Clear Context

Table of Contents

• Quick Start
• When to Use
• The Auto-Clear Pattern
• Thresholds
• Auto-Clear Workflow
• Integration with Existing Hooks
• Self-Monitoring Pattern

Clear Context Skill

Quick Start

When context pressure reaches critical levels (80%+), invoke this skill to: 1. Save current session state 2. Delegate continuation to a fresh subagent 3. Continue work without manual intervention

Skill(conserve:clear-context)

When To Use

• Proactively: Before starting large multi-chained tasks
• Reactively: When context warning indicates 80%+ usage
• Automatically: Integrated into long-running workflows

When NOT To Use

• Context usage is under 50% - continue working normally
• Mid-critical-operation where handoff would lose state
• Consider "Summarize from here" first (Claude Code 2.1.32+): Before full auto-clear,

try partial summarization via the message selector. This compresses older context while preserving recent work, often sufficient to relieve pressure without a full handoff.

The Auto-Clear Pattern

Since /clear requires user action, we achieve automatic context clearing without interruption through subagent delegation:

Main Agent (high context)
    ↓
    Saves state to .claude/session-state.md
    ↓
    Spawns continuation subagent (fresh context)
    ↓
    Subagent reads state, continues work

Thresholds

Configuration (environment variables):

• CONSERVE_EMERGENCY_THRESHOLD: Override 80% default (e.g., 0.75 for 75%)
• CONSERVE_SESSION_STATE_PATH: Override .claude/session-state.md default

Auto-Clear Workflow

Step 1: Assess Current State

Before triggering auto-clear, gather:

• Current task/goal description
• Progress made so far
• Key decisions and rationale
• Files being actively worked on
• Open TodoWrite items

Step 1.5: Finalize Task List Before Handoff

Important: Before saving state or spawning a continuation agent, reconcile the task list:

1. Review all tasks via TaskList 2. Mark completed tasks as completed via TaskUpdate: do NOT leave done work as in_progress 3. Record existing task IDs: collect all task IDs (pending and in_progress) to pass in the session state so the continuation agent references them instead of creating duplicates 4. Include task IDs in session state under the existing_task_ids field (see Step 2)

This prevents the continuation agent from creating duplicate tasks.

Step 2: Save Session State

Important: If .claude/session-state.md already exists, always Read it first before writing (Claude Code requires reading existing files before overwriting). Create the .claude/ directory if it doesn't exist.

Write to .claude/session-state.md (or $CONSERVE_SESSION_STATE_PATH):

# Session State Checkpoint
state_version: 1
Generated: [timestamp]
Reason: Context threshold exceeded (80%+)

## Execution Mode

**Mode**: [unattended | interactive | dangerous]
**Auto-Continue**: [true | false]
**Source Command**: [do-issue | execute-plan | etc.]
**Remaining Tasks**: [list of pending items]

> **Important**: If `auto_continue: true` or mode is `dangerous`/`unattended`,
> the continuation agent should not pause for user confirmation.
> Continue executing all remaining tasks until completion.

## Current Task
[What we're trying to accomplish]

## Progress Summary
[What's been done so far]

## Key Decisions
- Decision 1: [rationale]
- Decision 2: [rationale]

## Active Files
- path/to/file1.py - [status]
- path/to/file2.md - [status]

## Pending TodoWrite Items
- [ ] Item 1
- [ ] Item 2

## Existing Task IDs
[List task IDs from TaskList so the continuation agent can reference them
instead of creating duplicates. Example:]
- Task #1: "Implement feature X" (in_progress)
- Task #2: "Write tests for feature X" (pending)

## Continuation Instructions
[Specific next steps for the continuation agent]

Execution Mode Detection:

Before writing state, detect the execution mode:

# Detect execution mode from environment/context
execution_mode = {
    "mode": "interactive",  # default
    "auto_continue": False,
    "source_command": None,
    "remaining_tasks": [],
    "dangerous_mode": False
}

# Check for dangerous/unattended mode indicators
if os.environ.get("CLAUDE_DANGEROUS_MODE") == "1":
    execution_mode["mode"] = "dangerous"
    execution_mode["auto_continue"] = True
    execution_mode["dangerous_mode"] = True
elif os.environ.get("CLAUDE_UNATTENDED") == "1":
    execution_mode["mode"] = "unattended"
    execution_mode["auto_continue"] = True

# Inherit from parent session state if exists
if parent_state and parent_state.get("execution_mode"):
    execution_mode = parent_state["execution_mode"]

Step 2.5: Verify Session-State Clarity Before Handoff

A continuation agent inherits only what session-state.md says. If the draft is ambiguous, the new agent starts from corrupted task state. Before spawning, gate the handoff on the belief-clarity check (the belief-clarity module of Skill(conserve:context-optimization)), which asks two anchor questions against the draft state:

1. Progress probe: what is the current task progress: what is done, and what state is the task in now? 2. Gap probe: what information is still needed to finish: a bounded list of concrete open items, not generic categories.

