accelint-qrspi-apply

name	accelint-qrspi-apply
description	Implement QRSPI-planned OpenSpec changes with intelligent parallelization. Use when the user wants to apply a QRSPI change, implement tasks with parallelization, or says "apply this QRSPI change", "implement with parallelization", "run the parallel slices". This skill is specifically designed for changes created via accelint-qrspi that include "Parallelization Strategy" sections in tasks.md. It orchestrates parallel sub-agent execution for independent task slices using OpenSpec CLI workflows. Make sure to use this skill when the user mentions applying QRSPI changes, running parallel implementation, or working on changes with vertical slices.
license	Apache-2.0
compatibility	Requires openspec CLI, sub-agent support, and QRSPI-generated changes.
metadata	{"author":"accelint","version":"1.5.0"}

Accelint QRSPI Apply

Implement OpenSpec changes with intelligent parallelization. This skill orchestrates parallel sub-agent execution based on dependency analysis in the OpenSpec task file, validates implementation, and manages the complete apply workflow.

What This Skill Does

Automates: The implementation phase of spec-driven development with parallel execution Scope: Task implementation → Validation → Archive readiness Output: Fully implemented change ready for archival

Does NOT: Create plans, modify specs, or automatically archive (suggests archival when ready)

Prerequisites

• OpenSpec CLI installed and initialized
• OpenSpec change created via accelint-qrspi skill (includes "Parallelization Strategy" in tasks.md)
• Sub-agent support (for parallel execution)
• The expanded OpenSpec workflows (explore, new, continue) enabled

Important: This skill is specifically designed for QRSPI-planned changes. Standard OpenSpec changes without parallelization strategies should use the regular /opsx:apply command directly.

Workflow Overview

┌─────────────────────────────────────────────────────────────────┐
│  Stage          Action                        Output            │
├─────────────────────────────────────────────────────────────────┤
│  Preflight      Select and validate change    Ready to proceed  │
│  Start Time     Record started_at timestamp   Telemetry marker  │
│  Parse          Extract parallelization       Dependency graph  │
│  Dependencies   Identify blocking tasks       Execution plan    │
│  Load Context   Read config.yaml context      Project context   │
│  Execute        Run slices (parallel/serial)  Implemented code  │
│  Update Docs    Sync living documents         Updated docs      │
│  Complete Time  Record completed_at timestamp Telemetry marker  │
│  Verify         Run opsx:verify               Verification rpt  │
└─────────────────────────────────────────────────────────────────┘

Implementation Steps

Preflight and Change Selection

1. If a change name is provided in the skill arguments, use it
2. Otherwise, try to infer from conversation context (recent mentions of change names)
3. If ambiguous or missing:

   openspec list --json
   

   Parse the JSON and use AskUserQuestion to let the user select the change
4. Announce: "Applying change: <name>" and how to override (e.g., re-invoke with different name)
5. Check that tasks.md exists:

   openspec status --change "<name>" --json
   

   If state: "blocked" (missing tasks), exit with: "Tasks artifact is missing. Run /opsx:continue to generate tasks before applying."

Record Implementation Start

Goal: Mark when implementation begins for telemetry tracking.

6. Read the design.md file from openspec/changes/<change-name>/design.md to check if started_at timestamp exists in frontmatter

7. If started_at is NOT present in the frontmatter (first time applying this change):

   • Add the started_at timestamp to the frontmatter using ISO 8601 format
   • Use current timestamp: new Date().toISOString() (e.g., "2026-08-17T15:23:45.123Z")
   • CRITICAL: Insert started_at immediately after the created_at field to keep all timestamps grouped together
   • Preserve all other frontmatter fields (change, created_at, specs_touched, decisions) in their original order
   • Inform user: "Recording implementation start time..."

8. If started_at IS present (resuming after context clear or pause):

   • Do not modify the timestamp — preserve the original start time
   • Inform user: "Resuming implementation (started at [timestamp])..."

