Feature: Automatic effort level switching based on task phase (plan/implement/review)

Feature Request: Automatic Effort Level Switching Based on Task Phase

Problem

When working iteratively on multi-step implementations, there's a natural cycle:

1. Plan (needs high effort — architecture decisions, trade-off analysis)
2. Implement (needs medium effort — focused code writing)
3. Debug/Review (can use subagents at lower effort, but main context coordination still needed)
4. Feedback (needs high effort — incorporating review, adjusting approach)

Currently, effort level (/effort high, /effort medium) must be manually changed by the user between these phases. In fast iterative cycles (plan → build → test → feedback → plan → build → test), this creates friction. The user has to remember to switch, and switching mid-conversation resets the effort for all subsequent work until changed again.

Proposed Solution

Allow Claude Code to programmatically suggest or set effort levels based on task context. Options:

Option A: Task-phase-aware effort (automatic)

Claude detects the current phase (planning, implementing, reviewing) and adjusts effort automatically. For example:

• Entering plan mode → effort high
• Exiting plan mode to implement → effort medium
• Spawning review/debug agents → already uses model parameter, no change needed

Option B: Effort presets per phase (user-configured)

Allow users to configure effort presets in settings:

{
  "effortPresets": {
    "planning": "high",
    "implementation": "medium", 
    "review": "low"
  }
}

Option C: Inline effort hints (programmatic)

Allow the AI to include an effort hint in its response that the CLI applies:

<effort-hint level="medium" reason="switching to implementation" />

Use Case

A developer working on a 6-step phased implementation with an AI advisor (GPT/Gemini) reviewing each step. The natural cycle is:

• High effort: discuss plan with advisor, incorporate feedback, make architecture decisions
• Medium effort: write the code, run builds, make targeted edits
• High effort: review results, adjust plan, handle advisor feedback
• Repeat for each step

Manually toggling /effort 10-20 times per session is friction that could be eliminated.

Current Workaround

User manually types /effort high and /effort medium between phases. Subagent model selection (model: "sonnet") partially helps for delegated work but doesn't affect the main conversation context.