Run a staged software-engineering delivery feedback loop from bootstrap through investigation, requirements, design, runtime review, implementation, API/E2E and executable validation, code review, docs sync, and final handoff with durable artifacts and explicit re-entry.
Run a staged software-engineering delivery feedback loop from bootstrap through investigation, requirements, design, runtime review, implementation, API/E2E and executable validation, code review, docs sync, and final handoff with durable artifacts and explicit re-entry.
Software Engineering Workflow Skill
Overview
Run a staged software-engineering delivery workflow for software changes: bootstrap ticket context, investigate and refine requirements, build design and future-state runtime artifacts, drive implementation with one implementation artifact that carries both a stable baseline and live progress tracking, validate behavior with API/E2E and other executable validation evidence appropriate to the system, apply independent code review, synchronize long-lived docs, and finish with explicit user-verified handoff and repository finalization. For medium/large scope, include a full proposed design document organized by data-flow spine inventory, ownership, off-spine concerns, and derived separation of concerns.
This workflow is stage-gated. Do not batch-generate all artifacts by default.
In this skill, future-state runtime call stacks are future-state (to-be) execution models. They are not traces of current (as-is) implementation behavior.
Skill Layout
SKILL.md is the workflow router. It defines the stage rules and points each stage to its owned templates/references.
shared/ stores cross-stage references that multiple stages reuse:
shared/design-principles.md
shared/workflow-state-template.md
stages/ stores stage-owned templates and references:
stages/00-bootstrap/
stages/01-investigation/
stages/02-requirements/
stages/03-design/
stages/04-future-state-runtime-call-stack/
stages/05-future-state-runtime-call-stack-review/
stages/06-implementation/
stages/07-api-e2e/
stages/08-code-review/
stages/09-docs-sync/
stages/10-handoff/
Keep stage-specific material in the matching stage folder. Use shared/ only for genuinely cross-stage references.
When a stage has a local guide or checklist, use that stage-owned file first before falling back to generic workflow prose.
Terminology
Subsystem / capability area: a larger functional area that owns a broader category of work and may contain multiple files plus optional module groupings.
Module: an optional intermediate grouping inside a subsystem when the codebase benefits from it. In this skill, module is not a synonym for one file or the default ownership term.
Folder / directory: a physical grouping used to organize files and any optional module groupings.
File: one concrete source file and the primary unit where one concrete concern should land.
Workflow
Ticket Folder Convention (Project-Local)
For each task, create/use one ticket folder under tickets/in-progress/.
Folder naming: use a clear, short kebab-case name (no date prefix required).
Write all task planning artifacts into the in-progress ticket folder while work is active.
Standard states:
active work path: tickets/in-progress/<ticket-name>/
Move rule (mandatory): move a ticket from in-progress to done only when the user explicitly confirms completion (for example: "done", "finished", or "verified") or explicitly asks to move it.
Final archive ordering rule (mandatory): when explicit user completion/verification also triggers repository finalization, move the ticket folder to tickets/done/<ticket-name>/ before the final commit so the committed state includes the archived ticket path.
Reopen rule (mandatory): if the user asks to continue/reopen a completed ticket, move it from tickets/done/<ticket-name>/ back to tickets/in-progress/<ticket-name>/ before making new updates.
Never auto-move a ticket to done based only on internal assessment.
If the user specifies a different location, follow the user-specified path.
Ticket + Worktree Bootstrap (Mandatory First Action)
Before investigation, bootstrap work context in this order:
create/use tickets/in-progress/<ticket-name>/,
if the project is a git repository:
resolve the bootstrap base branch from explicit user instruction when provided; otherwise infer the tracked remote default/integration branch with highest confidence,
when creating a new ticket worktree/branch, refresh tracked remote refs first so bootstrap starts from the latest remote state instead of a stale local head,
create/reuse a dedicated ticket worktree for the ticket branch before writing artifacts,
when creating a new ticket branch, create codex/<ticket-name> from the latest tracked remote base branch,
create/update requirements.md with status Draft from user-provided requirement intent.
Investigation must not start before the ticket bootstrap and requirements.mdDraft are physically written.
If a dedicated worktree already exists for the ticket, reuse it instead of creating a new one.
If the user specifies a base branch, always use the latest tracked remote state of that branch rather than guessing from a local copy.
If remote refresh or base-branch resolution fails, keep Stage 0 Blocked and record the blocker before investigation.
If the environment is not a git repository, continue without worktree setup and still enforce ticket-folder + Draft requirement capture.
Workflow State File (Mandatory Enforcement Artifact)
Create and maintain tickets/in-progress/<ticket-name>/workflow-state.md as the canonical stage-control artifact.
Initialize it during Stage 0 immediately after ticket bootstrap with:
Current Stage = 0,
Code Edit Permission = Locked,
Stage 0 Bootstrap Record filled with bootstrap mode plus, when git repo, requested base branch if any, resolved base remote/base branch, remote-refresh result when performed, worktree path, and ticket branch,
stage gate rows in Not Started/In Progress state.
Update model (mandatory):
rewrite Current Snapshot in place on every stage transition,
append one row to Transition Log for every transition/re-entry,
keep Stage Gates rows current with evidence links/paths.
Source-code edit lock (hard rule):
no source code edits are allowed unless workflow-state.md explicitly shows Code Edit Permission = Unlocked,
default state is Locked; unlock source-code edits only when Stage 6 prerequisites are satisfied.
Re-entry lock rule:
on any Stage 7/8 failure, set Code Edit Permission = Locked before re-entry actions,
record trigger/classification/return path in workflow-state.md before proceeding.
Violation protocol:
if source code is edited while Code Edit Permission = Locked, record a violation entry in workflow-state.md,
pause further source edits, declare re-entry, and return to the required upstream stage path.
Audible Notifications (Speak Tool, Required)
Use the Speak tool for workflow-state transition updates so the user can follow where execution is and what is next.
Playback rule (mandatory): for required audible notifications, call Speak with play=true explicitly.
Do not set play=false by default for required transition notifications.
Exception: set play=false only when the user explicitly requests silent mode.
Transition-driven speak rule (mandatory):
speak only when workflow-state.md is updated for a stage transition, gate decision, re-entry decision, or code-edit lock/unlock change,
do not speak for low-level command execution, intermediate analysis notes, or partial drafts.
Required audible events:
workflow kickoff (task accepted, next stage),
every stage transition (From Stage -> To Stage) after workflow-state.md transition log is appended,
every gate decision (Pass/Fail/Blocked) after gate evidence is written,
every re-entry declaration (classification + return path) after workflow-state.md re-entry section is updated,
every Code Edit Permission change (Locked/Unlocked) after snapshot update.
Speak ordering rule:
update workflow-state.md first,
then emit audible notification reflecting the persisted state.
Spoken message content (mandatory):
current stage,
what just completed/changed (transition or gate result),
next stage/action,
code-edit lock state when it changed.
If multiple transitions happen close together, batch them into one short message after the final persisted update.
Keep each spoken message short (1-2 sentences), status-first, and action-oriented.
If the Speak tool fails or is unavailable, continue the workflow and provide the same update in text.
Do not speak secrets, tokens, or full sensitive payloads.
Execution Model (Strict Stage Gates)
Work in explicit stages and complete each gate before producing downstream artifacts.
