Subagent Context Isolation

Enforce independent, context-isolated sub-agent invocations for unbiased validation.

Core Principle

CRITICAL: In refine-loop patterns, creator and validator sub-agents MUST be completely independent. The validator has NO knowledge of the creator's reasoning, only the artifact itself.

Why Context Isolation Matters

With Isolation	Without Isolation
Validator sees only artifact	Validator influenced by creator's reasoning
Unbiased critique	Confirmation bias
Catches actual issues	Misses issues creator justified

Invocation Protocol

1. Creator Sub-agent

Context: Full discovery/requirements + template

2. Validator Sub-agent (Independent)

Context: ONLY the artifact (PRD/Plan file) + requirements reference

javascript

// Validator: ONLY artifact + requirements
@agent prd-validator
PRD_FILE: .tot-docs/prd/001-feature.md

// NO access to:
// - Creator's reasoning
// - Creator's chat history
// - Creator's todo list

3. Creator Sub-agent (Refinement)

Context: Artifact + Validator feedback ONLY

File-Based Communication

RULE: Sub-agents communicate ONLY through files, never through shared memory/context.

Artifact	Who Reads	Who Writes
PRD file	Validator, Planner	Creator
Feedback file	Creator (refine)	Validator
Plan file	Validator, Executor	Planner
Status file	All	Coordinator

Anti-Patterns

•❌ Passing coordinator context to sub-agents
•❌ Validator sees creator's reasoning
•❌ Reusing same agent session
•❌ Feedback loop contamination

Validation Loop Isolation

During parent task completion in process-task-list, a validation loop runs with strict isolation:

Validation SubAgent

Context: ONLY staged changes + lint/test commands NO Access To:

•Implementation task details
•Why code was written this way
•Previous validation attempts

javascript

// Validation SubAgent receives ONLY:
runSubagent({
  prompt: "Review staged changes: `git diff --cached`",
  // NO task context, NO PRD, NO previous reasoning
})

Remediation SubAgent

Context: ONLY validation failures + file paths NO Access To:

•Original implementation reasoning
•Why validation failed previously
•Task acceptance criteria

javascript

// Remediation SubAgent receives ONLY:
runSubagent({
  prompt: "Fix issues: [list from validation]",
  // NO implementation context, NO validation logic
})

Why This Matters

With Isolation	Without Isolation
Finds real issues	Overlooks "intentional" issues
Unbiased fixes	Confirmation bias
Complete remediation	Partial fixes

Isolation Rules Table

SubAgent	Receives ✅	Does NOT Receive ❌
Implementation	Task detail file, files to modify, 1 pattern example	Full PRD, other tasks, previous implementation
Validation	Staged diff (`git diff --cached`), lint/test commands	Implementation reasoning, task context, why changes made
Remediation	Validation failures list, file paths with issues	Implementation intent, validation logic, previous fixes

Enforcement Checklist

Before invoking each SubAgent, verify:

• Implementation: Only sees its own task detail file
• Validation: Cannot infer why code was written
• Remediation: Cannot see original task or validation reasoning
• All: No shared context between invocations

Verification Questions

Before invoking a sub-agent:

•Does this sub-agent see ONLY its required inputs?
•Could this sub-agent be influenced by previous agents?
•Can I explain the input in one sentence?