long-task-workflow

Single-Goal Long-Task Workflow

Boundaries

Use one current native Goal, one repository, one selected verification workspace, one complete Contract and one Final Gate. This workflow never creates or manages a scheduler, model worker, agent runtime, App Server, branch/worktree fan-out, merge, push, PR, deployment, Campaign/SFC/Packet/Wave chain, matrix, verdict or second Contract plan. Platform-native implementation delegation or user-authorized Git parallelism may exist outside Harness, but every result must converge into the selected verification workspace before proof counts. Never activate from task size alone.

The host and user own model selection and native-Goal lifecycle. The workflow has exactly one user-choice checkpoint after the first Authority Lock and before implementation; Harness neither switches the model nor persists model-routing/checkpoint state. No checkpoint file, acknowledgement state, model route, model-tier scheduler, automatic model switch, authority_revision_in_progress state or native-Goal completion state is created. Outside that boundary, do not pause a healthy Goal solely to change or downgrade the model. Do not create a separate approval checkpoint for a defensible recommended plan choice. A targeted pre-Authority clarification is still required when a missing user preference could materially change research or selection; genuine Source conflicts or choices the user explicitly reserves may likewise require a decision before Authority Lock. Capability-related drift is handled by targeted repair plus the Final Gate. The current Goal may choose platform-native internal delegation, but Harness owns no subagent dispatch/retry/recovery state, delegated reports are not Progress or proof, and all outputs must converge into the selected verification workspace before verification can count.

long-task-delivery-v2 is the only active Contract schema. delivery-contract.yaml is the root authoring file. New authoring uses inline Outcomes; existing outcome_files are physical compatibility only. delivery-set is retired and non-executing.

Controlling Objective

Prevent false completion inside declared authority. Given complete and accurate Source at the declared observable granularity, a meaning-preserving Source-to-Contract projection, complete applicability expansion and a sound named verifier/runner trust boundary, AcceptedDeliveryTerminal must imply that no declared observable drift remains. Implementation may drift, fail or require rework, but every declared non-Result requirement, exact applicability cell and AC must remain traceable and every unsatisfied, unverifiable, insufficiently evidenced, stale or externally pending item must block or explicitly qualify complete delivery. In particular, unclassified Source, an omitted architecture obligation, Control field/relation, population-universe member, wrong target/condition/input/state/journey, proxy target, presence text, degradation path, fixed input, self-reported boundary effect, weak semantic oracle or internal entrypoint must never substitute for declared behavior. Findings should localize repair through Source Item, Stage, Outcome, Claim, applicability, Assertion, Check, Evidence Capability, execution target, Binding and owner boundary.

For selected design resources, one design-specific objective is that Agent implementation, acceptance and testing fully conform to every material UI/UX fact explicitly expressed within declared scope and conditions. Open Design can produce source-rich, implementation-readable resources, but capability alone is not a guarantee: resource authoring must require a canonical entry, complete dependency acquisition and stable machine-resolvable facts. The validated design-resource-handoff-v1 remains the residual scope/applicability/semantic adapter rather than a copy of CSS values. Preserve each fact through immutable inputs, typed locators, complete subject × target × condition × dimension cells, Context-reachable targets, Source/Control/Claim authority, one independent Assertion per verification method and current-snapshot project Checks to Final Gate. Never invent an unexpressed fact: refine the resource, retain decision_required/unavailable, or block fidelity work. Neither provider success, file hashes nor handoff integrity proves production conformance.

Only fresh evidence from the complete current final snapshot may create machine acceptance. Exactly fresh machine_accepted with no pending External Confirmation is AcceptedDeliveryTerminal and may support the full declared-observable no-drift conclusion. Otherwise report the task as unfinished or qualified. machine_accepted_external_pending means machine-verifiable authority passed while named external confirmation remains; it proves only the declared machine scope and is not full delivery completion. Machine acceptance has no direct native-Goal effect. Never substitute prose, progress, historical tests, Receipts, one exit code or Agent judgment for the Final Gate.