Gate logic:

• Both answers specific and bounded: save and proceed to Step 3.
• Gap probe open-ended or progress probe hedging: append the failing

probe's answer to session-state.md as explicit Current state: and Still needed: bullets, then re-score.

• Progress probe vague or empty: do not hand off. Confirm current

state with the user (or imbue:proof-of-work in unattended mode) before writing state and retrying.

When memory-palace:memory-clarity-probe is installed, delegate the dual-probe evaluation to it and use its "Proceed" composite as the gate. This check is qualitative: it catches drift and omission, not confidently-wrong state, so pair it with task-state verification for high-stakes handoffs.

Step 3: Spawn Continuation Agent

Use the Task tool to delegate. Important: Include execution mode in the task prompt:

Task: Continue the work from session checkpoint

Instructions:
1. Read .claude/session-state.md for full context
2. Check the "Execution Mode" section FIRST
3. If `auto_continue: true` or mode is `dangerous`/`unattended`:
   - DO NOT pause for user confirmation
   - Continue executing ALL remaining tasks until completion
   - Only stop on actual errors or when all work is done
4. **TASK LIST**: Do NOT create new tasks via TaskCreate. The parent agent
   already created the task list. Use TaskList to see existing tasks, and
   TaskUpdate to mark them in_progress/completed. Check the "Existing Task IDs"
   section in the session state for the authoritative list.
5. Verify understanding of current task and progress
6. Continue from where the previous agent left off
7. If you also approach 80% context, repeat this handoff process
   - PRESERVE the execution mode when creating your own checkpoint

The session state file contains all necessary context to continue without interruption.

**Execution mode inheritance**: Always inherit and propagate the execution
mode from the session state. If the parent was in dangerous/unattended mode,
you are also in that mode. Do not ask the user for confirmation.

**Task deduplication**: Do not create duplicate tasks. The parent has already
populated the task list. Use TaskUpdate on existing task IDs only.

For batch/multi-issue workflows (e.g., /do-issue 42 43 44):

Task: Continue batch processing from session checkpoint

Instructions:
1. Read .claude/session-state.md for full context
2. EXECUTION MODE: This is a batch operation with auto_continue=true
3. Process ALL remaining tasks in the queue:
   - Remaining: [issue #43, issue #44]
4. DO NOT stop between tasks - continue until all are complete
5. If you hit 80% context, hand off with the same execution mode
6. Only pause for:
   - Actual errors requiring human judgment
   - Completion of ALL tasks

This is an unattended batch operation. Continue without user prompts.

Task Tool Details:

• Spawns subagent with fresh 1M context window
• Up to 10 parallel agents supported
• ~20k token overhead per subagent

Step 3 Fallback: Graceful Wrap-Up

If Task tool is unavailable (permissions, context restrictions):

1. Complete current in-progress work (finish edits, commits) 2. Summarize remaining tasks in your response 3. Let auto-compact handle continuation - Claude Code compresses context automatically 4. Manual continuation options:

• claude --continue to resume session
• New session and /catchup to understand changes
• Read .claude/session-state.md for saved context

Fixed in 2.1.63: /clear now properly resets cached skills. Previously, stale skill content could persist into the new conversation. The /clear and /catchup pattern is now fully reliable.

Fixed in 2.1.72: /clear now only clears foreground tasks. Background agent and bash tasks continue running. Previously, /clear would kill all tasks including background ones, which was problematic for long-running background agents that should survive context resets.

