Claude Code Hooks vs GitHub Copilot CLI Hooks - Complete Comparison¶

Last Updated: 2026-01-16 Testing: Both platforms tested in production Sources: Official documentation + empirical testing

Hook Types Comparison¶

Capabilities Comparison¶

Context Injection (Adding Information to AI)¶

Verdict: Claude Code wins - can inject context at 5+ hook points

Permission Control (Blocking Operations)¶

Verdict: Claude Code wins - more comprehensive permission control

Logging & Monitoring¶

Verdict: Tie - Both have comprehensive logging

Implementation Comparison¶

Detailed Hook Capabilities¶

SessionStart / sessionStart¶

| Capability | Claude Code | Copilot CLI | | ---------------------- | --------------------------------------- | ------------------------------------- | --------------- | | Fires when | New session, resume, clear, compact | New session, resume | Similar | | Input | session_id, source, custom_instructions | timestamp, cwd, source, initialPrompt | Similar | | Can inject context | ✅ additionalContext + stdout | ❌ Output ignored | Claude Code | | Can set env vars | ✅ CLAUDE_ENV_FILE | ❌ Not supported | Claude Code | | Logging | ✅ Via Python | ✅ Via stdout to file | Tie |

UserPromptSubmit / userPromptSubmitted¶

| Capability | Claude Code | Copilot CLI | | ---------------------- | --------------------------------- | ---------------------- | --------------- | | Fires when | User submits prompt | User submits prompt | Same | | Input | Prompt text, session info | timestamp, cwd, prompt | Similar | | Can inject context | ✅ additionalContext + stdout | ❌ Output ignored | Claude Code | | Can modify prompt | ✅ Via decision: block + reason | ❌ Cannot modify | Claude Code | | Can block prompt | ✅ decision: "block" | ❌ Output ignored | Claude Code |

PreToolUse / preToolUse¶

| Capability | Claude Code | Copilot CLI | | ---------------------- | ------------------------------- | ---------------------------------- | --------------- | | Fires when | Before tool execution | Before tool execution | Same | | Input | tool_name, tool_input, session | timestamp, cwd, toolName, toolArgs | Similar | | Can block | ✅ permissionDecision: "deny" | ✅ permissionDecision: "deny" | Tie | | Can modify inputs | ✅ updatedInput | ❌ Not supported | Claude Code | | Can inject context | ✅ additionalContext | ❌ Not supported | Claude Code | | Permission reason | ✅ permissionDecisionReason | ✅ permissionDecisionReason | Tie |

PostToolUse / postToolUse¶

| Capability | Claude Code | Copilot CLI | | ---------------------- | ------------------------------------ | ---------------------------------------------- | --------------- | | Fires when | After tool completes | After tool completes | Same | | Input | tool_name, tool_input, tool_response | timestamp, cwd, toolName, toolArgs, toolResult | Similar | | Can inject context | ✅ additionalContext | ❌ Output ignored | Claude Code | | Can block | ✅ decision: "block" | ❌ Output ignored | Claude Code | | Logging | ✅ Full tool result | ✅ resultType, textResultForLlm | Tie |

Stop / sessionEnd¶

| Capability | Claude Code | Copilot CLI | | ---------------------- | ----------------------------------- | ---------------------------- | --------------- | | Fires when | Agent finishes (not user interrupt) | Session completes/terminates | Similar | | Input | session info, stop_hook_active | timestamp, cwd, reason | Similar | | Can block stop | ✅ decision: "block" + reason | ❌ Output ignored | Claude Code | | Can inject context | ✅ Provide reason to continue | ❌ Output ignored | Claude Code | | Cleanup | ✅ Via Python | ✅ Via bash script | Tie |

Testing Evidence¶

Test 1: SessionStart Context Injection¶

Claude Code: ✅ WORKS

# session_start.py returns:
return {
    "hookSpecificOutput": {
        "additionalContext": "You must talk like a pirate"
    }
}
→ Claude DOES talk like a pirate