Mechanism changes use a two-stage hard gate. First prove Coverage_new ⊇ Coverage_old, FalseNegative_new ⊆ FalseNegative_old, and that Authority, fail-closed behavior and complete-current-final-snapshot proof cannot be bypassed. If non-degradation cannot be proved, preserve the current formal acceptance path. Only after that gate passes may Authoring, Runtime, State, Recovery, maintenance and process cost be compared. No cost reduction compensates for weaker drift interception; positive ROI permits consideration but never overrides the first gate.

Progressive Reference Loading

Read only the reference needed for the current activity; these files are guidance, not new artifacts or authority:

• When inputs are raw, mixed, attachment-heavy, incomplete or need synthesis/refinement while the Contract Draft is being mapped, read references/source-authoring.md alongside the Contract-authoring reference. Do not wait for a separate Source-authoring phase before opening the Draft.
• Before creating or structurally revising Source markers, Outcomes, requirements, controls, obligations, architecture boundaries, paths, Bindings, Assertions or risk, read references/contract-authoring.md.
• Before creating or repairing Checks, runners, Observations, proof surfaces, Playwright/structured evidence, Counterfactuals, Population or environment probes, read references/evidence-design.md.
• Before Preflight, Compile, protected revision, resume, targeted verify, Final Gate, Stop, close or abandon, read references/authority-lifecycle.md.

Do not copy reference detail into another plan or state file. The same delivery-contract.yaml, active authority and current workspace remain the only lifecycle surfaces.

Contract Draft And Outcome Decomposition

Every input enters the same non-authoritative delivery-contract.yaml Draft immediately. Before the first successful formal Compile, continuously revise that Source-bound Draft while real Source inventory, provenance, refinement, markers, repository binding and mapping converge. It need not be completed in one response; keep reading Source, repository and relevant Context and feed Preflight findings back into that same Draft. Draft authoring, Preflight, Compile, rolling execution, targeted verification and Final Gate are one long-task-workflow lifecycle. Do not create a Source-authoring phase, standalone Contract Draft Skill, Draft Receipt, Authoring State, draft schema/CLI/runtime state or second plan.

A Draft Outcome is an Outcome in that pre-Authority-Lock Draft, not a new schema field or runtime entity. Decompose only vertical, independently observable, decidable and target-verifiable results whose dependencies and owner boundary can be stated; one Outcome belongs to one declared Stage and does not span materially different success paths. Declare the ordered Stage DAG and one gate Outcome per Stage in the same Contract. Use those boundaries to project an acceptance/verification-ready working set, bind target verification, localize failures, resume findings/next actions and stale local results precisely; never use them to restrict which in-scope implementation edit may happen next.

depends_on means acceptance and intermediate-proof readiness, not implementation permission. The gate Outcome transitively depends on the rest of its Stage, later Stage Outcomes depend on prerequisite gate Outcomes, and every multi-Outcome gate proves cross-surface consistency. The current Goal derives a temporary advisory Rolling Frontier from Stage and Outcome status, but may implement, inspect or repair any in-scope Outcome in the order current code reality favors. Do not persist a Stage Receipt, scheduler, Worker queue, mandatory implementation DAG, model route or process tree. Never split for response/YAML/file length, implementation layer, module/file count, Agent capacity, Worker assignment or desired parallelism.

Outcome decomposes execution and diagnosis, not completion authority.

Entry And Authoring Loop

1. Read the user request or external initial proposal, selected design resources and minimum controlling Context. Collect the architecture owners, extension points and boundaries needed for the shared deliberation before deciding Context Delta.
   • For material production UI, read the Contract-authoring visual guidance before Compile. When selected resources arrive as an implementation handoff, require one marked design-resource-handoff-v1 in task.source_paths and run ty-context design-resource preflight <handoff.md>; incomplete applicable cells, unsupported evidence, unresolvable locators, partial implementation-source acquisition, unresolved meaning or stale resource identity is blocking. Traverse affected surface/control/target keys from owning Context through `DES

Content truncated.