terryyin

dough-slice-plan-refinement

Refines an executable plan in place into proportionate, proof-owned Behavior/Structure slices. Use during or after dough-slice-planning when boundaries fragment cohesive work, combine independent concerns, accumulate special cases, or overrun. Creates no new plan and preserves the selected story or bounded correction outcome. After rewriting the plan, reassess readiness through the shared preparation procedure without granting Take or execution.

terryyin 4 Actualitzat fa 2 setm.
GitHub

Instal·la

npx skillscat add terryyin/open-dough/src-skills-dough-slice-plan-refinement

Instal·la des del registre de SkillsCat.

SKILL.md

Slice-plan refinement

Refine an existing executable plan in place. Do not create another plan,
implement product code, or change its selected story or bounded correction
outcome.

Require a refinable plan

Require an existing executable plan and this project's context required by
dough-slice-planning, especially any supplied
target, hard limit, exceptions, overrun policy, and the plan lifecycle.

Identify whether the plan's source is a selected feature story or a bounded
retrospective correction under
planning scope and lifecycle.
For a complete correction, use its correction input directly from its story
and plan, or from a plan-homed correction's plan alone; create no seed. Name
whichever required correction field is missing and stop. A complete plan permits
refinement, not execution; preserve the invoking instruction's authorization
boundary. Before revising boundaries, apply the shared
reassessment decision.

  • If no plan exists, use dough-slice-planning.
  • If a plan is marked as awaiting story refinement after resplitting, use
    dough-story-refinement on its mapped
    story and realign the plan before treating it as refinable or executable.
  • If the selected story's goal, scope, or examples must change, use
    dough-story-refinement.
  • If a correction's beneficiary, bounded outcome, scope, or proof must change,
    stop through the shared execution-decision handoff; do not turn it into a
    feature story. A disputed product constraint uses the plan-conflict handoff
    below and remains human-owned.
  • If the parent problem, candidate selection, or sibling ordering must change,
    use dough-story-decomposition.
  • Report ready for direct execution only when cumulative design is supported,
    every boundary preserves useful progress or learning, and all slices are
    cohesive, single-proof-loop, meet supplied targets, and have no unexplained
    path beyond a hard limit. Record that outcome through
    assess readiness at preparation completion
    when the work item has a recorded identity. Execution still requires separate
    authorization from the invoking workflow.

Refine the plan

Before writing to the plan, establish or reuse the required workspace under
preparation workspace,
then, for a queued story,
announce the preparation assignment;
reading the plan and the code and tests needed to judge execution boundaries
needs neither on its own. Read and apply:

When refinement reconsiders the plan's selected existing solution or the
architectural direction supporting it, reapply
architectural thinking,
including its authoritative PFE handoff. Otherwise carry the decision forward;
do not repeat PFE or create a new direction topic merely because slices are
being refined.

Preserve completed slices, applicable evidence, and the selected story's goal
and scope or the correction's bounded outcome and scope.

Assess the cumulative design, including how remaining examples build on
completed slices, then classify each remaining boundary:

Result Decision
Retain One gate and proof loop yield cohesive progress within supplied limits
Consolidate Adjacent slices are useful only together; combining hides no consequential risk and preserves safe recovery
Split One slice has independent outcomes or proof loops, hidden preparation, credible risk, special-case design, or a sizing concern
Escalate Learning requires review of the source outcome, story, or parent problem

A suspected accidental contractual restriction uses the shared
plan-conflict handoff
before a conflicting plan edit; classify it as Escalate, not an automatically
correctable special case. Route other Escalate findings through the input gate.
Apply every Consolidate or Split decision.

After an overrun, confirm attempt-owned work is safely parked or reverted before
editing the plan. Stop for human judgment when ownership is unclear. Do not
commit, push, implement, or verify product behavior unless the invoking workflow
separately authorizes it.

Count the resulting plan's slices, including completed slices but excluding
obsolete replaced slices. If the count is greater than 15, report
story resplit recommended: <count> slices; use dough-resplit-story and link to
dough-resplit-story. This is a recommendation,
not an automatic invocation or a new execution-readiness gate. Do not resplit
the story or change its backlog placement during slice-plan refinement.

Report the plan path, replaced slices, resulting slice count, sizing exceptions,
any resplit recommendation, whether execution can resume, and — when the work
item has a recorded identity — the readiness reassessment through
assess readiness at preparation completion:
remaining Refine or Escalate findings, or any other blocking concern, become
not-ready reasons; when the cumulative design is supported and remaining
slices are ready under the table above, record ready. Do not Take, move the
queue, or start execution from this reassessment.

When this session ends, follow
land at the end of preparation,
then close or retain the workspace under
preparation workspace.
An execution instruction for this story that arrives while its preparation is
unlanded follows
hand off to execution in the same session
instead.
End with:

## SLICE PLAN REFINED