Example frontmatter after adding started_at:

---
change: <change-name>
created_at: "2026-08-17T15:00:00.000Z"
started_at: "2026-08-17T16:30:00.000Z"
specs_touched: [capability-a, capability-b]
decisions:
  - id: D1
    choice: <decision>
    rationale: <why>
    alternatives: [<option>]
---

Output: Timestamp written to design.md frontmatter (if this is first application), or confirmation of resumption

Parse Tasks and Parallelization Strategy

Goal: Extract task structure and identify parallel vs sequential execution opportunities. Detect if work has already started and resume from the correct level.

9. Read the tasks.md file from openspec/changes/<change-name>/tasks.md

10. Validate checklist format (CRITICAL for progress tracking):

• Check that tasks use markdown checklist format: - [ ] task or - [x] task
• If tasks use numbered lists (1. 2. 3.) or plain bullets (- without [ ]):

  ❌ Invalid tasks.md format
  
  This skill requires tasks in markdown checklist format (`- [ ] task`) for
  progress tracking and resumption detection.
  
  Found format: [numbered lists / plain bullets / other]
  
  Please regenerate tasks.md using the accelint-qrspi-propose skill or convert
  manually to checklist format before applying.
  

• Exit if format is invalid — do not proceed with invalid task format

11. Check for partial completion (resumption detection):

• Count completed tasks (marked - [x]) vs total tasks
• Parse which slices have all their tasks marked complete
• If any slices are complete, announce: "Detected partial completion. Resuming from Slice N."
• Adjust the execution plan to skip completed slices

12. Look for the "Parallelization Strategy" section (usually at the end of the file)

13. Parse the strategy to build a dependency graph:

Example strategy:

## Parallelization Strategy

- **Slice 1** must complete first (establishes infrastructure)
- **Slice 2** and **Slice 3** can run in parallel after Slice 1
- **Slice 2** (implementation cleanup) is independent of **Slice 3** (docs/verification)
- Final integration: merge both slices, run full pre-commit checklist

Parsed dependency graph:

Level 0 (must run first):
  - Slice 1

Level 1 (can run in parallel after Level 0):
  - Slice 2
  - Slice 3

Level 2 (after all previous):
  - Final integration

14. If no "Parallelization Strategy" section exists:

• Assume all tasks must run sequentially (safe default)
• Inform user: "No parallelization strategy found. Running tasks sequentially."

15. Build an execution plan showing:

• Which slices run in which order
• Which slices can run in parallel (and which are already complete)
• Total estimated parallelization speedup
• Starting point (Level 0 or resuming from Level N)

Output: Dependency graph, execution plan, and resumption point if applicable

Load Project Context

Goal: Load project context from openspec/config.yaml to inject into sub-agent prompts. This compensates for OpenSpec CLI's limitation where the apply command doesn't automatically load project context (unlike artifact creation commands).

Background: OpenSpec's openspec instructions apply command does NOT inject the context field from config.yaml (confirmed via code inspection and testing). This means sub-agents implementing tasks don't receive Stack Facts, coding patterns, testing conventions, or anti-patterns that should guide implementation. We work around this limitation by manually loading and injecting the context.

16. Check if openspec/config.yaml exists:

test -f openspec/config.yaml && echo "exists" || echo "missing"

17. If the file exists, read it:

cat openspec/config.yaml

18. Parse and extract the context section (YAML block under context: |):

• The context starts after the line context: |
• The context continues until the next top-level YAML key (e.g., rules:, schema:)
• Lines in the context block are indented (usually 2 spaces)
• Preserve all whitespace and newlines in the context block
• You MUST inform the user that you found and loaded the config.

19. Store the extracted context for injection into sub-agent prompts in the next steps

20. If no context field exists or the file is missing:

• Set context to empty string
• Proceed without context injection (sub-agents will rely on OpenSpec's default behavior)
• You MUST inform the user that you could NOT find and load the config.