Before every stage transition, update workflow-state.md first (snapshot + transition log + gate statuses), then proceed.
Treat workflow-state.md as an execution lock controller, not optional documentation.
Transition authority rule (mandatory): stage movement is controlled by the Stage Transition Contract + Transition Matrix. When a trigger condition is met, transition immediately to the mapped path; do not continue in the current stage by preference.
Transition Execution Rule (mandatory): whenever workflow-state.md is updated to declare a stage transition or classified re-entry, immediately enter and execute the target stage by default, without waiting for another user message. Recording the transition and describing the next step is not sufficient. The transition is not complete until work has actually resumed in the target stage. Stop only for a real blocker or an explicit user-only gate.
Requirements can start as rough Draft from user input/bug report artifacts before deep analysis.
Do not start investigation until ticket/worktree bootstrap is complete and requirements.md status Draft is physically written.
Do not mark understanding pass complete until investigation-notes.md is physically written and current for the ticket.
Do not draft design artifacts (proposed-design.md or small-scope solution sketch in implementation.md) until deep understanding pass is complete and requirements.md reaches Design-ready.
Do not finalize implementation.md baseline sections or start execution tracking sections until the future-state runtime call stack review gate is fully satisfied for the current scope.
Do not start implementation execution until implementation.md baseline is finalized and its execution-tracking sections are initialized.
Do not start source-code edits until all of the following are true in workflow-state.md:
Current Stage = 6,
Code Edit Permission = Unlocked,
Stage 5 gate is Go Confirmed,
required upstream artifacts are marked Pass with evidence.
Do not start Stage 7 executable-validation implementation and execution until implementation execution is complete with required unit/integration verification and Stage 6 modernization/ownership-dependency checks are satisfied.
Canonical Stage Sequence (Quick Map)
Stage
Name
Core Deliverable/Gate
Code Edit Permission
0
Bootstrap + Draft Requirement
Ticket/worktree bootstrap complete (ticket path + if git repo: base branch resolved from tracked remote, remote freshness handled for new bootstrap, dedicated ticket worktree/branch created or reused) + requirements.md = Draft
Locked
1
Investigation + Triage
investigation-notes.md current + scope triage complete
Locked
2
Requirements Refinement
requirements.md reaches Design-ready/Refined
Locked
3
Design Basis
implementation.md solution sketch (Small) or proposed-design.md (Medium/Large)
Locked
4
Future-State Runtime Call Stack
future-state-runtime-call-stack.md current
Locked
5
Future-State Runtime Call Stack Review
Go Confirmed (two consecutive clean rounds with no blockers/persisted updates/new use cases)
Code review decision recorded (Pass/Fail) with <=500 effective-line hard-limit on changed source files only + required >220 delta-gate assessments on changed source files only + data-flow spine inventory/ownership/off-spine concern checks + existing-capability reuse + reusable-owned-structure extraction + shared-structure/data-model tightness + shared-base coherence + repeated-coordination ownership + empty-indirection + scope-appropriate separation of concerns + ownership-driven dependency quality + Authoritative Boundary Rule preservation + file placement within the correct subsystem and folder, with any optional module grouping justified + flat-vs-over-split layout judgment + interface-boundary clarity + naming quality across files/folders/APIs/types/functions/parameters/variables + naming-to-responsibility alignment + no unjustified duplication of code/repeated structures in changed scope + patch-on-patch complexity control + test quality + test maintainability + validation-evidence sufficiency + no-backward-compat/no-legacy checks
Stage Transition Contract (Enforcement)
Stage
Exit Condition (Must Be True To Transition)
On Fail/Blocked
Next Stage On Pass
0 Bootstrap + Draft Requirement
Ticket/worktree bootstrap is complete, the base branch/worktree decision is recorded, and requirements.md status is Draft
Stay in 0 until bootstrap is complete
1
1 Investigation + Triage
investigation-notes.md is current and scope triage (Small/Medium/Large) is recorded
Stay in 1 until investigation evidence is complete
2
2 Requirements
requirements.md is Design-ready (or Refined) with requirement/acceptance-criteria coverage maps
Stay in 2 until requirements are design-ready
3
3 Design Basis
Design basis artifact is current (implementation.md solution sketch for Small, proposed-design.md for Medium/Large)
Stay in 3 and revise design basis
4
4 Future-State Runtime Call Stack
future-state-runtime-call-stack.md is current for in-scope use cases
Stay in 4 and regenerate the future-state runtime call stack
5
5 Future-State Runtime Call Stack Review
Future-state runtime call stack review reaches Go Confirmed (two consecutive clean rounds with no blockers, no required persisted artifact updates, and no newly discovered use cases)
Source implementation complete, required unit/integration checks pass, no backward-compatibility/legacy-retention paths remain in scope, ownership-driven dependencies remain valid (no new unjustified cycles/tight coupling), the Authoritative Boundary Rule remains valid (no new bypass of authoritative boundaries), and touched files sit in the correct folder under the correct owning subsystem
Re-open Stage 0 controls in the same ticket context, then rerun full chain before continuing implementation.
Stage 7 failure classified Local Fix
6 -> 7
Update artifacts first, then code fix, then rerun Stage 7 scenarios.
Stage 7 failure classified Design Impact
1 -> 3 -> 4 -> 5 -> 6 -> 7
Re-open investigation, then re-enter design/runtime chain before retrying Stage 7.
Stage 7 failure classified Requirement Gap
2 -> 3 -> 4 -> 5 -> 6 -> 7
Update requirements first, then rerun downstream chain.
0) Bootstrap Ticket + Capture Draft Requirement
Use stages/00-bootstrap/bootstrap-checklist.md.
Run mandatory first-action sequence:
create/use tickets/in-progress/<ticket-name>/,
if git repo, resolve bootstrap base branch from explicit user instruction when provided; otherwise infer the tracked remote default/integration branch with highest confidence,
if git repo and a new ticket worktree/branch is needed, refresh tracked remote refs first,
if git repo, create/reuse dedicated ticket worktree for codex/<ticket-name>; when creating a new ticket branch, branch it from the latest tracked remote base,
create/update tickets/in-progress/<ticket-name>/workflow-state.md from shared/workflow-state-template.md, set Current Stage = 0, Code Edit Permission = Locked, and fill Stage 0 Bootstrap Record,
capture initial requirement snapshot (requirements.md status Draft) from user input/bug report evidence first (text, images, logs, repro notes, constraints).
Do not run deep investigation before Stage 0 bootstrap and requirements.mdDraft are physically written.
If remote refresh, base-branch resolution, or ticket-worktree creation fails, keep Stage 0 blocked and record the failure in workflow-state.md instead of falling back to a stale local branch.
Before transitioning to Stage 1, update workflow-state.md snapshot + transition log + stage gate evidence.
1) Investigation + Understanding Pass + Triage
Use stages/01-investigation/investigation-guide.md.
Create/update tickets/in-progress/<ticket-name>/investigation-notes.md continuously during investigation. Treat it as a durable evidence log, not a brief summary. Do not keep investigation results only in memory.
Investigation is not read-only: when needed to establish current behavior or isolate root cause, actively reproduce the issue, run focused commands/tests, write small diagnostic scripts, add temporary probes/instrumentation, perform the minimal environment setup or mocks needed to observe the problem, read public APIs/specs/issues, and inspect or clone upstream/vendor/sample repositories when that is the practical way to gather evidence.
