[FEATURE] Add setting to disable context window compacting reservation

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

Claude Code reserves 10% of the context window to accommodate auto-compacting and maintain a hard blocking limit. This reservation persists even when auto-compact is disabled in settings, preventing users from reclaiming unused capacity.

Current impact:

• Context window: 200k tokens
• System tools & prompt: ~10% (20k tokens)
• Compacting reservation (hard blocking limit): ~10% (20k tokens)
• Effective available context: 160k tokens (instead of 180k)

Users who disable auto-compact and prefer session handoff workflows cannot reclaim this 20k token reservation, even though it serves no functional purpose for their workflow.

Technical Analysis

The 10% hard cutoff originates from the interaction of three components:

1. Auto-compact threshold: 95% (default, configurable via CLAUDE_AUTOCOMPACT_PCT_OVERRIDE)
2. Compact buffer: Fixed 3k token allocation (verified through testing)
3. Processing overhead: Small buffer for accepting and processing the next request

Combined, these reserve approximately 10% of the context window, creating an effective hard blocking limit at ~90% utilization. When auto-compact is disabled, this reservation becomes unused overhead.

Verification: Testing confirmed the compact buffer remains fixed at 3k tokens across context sizes, indicating it is not an active parallel summarization process but a reserved allocation.

Why It Matters

The 20k token reservation represents significant unused capacity:

• Last-minute actions: Users approaching context limits lose 20k tokens for final decisions, writes, or critical actions before context limit
• Agent task completion: Long-running agent tasks that reach 180k capacity get blocked at 160k, preventing task completion when that final 10% could be sufficient
• Session handoff workflows: Users who prefer detailed handoff protocols and starting fresh sessions in new contexts lose 20k tokens that could be used for comprehensive documentation
• Resource efficiency: For users who don't use auto-compact or /compact, this 10% represents pure waste with no corresponding benefit

Proposed Solution

Add a new setting "Disable compacting reservation" (boolean toggle) that:

• Aligns with existing auto-compact setting granularity (same scope: global or per-session)
• When disabled (default): Maintains current 10% reservation behavior
• When enabled: Removes the 10% reservation and /compact ability if limit is reached, increasing effective available context to 180k tokens
• Implements same hard rejection behavior as current 90% limit—user decides whether to accept this trade-off (session cannot continue past true 100% context, but user has full capacity available)

Technical Considerations

Implementation is straightforward:

if (autoCompactEnabled OR compactingReservationDisabled) {
  reserve 10%
} else {
  // Use full capacity up to hard context limit
}

• Can likely be implemented as runtime check before message acceptance
• May require restart depending on architecture
• No breaking changes to existing behavior (disabled by default)

Acceptance Criteria

• [ ] New boolean setting added: disableCompactingReservation (default: false)
• [ ] Setting location aligns with auto-compact in settings hierarchy
• [ ] When enabled, 10% reservation is not subtracted from available context
• [ ] Context rejection behavior remains consistent (respects user's choice)
• [ ] Setting persists across sessions
• [ ] When disabled (default), existing behavior unchanged
• [ ] Documentation clarifies the trade-off: full capacity available but no emergency reserve for /compact
• [ ] No breaking changes to existing workflows

Context Window Math

Current behavior (auto-compact disabled):

• Total: 200k tokens
• System + tools: 20k tokens (10%)
• Compacting reservation: 20k tokens (10%)
• Effective available: 160k tokens (80%)
• Wasted: 20k tokens

Desired behavior (with setting enabled):

• Total: 200k tokens
• System + tools: 20k tokens (10%)
• Compacting reservation: 0 tokens (disabled by user)
• Effective available: 180k tokens (90%)
• Wasted: 0 tokens (user's choice to not reserve)

Alternative Solutions

_No response_

Priority

High - Significant impact on productivity

Feature Category

Configuration and settings

Use Case Example

1. Session handoff strategy: User disables auto-compact, prefers to write comprehensive handoff protocol documenting context, decisions, and next steps before starting a fresh session. The final 10% capacity would be available for detailed handoff documentation.

1. Agent task completion: Long-running agent task approaches 180k context limit but is blocked at 160k. The agent has insufficient capacity to complete the final steps that would fit in the remaining 10%.

1. Final user decisions: User approaches context limit with a critical decision or code change. The 20k reservation prevents them from acting with the available capacity.

1. Memory-intensive workflows: Users working with large conversations who deliberately avoid auto-compact or /compact because summarization loses important context details which leads to lower quality output.

Additional Context

_No response_