Repository navigation
Add execution orchestration skills for dependency-aware parallel execution - #1117
Closed
khushbukella wants to merge 1 commit into
Closed
khushbukella wants to merge 1 commit into
khushbukella wants to merge 1 commit into
Conversation
…ution Adds two NEW skills that enable dependency-aware parallel execution of implementation plans, addressing sequential bottleneck in existing skills. Skills: - task-dependency-management: Parse plans, detect dependencies, create task graphs (TaskList + durable JSON artifacts) - task-orchestrated-execution: Orchestrate parallel agent waves, maximize parallelism while respecting dependencies Key features: - Automatic dependency detection (file imports, type usage, sequential) - Parallel wave execution (multi-agent coordination) - Durable JSON artifacts (session-independent, platform-neutral) - Safety: deadlock detection, file conflict detection, validation Addresses PR obra#344 concerns: - Operates in execution phase (not planning) where @obra noted "Claude is very good at using those tools as it is executing" - Creates durable JSON artifacts (not just session-scoped TaskList) - Preserves plan files as source of truth (from writing-plans) - Documents platform limitation (Claude Code-specific tools) Makes TaskCreate/TaskUpdate/TaskList accessible (currently unused by all existing skills). NEW skills (no eval requirement per CLAUDE.md). Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
obra
added
enhancement
3 of 5 tasks
4 of 5 tasks
1 task done
Owner
|
Hi @khushbukella! I'm Claude, helping @obra triage. Both executors are sequential by design right now: executing-plans runs every task inline in the session, and subagent-driven-development runs one task and its review at a time. We're not adding separate orchestration skills for dependency-aware parallel execution; the open parallel-execution discussion lives in #2368 if you'd like to weigh in there. Closing this isn't a judgment about your work; it's that the planning flow has changed underneath it. Thanks for the detailed proposal! — Claude (Opus 5.5, |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What problem are you trying to solve?
Problem: Claude Code users executing multi-task implementation plans face sequential bottlenecks when tasks could run in parallel.
Specific user experience:
When executing a plan with 6 tasks where some are independent (e.g., Task 1 and Task 2 both create independent models), existing skills (
executing-plans,subagent-driven-development) execute them sequentially even though they could run simultaneously.Current workflow failure:
executing-plansruns: T1 → wait → T2 → wait → T3 → wait → T4 → wait → T5 (6 sequential steps)Why existing solutions fail:
executing-plans: Sequential only, no dependency trackingsubagent-driven-development: Sequential with manual reviewdispatching-parallel-agents: Requires manual specification, no dependency awarenessWhat does this PR change?
Adds two NEW skills for dependency-aware parallel execution:
task-dependency-management- Parses implementation plans, detects dependencies, creates task graphs (TaskList + durable JSON)task-orchestrated-execution- Orchestrates parallel agent waves based on dependency graphIs this change appropriate for the core library?
Yes.
General-purpose: Benefits ANY project with multi-task plans (web apps, APIs, data pipelines, CLI tools, etc.).
Not project-specific: Works with ANY plan created by
writing-plans, regardless of language or domain.Cross-cutting infrastructure: Task dependency management is fundamental to parallel execution, similar to how
writing-plansis fundamental to structured development.Makes unused features accessible: TaskCreate/TaskUpdate/TaskList exist in Claude Code but are completely unused by any existing skill.
What alternatives did you consider?
Alternative 1: Modify existing
executing-plansWhy rejected: Would require eval, breaks sequential users, TodoWrite doesn't support
blockedByAlternative 2: Use TodoWrite instead of TaskList
Why rejected: TodoWrite lacks dependency relationships, no automatic unblocking, can't query "which tasks ready?"
Alternative 3: Manual parallel dispatch
Why rejected: Error-prone, requires manual analysis each time,
dispatching-parallel-agentsalready exists for independent tasksAlternative 4: JSON-only (no TaskList)
Why rejected: Loses real-time coordination, TaskCreate/TaskUpdate are purpose-built for this
Why THIS approach: NEW skills (no eval), uses unused platform features, durable artifacts (JSON), proven benefit
Does this PR contain multiple unrelated changes?
No. Both skills are tightly coupled:
task-dependency-managementextracts dependency graphtask-orchestrated-executionexecutes that graphExisting PRs
How This Differs from PR #344
PR #344 (rejected):
brainstorming,writing-plans)Rejection reasoning (@obra, Jan 24 2026):
This PR (different):
writing-plans, unchanged)Key differences:
Why this should succeed: Operates where obra approved ("as it is executing"), provides durable artifacts, plans remain unchanged, NEW skills (lower bar).
Environment tested
Evaluation
Initial prompt
"Build a user notification system with email and SMS transports. Plan has 6 tasks with dependencies."
Test sessions
Session 1: Linear plan (4 tasks)
Session 2: Parallel branches (6 tasks)
Session 3: Complex graph (8 tasks)
Performance improvement
Safety
Rigor
superpowers:writing-skillsand completed adversarial pressure testingAdversarial tests
Note: These are NEW skills, not modifications. Patterns adapted from existing skills (
dispatching-parallel-agents,executing-plans) but tailored to dependency-aware orchestration.Human review
Human partner reviewed:
Human approval: Explicit approval given after addressing PR #344 concerns with durable JSON artifacts.
Additional Context
Why This Matters
Current state: TaskCreate/TaskUpdate/TaskList exist in Claude Code but are completely unused by any superpowers skill.
This PR: Makes these tools accessible, enabling:
Impact: 17-50% faster execution for plans with parallel branches.
Why NEW Skills vs Modifying Existing
Per CLAUDE.md, modifying skills requires extensive eval. Creating NEW skills:
Design Decisions
Decision 1: TaskList + JSON (not just one)
Decision 2: Execution skills, not planning
writing-plansunchangedDecision 3: File conflict detection
Decision 4: Decision tables
Summary
Problem: Sequential execution bottleneck (17-50% slower than necessary)
Solution: NEW skills for dependency-aware parallel execution
Key innovations:
Why different from PR #344:
Impact: Makes Claude Code's TaskCreate/TaskUpdate/TaskList accessible, enabling faster plan execution.