Treat Stage 1 methods as problem-dependent and non-exhaustive: use whatever bounded evidence-gathering work is needed to reduce uncertainty and verify assumptions before design.
Minimum codebase understanding pass before design:
identify entrypoints and execution boundaries for in-scope flows,
identify touched files, affected modules/subsystems, and owning concerns,
identify expected canonical file locations/folder owners for touched concerns (for example: platform-specific vs shared),
identify current naming conventions (subsystem names, file names, optional module names when relevant, API style),
identify unknowns that could invalidate design assumptions.
In investigation-notes.md, record in enough detail that later stages can reuse the investigation without repeating the same major search/research work unless the facts have changed:
investigation goals/questions,
detailed source log (local file paths, web links, open-source references, papers, exact commands run, and search queries when those materially shaped the result),
key findings, constraints, and observed behavior tied back to the exact source used,
codebase findings with exact files/symbols/entrypoints/owners where relevant,
file-placement observations (which files are already under the right owning folder, which are misplaced, and what the canonical location should be),
runtime/probe findings when reproductions, traces, or scripts were used,
reproduction prerequisites and environment/setup details when those materially affected the investigation outcome,
external dependency, public API, upstream-source, issue-tracker, or sample-app findings when those materially shaped the understanding,
open unknowns/questions,
implications for requirements/design.
Prefer structured, detailed notes over compressed summaries. Use short synthesis sections for readability, but keep enough concrete evidence in the artifact that later stages do not need to rediscover the same facts from scratch.
2) Refine Requirements Document To Design-Ready (Mandatory)
Use stages/02-requirements/requirements-refinement-guide.md.
Create/update tickets/in-progress/<ticket-name>/requirements.md for all sizes (Small, Medium, Large).
Requirement writing is mandatory even for small tasks (small can be concise).
Use one canonical file path only: update requirements.md in place. Do not create versioned duplicates such as requirements-v2.md.
Requirements maturity flow:
Draft: rough capture from report/input evidence,
Design-ready: refined after deep understanding pass,
Refined: further updates when design/review/implementation feedback reveals gaps.
Minimum required sections in requirements.md:
status (Draft/Design-ready/Refined),
goal/problem statement,
in-scope use cases,
acceptance criteria,
constraints/dependencies,
assumptions,
open questions/risks.
Requirements quality rule (mandatory): requirements must be verifiable behavior specifications, not only descriptive narratives.
For each requirement, include a stable requirement_id and explicit expected outcome.
For each acceptance criterion, include a stable acceptance_criteria_id (for example: AC-001) and an explicit measurable expected outcome.
Acceptance criteria must be testable and cover primary behavior plus relevant edge/error behavior.
Include a requirement coverage map to call-stack use cases (all requirements must map to at least one use case).
Include an acceptance-criteria coverage map to Stage 7 scenarios (all acceptance criteria must map to at least one executable validation scenario).
Confirm the triage result (Small vs Medium vs Large) and rationale in the requirements doc.
Refine requirements from the latest investigation-notes.md; do not derive requirements from memory-only investigation.
Design-ready requirement gate must make expected behavior clear enough to draft design and runtime call stacks.
If understanding is not sufficient to reach Design-ready, continue understanding pass and refine requirements first.
Core Modernization Policy (Mandatory)
Mandatory stance: no backward compatibility and no legacy code retention.
Do not preserve legacy behavior, legacy APIs, compatibility wrappers, dual-write paths, or fallback branches kept only for old flows.
Prefer clean replacement and explicit deletion of obsolete code, files, tests, flags, and adapters in the same ticket.
Do not add compatibility exceptions in this workflow.
Design and review must use the same principles and vocabulary. Review is a check of design quality, not a different rule system.
Use shared/design-principles.md as the shared reference for Stage 3 design work, Stage 5 review, and Stage 8 structural review.
shared/design-principles.md also carries the derived applied workflow, structural triggers, local mechanisms, and example shapes that were previously split into a second helper file.
Use stages/08-code-review/code-review-principles.md as the Stage 8 review reference.
The core principles are:
data-flow spine inventory and clarity,
ownership clarity and boundary encapsulation,
off-spine concerns around the spine.
Separation of concerns is still mandatory, and in this workflow it should become stronger, not weaker. The difference is ordering: start with spine and ownership, then decompose concerns around that structure so each boundary stays coherent.
Ownership is the concrete form of separation of concerns: if a concern has no clear owner, the structure is wrong.
Dependency direction follows ownership. State allowed directions and forbidden shortcuts explicitly when design risk exists.
Authoritative Boundary Rule (mandatory): when one boundary is the authoritative public entrypoint for a domain subject, callers above it should depend on that boundary rather than on both the boundary and one of its internal owned mechanisms at the same time. This is the no boundary bypass / no mixed-level dependency rule.
If callers only bypass an internal concern because the outer boundary does not expose enough usable API, strengthen that authoritative boundary or redesign the ownership split explicitly instead of normalizing the bypass.
Off-spine concern reuse follows ownership too. Before adding a new helper or off-spine concern, check whether an existing capability area or subsystem already owns that responsibility.
Spine, owner, and off-spine concern are architecture relationship terms, not naming templates. Do not literalize them into file, folder, class, service, or type names.
Module placement and file placement follow ownership. If a file's path no longer matches the real concern, move, split, or rename it instead of rationalizing the mismatch.
Subsystem, folder, and file mapping should be spine-led and ownership-led, but not mechanical. Optional module groupings are secondary structure only when they make boundaries easier to read.
Distinct structural depths often deserve distinct folders, but do not force artificial over-splitting. If a flatter layout is clearer, justify it explicitly.
3) Draft The Proposed Design Document
Required for Medium/Large. Optional for Small.
Prerequisite: requirements.md is Design-ready (or Refined) and current for this ticket.
For Small, do not require a full proposed design doc; use the draft implementation.md solution sketch as the lightweight design basis for runtime call stacks.
Architecture-first rule: define the target architecture shape before mapping work onto existing files.
Design from the data-flow spine first:
identify the primary execution/data-flow spine(s),
enumerate all relevant spines that matter to understanding the design,
name the main domain subject nodes on each spine,
define what each main-line node owns,
identify off-spine concerns that serve those owners,
check whether those off-spine needs should reuse or extend existing capability areas/subsystems before creating anything new,
record the return/event spine(s) when the change is async or event-driven,
record bounded local/internal spines when a loop, worker cycle, state machine, or dispatcher materially shapes one owner's behavior.
Separation-of-concerns rule (mandatory): once the spine and owners are clear, split off-spine responsibilities so each main node and each off-spine concern remains self-encapsulated. Do not overload one spine node with every concern just because it sits on the main line.
Authoritative Boundary Rule (mandatory): identify which boundaries are authoritative public entrypoints versus internal owned sub-layers. Callers above an authoritative boundary should not depend on both that boundary and one of its internal managers, repositories, helpers, or lower-level concerns at the same time.
Do not anchor design to current file layout when the layout is structurally wrong for the target behavior.
File-placement rule (mandatory): target file paths must match owning concern/boundary/platform; if a file is codex-specific, cloud-specific, UI-specific, infra-specific, or otherwise scope-bound, place it in that canonical area instead of near whichever caller currently imports it.