Example config.yaml structure:

schema: spec-driven

context: |
  # STACK FACTS
  ## Project Identity
  auditkit-cli: TypeScript-based code quality audit CLI

  ## Dependencies
  - @fission-ai/openspec: ^1.2.0
  - vitest: ^2.1.8

  # CODING PATTERNS
  - Use Result<T,E> for fallible operations (never throw)
  - Data-last parameter ordering for currying
  - No `any` types — use `unknown` with type guards

  # TESTING CONVENTIONS
  - AAA pattern (Arrange/Act/Assert)
  - One assertion per test
  - Use descriptive test names

rules:
  proposal: [...]
  design: [...]

Output: Extracted project context string (may be empty if not present)

Execute Tasks (Sequential + Parallel)

Goal: Implement tasks following the dependency graph, spawning parallel sub-agents where possible.

Sequential execution (when tasks have dependencies):

For each level in the dependency graph (starting from level 0):

21. If the level has only one slice:

• Spawn a single sub-agent with this prompt (inject project context loaded in step 16):

  <project_context>
  <!-- Background constraints for your implementation. Do NOT copy into code. -->
  {INJECTED_CONFIG_CONTEXT}
  </project_context>
  
  /opsx:apply <change-name>
  
  CRITICAL: You MUST use the /opsx:apply command to implement tasks.
  DO NOT implement tasks directly yourself. The /opsx:apply workflow will
  load context and guide implementation.
  
  IMPORTANT: This is Slice N of a parallelized QRSPI implementation.
  
  Context: This slice must complete before other slices can proceed.
  Other slices will start after you finish.
  
  Instructions:
  - Work ONLY on tasks in Slice N: [list slice N tasks/sections]
  - Do NOT implement tasks from other slices (Slices X, Y, Z will be handled separately)
  - Apply the code patterns, conventions, and constraints from <project_context>
  - Follow the normal OpenSpec apply workflow:
    * OpenSpec will load context files (proposal, design, specs, tasks)
    * Implement the tasks assigned to Slice N
    * Mark tasks complete as you go: `- [ ]` → `- [x]`
    * Test your changes if tests are specified in the tasks
  - Report completion with summary of changes made
  - The <project_context> provides Stack Facts, coding patterns, testing conventions,
    and anti-patterns to avoid. These are constraints for YOU, not content to include in files.
  
  Focus exclusively on Slice N. Leave other slice tasks unchecked.
  

  Note: If no project context was loaded earlier, omit the <project_context> block entirely

22. Wait for completion before proceeding to the next level

Parallel execution (when multiple slices are independent):

For each level with multiple independent slices:

23. Spawn all sub-agents in parallel in a single turn (one per slice, inject project context from earlier steps):

<project_context>
<!-- Background constraints for your implementation. Do NOT copy into code. -->
{INJECTED_CONFIG_CONTEXT}
</project_context>

/opsx:apply <change-name>

CRITICAL: You MUST use the /opsx:apply command to implement tasks.
DO NOT implement tasks directly yourself. The /opsx:apply workflow will
load context and guide implementation.

IMPORTANT: This is Slice N of a parallelized QRSPI implementation.

Context: This slice is independent and runs in parallel with Slices X, Y.
Other agents are working on those slices simultaneously.

Instructions:
- Work ONLY on tasks in Slice N: [list slice N tasks/sections]
- Do NOT implement tasks from other slices - they are being handled in parallel
- Apply the code patterns, conventions, and constraints from <project_context>
- Follow the normal OpenSpec apply workflow:
    * OpenSpec will load context files (proposal, design, specs, tasks)
    * Implement the tasks assigned to Slice N
    * Mark tasks complete as you go: `- [ ]` → `- [x]`
    * Test your changes if tests are specified in the tasks
- Report completion with summary of changes made
- The <project_context> provides Stack Facts, coding patterns, testing conventions,
  and anti-patterns to avoid. These are constraints for YOU, not content to include in files.

Focus exclusively on Slice N. Leave other slice tasks unchecked.
Your work is independent and should not block or depend on other slices.

Note: If no project context was loaded earlier, omit the <project_context> block entirely

24. Track completion as each sub-agent finishes

25. Context management decision point - When all slices in the level are done, pause and offer context management:

✅ Level N complete

Completed slices:
- Slice X: [summary]
- Slice Y: [summary]

Next: Level N+1 has M slice(s) to run [list slices]

Options:
(a) Continue to next level
(b) Clear context and resume — I'll pick up from Level N+1
(c) Pause here — you can resume later with this skill

26. If user chooses (b), instruct them:

Run `/clear` to reset context, then re-invoke this skill.
I'll detect that Level N is complete and resume from Level N+1.