Integration with Existing Hooks

This skill works with context_warning.py hook:

1. Hook fires on PreToolUse 2. At 80%+ threshold, hook injects emergency guidance 3. Guidance recommends invoking this skill 4. Skill executes auto-clear workflow

Module Loading

For detailed session state format and examples:

• See modules/session-state.md for checkpoint format and handoff patterns
• See modules/session-state-schema.md for versioned schema and migration logic

Self-Monitoring Pattern

For workflows that might exceed context, add periodic checks:

# Pseudocode for context-aware workflow
def long_running_task():
    for step in task_steps:
        execute_step(step)

        # Check context after each major step
        if estimate_context_usage() > 0.80:
            invoke_skill("conserve:clear-context")
            return  # Continuation agent takes over

Context Measurement Methods

Precise (Headless/Batch)

For accurate token breakdown in automation:

claude -p "/context" --verbose --output-format json

See /conserve:optimize-context for full headless documentation.

Fast Estimation (Real-time Hooks)

For hooks where speed matters, use heuristics:

1. JSONL file size: ~800KB ≈ 100% context (used by context_warning hook) 2. Turn count: ~5-10K tokens per complex turn 3. Tool invocations: Heavy tool use = faster growth

Example: Brainstorm with Auto-Clear

## Brainstorm Session with Context Management

1. Before starting, note current context level
2. Set checkpoint after each brainstorm phase:
   - Problem definition checkpoint
   - Constraints checkpoint
   - Approaches checkpoint
   - Selection checkpoint

3. If context exceeds 80% at any checkpoint:
   - Save brainstorm state
   - Delegate to continuation agent
   - Agent continues from checkpoint

Best Practices

1. Checkpoint Frequently: During long tasks, save state at natural breakpoints 2. Clear Instructions: Continuation agent needs specific, concrete guidance 3. Verify Handoff: Ensure state file is written before spawning subagent 4. Monitor Recursion: Continuation agents can also hit limits - design for chaining

Troubleshooting

Continuation agent doesn't have full context

• Ensure session-state.md is complete
• Include all relevant file paths
• Document implicit assumptions

Subagent fails to continue properly

• Check that state file path is correct
• Verify file permissions
• Add more specific continuation instructions

Context threshold not detected

• CLAUDE_CONTEXT_USAGE may not be set
• The context_warning hook uses fallback estimation from session file size
• Manual invocation always works

Hook Integration

This skill is triggered automatically by the context_warning hook (hooks/context_warning.py):

• 40% usage: WARNING - plan optimization soon
• 50% usage: CRITICAL - immediate optimization required
• 80% usage: EMERGENCY - this skill should be invoked immediately

The hook monitors context via: 1. CLAUDE_CONTEXT_USAGE environment variable (when available) 2. Fallback: estimates from session JSONL file size (~800KB = 100%)

Configure thresholds via environment:

• CONSERVE_EMERGENCY_THRESHOLD: Override 80% default (e.g., "0.75")
• CONSERVE_CONTEXT_ESTIMATION: Set to "0" to disable fallback
• CONSERVE_CONTEXT_WINDOW_BYTES: Override 800000 byte estimate

Exit Criteria

• [ ] The task list is reconciled (completed tasks marked, open task

IDs recorded) before any state is written

• [ ] .claude/session-state.md exists and records current task,

progress, execution mode, and continuation instructions

• [ ] The session-state clarity gate (Step 2.5) passed, or the failing

probe's answer was appended and re-scored, before handoff

• [ ] A continuation agent was spawned with the execution mode and the

existing task IDs passed through (no duplicate task creation)

• [ ] Handoff is refused when the progress probe is vague or empty

Session State Schema

Current Version: 1

All session state files MUST include a state_version field on the first content line after the heading. This enables continuation agents to detect the format and apply migration logic when needed.

Version 1 Schema

Required Header

# Session State Checkpoint
state_version: 1
Generated: YYYY-MM-DD HH:MM:SS
Reason: [Context threshold | Manual checkpoint | Task boundary]

Required Sections

Optional Sections

Version 0 (Unversioned Legacy)

Any session state file that lacks the state_version field is v0. These files were written before versioning was introduced.

Identifying v0 Files

A v0 file typically starts with:

# Session State Checkpoint
Generated: YYYY-MM-DD HH:MM:SS

No state_version line is present.

V0 Field Mapping

V0 files use the same section names as v1. The only difference is the missing version header. All sections are compatible.

Migration: V0 to V1

When a continuation agent encounters a v0 file, apply this migration:

1. Treat it as v1. The content is structurally identical. 2. Do not rewrite the file just to add the version header. Only add state_version: 1 if you are already updating the file for other reasons (e.g., progress checkpoint). 3. Log the migration. Note in your handoff summary: "Migrated session state from v0 (unversioned) to v1."

