Skip to content

T-U3NR-task-write-set-conflicts

Status: closed/superseded · Impact: high · Complexity: medium

The orchestrator dispatches up to max_implementations parallel task-work runs with no awareness of file overlap — two ready tasks touching the same files discover it as a merge conflict after both burn a run. Derive each task’s write-set from its ## Files to touch table and surface conflicts before dispatch (D-3PDM-cross-project-data-model-adoption P1).

  • A task’s write-set = its Files to touch rows with Kind ∈ {new, modify, delete} (read-only Today rows don’t count). No new frontmatter field — the body table stays the single source (parse-touchpoints already reads it).
  • DarkFactory’s rules adopted wholesale: leaf-only declaration (a parent task’s effective set is the union of its leaf descendants; parent/child overlap is exempt — containment is not conflict); literal paths conflict even when the file doesn’t exist yet.
  • Consumers: a sdlc task conflicts op; an audit warning for overlapping ready/in-flight pairs with no depends_on between them; a graph-scheduler filter rule at dispatch (with T-8I15-adapt-task-next-onto-graph-scheduler); a worktree-scope gate extension comparing actual vs declared touches.
  • AC-1: sdlc task conflicts lists overlapping write-set pairs among open/ready and in-flight tasks, naming the overlapping locations.
  • AC-2: entities audit emits a warning-level finding for each overlapping pair that has no depends_on edge in either direction.
  • AC-3: a parent task’s effective write-set is the union of its leaf descendants; parent/child overlaps are not flagged.
  • AC-4: two tasks declaring the same new file conflict before the file exists.
  • AC-5: the worktree-scope gate reports files touched in a task worktree that its write-set never declared (report-only).
  • Wiring the graph-scheduler filter rule — that rides T-8I15-adapt-task-next-onto-graph-scheduler once the scheduler lands.
  • Auto-inserting depends_on edges between conflicting tasks (surface only).
  • Any frontmatter mirror of the write-set (open question in D-3PDM-cross-project-data-model-adoption; only a generated mirror would be acceptable).
  • D-3PDM-cross-project-data-model-adoption P1 / finding F4: DarkFactory’s impacts was its most-adopted optional field (196/232 PRDs) with a validator-enforced leaf-only rule; ontological carried the same idea as agent sandbox envelopes. SDLC had the substrate (Location grammar, parse-touchpoints, worktree-scope) but no conflict awareness.

← Back to Tasks