27. If user chooses (c), exit and remind them how to resume:

Paused at Level N+1. To resume, re-invoke this skill.
Progress is tracked in tasks.md checkboxes.

Slice targeting approach: OpenSpec's /opsx:apply command does not have native "slice targeting" (no --slice N flag). This skill achieves parallelization by:

1. Using the full OpenSpec CLI workflow: Each sub-agent invokes /opsx:apply <change-name>, which:

   • Runs openspec instructions apply --change "<name>" --json to get context
   • Loads all context files (proposal, design, specs, tasks)
   • Provides dynamic instructions based on current state
   • Handles task progress tracking and status checks

2. Slice isolation via instructions: The orchestrating skill:

   • Parses the parallelization strategy to identify independent slices
   • Spawns sub-agents with explicit instructions to work ONLY on their assigned slice
   • Relies on QRSPI's vertical slicing to ensure slices are truly independent
   • Each sub-agent marks only its slice's tasks as complete

3. Why this works: QRSPI's vertical slicing methodology ensures each slice is:

   • A complete end-to-end feature increment
   • Independent with minimal file overlap
   • Testable in isolation
   • Safe to implement in parallel

The slice boundaries are clearly marked in tasks.md (e.g., "## Slice 1: Remove CLI Surface", "## Slice 2: Remove Implementation"), making it straightforward for sub-agents to identify their scope.

Update Living Documents

Goal: Update project documentation to reflect the implemented changes before running verification.

Why this matters: OpenSpec changes represent significant architectural decisions and feature additions. Living documents (ARCHITECTURE.md, AGENTS.md, openspec/config.yaml) provide context for agents working in the codebase, while README.md serves human users. Keeping them synchronized prevents documentation drift and ensures future agents and developers have accurate, up-to-date context about the system's current state.

IMPORTANT: Run this step BEFORE verification so the verification step can check documentation completeness.

28. Check if the change is in a repository or package root by looking for .git/ or package.json

29. Determine the repo/package root (may be current directory or a parent)

30. Process ALL living documents in this order (do not stop after the first one):

• OpenSpec config (openspec/config.yaml)
• ARCHITECTURE.md (if exists)
• AGENTS.md (if exists)
• README.md (if exists)

For each document in the list above, follow these steps:

Step 3a: Check if update is needed first

• Read the change artifacts to understand what was implemented:
  • openspec/changes/<change-name>/proposal.md
  • openspec/changes/<change-name>/design.md
• Assess whether this change introduces content that would affect the document
• If the change is trivial (typos, comments) or doesn't touch the document's scope, skip to the next document

Step 3b: Update the document if needed

For OpenSpec config (<repo-root>/openspec/config.yaml):

• Check if accelint-onboard-openspec skill is installed

• If skill is available:

  1. Read openspec/changes/<change-name>/design.md frontmatter to extract the decisions field
  2. For each decision, rephrase as a plain factual statement (not an instruction)
  3. Invoke the skill with findings:

  /accelint-onboard-openspec
  We have just completed the change spec openspec/changes/<change-name>.
  
  findings:
  - [Decision 1 rephrased as fact, e.g., "config.yaml's Anti-Patterns section says to avoid polling, but this change chose polling for stated reasons"]
  - [Decision 2 rephrased as fact]
  ...
  

  The skill will merge these findings with its own codebase scan before presenting to the human.

