fresh-context

Fresh Context Subagent Pattern

Problem: Context rot degrades quality as session accumulates history.

Solution: Fresh isolated context per subagent dispatch.

---

When to Use

Dispatching subagent for implementation task?
        ↓ YES
FRESH CONTEXT PROTOCOL ← APPLY

Use when:

• Implementation tasks requiring isolated focus
• Multi-file changes where context pollution risk exists
• Quality-critical tasks requiring consistent reasoning
• Tasks where parent session history could mislead

Don't use when:

• Simple single-file edits (copy, rename, trivial changes)
• Quick lookups (read-only, no implementation)
• Context-independent tasks (no dependencies on session state)

---

Context Budget

Task Type	Thinking Budget	Context Budget	Purpose
Implementation	~20k tokens	~200k tokens	Pure code generation
Review	~15k tokens	~50k tokens	Focused quality check
Research	~10k tokens	~100k tokens	Investigation tasks
Debug	~15k tokens	~80k tokens	Root cause analysis

---

Dispatch Protocol

Pre-Dispatch Checklist

Canonical reference: See rules/AGENTS.md — DELEGATION QUALITY CHECKLIST

This is the authoritative checklist. For quick reference:

□ Is runSubagent/Agent tool available? → TIER 1 (MANDATORY)
□ Task complexity assessed? → Simple/Standard/Complex
□ Specific file/directory boundaries assigned?
□ Acceptance criteria explicit in spawn prompt?
□ Subagent has MINIMAL context (not everything)?
□ Reviewer is DIFFERENT from executor?
□ File ownership is EXCLUSIVE (no conflicts)?
□ Wave grouping planned if multiple parallel tasks?
□ Token budget estimated for this phase?
□ Mailbox path confirmed and writable?
□ Confidence self-assessment completed?

Context Block Template

## SUBAGENT DISPATCH: {task_name}

### ISOLATED CONTEXT
{self-contained task description — what to do, NOT why}

### TARGET FILES
{file paths — exact locations to modify}

### ACCEPTANCE CRITERIA
1. {criterion_1 — verifiable outcome}
2. {criterion_2}
3. {criterion_n}

### CONTEXT ISOLATION
- NO parent session history
- NO previous attempts
- NO reasoning traces
- NO failed attempts
- Fresh context ONLY

### VERIFICATION STEPS
{how to verify task completion}