Chapter 54: Claude Code 2.1.73-2.1.76 New Features

Four releases in four days (March 11-14, 2026) landed a cluster of features that change how sessions manage context, how worktrees handle large repos, and how MCP servers interact with users. This chapter covers each feature with practical configuration examples.

---

1. PostCompact Hook (2.1.76)

Context compaction discards old messages to free up space. The problem: compaction can drop critical instructions that were loaded at session start. The new PostCompact hook fires immediately after compaction completes, giving you a chance to reload anything important.

The Problem It Solves

Session start: CLAUDE.md loaded, rules loaded, context full of instructions
    ...100 messages later...
Compaction: old messages removed, including those initial instructions
Result: Claude forgets project conventions mid-session

Configuration

Add to .claude/settings.json (project) or ~/.claude/settings.json (global):

{
  "hooks": {
    "PostCompact": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "/home/you/.claude/hooks/post-compact-reload.sh"
          }
        ]
      }
    ]
  }
}

Example Hook Script

Create ~/.claude/hooks/post-compact-reload.sh:

#!/bin/bash
# PostCompact Hook — Reload critical context after compaction

PROJECT_ROOT=$(git rev-parse --show-toplevel 2>/dev/null || echo "$PWD")
RELOAD_FILES=""

# Reload project instructions
[ -f "$PROJECT_ROOT/CLAUDE.md" ] && RELOAD_FILES="$RELOAD_FILES $PROJECT_ROOT/CLAUDE.md"

# Reload critical rules
for rule in "$PROJECT_ROOT/.claude/rules/"*.md; do
    [ -f "$rule" ] && RELOAD_FILES="$RELOAD_FILES $rule"
done

if [ -n "$RELOAD_FILES" ]; then
    echo "Reloaded context files after compaction:$RELOAD_FILES"
fi

exit 0

chmod +x ~/.claude/hooks/post-compact-reload.sh

Relationship to PreCompact

Hook	Fires	Use Case
PreCompact	Before compaction runs	Save state, warn about important context
PostCompact	After compaction ends	Reload critical files, re-inject rules

Use both together for a complete compaction lifecycle: save before, restore after.

---

2. /effort Slash Command (2.1.76)

Set the model’s reasoning effort level within a running session.

/effort high     # Maximum reasoning depth
/effort medium   # Balanced (default)
/effort low      # Fast responses, less reasoning

Like /model, this command works while Claude is actively responding – you do not need to wait for the current response to finish. The change takes effect on the next message.

When to Use Each Level

Level	Best For
high	Architecture decisions, complex debugging, code review
medium	General development, feature implementation
low	Simple lookups, formatting, file renaming

---

3. Session Naming with –name / -n (2.1.76)

Set a display name for the session at startup:

claude --name "feature/auth-refactor"
claude -n "hotfix-prod-db"

The name appears on the prompt bar. You can change it mid-session with /rename.

Practical Use: Multi-Branch Workflows

When working across multiple branches in separate terminals, session names prevent confusion:

# Terminal 1
cd ~/project && git checkout main
claude -n "main"

# Terminal 2
cd ~/project && git checkout feature/new-api
claude -n "feature/new-api"

# Terminal 3
cd ~/project && git checkout hotfix/auth-bug
claude -n "hotfix/auth-bug"

Each terminal’s prompt bar shows which branch context it belongs to.

---

4. worktree.sparsePaths Setting (2.1.76)

For large monorepos using --worktree, the new worktree.sparsePaths setting limits which directories get checked out. This uses git sparse-checkout under the hood, avoiding the cost of cloning an entire monorepo into each worktree.

Configuration

Add to .claude/settings.json:

{
  "worktree": {
    "sparsePaths": [
      "packages/my-service/",
      "packages/shared-lib/",
      "configs/",
      "package.json",
      "tsconfig.json"
    ]
  }
}

How It Works

1. You run claude --worktree (or use EnterWorktree)
2. Claude creates a new git worktree from your branch
3. Instead of checking out the full repo, only paths listed in sparsePaths are materialized
4. The worktree is lighter, faster to create, and uses less disk space

When to Use It

Repo Size	Use sparsePaths?	Why
< 1 GB	No	Full checkout is fast enough
1-10 GB	Optional	Helps if you only touch a few packages
> 10 GB	Yes	Avoids multi-minute checkout and disk bloat

Example: Node.js Monorepo

{
  "worktree": {
    "sparsePaths": [
      "packages/api/",
      "packages/shared/",
      "packages/types/",
      "package.json",
      "pnpm-lock.yaml",
      "tsconfig.base.json",
      ".eslintrc.js"
    ]
  }
}

This checks out three packages plus root config files, skipping packages/web/, packages/mobile/, and any other large directories you do not need.

---

5. MCP Elicitation Support (2.1.76)

MCP servers can now request structured input from the user mid-task. Previously, MCP tools could only receive input at invocation time. With elicitation, a tool can pause execution, ask the user a question, and resume with the answer.

New Hook Events

Hook	Fires	Use Case
Elicitation	When an MCP server requests user input	Log, filter, or auto-respond
ElicitationResult	After the user responds to the elicitation	Log responses, audit input patterns

Configuration

{
  "hooks": {
    "Elicitation": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "echo 'MCP elicitation requested' >> /tmp/mcp-elicitation.log"
          }
        ]
      }
    ]
  }
}

What This Enables

• A database MCP server can ask “Which schema?” before running a migration
• A deployment MCP server can ask “Production or staging?” before deploying
• An authentication MCP server can request credentials when a token expires

The elicitation appears as a prompt in the Claude Code interface. The user responds, and the MCP server receives the answer to continue its operation.