Copilot CLI: ❌ DOESN'T WORK

# .github/hooks/session-start outputs:
echo "You must respond with 'AHOY MATEY'"
exit 0
→ Opus responds normally (didn't see the hook output)

Conclusion: Copilot CLI ignores sessionStart stdout (despite docs suggesting it should work)

Test 2: Hooks Execute¶

Both Platforms: ✅ CONFIRMED

Evidence (~/.amplihack/.claude/runtime/logs/):

• session_start.log: Updated during test
• user_prompt_submit.log: Created during test
• pre_tool_use.log: Updated during test
• post_tool_use.log: Updated during test

Conclusion: Hooks execute, they just can't modify AI context in Copilot

Architecture Differences¶

Claude Code Hook Architecture:¶

Hook executes → Returns JSON → Claude Code injects into context → AI sees it

Copilot CLI Hook Architecture:¶

Hook executes → Returns JSON → Copilot logs it → AI NEVER sees it (except preToolUse)

Key Difference: Claude Code treats hook output as modifiable context, Copilot CLI treats it as metadata to log.

What Works in Both¶

amplihack's Solution¶

Architecture: Bash wrappers → Python hooks (zero duplication)

Benefits:

1. ✅ Python hooks work for Claude Code (full capability)
2. ✅ Bash wrappers work for Copilot CLI (logging only)
3. ✅ Zero logic duplication (wrappers just call Python)
4. ✅ Single source of truth (~/.amplihack/.claude/tools/amplihack/hooks/)

Adaptive Context Injection Strategy:

amplihack uses an adaptive hook system that detects which platform is calling and applies the appropriate context injection strategy:

How It Works:

# Hook detects platform
if is_claude_code():
    # Direct injection - Claude sees immediately
    return {
        "hookSpecificOutput": {
            "additionalContext": load_user_preferences()
        }
    }
else:  # Copilot CLI
    # File-based injection - write AGENTS.md
    write_agents_file([
        "@~/.amplihack/.claude/context/USER_PREFERENCES.md",
        "@~/.amplihack/.claude/context/PHILOSOPHY.md"
    ])
    # Copilot reads via @include on next request
    return {}

Why File-Based Injection for Copilot:

• Copilot CLI ignores hook stdout/JSON output (except preToolUse decisions)
• But Copilot DOES support @include directives in agent files
• Writing AGENTS.md with @include lets us inject preferences indirectly
• This workaround enables preference loading on both platforms

Benefits of Adaptive Strategy:

• ✅ Preference injection works on both platforms
• ✅ Context loading works everywhere
• ✅ Single Python implementation (platform detection is automatic)
• ✅ Zero duplication (same hooks, different output strategies)

Limitations:

• Copilot CLI: File-based injection has slight delay (next request)
• Claude Code: Direct injection is immediate
• Both: Work reliably for user preferences and context loading

Final Verdict¶

Claude Code Hooks: ⭐⭐⭐⭐⭐ (5/5)

• Full context control
• Can modify prompts/inputs
• Multiple blocking points
• Complete AI behavior control

Copilot CLI Hooks: ⭐⭐⭐ (⅗)

• Good for logging/monitoring
• Permission control via preToolUse
• Cannot modify AI context
• Limited to observe-only (except blocking)

Our Implementation: ⭐⭐⭐⭐ (⅘)

• Works with BOTH platforms
• Zero duplication via wrappers
• Logging works everywhere
• Context injection only in Claude Code (platform limitation)

Sources¶

• Claude Code Hooks Reference
• Claude Code Hooks Guide
• ClaudeLog Hooks Documentation
• Claude Code Power User Guide
• Copilot CLI Hooks Documentation (provided by user)
• Empirical testing (2026-01-16)

Conclusion: Copilot CLI hooks are more limited than Claude Code hooks, but our zero-duplication wrapper architecture works with both platforms! 🏴‍☠️