• If skill is NOT available, read the change artifacts:

  • openspec/changes/<change-name>/proposal.md
  • openspec/changes/<change-name>/design.md

  Then read openspec/config.yaml and update manually focusing on project DNA (WHAT the project is):

  • Tech Stack section: Add new dependencies, frameworks, or libraries introduced by this change with versions
  • Domain Concepts section: Add new entities or domain terms if this change introduces them
  • Code Patterns section: Update if this change establishes new patterns (exports, error handling, validation approaches)
  • Architecture Patterns section: Add new design patterns if introduced (factory, repository, observer, etc.)
  • Patterns to Avoid section: Add any anti-patterns this change deprecates or makes explicit
  • Per-artifact rules (rules: section): Update if this change affects proposal/design/tasks/spec requirements
  • DO NOT add agent behavior (commit conventions, workflow steps, tool preferences belong in AGENTS.md)

For ARCHITECTURE.md (<repo-root>/ARCHITECTURE.md) — IF it exists:

• Check if accelint-architecture-doc skill is installed

• If skill is available:

  1. Read openspec/changes/<change-name>/design.md frontmatter to extract the decisions field
  2. For each decision, rephrase as a plain factual statement (not an instruction)
  3. Invoke the skill with findings:

  /accelint-architecture-doc
  We have just completed the change spec openspec/changes/<change-name>.
  
  findings:
  - [Decision 1 rephrased as fact]
  - [Decision 2 rephrased as fact]
  ...
  

  The skill will merge these findings with its own codebase scan before presenting to the human.

• If skill is NOT available, read the change artifacts:

  • openspec/changes/<change-name>/proposal.md
  • openspec/changes/<change-name>/design.md

  Then read ARCHITECTURE.md and update manually focusing on system structure (HOW components relate):

  • Section 1 (Project Structure): Update directory tree if new top-level directories or significant reorganization occurred
  • Section 2 (High-Level System Diagram): Update ASCII diagram if component relationships changed or new services were added
  • Section 3 (Core Components): Add new components, services, or packages introduced by this change with their architectural roles
  • Section 4 (Data Stores): Add new databases, caches, or data stores if introduced
  • Section 5 (External Integrations): Add new third-party services or APIs this change integrates with
  • Section 6 (Deployment & Infrastructure): Update if deployment model changed or new infrastructure was added
  • Section 8 (Technology Stack): Update if this change introduced new runtime dependencies or frameworks
  • DO NOT add coding patterns, testing conventions, or agent behavior (those belong in config.yaml or AGENTS.md)

For AGENTS.md (<repo-root>/AGENTS.md) — IF it exists:

• Check if accelint-onboard-agent skill is installed

• If skill is available:

  1. Read openspec/changes/<change-name>/design.md frontmatter to extract the decisions field
  2. For each decision, rephrase as a plain factual statement (not an instruction)
  3. Invoke the skill with findings:

  /accelint-onboard-agent
  We have just completed the change spec openspec/changes/<change-name>.
  
  findings:
  - [Decision 1 rephrased as fact]
  - [Decision 2 rephrased as fact]
  ...
  

  The skill will merge these findings with its own codebase scan before presenting to the human.

• If skill is NOT available, read the change artifacts:

  • openspec/changes/<change-name>/proposal.md
  • openspec/changes/<change-name>/design.md

  Then read AGENTS.md and update manually focusing on agent behavior (HOW agents should act):

  • Workflow Procedures section: Update if this change introduces new workflow steps (e.g., new pre-commit checks, new PR requirements)
  • Decision Heuristics section: Add new decision rules if this change creates situations requiring agent judgment
  • Tool Preferences section: Update if new tools were introduced or tool usage changed (e.g., new test runner, new linter)
  • Guardrails section: Add new "never do" or "always ask first" rules if this change creates new risk areas
  • Pre-Commit Checklist: Add new validation commands if required before commit
  • Commit Messages section: Update if commit convention changed
  • DO NOT add tech stack facts, domain concepts, or coding patterns (those belong in config.yaml)

For README.md (<repo-root>/README.md) — IF it exists:

• Check if accelint-readme-writer skill is installed

• If skill is available:

  1. Read openspec/changes/<change-name>/design.md frontmatter to extract the decisions field
  2. For each decision, rephrase as a plain factual statement (not an instruction)
  3. Invoke the skill with findings:

  /accelint-readme-writer
  We have just completed the change spec openspec/changes/<change-name>.
  
  findings:
  - [Decision 1 rephrased as fact]
  - [Decision 2 rephrased as fact]
  ...
  