Folder-layout rule (mandatory): map subsystems/folders/files from the spine and ownership model, and add optional module groupings only when they materially help readability rather than by default. Do it with judgment rather than as a rigid one-folder-per-step projection. If a flatter layout is clearer, justify it. If a split is proposed, it should reflect a real owner or boundary.
Interface-boundary rule (mandatory): design APIs, queries, commands, and reused service methods around explicit subject ownership and explicit identity shape. Do not use one generic boundary that accepts an ambiguous ID or generic selector and guesses whether the subject is, for example, an agent run, team run, or team member run.
4) Build Future-State Runtime Call Stacks Per Use Case
Required for all sizes (Small, Medium, Large).
For Small, keep it concise but still cover each in-scope use case with primary path plus key fallback/error branch when relevant.
Basis by scope:
Small: use the draft implementation.md solution sketch as the design basis.
Medium/Large: use the proposed design document as the design basis.
For each use case, write a future-state runtime call stack from entry point to completion in a debug-trace format.
Use stable use_case_id values and ensure IDs match the design coverage matrix and review artifact.
For each use case, record mapped spine ID(s), spine scope, and governing owner from the design basis.
Use two allowed use-case source types:
Requirement: derived from explicit requirement behavior.
Design-Risk: derived from architecture or technical risk that must be validated.
Coverage rule (mandatory): every in-scope requirement must be covered by at least one Requirement use case.
For each Design-Risk use case, record the technical objective/risk justification and expected observable outcome.
Treat this artifact as to-be architecture behavior derived from the proposed design (or small-scope solution sketch), not as-is code tracing.
If current code differs from target design, model the target design behavior and record migration/transition notes separately.
Include file and function names at every frame using path/to/file.ts:functionName(...).
Show architectural boundaries explicitly (e.g., controller -> service -> repository -> external API).
Include primary path and fallback/error paths, not only happy path.
Include explicit use-case coverage status per use case (primary/fallback/error) and mark intentional N/A branches.
If a use case exercises a bounded local spine, make that explicit instead of hiding it inside a generic end-to-end stack.
Mark async/event boundaries (await, queue enqueue/dequeue, worker loop handoff).
Mark state mutations and persistence points (in-memory state, cache write, DB/file write).
Capture decision gates and conditions that choose one branch over another.
Create future-state-runtime-call-stack-review.md as a mandatory review artifact.
Review focus is future-state correctness and implementability against the target design basis (proposed-design.md for Medium/Large, small-scope solution sketch in implementation.md for Small), not parity with current code structure.
Review must challenge architecture choice itself (spine/ownership/off-spine concerns/allocation), not only file-level separation of concerns.
Review must reuse the same shared design principles from Stage 3; do not apply a different principle set in review.
Review must explicitly verify file placement against the Stage 3 design basis; wrong folder placement is a structural defect, not a cosmetic nit.
Review must explicitly judge whether the proposed layout is readable for the scope or whether it is either too flat or too artificially fragmented.
Review must explicitly check example-based clarity when the design uses examples or when the design would otherwise stay too abstract.
Run review in explicit rounds and record each round in the same review artifact.
In every round, run a dedicated missing-use-case discovery sweep before verdicting the round.
Review each use case against these criteria:
architecture fit check (Pass/Fail): chosen architecture shape is appropriate for this use case and expected growth,
data-flow spine clarity check (Pass/Fail): the main end-to-end motion and any bounded local spines are readable and not fragmented across peer coordinators,
spine inventory completeness check (Pass/Fail): all relevant spines are explicitly named with clear start/end boundaries and owner relationships,
ownership clarity check (Pass/Fail): main-line nodes and off-spine concerns have clear owners with concrete authority,
off-spine concern clarity check (Pass/Fail): off-spine concerns serve the spine without competing with it or owning hidden sequencing,
boundary placement check (Pass/Fail): responsibilities are assigned to the right owner or subsystem boundary,
6) Implement Source + Unit/Integration And Track Progress
Use a bottom-up, test-driven approach: implement foundational dependencies first.
Sequence is mandatory:
first ensure implementation.md baseline is finalized,
then initialize/update the execution-tracking sections in implementation.md,
then execute source implementation with required unit/integration verification.
Stage separation rule (mandatory):
Stage 6 implements source code and verifies file and service boundaries, while preserving readable subsystem grouping, with unit/integration tests.
Stage 7 implements executable validation assets and runs the mapped Stage 7 scenarios against acceptance criteria.
Stage 8 runs the code review gate after Stage 7 passes.
Include requirement traceability in plan/progress (requirement -> design section -> call stack/use_case -> implementation tasks/tests).
Include spine traceability in plan/progress (spine -> owner -> use case -> implementation tasks -> validation scenarios).
Integration test coverage is required for behavior that crosses module boundaries, process boundaries, storage boundaries, or external API boundaries. If any such behavior is not covered, record a concrete rationale.
Finalize planning artifacts before kickoff:
Small: refine the draft implementation.md solution sketch until future-state runtime call stack review passes final stability gate (Go Confirmed).
Medium/Large: create implementation.md only after future-state runtime call stack review passes final stability gate (Go Confirmed).
Initialize the execution-tracking sections in implementation.md at implementation kickoff, after required pre-implementation artifacts are ready (including the proposed design document for Medium/Large).
One merged implementation artifact with both baseline planning and live tracking is required for all sizes (Small, Medium, Large).
Keep detailed Stage 7, Stage 8, and Stage 9 records in their own canonical artifacts. implementation.md should carry only implementation-owned planning/tracking plus downstream handoff inputs and short status pointers.
Treat future-state runtime call stack + review as a pre-implementation verification gate: ensure each use case is represented and reviewed before coding starts.
Start implementation only after the review gate says implementation can start and all in-scope use cases are .
7) Implement API/E2E And Other Executable Validation And Run Stage 7 Gate (Mandatory)
Run this stage immediately after Stage 6 completes.
At Stage 7 entry, update workflow-state.md and keep Code Edit Permission = Unlocked while Stage 7 validation artifacts are being implemented/executed.
Create/update tickets/in-progress/<ticket-name>/api-e2e-testing.md as the canonical scenario + result artifact.
Multi-round artifact rule (mandatory): keep one canonical api-e2e-testing.md path for the ticket. Do not create versioned copies by default. On every rerun, check prior unresolved failures first, then continue validation. The latest round result is authoritative; earlier rounds remain history.
Round-resolution rule (mandatory): on Stage 7 round >1, update the prior-failure resolution check before declaring the new gate result.
Stage 7 scope includes:
implementing durable repo-resident validation assets as needed,
implementing API, browser/UI, native desktop/UI, CLI, process/lifecycle, integration, multi-service, or multi-node harnesses as needed,
executing Stage 7 scenarios mapped from acceptance criteria.
Durable validation-first rule (mandatory): first implement or update the validation tests, harnesses, and assets that should live in the repository and govern future changes.
이 SKILL.md는 매우 커서 SkillsMP가 여기에는 첫 섹션만 미리 보여줍니다.GitHub에서 보기
Do not start code review until the Stage 7 executable-validation gate is Pass.
