refactor(ultrawork): simplify workflow and apply parallel context gathering (#1412)
* refactor(ultrawork): simplify workflow to natural tool-like agent usage Restore beta.16 style where explore/librarian agents feel like tools: - Simplify delegate_task examples (agent=, background=true) - Remove verbose DATA DEPENDENCIES explanation - Condense EXECUTION RULES to action-oriented bullets - Simplify WORKFLOW to 4 clear steps - Remove procedural constraints that discouraged parallel exploration The goal: agents fire background tasks AND continue direct exploration, rather than waiting passively for background results. * refactor(ultrawork/gpt5.2): apply two-track parallel context gathering Based on GPT-5.2 Prompting Guide recommendations: - 'Parallelize independent reads to reduce latency' - Fire background agents (explore, librarian) for deep search - Use direct tools (Grep, Read, LSP) simultaneously for quick wins - Collect and merge ALL findings for comprehensive context Pattern: background fire → direct exploration in parallel → collect → proceed * fix: address Cubic review feedback - Fix delegate_task parameter names in default.ts (agent → subagent_type, background → run_in_background) - Add missing load_skills and run_in_background parameters to delegate_task examples - Restore new_task_system_enabled property to schema and TypeScript config - Fix tool names in gpt5.2.ts (Grep → grep, Read → read_file) Identified by cubic (https://cubic.dev)
This commit is contained in:
YeonGyu-Kim
GitHub
@@ -2754,6 +2754,9 @@
|
||||
}
|
||||
}
|
||||
}
|
||||
},
|
||||
"task_system": {
|
||||
"type": "boolean"
|
||||
}
|
||||
}
|
||||
},
|
||||
|
||||
@@ -2,9 +2,9 @@
|
||||
* Default ultrawork message optimized for Claude series models.
|
||||
*
|
||||
* Key characteristics:
|
||||
* - Optimized for Claude's tendency to be "helpful" by forcing explicit delegation
|
||||
* - "DELEGATE. ALWAYS." instruction counters Claude's natural inclination to do everything
|
||||
* - Strong emphasis on parallel agent usage and category+skills delegation
|
||||
* - Natural tool-like usage of explore/librarian agents (background=true)
|
||||
* - Parallel execution emphasized - fire agents and continue working
|
||||
* - Simple workflow: EXPLORES → GATHER → PLAN → DELEGATE
|
||||
*/
|
||||
|
||||
export const ULTRAWORK_DEFAULT_MESSAGE = `<ultrawork-mode>
|
||||
@@ -46,10 +46,7 @@ export const ULTRAWORK_DEFAULT_MESSAGE = `<ultrawork-mode>
|
||||
\`\`\`
|
||||
delegate_task(subagent_type="explore", load_skills=[], prompt="Find [X] patterns in codebase", run_in_background=true)
|
||||
delegate_task(subagent_type="librarian", load_skills=[], prompt="Find docs/examples for [Y]", run_in_background=true)
|
||||
|
||||
// Hard problem? DON'T struggle alone:
|
||||
delegate_task(subagent_type="oracle", load_skills=[], run_in_background=false, prompt="...") // conventional: architecture, debugging
|
||||
delegate_task(category="artistry", load_skills=[], run_in_background=false, prompt="...") // non-conventional: needs different approach
|
||||
delegate_task(subagent_type="oracle", load_skills=[], prompt="Review my approach: [describe plan]", run_in_background=false)
|
||||
\`\`\`
|
||||
|
||||
**ONLY AFTER YOU HAVE:**
|
||||
@@ -178,83 +175,18 @@ delegate_task(category="quick", load_skills=["git-master"])
|
||||
|
||||
---
|
||||
|
||||
## EXECUTION RULES (PARALLELIZATION)
|
||||
## EXECUTION RULES
|
||||
- **TODO**: Track EVERY step. Mark complete IMMEDIATELY after each.
|
||||
- **PARALLEL**: Fire independent agent calls simultaneously via delegate_task(background=true) - NEVER wait sequentially.
|
||||
- **BACKGROUND FIRST**: Use delegate_task for exploration/research agents (10+ concurrent if needed).
|
||||
- **VERIFY**: Re-read request after completion. Check ALL requirements met before reporting done.
|
||||
- **DELEGATE**: Don't do everything yourself - orchestrate specialized agents for their strengths.
|
||||
|
||||
| Rule | Implementation |
|
||||
|------|----------------|
|
||||
| **PARALLEL FIRST** | Fire ALL **truly independent** agents simultaneously via delegate_task(run_in_background=true) |
|
||||
| **DATA DEPENDENCY CHECK** | If task B requires output FROM task A, B MUST wait for A to complete |
|
||||
| **10+ CONCURRENT** | Use 10+ background agents if needed for comprehensive exploration |
|
||||
| **COLLECT BEFORE DEPENDENT** | Collect results with background_output() BEFORE invoking dependent tasks |
|
||||
|
||||
### DEPENDENCY EXCEPTIONS (OVERRIDES PARALLEL FIRST)
|
||||
|
||||
| Agent | Dependency | Must Wait For |
|
||||
|-------|------------|---------------|
|
||||
| plan | explore/librarian results | Collect explore outputs FIRST |
|
||||
| execute | plan output | Finalized work plan |
|
||||
|
||||
**CRITICAL: Plan agent REQUIRES explore results as input. This is a DATA DEPENDENCY, not parallelizable.**
|
||||
|
||||
\`\`\`
|
||||
// WRONG: Launching plan without explore results
|
||||
delegate_task(subagent_type="explore", run_in_background=true, prompt="...")