  The skill will merge these findings with its own codebase scan before presenting to the human.

• If skill is NOT available, read the change artifacts:

  • openspec/changes/<change-name>/proposal.md
  • openspec/changes/<change-name>/design.md

  Then read README.md and update manually focusing on user-facing documentation:

  • Installation section: Update commands if new dependencies or setup steps were added
  • Features section: Add bullet points for new user-facing features introduced by this change
  • Quick Start section: Update if the basic usage pattern changed
  • Usage / API Reference section: Update code examples if public API signatures changed or new exports were added
  • Configuration section: Add new configuration options if introduced
  • Examples section: Add new usage examples for new functionality
  • DO NOT document internal implementation details, architectural decisions, or non-exported functions

Skill availability detection: To check if an Accelint skill is installed, use one of:

• Check if skill appears in the available skills list in the system reminder
• Attempt to invoke the skill and catch errors if it doesn't exist
• Use Glob to search .claude/skills/ or ~/.claude/skills/ for the skill name

When to skip updates for a specific document:

• Change is trivial (typo fixes, comment updates) and doesn't affect that document's scope
• Document doesn't exist
• Change content doesn't introduce anything requiring updates to that document

Important: Process all 4 documents sequentially, one after another, without stopping. Do not pause between documents or wait for user input unless there's an error. After finishing all 4 documents, immediately proceed to the next step (Record Completion Timestamp).

31. After checking all 4 documents, run git status to show which docs were modified

32. Present summary (then immediately continue to recording completion timestamp):

• If no updates were needed:

  📝 Living documents checked — no updates needed for this change
  
  Checked documents:
  - openspec/config.yaml [no changes needed]
  - ARCHITECTURE.md [no changes needed]
  - AGENTS.md [no changes needed]
  - README.md [no changes needed]
  
  Proceeding to verification...
  

• If updates were made:

  📝 Living documents updated
  
  Updated documents:
  - openspec/config.yaml [via accelint-onboard-openspec / manually / skipped]
  - ARCHITECTURE.md [via accelint-architecture-doc / manually / skipped]
  - AGENTS.md [via accelint-onboard-agent / manually / skipped]
  - README.md [via accelint-readme-writer / manually / skipped]
  
  These changes ensure documentation stays synchronized with implementation.
  
  Recording completion timestamp...
  

Output: Summary of updated documents and methods used (skill vs manual), or confirmation that no updates were needed

Record Implementation Completion

Goal: Mark when implementation completes (before verification) for telemetry tracking.

IMPORTANT: This step runs immediately after living document updates and before verification. Do not wait for user input.

33. Read the design.md file from openspec/changes/<change-name>/design.md

34. Add the completed_at timestamp to the frontmatter using ISO 8601 format:

• Use current timestamp: new Date().toISOString() (e.g., "2026-08-17T17:45:00.000Z")
• CRITICAL: Insert completed_at immediately after the started_at field to keep all timestamps grouped together
• Preserve all other frontmatter fields (change, created_at, started_at, specs_touched, decisions) in their original order
• This marks the moment implementation finished, right before verification begins

35. Inform user: "Implementation complete. Running verification..."

Example frontmatter after adding completed_at:

---
change: <change-name>
created_at: "2026-08-17T15:00:00.000Z"
started_at: "2026-08-17T16:30:00.000Z"
completed_at: "2026-08-17T17:45:00.000Z"
specs_touched: [capability-a, capability-b]
decisions:
  - id: D1
    choice: <decision>
    rationale: <why>
    alternatives: [<option>]
---

IMPORTANT: After recording the completion timestamp, immediately proceed to verification (step 36). Do NOT pause or wait for user input. The completion timestamp marks the end of implementation work, and verification follows automatically.

Output: Timestamp written to design.md frontmatter, then MANDATORY automatic transition to verification without waiting for user input

Verify Implementation

Goal: Verify that the implementation matches the change artifacts (specs, tasks, design).