Do not start post-testing docs/ synchronization until code review is complete (for infeasible acceptance criteria in Stage 7, explicit user waiver + constraints + compensating evidence + residual risk must be recorded).
Do not close the task until post-testing docs/ synchronization is completed (or explicit no-impact decision is recorded with rationale), and do not mark final completion until any required Stage 10 user-verification/archive/finalization work is complete.
User-verification hold rule (mandatory): after Stage 9 passes, persist the handoff summary and keep Stage 10 open until the user explicitly confirms completion/verification (for example after manual testing). Do not commit, push, merge, run release/publication/deployment work, or move the ticket to done before that user signal.
Release-notes artifact rule (mandatory when applicable): if the ticket leads to a user-facing app release or any GitHub Release body, Stage 10 must also persist tickets/in-progress/<ticket-name>/release-notes.md with short functional user-facing notes before final release. If not applicable, record an explicit release-notes not required rationale in the handoff summary.
Repository finalization rule (mandatory for git repositories): after the explicit user completion/verification signal is received and before Stage 10 is marked complete, first move the ticket folder to tickets/done/<ticket-name>/, then commit all in-scope changes on the ticket branch (including the moved ticket files), push the ticket branch to remote, update the resolved finalization target branch from remote, merge the ticket branch into that updated target branch locally, and push the updated target branch.
Release/publication/deployment rule (conditional): after repository finalization, if the project has an applicable documented release, publication, tagging, or deployment step, run that project-specific method. This may be a release script, a documented command, a git tag workflow, GitHub Release creation, or another documented deployment/publication path. If no such step is applicable, record release/publication/deployment not required (or equivalent rationale) in the handoff summary instead of blocking repository finalization.
Release publication handoff rule (mandatory when release notes are required): after the ticket is moved to tickets/done/<ticket-name>/, pass the archived ticket-local tickets/done/<ticket-name>/release-notes.md artifact into the project release/publication path when such a step is applicable (for example via a release script, a documented command, a tag/release workflow, or the repo's release-body source file).
Finalization-target rule (mandatory): use the Stage 0 Resolved Base Remote and Resolved Base Branch as the default Stage 10 merge target unless the user explicitly overrides that target later. If the target cannot be derived with high confidence, pause Stage 10 and ask once before merge/finalization instead of guessing.
Post-finalization cleanup rule (mandatory when a dedicated ticket worktree/branch exists): after repository finalization and any applicable release/publication/deployment work are complete (or explicitly recorded as not required), remove the dedicated ticket worktree recorded in Stage 0, run worktree-prune cleanup, and, when the local ticket branch is fully merged into the resolved finalization target and no longer needed, delete that local ticket branch. Do not delete remote branches unless explicit user instruction or documented project policy requires it. If cleanup requires stepping out of the ticket worktree, execute it from a safe parent repo checkout instead of skipping it.
Stage 10 blockage rule (mandatory): if the move to tickets/done/, commit, push, target-branch update, merge, or required post-finalization cleanup fails after user confirmation, keep Stage 10 In Progress/Blocked, record the blocker in workflow-state.md, and do not mark final handoff complete. If a release/publication/deployment step is applicable and later fails or is undocumented, record that blocker too, but do not undo completed repository finalization.
Keep the ticket folder under tickets/in-progress/ until explicit user completion confirmation is received.
Treat engineering completion, user verification, ticket archival, repository finalization, release/publication/deployment, and post-finalization cleanup as separate gates: engineering completion ends at implementation + Stage 7 executable-validation gate + code review + docs sync; after that, wait for explicit user completion/verification; only then move the ticket to tickets/done/, complete required Stage 10 repository finalization, run any applicable release/publication/deployment work, and complete any required worktree/branch cleanup.
Small scope exception: drafting implementation.md (solution sketch section only) before review is allowed as design input, but this draft does not unlock implementation kickoff.
Future-state runtime call stack review must run as iterative deep-review rounds (not one-pass review).
Go Confirmed cannot be declared immediately after required upstream artifact updates from a blocking round.
Stability rule (mandatory): unlock Go Confirmed only after two consecutive deep-review rounds report no blockers, no required persisted artifact updates, and no newly discovered use cases.
First clean round is provisional (Candidate Go), second consecutive clean round is confirmation (Go Confirmed).
Missing-use-case discovery rule (mandatory): every Stage 5 round must run a dedicated missing-use-case discovery sweep (requirement coverage, boundary crossings, fallback/error branches, and design-risk scenarios).
A Stage 5 round is not clean if it discovers new use cases, requires persisted artifact updates, or finds blockers.
Any review finding with a required design/call-stack update is blocking; regenerate affected artifacts and re-review before proceeding.
If design/review reveals missing understanding or requirement ambiguity, return to understanding + requirements stages, update requirements.md, then continue design/review.
Stage 5 classified re-entry mapping (mandatory):
Design Impact (clear spine/ownership/boundary/naming/architecture issue with high confidence): return to Stage 3 -> Stage 4 -> Stage 5.
Requirement Gap (missing/ambiguous requirement or acceptance criterion): return to Stage 2 -> Stage 3 -> Stage 4 -> Stage 5.
For any Stage 5 blocking round, record classification + return path in future-state-runtime-call-stack-review.md and workflow-state.md before starting the next round.
Stage 6 non-local re-entry rule (mandatory): if a Stage 6 issue is classified as Design Impact, Requirement Gap, or Unclear, record classification + return path in workflow-state.md, set Code Edit Permission = Locked, and transition to the mapped upstream stage path before further source edits.
Re-entry declaration rule (mandatory): when a post-implementation gate (Stage 7 executable validation or Stage 8 code review) finds issues, explicitly record:
Re-entry declaration must be recorded in workflow-state.md before any artifact/code update work begins.
Re-entry Completion Rule (mandatory): after a classified re-entry is recorded and the target return path is persisted in workflow-state.md, immediately resume the first returned stage by default, without waiting for another user message. Do not stop after only updating state and describing what should happen next. Re-entry is not complete until work has actually resumed in that returned stage.
No-direct-patch rule (mandatory): for post-implementation gate findings, do not edit source code first. Update required upstream artifacts first based on classification path.
Re-entry mapping (mandatory):
Local Fix: update implementation artifacts (implementation.md / api-e2e-testing.md / code-review.md as applicable), then implement fix, then rerun Stage 6 -> Stage 7; once Stage 7 passes, continue to Stage 8.
Design Impact: return to Stage 1 -> Stage 3 -> Stage 4 -> Stage 5 -> Stage 6 -> Stage 7; once Stage 7 passes, continue to Stage 8.
Requirement Gap: return to Stage 2 -> Stage 3 -> Stage 4 -> Stage 5 -> Stage 6 -> Stage 7; once Stage 7 passes, continue to Stage 8.
Stage 0 in a re-entry path means re-open bootstrap controls in the same ticket/worktree (workflow-state.md, lock state, artifact baselines); do not create a new ticket folder.
If the user asks for all artifacts in one turn, still enforce stage gates within that turn (iterate review rounds first; only then produce implementation artifacts).
Stage-first persistence rule: if a Stage 5 round finds issues, first record classification + return path and transition to the required upstream stage in workflow-state.md, then persist required artifact updates in that stage.
No mental-only review refinements: do not carry unresolved updates in memory across rounds.
