[projman] Executor task scoping is too broad #238

New Issue

lmiranda · 2026-01-28T15:23:28Z

lmiranda commented

2026-01-28 15:23:28 +00:00

Problem

Executor agents were given "implement entire hook" which was too large a scope. They got lost in testing loops and ran for 400+ tool calls.

Current Behavior

Executor gets full issue as prompt
"Implement X" is interpreted as: read, analyze, create, test, validate, iterate
Agent tries to be thorough, ends up in loops
No clear stopping point

Expected Behavior

Tasks broken into smaller sub-tasks
Each sub-task has clear completion criteria
Agent completes one sub-task, reports, moves to next
Orchestrator coordinates the sequence

Suggested Task Breakdown

Instead of: "Implement commit message enforcement hook"

Break into:

Create hooks directory and hooks.json (verify: file exists)
Create shell script skeleton (verify: file exists, is executable)
Implement validation logic (verify: script runs without error)
Test with sample inputs (verify: passes test cases)
Commit changes (verify: clean git status)

Acceptance Criteria

Tasks are scoped to complete in < 50 tool calls
Clear verification at each step
Progress visible between steps

## Problem Executor agents were given "implement entire hook" which was too large a scope. They got lost in testing loops and ran for 400+ tool calls. ## Current Behavior - Executor gets full issue as prompt - "Implement X" is interpreted as: read, analyze, create, test, validate, iterate - Agent tries to be thorough, ends up in loops - No clear stopping point ## Expected Behavior - Tasks broken into smaller sub-tasks - Each sub-task has clear completion criteria - Agent completes one sub-task, reports, moves to next - Orchestrator coordinates the sequence ## Suggested Task Breakdown Instead of: "Implement commit message enforcement hook" Break into: 1. Create hooks directory and hooks.json (verify: file exists) 2. Create shell script skeleton (verify: file exists, is executable) 3. Implement validation logic (verify: script runs without error) 4. Test with sample inputs (verify: passes test cases) 5. Commit changes (verify: clean git status) ## Suggested Solutions 1. **Sub-task generation** - Planner creates sub-tasks, not monolithic tasks 2. **Step-by-step executor** - Executor does one step, checkpoints, continues 3. **Verification gates** - Each step has explicit verification before proceeding 4. **Smaller max_turns** - Force agents to complete quickly or checkpoint ## Acceptance Criteria - [ ] Tasks are scoped to complete in < 50 tool calls - [ ] Clear verification at each step - [ ] Progress visible between steps

lmiranda added the Component/Backend Type/Feature Priority/High labels 2026-01-28 15:23:45 +00:00

lmiranda referenced this issue

2026-01-28 15:53:12 +00:00

feat(projman): implement 8 improvement issues (#231-#238) #239

lmiranda referenced this issue from a commit

2026-01-28 15:53:51 +00:00

feat(projman): add strict task sizing rules to prevent runaway agents (#238)

lmiranda closed this issue

2026-01-28 16:02:48 +00:00

Sign in to join this conversation.