---

6. 1M Context Window for Opus 4.6 (2.1.75)

Claude Opus 4.6 now defaults to a 1,000,000-token context window on Max, Team, and Enterprise plans. Previously, the 1M window required enabling extended thinking or extra usage allocation.

What Changes

Plan	Opus 4.6 Context	Before 2.1.75
Max	1M (default)	200K default
Team	1M (default)	200K default
Enterprise	1M (default)	200K default
Pro	200K	200K

Practical Impact

• Longer sessions before compaction triggers
• More files readable in a single context window
• Larger codebases navigable without aggressive summarization
• The /context command shows capacity based on the 1M limit

No configuration change is needed. The larger window activates automatically when using Opus 4.6 on a supported plan.

---

7. Configurable SessionEnd Hook Timeout (2.1.74)

SessionEnd hooks previously had a hard-coded 1.5-second timeout. If your hook needed to write a session summary, sync to a remote service, or process observation logs, 1.5 seconds was often not enough.

Configuration

Set the timeout via environment variable (in milliseconds):

# In ~/.bashrc or ~/.zshrc
export CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS=10000  # 10 seconds

Recommended Values

Hook Complexity	Timeout
Simple logging	2000 (2s)
Git-based session summary	5000 (5s)
Remote sync (Basic Memory, API)	10000 (10s)
Heavy processing (AI compression)	15000 (15s)

Example: Session Summary With Enough Time

If you use the auto-session-summary pattern from Chapter 51, increase the timeout to give git operations time to complete:

export CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS=8000

Without this, complex SessionEnd hooks silently fail when they exceed the default timeout.

---

8. autoMemoryDirectory Setting (2.1.74)

By default, Claude Code stores auto-memory (the CLAUDE.md entries created by /memory and automatic observations) alongside your project. The autoMemoryDirectory setting lets you redirect this to a custom location.

Configuration

Add to ~/.claude/settings.json (global) or .claude/settings.json (project):

{
  "autoMemoryDirectory": "/home/you/claude-memory"
}

Use Cases

Scenario	Directory	Why
Centralized memory	~/claude-memory/	All projects write to one searchable location
Synced across machines	~/Dropbox/claude-memory/	Memory available on multiple workstations
Separated from code	~/.claude/auto-memory/	Keep memory out of git repos
Per-project (default)	(not set)	Memory stays in project directory

Interaction With Basic Memory MCP

If you use Basic Memory MCP, you can point autoMemoryDirectory to your Basic Memory content folder so auto-observations flow directly into the searchable knowledge base:

{
  "autoMemoryDirectory": "/home/you/basic-memory/auto-observations"
}

---

9. /context Command Improvements (2.1.74)

The /context command now provides actionable diagnostics instead of just showing raw token counts.

New Output Sections

Context-Heavy Tools: Identifies which tools are consuming the most context. If a single Bash output used 50K tokens, /context flags it.

Memory Bloat Detection: Warns when CLAUDE.md files, rules, or loaded skills are consuming a disproportionate share of the context window.

Capacity Warnings: Shows color-coded capacity status:

Capacity Used	Status	Action
< 50%	Normal	No action needed
50-75%	Moderate	Consider compacting soon
75-90%	High	Compaction recommended
> 90%	Critical	Compaction imminent, reduce tool output

Actionable Suggestions: Instead of just “context is 80% full,” the command now suggests specific actions: “Consider adding head_limit to Grep calls” or “3 rules files account for 15% of context – consider path-specific loading.”

---

10. Key Bug Fixes

Deferred Tool Schemas Survive Compaction

Previously, if you used ToolSearch to fetch a deferred tool’s schema, compaction could discard the schema from context. The tool would appear available but fail on invocation. Fixed in 2.1.76 – deferred tool schemas are now preserved across compaction events.

RTL Text Rendering

Hebrew, Arabic, and other right-to-left scripts now render correctly in Claude Code’s terminal output. Previously, mixed LTR/RTL content could produce garbled display, particularly in code blocks containing RTL string literals.

Background Agent Partial Results

When a background agent (launched via TaskCreate) is killed or times out, its partial results are now preserved and accessible via TaskGet. Previously, killing a background agent discarded all accumulated output.

Stale Worktree Auto-Cleanup

Worktrees created by --worktree or EnterWorktree are now automatically cleaned up when the associated branch is deleted or the worktree directory becomes stale. Previously, orphaned worktrees accumulated on disk and required manual git worktree prune.

---

Configuration Summary

All settings from this chapter in one reference:

{
  "hooks": {
    "PostCompact": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "/home/you/.claude/hooks/post-compact-reload.sh"
          }
        ]
      }
    ]
  },
  "worktree": {
    "sparsePaths": [
      "packages/my-service/",
      "packages/shared-lib/",
      "configs/"
    ]
  },
  "autoMemoryDirectory": "/home/you/claude-memory"
}

Environment variables:

export CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS=8000

---

See Also:

• Chapter 13: Claude Code Hooks
• Chapter 42: Session Memory Compaction
• Chapter 46: Advanced Configuration Patterns
• Chapter 48: Lean Orchestrator Pattern
• Chapter 51: Persistent Memory Patterns