For each review round, record explicit persisted artifact updates in future-state-runtime-call-stack-review.md:
updated files,
new artifact versions,
changed sections,
which findings were resolved.
A review round cannot be considered complete until its required file updates are physically written.
If announcing review-round status, do it only after the round record + required staged artifact updates are physically written and the related Stage 5 gate status is persisted in workflow-state.md.
Locked
9
Docs Sync
docs-sync.md current + docs/ updates complete or no-impact rationale recorded
Locked
10
Final Handoff
Delivery summary ready + explicit user verification -> move ticket to done -> repository finalization into resolved target branch (when git repo) + any applicable release/publication/deployment step completed or explicitly recorded as not required + required post-finalization worktree/branch cleanup complete when applicable + ticket state decision recorded
executable validation scenarios are implemented and all executable mapped acceptance criteria are Passed (or explicitly Waived by user for infeasible cases), and all relevant executable spines have passing scenario evidence (or explicit N/A rationale)
docs-sync.md is current and docs updates are completed, or explicit no-impact rationale is recorded
If docs cannot yet be made truthful, classify and re-enter (Local Fix: 6 -> 7 -> 8 -> 9, Requirement Gap: 2 -> 3 -> 4 -> 5 -> 6 -> 7 -> 8 -> 9, Unclear: 0 -> 1 -> 2 -> 3 -> 4 -> 5 -> 6 -> 7 -> 8 -> 9); stay in 9 (Blocked) only for external docs blockers that do not require upstream artifact changes
10
10 Final Handoff
Handoff summary is complete, explicit user completion/verification instruction is received, the ticket has been moved to tickets/done/<ticket-name>/, and, when in a git repository, repository finalization is complete, any applicable release/publication/deployment step is complete or explicitly recorded as not required, and required post-finalization worktree/branch cleanup is complete when applicable
Stay in Stage 10 until the user verifies completion and Stage 10 archival/finalization/cleanup is complete
Record the move/commit/git/release-publication-deployment/cleanup blocker, resolve it, then finish handoff.
Boundary rule: narrowly scoped repro scripts/tests, temporary logs, investigation-only setup, disposable cloned repos, and throwaway harnesses remain Stage 1 work when they exist to gather evidence; durable production fixes and durable validation assets still belong to later stages.
Re-entry rule: when later implementation/testing uncovers large or unclear issues, reopen this understanding stage and append new evidence, unknowns, and implications before changing requirements/design artifacts.
Do not draft design artifacts or runtime call stacks until this understanding pass is complete and requirements.md reaches Design-ready.
Classify as Small, Medium, or Large using practical signals:
Estimated files touched (roughly <= 3 is usually Small if not cross-cutting).
New/changed interface boundaries, schema/storage changes, or cross-cutting behavior.
Multi-boundary impact (API + service + persistence + runtime flow) or architectural impact.
Choose workflow depth:
Small: create a draft implementation.md solution sketch, build per-use-case future-state runtime call stacks from that basis, review them, then refine until stability gate Go Confirmed and continue in the same implementation.md artifact with live execution tracking.
Medium: create proposed design doc first, build future-state runtime call stacks from the proposed design doc, run iterative deep-review rounds until stability gate Go Confirmed, and only then create implementation.md as the combined implementation baseline + live execution tracker.
Large: create proposed design doc first, build future-state runtime call stacks from the proposed design doc, run iterative deep-review rounds until stability gate Go Confirmed, and only then create implementation.md as the combined implementation baseline + live execution tracker.
Re-evaluate during implementation; if scope expands or smells appear, escalate from Small to full workflow.
Before transitioning to Stage 2, update workflow-state.md snapshot + transition log + stage gate evidence.
After triage depth is finalized (Small/Medium/Large) and investigation-notes.md is current, announce only with the persisted workflow-state.md transition/gate update.
After requirements.md reaches Design-ready and is confirmed as design input, announce only with the persisted workflow-state.md transition/gate update.
Interface-boundary rule: APIs, queries, commands, and reused service methods must also follow ownership and separation of concerns. Prefer one boundary per subject/responsibility with explicit identity shape; avoid generic boundaries that infer subject type from ambiguous IDs or generic selectors.
Use examples when they materially improve clarity. Do not keep a non-obvious design fully abstract when a short example would explain the intended shape faster.
Keep is valid when the current structure already preserves a readable spine, clear ownership, and off-spine concerns that do not compete with the spine.
Current-state grounding rule (mandatory): before finalizing the target design, read the current implementation enough to understand the existing main flow, current owners, coupling problems, file placement drift, and real migration constraints.
Explicitly evaluate whether new subsystem-internal module groupings, boundary interfaces, or orchestration owners should be introduced.
Explicitly evaluate whether existing subsystems/owners should be split, merged, moved, or removed.
Explicitly evaluate whether an existing capability area or subsystem should absorb the new off-spine responsibility instead of introducing a fresh helper beside the spine.
When coordination policy is repeated across callers (provider selection, fallback, retry, aggregation, routing, fan-out), extract that policy into a clear owner instead of leaving it fragmented.
Reject empty indirection: if a proposed layer or optional module grouping only passes work through and owns no real policy, boundary, or sequencing, keep the structure flatter.
Record the architecture direction decision and rationale (complexity, testability, operability, evolution cost).
For straightforward local changes, one concise decision is enough; alternatives are optional.
For non-trivial or uncertain architecture changes, include a small alternatives comparison before deciding.
Choose the proper structural change for the problem (Keep, Add, Split, Merge, Move, Remove) without bias toward minimal edits.
Removal-first tightening rule (mandatory): when the target design introduces a clearer owner, reusable structure, or better file boundary, explicitly identify what redundant/fragmented pieces become unnecessary and remove/decommission them in scope when possible.
Keep is a valid outcome when the current data-flow spine, ownership boundaries, and off-spine concerns are already coherent.
If a file's concern and folder disagree, Keep is usually invalid; prefer explicit Move, Split, or justified Promote Shared.
Functional/local correctness is not sufficient: if a bug fix "works" but degrades the data-flow spine, ownership boundaries, or off-spine concerns, redesign the structure instead of accepting the patch.
Reject patch-on-patch hacks that bypass clear boundaries just to make a local change compile.
Follow separation of concerns after the target architecture direction is chosen: each file should own a clear responsibility, and each subsystem should own a coherent category of work. Optional module groupings may exist inside a subsystem when they make the structure easier to read. This is a stronger SoC check, not a weaker one, because off-spine concerns must also land in explicit owners instead of drifting into generic blobs.
Ownership-driven dependency rule (mandatory): define and preserve clear subsystem, service, and component boundaries so allowed dependencies follow ownership and forbidden shortcuts are explicit.
Existing-capability reuse rule (mandatory): do not create a new helper/off-spine concern by default. Reuse or extend an existing well-owned subsystem when the fit is natural, and justify Create New when no current area is appropriate.
Reusable-owned-structure rule (mandatory): when repeated data structures, types, normalizers, converters, mappers, or schemas appear across several files, extract them into reusable owned files under the correct subsystem instead of duplicating them.
Data-model-tightness rule (mandatory): when a shared structure is extracted or revised, tighten it. Remove redundant attributes, avoid overlapping parallel representations for the same subject, and keep each field's meaning singular and explicit.