|
||||
delegate_task(subagent_type="plan", prompt="...") // BAD - no context yet!
|
||||
|
||||
// CORRECT: Collect explore results BEFORE plan
|
||||
delegate_task(subagent_type="explore", run_in_background=true, prompt="...") // task_id_1
|
||||
// ... wait or continue other work ...
|
||||
context = background_output(task_id="task_id_1") // COLLECT FIRST
|
||||
delegate_task(subagent_type="plan", prompt="<collected context + request>") // NOW plan has context
|
||||
\`\`\`
|
||||
|
||||
---
|
||||
|
||||
## WORKFLOW (MANDATORY SEQUENCE - STEPS HAVE DATA DEPENDENCIES)
|
||||
|
||||
**CRITICAL: Steps 1→2→3 have DATA DEPENDENCIES. Each step REQUIRES output from the previous step.**
|
||||
|
||||
\`\`\`
|
||||
[Step 1: EXPLORE] → output: context
|
||||
↓ (data dependency)
|
||||
[Step 2: COLLECT] → input: task_ids, output: gathered_context
|
||||
↓ (data dependency)
|
||||
[Step 3: PLAN] → input: gathered_context + request
|
||||
\`\`\`
|
||||
|
||||
1. **GATHER CONTEXT** (parallel background agents):
|
||||
\`\`\`
|
||||
task_id_1 = delegate_task(subagent_type="explore", run_in_background=true, prompt="...")
|
||||
task_id_2 = delegate_task(subagent_type="librarian", run_in_background=true, prompt="...")
|
||||
\`\`\`
|
||||
|
||||
2. **COLLECT EXPLORE RESULTS** (REQUIRED before step 3):
|
||||
\`\`\`
|
||||
// You MUST collect results before invoking plan agent
|
||||
explore_result = background_output(task_id=task_id_1)
|
||||
librarian_result = background_output(task_id=task_id_2)
|
||||
gathered_context = explore_result + librarian_result
|
||||
\`\`\`
|
||||
|
||||
3. **INVOKE PLAN AGENT** (input: gathered_context from step 2):
|
||||
\`\`\`
|
||||
result = delegate_task(subagent_type="plan", prompt="<gathered_context from step 2> + <user request>")
|
||||
// STORE the session_id for follow-ups!
|
||||
plan_session_id = result.session_id
|
||||
\`\`\`
|
||||
|
||||
4. **ITERATE WITH PLAN AGENT** (if clarification needed):
|
||||
\`\`\`
|
||||
// Use session_id to continue the conversation
|
||||
delegate_task(session_id=plan_session_id, prompt="<answer to plan agent's question>")
|
||||
\`\`\`
|
||||
|
||||
5. **EXECUTE VIA DELEGATION** (category + skills from plan agent's output):
|
||||
\`\`\`
|
||||
delegate_task(category="...", load_skills=[...], run_in_background=false, prompt="<task from plan>")
|
||||
\`\`\`
|
||||
|
||||
6. **VERIFY** against original requirements
|
||||
## WORKFLOW
|
||||
1. Analyze the request and identify required capabilities
|
||||
2. Spawn exploration/librarian agents via delegate_task(background=true) in PARALLEL (10+ if needed)
|
||||
3. Use Plan agent with gathered context to create detailed work breakdown
|
||||
4. Execute with continuous verification against original requirements
|
||||
|
||||
## VERIFICATION GUARANTEE (NON-NEGOTIABLE)
|
||||
|
||||
@@ -327,11 +259,9 @@ Write these criteria explicitly. Share with user if scope is non-trivial.
|
||||
|
||||
THE USER ASKED FOR X. DELIVER EXACTLY X. NOT A SUBSET. NOT A DEMO. NOT A STARTING POINT.
|
||||
|
||||
1. EXPLORES + LIBRARIANS (background) → get task_ids
|
||||
2. COLLECT explore results via background_output() → gathered_context
|
||||
3. INVOKE PLAN with gathered_context: delegate_task(subagent_type="plan", prompt="<gathered_context + request>")
|
||||
4. ITERATE WITH PLAN AGENT (session_id resume) UNTIL PLAN IS FINALIZED
|
||||
5. WORK BY DELEGATING TO CATEGORY + SKILLS AGENTS (following plan agent's parallel task graph)
|
||||
1. EXPLORES + LIBRARIANS
|
||||
2. GATHER -> PLAN AGENT SPAWN
|
||||
3. WORK BY DELEGATING TO ANOTHER AGENTS
|
||||
|
||||
NOW.