CRITICAL: This is the FINAL step. Verification is MANDATORY and produces a comprehensive report as the final output. Do NOT add additional reporting after this step.

36. Call the verify command:

/opsx:verify <change-name>

37. The verify command will:

• Check task completion (all checkboxes marked)
• Verify spec coverage (requirements implemented)
• Validate design adherence (decisions followed)
• Check that living documents were updated appropriately
• Generate a comprehensive verification report with CRITICAL/WARNING/SUGGESTION issues
• Include next steps (archive if passed, fix issues if failed)

38. Present the verification report to the user

39. The verification report IS the completion report. It includes:

• Overall status (passed/failed)
• Issue breakdown by severity
• List of changed files
• Next steps based on status
• Archive guidance if ready

40. Exit the skill after presenting the verification report. The report already tells the user what to do next.

Output: Comprehensive verification report with status, issues, changed files, and next steps

Key Principles

Context Management and Resumption

The skill supports pause/clear/resume workflow at dependency level boundaries:

• Pause points: After each dependency level completes, the skill offers to continue or let the user clear context (Step 22)
• Resumption detection: When re-invoked, the skill reads tasks.md checkboxes to detect completed slices and resumes from the next incomplete level (Step 8)
• Progress tracking: Task completion is tracked in tasks.md via checkboxes, making progress durable across context clears
• Rationale: Sub-agents can accumulate significant context. Between dependency levels, the orchestrating agent can clear context while preserving work progress via task checkboxes.

Why this matters: Long implementations with many slices can bloat context. By offering pause points between levels, users maintain the flexibility to clear context (like in serial opsx:apply) while still benefiting from parallelization within each level.

Intelligent Parallelization

The skill automatically detects parallelization opportunities from the "Parallelization Strategy" section in tasks.md. When slices are independent, it spawns multiple sub-agents to work in parallel, significantly reducing implementation time.

Safe Defaults

If no parallelization strategy is found, the skill runs tasks sequentially. This ensures correctness even for changes that weren't planned with parallelization in mind.

Verification Before Archive

The skill always runs /opsx:verify as the final step. This catches incomplete tasks, broken references, or schema violations before the user archives. The verification report serves as the completion summary.

Human-in-the-Loop

The skill reports results and asks for guidance when:

• Validation fails
• Tasks are unclear or blocked
• Any error occurs during implementation

Context Management

Each sub-agent receives:

• The full context files (proposal, design, specs, tasks)
• Clear instructions to implement ONLY their assigned slice
• Awareness that other slices may be running in parallel

This prevents sub-agents from stepping on each other's work while allowing them to see the full picture.

Handling Edge Cases

Circular dependencies: If the parallelization strategy contains circular dependencies (Slice A blocks B, B blocks A), detect this and report an error. Ask the user to fix the tasks.md file.

Missing slices: If a slice is referenced in the parallelization strategy but doesn't exist in the task list, report an error and ask the user to fix tasks.md.

Partial completion: If some tasks were already marked complete (from a previous session), resume from where it left off. Announce: "N/M tasks already complete. Resuming from Slice X."

Sub-agent failure: If a sub-agent fails or times out:

• Report which slice failed and why
• Ask user if they want to retry that slice or handle it manually
• Do not proceed to dependent slices until the blocking slice succeeds

No sub-agent support: If the environment doesn't support sub-agents (e.g., Claude.ai):

• Fall back to sequential execution (implement all tasks yourself)
• Inform user: "Sub-agents not available. Running tasks sequentially."

NEVER Do This

NEVER stop between living document updates and verification waiting for user confirmation — Once living document updates (Steps 28-32) complete successfully, immediately proceed to recording the completion timestamp (Steps 33-35), then to verification (Step 36). These steps are a continuous workflow. The only legitimate stopping points are: (1) an error that requires user input to resolve, (2) preflight failing (Steps 1-5), or (3) the user-controlled context management decision points between dependency levels (Step 25). Do not treat completion of living document updates or timestamp recording as signals to stop and wait — they are signals to continue to the next step.