Shared-base coherence rule (mandatory): use a shared base/core only when it represents a real common shape. If one case needs extra fields or behavior, prefer meaningful specialized variants or composition instead of growing a one-for-all structure full of optional attributes.
For Small, the solution sketch in implementation.md must still include a concise architecture sketch (target ownership boundaries and any new files, plus optional module groupings only when they help readability).
Apply separation-of-concerns at the correct technical boundary for the stack:
frontend/UI scope: evaluate responsibility at view/component level (each component should own a clear concern),
non-UI scope (backend/service/worker/domain): evaluate responsibility at file and service level, while keeping subsystem grouping readable,
integration/infrastructure scope: each adapter or integration-focused file/group should own one integration concern with clear contracts.
Make the proposed design doc delta-aware, not only target-state:
include current-state summary (as-is),
include target-state summary (to-be),
include explicit change inventory rows for Add, Modify, Rename/Move, Remove.
Proposed design document organization rule (mandatory): write the design spine-first, not file-first. The main structure of the document should be the spine inventory, main domain subject nodes, ownership, and off-spine concerns.
Design concretion order (mandatory): after the spine and ownership model are clear, move through subsystem/capability-area allocation, then draft file responsibilities, then extract reusable owned structures where repetition appears, then finalize file responsibilities, then folder/path mapping.
File sections are required, and subsystem shape plus optional module groupings may be included where they add clarity, but only as derived implementation mapping after the spine-led explanation is clear.
Example rule (mandatory when needed): if a short good-shape example or bad-shape anti-example would make a non-obvious design materially clearer, include it in the proposed design doc instead of leaving the point abstract.
For each file, state mapped spine, target owner/boundary placement, responsibility, key APIs, inputs/outputs, dependencies, and change type. Mention subsystem placement and optional module grouping where relevant.
For each new off-spine concern, state whether it reuses/extends an existing capability area or why a new area is required.
For each file, state whether its target path matches the owning concern and record the move/split rationale when placement changes are required.
Include a naming decisions section:
proposed file names, subsystem names, optional module names when relevant, and API names,
rationale for each naming choice,
mapping from old names to new names when renaming.
Use natural, unsurprising, implementation-friendly names; naming choices should be understandable without domain-specific insider context.
Add a naming-drift check section in the design doc:
verify each file name and any subsystem/module name still matches its current responsibility after scope expansion,
identify drifted names (name no longer represents real behavior),
choose corrective action per drifted item (Rename, Split, Move, or N/A with rationale),
map each corrective action to change inventory rows and implementation tasks.
Document dependency flow and cross-reference risk explicitly (including how cycles are avoided or temporarily tolerated).
Document allowed dependency directions between ownership boundaries and note any temporary violations with removal plan.
For Rename/Move and Remove, include decommission/cleanup intent (import cleanup and dead-code removal).
Do not produce an addition-only design when the real improvement comes from both adding a clearer owner and removing redundant fragments that the new owner makes unnecessary.
Capture data models and error-handling expectations if relevant.
Add a use-case coverage matrix in the design doc with at least:
use_case_id,
primary path covered (Yes/No),
fallback path covered (Yes/No/N/A),
error path covered (Yes/No/N/A),
mapped sections in runtime call stack doc.
Version the design during review loops (v1, v2, ...) and record what changed between rounds.
Use the template in stages/03-design/proposed-design-template.md as a starting point.
No-backward-compat design gate (mandatory): proposed design must not introduce compatibility wrappers, dual-read/dual-write flows, legacy adapters, or fallback branches kept only for old behavior.
If backward-compatibility or legacy-retention mechanisms are required to make the design work, classify as Fail and redesign the architecture direction before continuing.
Do not speak solely because proposed-design.md changed; announce only when the related workflow-state.md transition/gate update is persisted.
Note key data transformations (input schema -> domain model -> output payload).
Version call stacks to match design revisions from review loops (v1, v2, ...).
Use the template in stages/04-future-state-runtime-call-stack/future-state-runtime-call-stack-template.md.
Do not speak solely because future-state-runtime-call-stack.md changed; announce only when the related workflow-state.md transition/gate update is persisted.
Authoritative Boundary Rule check (Pass/Fail): callers use the authoritative boundary instead of depending on both an outer owner and one of its internal lower-level concerns,
file-placement alignment check (Pass/Fail): each file path matches the owning concern/boundary/platform, and any shared placement is explicitly justified,
flat-vs-over-split layout judgment (Pass/Fail): layout is neither too flat nor artificially fragmented for the scope,
interface/API/service-method boundary clarity (Pass/Fail): APIs, queries, commands, and reused service methods expose one clear subject/responsibility with explicit identity shape instead of generic ambiguous IDs or mixed-subject selectors,
existing-structure bias check (Pass/Fail): design is not forced to mirror current files when that harms target architecture,
anti-hack check (Pass/Fail): no patch-on-patch tricks that hide architecture issues behind local fixes,
local-fix degradation check (Pass/Fail): a functionally working fix does not degrade the data-flow spine, ownership boundaries, or off-spine concerns,
example-based clarity (Pass/Fail/N/A): non-obvious areas use examples when needed, and the examples actually clarify the target shape,
terminology and concept vocabulary is natural and intuitive (Pass/Fail),
file and API naming is clear and unsurprising for implementation mapping (Pass/Fail),
name-to-responsibility alignment under scope drift (Pass/Fail),
future-state alignment with target design basis (Pass/Fail),
use-case source traceability (Pass/Fail) (Requirement or Design-Risk, with source reference),
requirement coverage closure (Pass/Fail) (all requirements mapped to at least one use case),
design-risk use-case justification quality (Pass/Fail) (clear objective/risk and expected outcome),
business flow completeness (Pass/Fail),
gap findings,
ownership-appropriate separation-of-concerns check (Pass/Fail) (frontend/UI: view/component boundary; non-UI: file and service boundaries with readable subsystem grouping),
dependency flow smells,
redundancy/duplication check (Pass/Fail),
simplification opportunity check (Pass/Fail),
remove/decommission and obsolete/deprecated/dead-path cleanup completeness for impacted changes (Pass/Fail),
legacy-retention cleanup check (Pass/Fail) (obsolete old-behavior/deprecated/dead paths are removed for impacted scope),
backward-compatibility mechanism check (Pass/Fail) (no compatibility wrappers, dual-path behavior, or legacy fallback branches retained),
overall verdict (Pass/Fail).
Round policy:
use deep review for every round (challenge assumptions, edge cases, and cleanup quality),
if a round finds blockers, requires persisted artifact updates, or discovers new use cases, classify the round (Design Impact/Requirement Gap/Unclear), apply required upstream updates through the classified stage path, and reset clean-review streak to 0,
if a round finds no blockers, no required persisted artifact updates, and no newly discovered use cases, mark Candidate Go and increment clean-review streak,
open Go only when clean-review streak reaches 2 consecutive deep-review rounds.
Across rounds, track trend quality: issues should decline in count/severity or become more localized; otherwise escalate design refinement before proceeding.
Round staged re-entry discipline (mandatory):
Step A: run review + missing-use-case discovery and record findings in the current round.
Step B: classify blockers as exactly one of Design Impact/Requirement Gap/Unclear.
