Skip to main content
trellis is the repository’s executable Dynamic Delegation v2 workflow. It is an experimental local workflow, not yet part of the default bunx smithers-orchestrator init pack. It maps Sol/Fable authors, a Terra validator, and a Luna implementer into the <Trellis> component and declares the real Smithers Trellis custom UI.
The explicit flag is required even at the default value: Trellis verifies that its component cap and the persisted run cap are identical. To use another value, pass it both as workflow input and as the run option:

​
Inputs

The optional policy contains allowedCategories, normalized workspace-relative allowedPathPrefixes, and maxChangedLines (1..500). It is a trusted per-execution admission ceiling; Phase A does not claim measured diff or filesystem-sandbox enforcement. For example:

​
UI

Open the workflow UI to launch and inspect runs:
The launcher submits the same selected concurrency as workflow input and run policy. The live surface shows the canonical final outcome, author fuel, accepted or rejected validation generations, the real execution tree, exact node outputs, and durable events. It discovers the semantic final node from trusted Trellis metadata rather than parsing hashed physical IDs. The controls are labeled as defaults for the next launch; they are not presented as configuration recovered from whichever historical run is selected. Each new run instead persists its pinned root concurrency, total author-turn cap, generation/depth limits, and per-invocation allocated/remaining author fuel in trusted Trellis task metadata. The selected-run strip reads those fields from the selected run, including runs launched outside the UI. A root-local remaining value is one immutable allocation, not a mutable global fuel counter; older runs without these fields are labeled as not persisted. The capability strip deliberately reports Phase A gaps: questions are unavailable, hierarchical token/USD budgets are not enforced, and tool or filesystem authority depends on the configured agent adapter and sandbox. Prompt restrictions alone are advisory. Rejected IR is visible but never described as executed. The current UI launcher omits a critical-execution policy, so its runs force implementation down to Terra/Luna; policy-enabled runs can still be launched through workflow input and inspected in the UI.

​
Agent mapping

  • Sol: synthesis, orchestration, refinement, and exceptional critical work.
  • Fable: alternate root or nested author.
  • Terra: bounded planning, validation, synthesis, and substantial worker goals.
  • Luna: concrete research, POC, execution, review, or preview goals.
Sol/Fable may recursively author subworkflows. Terra/Luna can only complete or block their assigned goal. Every child result passes assignment-aware settlement before it reaches an author continuation.