief contract now scopes the worker's tool
  surface to its task — a narrower surface means less confusion, misuse, and
  injection exposure.
• skills/verification: Diagnosis attributes a failure to its layer —
  reasoning, tool interface, context, or control flow — before patching, since a
  wrong-layer fix is a symptom patch by construction.
• README.md: the skills tree lists the harness-design reference.
• Version fields bumped to 5.7.0 across the four versioned manifests.

Motivation

• The doctrine governed how one job runs (shapes, typed edges, caps, anchors,
  cost) but said little about designing the harness that runs it. A coverage
  sweep against the two sources found the workflow-vs-agent ladder,
  voting-vs-sectioning, tool-surface (ACI) design, scoped tool access,
  server-side context injection, layer-attribution diagnosis, and the
  anti-framework stance absent, with the correlated-lens hazard present but
  voting unnamed. The reference adds the missing method without restating what
  the "Right-size, don't overengineer" banner and the Kirby Effect already
  encode — those are sharpened, not duplicated.