Step C: if findings require updates, transition to the classified upstream stage path first, then modify required artifacts there and bump versions (vN -> vN+1).
Step D: record an "Applied Updates" entry in the review artifact (what changed, where, and why), including classification + return path.
Step E: start the next round from updated files only; do not carry unresolved edits in memory.
Step F: record clean-review streak state in the review artifact (Reset, Candidate Go, or Go Confirmed).
Gate Go criteria (all required):
architecture fit check is Pass for all in-scope use cases,
data-flow spine clarity check is Pass for all in-scope use cases,
spine inventory completeness check is Pass for the design basis,
ownership clarity check is Pass for all in-scope use cases,
off-spine concern clarity check is Pass for all in-scope use cases,
boundary placement check is Pass for all in-scope use cases,
ownership-driven dependency check is Pass for all in-scope use cases,
Authoritative Boundary Rule check is Pass for all in-scope use cases,
file-placement alignment check is Pass for all in-scope use cases,
interface/API/service-method boundary clarity is Pass for all in-scope use cases,
existing-structure bias check is Pass for all in-scope use cases,
anti-hack check is Pass for all in-scope use cases,
local-fix degradation check is Pass for all in-scope use cases,
terminology/concept vocabulary is Pass for all in-scope use cases,
file and API naming clarity is Pass for all in-scope use cases,
name-to-responsibility alignment under scope drift is Pass for all in-scope use cases,
future-state behavior is consistent with target design basis across all in-scope use cases,
ownership-appropriate separation-of-concerns check is Pass for all in-scope use cases,
use-case coverage completeness is Pass for all in-scope use cases,
use-case source traceability is Pass for all in-scope use cases,
requirement coverage closure is Pass for the full in-scope requirement set,
design-risk use-case justification quality is Pass for all design-risk use cases,
redundancy/duplication check is Pass for all in-scope use cases,
simplification opportunity check is Pass for all in-scope use cases,
all in-scope use cases have overall verdict Pass,
no unresolved blocking findings (including any required design/call-stack updates),
for any impacted Add/Modify/Rename/Move/Remove scope items, decommission/cleanup and obsolete/deprecated/dead-path checks are Pass,
no new use cases are discovered in either of the two clean rounds,
legacy-retention cleanup check is Pass for all in-scope use cases,
backward-compatibility mechanism check is Pass for all in-scope use cases,
two consecutive deep-review rounds have no blockers, no required persisted artifact updates, and no newly discovered use cases.
If issues are found:
classify each blocking round as exactly one of Design Impact/Requirement Gap/Unclear and record it in future-state-runtime-call-stack-review.md and workflow-state.md.
Design Impact (clear and high-confidence design issue): update design basis in Stage 3 (Medium/Large: proposed-design.md; Small: design section in implementation.md), regenerate call stacks in Stage 4, then return to Stage 5.
Requirement Gap: update requirements.md first in Stage 2 (status Refined), then update design basis in Stage 3, regenerate call stacks in Stage 4, then return to Stage 5.
Unclear (cross-cutting/low-confidence): update investigation-notes.md in Stage 1 first, then proceed through 2 -> 3 -> 4 -> 5.
If naming drift is found, prefer explicit rename/split/move updates in the same review loop instead of carrying stale names forward.
Even when a round reports no findings, still complete the round record in-file and run another deep-review round until the two-consecutive-clean stability rule is satisfied.
Repeat until all gate Go Confirmed criteria are satisfied.
Use the template in stages/05-future-state-runtime-call-stack-review/future-state-runtime-call-stack-review-template.md.
Pass
Before first source-code edit in Stage 6, update workflow-state.md:
set Current Stage = 6,
set Code Edit Permission = Unlocked,
record transition evidence that Stage 5 gate is Go Confirmed and pre-edit checklist is satisfied.
Ensure traceability when a proposed design doc exists: every design change-inventory row (especially Rename/Move and Remove) maps to implementation tasks and verification steps.
Enforce clean-cut implementation: do not keep legacy compatibility paths, dead code, or dormant replaced paths in scope.
No-backward-compat implementation rule (mandatory): reject compatibility wrappers, dual-path reads/writes, and old-behavior fallback branches even if they make rollout easier.
Shared-principles implementation rule (mandatory): implementation must continue applying the shared design principles and common design practices independently at file-level detail. The reviewed design artifact is the current target, not permission to ignore tighter realities discovered during coding.
Ownership-dependency preservation rule (mandatory): implementation must not introduce new tight coupling, forbidden shortcuts, or cyclic dependencies across owners or subsystem boundaries.
File-placement preservation rule (mandatory): do not leave touched files in the wrong concern folder just to minimize edits; move/split them when the current path no longer matches ownership.
Implementation completeness rule (mandatory): implementation is not complete until obsolete code paths, dead code, unused helpers/tests/flags/adapters, dormant replaced paths, and compatibility shims in scope are removed.
Proactive source-file size rule (mandatory): treat the Stage 8 source-file size gates as active Stage 6 guardrails for changed source implementation files. Do not knowingly grow or leave a changed source implementation file above 500 effective non-empty lines. If a changed source implementation file trends toward that limit, or if one file's current diff exceeds 220 changed lines, split/refactor/escalate during implementation instead of waiting for Stage 8 to reject it. Test files stay outside that hard source-file limit.
Stage 6 failure classification rule (mandatory):
Local Fix: issue is bounded and does not require requirement/design/call-stack updates; remain in Stage 6.
If detailed implementation work reveals that the reviewed design is incomplete, weak, or wrong, do not patch around it locally. Record the issue and classify Design Impact so upstream design/runtime artifacts are corrected before further source edits.
If a change only "works" by leaving a file in the wrong folder or adding new misplaced files, do not classify it as Local Fix; treat it as Design Impact unless the real issue is missing requirements.
Use "one file at a time" as the default execution strategy, not an absolute rule.
When rare cross-referencing is unavoidable, allow limited parallel/incomplete implementation, but explicitly record:
the cross-reference smell,
the blocked dependency edge,
the condition to unblock,
the required proposed-design-document update.
Update progress in real time during implementation (immediately after file status changes, test runs, blocker discoveries, and design-feedback-loop updates).
Track change IDs and change types in progress (Add/Modify/Rename/Move/Remove) so refactor deltas remain explicit through execution.
Track file build state explicitly (Pending, In Progress, Blocked, Completed, N/A).
Track unit/integration test state separately (Not Started, In Progress, Passed, Failed, Blocked, N/A).
If a file is blocked by unfinished dependencies, mark it Blocked and record the dependency and unblock condition.
Mark a file Completed only when implementation is done and required tests are passing.
Mark Stage 6 complete only when:
implementation baseline scope is delivered (or deviations are explicitly documented),
required unit/integration tests pass,
no backward-compatibility shims or legacy old-behavior branches remain in scope,
dead code, obsolete files, unused helpers/tests/flags/adapters, and dormant replaced paths in scope are removed,
touched files either already have correct placement or are moved/split so their paths match owning concerns,
changed source implementation files have proactive size-pressure handling recorded (>500 avoided and >220 changed-line pressure assessed/acted on where needed).
Use stages/06-implementation/implementation-template.md.
Do not speak for routine implementation.md edits. Announce only for persisted workflow-state.md events (Stage 6 entry, lock/unlock change, gate/transition outcomes).