No field renaming or restructuring is needed. V0 and v1 are content-compatible.

Version Check Logic

Continuation agents MUST follow this sequence when reading a session state file:

1. Read the file
2. Look for "state_version: N" in the first 5 lines
3. If found:
   a. version == 1  -> proceed normally
   b. version > 1   -> warn "Unknown state version N, attempting to read"
                        then proceed (best-effort forward compatibility)
4. If not found:
   -> treat as v0, proceed normally (v0 is compatible with v1)

Forward Compatibility

Future versions should maintain backward-compatible section names where possible. If a breaking change is needed, the version number increments and this document gets a new migration section.

A continuation agent encountering an unknown future version should:

• Log a warning: "Session state version N is newer than expected (v1). Reading with best effort."
• Attempt to read all recognized sections
• Skip unrecognized sections without error
• Continue the work rather than failing

Version History

Session State Module

Overview

This module defines the session state format used for context handoffs. A well-structured state file enables continuation agents to seamlessly pick up where the previous agent left off.

State File Location

Default: .claude/session-state.md

Override: Set CONSERVE_SESSION_STATE_PATH environment variable

export CONSERVE_SESSION_STATE_PATH="/tmp/my-session-state.md"

State File Format

Full Template

# Session State Checkpoint
state_version: 1

**Generated**: YYYY-MM-DD HH:MM:SS
**Reason**: [Context threshold | Manual checkpoint | Task boundary]
**Context Level**: [Estimated percentage if known]

---

## Current Objective

[One clear sentence describing what we're trying to accomplish]

### Success Criteria
- [ ] Criterion 1
- [ ] Criterion 2
- [ ] Criterion 3

---

## Progress Summary

### Completed
- [x] Step 1: [Brief description]
- [x] Step 2: [Brief description]

### In Progress
- [ ] Step 3: [What's currently being worked on]

### Remaining
- [ ] Step 4: [Next step]
- [ ] Step 5: [Future step]

---

## Key Decisions Made

| Decision | Rationale | Alternatives Considered |
|----------|-----------|------------------------|
| Decision 1 | Why we chose this | Option A, Option B |
| Decision 2 | Why we chose this | Option C, Option D |

---

## Active Context

### Files Being Modified
| File | Status | Notes |
|------|--------|-------|
| path/to/file1.py | In progress | Adding function X |
| path/to/file2.md | Pending review | Needs formatting |

### Files Read (Reference)
- `path/to/reference1.py` - Used for pattern reference
- `path/to/reference2.md` - Contains requirements

### Open TodoWrite Items

[ {"id": "todo-1", "status": "pending", "description": "Item 1"}, {"id": "todo-2", "status": "in_progress", "description": "Item 2"} ]

### Existing Task IDs

**IMPORTANT**: List all task IDs from TaskList so the continuation agent
references them via TaskUpdate instead of creating duplicates.

| Task ID | Subject | Status |
|---------|---------|--------|
| #1 | Task subject | in_progress |
| #2 | Task subject | pending |

> The continuation agent MUST use TaskUpdate on these IDs.
> It MUST NOT create new tasks via TaskCreate unless genuinely new work is discovered.

---

## Continuation Instructions

### Immediate Next Step
[Specific action to take first]

### Context to Re-read
1. [File path 1] - [Why it's needed]
2. [File path 2] - [Why it's needed]

### Warnings/Gotchas
- [Any pitfalls the continuation agent should avoid]
- [Known issues encountered]

### If Blocked
[What to do if the immediate next step fails]

---

## Execution Mode

**CRITICAL**: Capture and propagate execution mode for unattended workflows.

Execution Mode

Mode: [unattended | interactive | dangerous] Auto-Continue: [true | false] Source Command: [do-issue | execute-plan | batch-process | etc.] Remaining Tasks: [list of pending task IDs or issue numbers]

Flags Inherited

• --dangerous: Continue executing without user prompts
• --no-confirm: Skip confirmation dialogs
• --batch: Processing multiple items

| Mode | Behavior | Use Case |
|------|----------|----------|
| `interactive` | Pause at checkpoints, ask user | Normal development |
| `unattended` | Continue automatically, log decisions | CI/CD, batch processing |
| `dangerous` | Like unattended + skip permissions | Fully automated pipelines |

