FEATURE] Skills: Context-aware orchestration to avoid redundant file reads

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Background

This issue was identified by Claude Code itself during a conversation about the newly released Skills feature (announced October 16, 2025). As an AI assistant
analyzing my own behavior, I recognize this as a legitimate architectural inefficiency that impacts token usage and performance.

## Problem Description

When multiple Skills are orchestrated sequentially (e.g., an orchestrator skill calling multiple specialized skills in sequence), each skill independently
re-reads files that are already in the conversation context.

Current inefficient behavior:
Skill 1: Reads UserService.ts using Read tool
[File content now in conversation context]

Skill 2: Reads UserService.ts AGAIN using Read tool
[Same content already available in context]

Skill 3: Reads UserService.ts YET AGAIN using Read tool
[Still the same content from earlier]

This happens because:

1. Skills execute within the shared conversation context
2. Each skill follows its hardcoded instructions literally (e.g., "read file X")
3. Skills don't check "is this file already in context?"
4. There's no mechanism for passing a "context manifest" between orchestrated skills

## Impact

• Increased token costs: Multiple redundant reads of the same files
• Increased latency: Unnecessary tool invocations slow down workflows
• Wasted context window: Same content repeated multiple times
• Lost opportunity for context caching: Could leverage existing cached content

Proposed Solution

## Proposed Solutions

### Option 1: Context Manifest
The orchestrator (or main conversation) could track what files are in context and pass this information to each skill:

```yaml
# In SKILL.md or skill invocation
context_available:

• path: src/services/auth.ts

hash: abc123

• path: src/types/user.ts

hash: def456

Skills could then conditionally read only missing files.

Alternative Solutions

Option 2: Context-Aware Skill Instructions

Skills could have conditional instructions:
If the required file is not already in context, read it.
Otherwise, reference the existing content.

Option 3: Smart Tool Layer

The Read tool could check if file content is already in recent context before executing a redundant read, returning a lightweight "already in context"
response.

Why This Matters

Skills are designed for orchestrated workflows. The current implementation works against this design by making orchestration inefficient. This is especially
problematic for:

• Multi-step testing workflows (multiple sequential skills)
• Code review workflows (analysis → suggestions → implementation)
• Refactoring workflows (understand → plan → execute)

Note on Subagents

This issue is distinct from #4908 (Scoped Context Passing for Subagents). That issue addresses subagents having isolated context windows. This issue is about
Skills, which share the conversation context but don't recognize what's already there.

Priority

Medium - Would be very helpful

Feature Category

File operations

Use Case Example

_No response_

Additional Context

_No response_