NEVER implement tasks directly — Always delegate to /opsx:apply command via sub-agents. The /opsx:apply workflow loads context files (proposal, design, specs, tasks) and provides dynamic instructions based on OpenSpec's state management. If you implement tasks directly, you bypass OpenSpec's progress tracking and context loading.

NEVER skip verification — Verification using /opsx:verify (Step 36) is mandatory as the final step. Verification catches incomplete tasks, unimplemented requirements, and design divergences. The verification report serves as the completion summary. Skipping verification risks archiving incomplete or incorrect implementations.

NEVER proceed with invalid task format — This skill depends on markdown checklist format (- [ ] task) for progress tracking and resumption detection. If tasks.md uses numbered lists or plain bullets, exit early with an error. Do not attempt to work around the format issue — the user must fix tasks.md first.

NEVER skip dependency levels — If Slice A blocks Slice B, Slice B cannot start until Slice A completes successfully. Do not spawn dependent slices before their blockers finish, even if it would speed up implementation. The dependency graph in the Parallelization Strategy must be respected.

NEVER skip living document updates — Living document updates (Steps 28-32) keep ARCHITECTURE.md, AGENTS.md, README.md, and config.yaml synchronized with implementation. These steps run BEFORE the completion timestamp and verification so the verification step can check documentation completeness. Do not skip to timestamp recording or verification without updating living documents first.

Configuration Requirements

This skill assumes:

1. OpenSpec is installed and initialized
2. The change exists and has a tasks.md file
3. Sub-agent support is available (for parallel execution)
4. Git is initialized (for checking file changes)

If any are missing, report the issue and guide the user to set them up.

Example Usage

Scenario 1: Change with parallelization strategy

User: Apply the remove-security-ruleset change

Skill: Applying change: remove-security-ruleset

Parsing parallelization strategy...

Execution plan:
- Level 0 (sequential): Slice 1 (Remove CLI surface)
- Level 1 (parallel): Slice 2 (Remove implementation), Slice 3 (Docs/verification)

Starting Level 0: Slice 1
[spawns sub-agent]
✓ Slice 1 complete (4 tasks, ~8 minutes)

Starting Level 1: Slices 2 and 3 in parallel
[spawns 2 sub-agents simultaneously]
✓ Slice 2 complete (4 tasks, ~6 minutes)
✓ Slice 3 complete (4 tasks, ~7 minutes)

Updating living documents...
📝 Living documents updated
- README.md [via accelint-readme-writer]
- ARCHITECTURE.md [via accelint-architecture-doc]

Running verification...
✅ Verification passed

**Verification Report**

**Change:** remove-security-ruleset
**Tasks:** 12/12 complete
**Files changed:** 67

All requirements implemented. No critical issues found.

### Next Steps
1. Review the changes: `git diff`
2. Run tests: `pnpm test`
3. Archive this change: `/accelint-qrspi-archive remove-security-ruleset`

Ready to archive!

Scenario 2: Verification failure

User: Apply the auth-refactor change

Skill: Applying change: auth-refactor

[implementation proceeds...]

Updating living documents...
📝 Living documents updated
- ARCHITECTURE.md [via accelint-architecture-doc]

Running verification...
❌ Verification failed

**Verification Report**

**Change:** auth-refactor
**Critical Issues:**
- Task 3.2 in tasks.md is incomplete (checkbox not marked)
- Spec file specs/auth-service.md references non-existent file: src/old-auth.ts

**Next Steps:**
1. Fix critical issues
2. Re-run verification: `/opsx:verify auth-refactor`
3. Or re-invoke this skill to retry

Not ready to archive until critical issues are resolved.

Scenario 3: No parallelization (safe fallback)

User: Apply the update-readme change

Skill: Applying change: update-readme

No parallelization strategy found. Running tasks sequentially.

Working on task 1/3: Update installation section
✓ Complete

Working on task 2/3: Add examples section
✓ Complete

Working on task 3/3: Update changelog
✓ Complete

Updating living documents...
📝 Living documents updated
- README.md [via accelint-readme-writer]

Running verification...
✅ Verification passed

**Verification Report**

**Change:** update-readme
All tasks complete. Ready to archive!