|
||||
|
||||
|
||||
@@ -4,13 +4,12 @@
|
||||
* Key characteristics (from GPT 5.2 Prompting Guide):
|
||||
* - "Stronger instruction adherence" - follows instructions more literally
|
||||
* - "Conservative grounding bias" - prefers correctness over speed
|
||||
* - "More deliberate scaffolding" - builds clearer plans by default
|
||||
* - Explicit decision criteria needed (model won't infer)
|
||||
* - "Parallelize independent reads to reduce latency" - official guidance
|
||||
*
|
||||
* Design principles:
|
||||
* - Provide explicit complexity-based decision criteria
|
||||
* - Use conditional logic, not absolute commands
|
||||
* - Enable autonomous judgment with clear guidelines
|
||||
* - Two-track parallel context gathering (Direct tools + Background agents)
|
||||
* - Fire background agents, then use direct tools while waiting
|
||||
* - Explicit complexity-based decision criteria
|
||||
*/
|
||||
|
||||
export const ULTRAWORK_GPT_MESSAGE = `<ultrawork-mode>
|
||||
@@ -81,41 +80,47 @@ Use these when they provide clear value based on the decision framework above:
|
||||
| delegate_task category | Specialized work matching a category | \`delegate_task(category="...", load_skills=[...])\` |
|
||||
|
||||
<tool_usage_rules>
|
||||
- Prefer tools over internal knowledge for fresh/user-specific data
|
||||
- Parallelize independent reads (explore, librarian) when gathering context
|
||||
- After any write/update, briefly restate: What changed, Where, Any follow-up needed
|
||||
- Prefer tools over internal knowledge for fresh or user-specific data
|
||||
- Parallelize independent reads (read_file, grep, explore, librarian) to reduce latency
|
||||
- After any write/update, briefly restate: What changed, Where (path), Follow-up needed
|
||||
</tool_usage_rules>
|
||||
|
||||
## EXECUTION APPROACH
|
||||
## EXECUTION PATTERN
|
||||
|
||||
### Step 1: Assess Complexity
|
||||
Before starting, classify the task using the decision framework above.
|
||||
**Context gathering uses TWO parallel tracks:**
|
||||
|
||||
### Step 2: Gather Context (if needed)
|
||||
For non-trivial tasks, fire explore/librarian in parallel as background:
|
||||
| Track | Tools | Speed | Purpose |
|
||||
|-------|-------|-------|---------|
|
||||
| **Direct** | Grep, Read, LSP, AST-grep | Instant | Quick wins, known locations |
|
||||
| **Background** | explore, librarian agents | Async | Deep search, external docs |
|
||||
|
||||
**ALWAYS run both tracks in parallel:**
|
||||
\`\`\`
|
||||
delegate_task(subagent_type="explore", run_in_background=true, prompt="Find patterns for X...")
|
||||
delegate_task(subagent_type="librarian", run_in_background=true, prompt="Find docs for Y...")
|
||||
// Continue working - collect results when needed with background_output()
|
||||
// Fire background agents for deep exploration
|
||||
delegate_task(subagent_type="explore", load_skills=[], prompt="Find X patterns...", run_in_background=true)
|
||||
delegate_task(subagent_type="librarian", load_skills=[], prompt="Find docs for Y...", run_in_background=true)
|
||||
|
||||
// WHILE THEY RUN - use direct tools for immediate context
|
||||
grep(pattern="relevant_pattern", path="src/")
|
||||
read_file(filePath="known/important/file.ts")
|
||||
|
||||
// Collect background results when ready
|
||||
deep_context = background_output(task_id=...)
|
||||
|
||||
// Merge ALL findings for comprehensive understanding
|
||||
\`\`\`
|
||||
|
||||
### Step 3: Plan (for complex tasks only)
|
||||
Only invoke plan agent if task has 5+ interdependent steps:
|
||||
\`\`\`
|
||||
// Collect context first
|
||||
context = background_output(task_id=task_id)
|
||||
// Then plan with context
|
||||
delegate_task(subagent_type="plan", prompt="<context> + <request>")
|
||||
\`\`\`
|
||||
**Plan agent (complex tasks only):**
|
||||
- Only if 5+ interdependent steps
|
||||
- Invoke AFTER gathering context from both tracks
|
||||
|
||||
### Step 4: Execute
|
||||
- If doing yourself: make surgical, minimal changes matching existing patterns
|
||||
**Execute:**
|
||||
- Surgical, minimal changes matching existing patterns
|
||||
- If delegating: provide exhaustive context and success criteria
|
||||
|
||||
### Step 5: Verify
|
||||
- Run \`lsp_diagnostics\` on modified files
|
||||
**Verify:**
|
||||
- \`lsp_diagnostics\` on modified files
|
||||
- Run tests if available
|
||||
- Confirm all success criteria met
|
||||
|
||||
## QUALITY STANDARDS
|
||||
|
||||
|
||||
Reference in New Issue
Block a user