**Detection**: Check environment variables and session context:
- `CLAUDE_DANGEROUS_MODE=1` → dangerous mode
- `CLAUDE_UNATTENDED=1` → unattended mode
- Presence of `--dangerous` in original command → dangerous mode
- Multiple issues/tasks in queue → likely batch mode

---

## Metadata

{ "checkpoint_version": "1.1", "parent_session_id": null, "handoff_count": 0, "estimated_remaining_work": "medium", "priority": "high", "execution_mode": { "mode": "interactive", "auto_continue": false, "source_command": null, "remaining_tasks": [], "dangerous_mode": false }, "existing_task_ids": [] }

Minimal Template

For quick checkpoints:

# Quick Checkpoint
state_version: 1

**Task**: [What we're doing]
**Progress**: [Where we are]
**Next**: [Immediate next step]
**Files**: [Key files to read]

Writing Session State

When to Write

1. Context threshold exceeded (80%+) 2. Natural task boundary (phase complete) 3. Before risky operation (safety checkpoint) 4. Manual checkpoint request

What to Include

Essential (always include):

• Current objective
• Progress summary
• Immediate next step
• Key files

Important (include when relevant):

• Decisions made with rationale
• Open TodoWrite items
• Warnings/gotchas

Optional (include for complex tasks):

• Alternatives considered
• Blocked items
• Detailed metadata

Writing Checklist

Before writing session state:

• [ ] Objective is clear and specific
• [ ] Progress accurately reflects completed work
• [ ] Next step is actionable (not vague)
• [ ] File paths are correct and accessible
• [ ] Decisions are documented with rationale
• [ ] No sensitive information included

Reading Session State

Continuation Agent Protocol

When a continuation agent starts:

1. Read state file first

   Read .claude/session-state.md

2. Verify understanding

• Summarize the objective
• Confirm progress status
• Identify immediate next step

3. Re-read critical files

• Files listed in "Context to Re-read"
• Files marked "In Progress"

4. Acknowledge handoff

• Confirm readiness to continue
• Note any questions or ambiguities

5. Execute continuation

• Start from "Immediate Next Step"
• Follow task workflow

Handling Ambiguity

If state file is unclear:

1. Check for related files

• Recent git changes
• Open PRs/issues
• Project documentation

2. Make reasonable assumptions

• Document assumptions made
• Proceed with caution

3. Flag uncertainties

• Note areas of confusion
• Ask for clarification if critical

State File Lifecycle

┌─────────────────┐
│ Task Started    │
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Work in Progress│◄──────────────┐
└────────┬────────┘               │
         │                        │
         ▼                        │
┌─────────────────┐    No        │
│ Context > 80%?  │──────────────┘
└────────┬────────┘
         │ Yes
         ▼
┌─────────────────┐
│ Write State File│
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Spawn Subagent  │
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Subagent Reads  │
│ State File      │
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Continue Work   │──────────────►(Back to Work in Progress)
└─────────────────┘

Multi-Handoff Chains

For very long tasks, multiple handoffs may occur:

## Metadata

{ "checkpoint_version": "1.0", "parent_session_id": "handoff-001", "handoff_count": 2, "handoff_history": [ {"id": "handoff-001", "reason": "context_threshold", "timestamp": "..."}, {"id": "handoff-002", "reason": "context_threshold", "timestamp": "..."} ] }

Each continuation agent increments handoff_count and adds to handoff_history.

Recommended max: Set a limit of 5-10 handoffs to prevent infinite chains. If handoff_count exceeds the limit, stop and report to the user rather than spawning another continuation agent.

Best Practices

1. Be Specific: "Add validation to parse_config()" not "Continue working on config"

2. Include Context: Don't assume continuation agent knows anything

3. Document Decisions: Future agents need to understand WHY, not just WHAT

4. Test File Paths: Ensure all referenced files exist and are readable

5. Keep It Updated: Stale state files cause confusion

6. Clean Up: Remove old state files after task completion

Example: Real Handoff

Original Agent State

# Session State Checkpoint
state_version: 1

**Generated**: 2025-01-15 14:30:00
**Reason**: Context threshold exceeded (82%)
**Context Level**: ~82%

---

## Current Objective

Implement the `clear-context` skill for the conserve plugin, including
the main skill file, session-state module, and hook integration.

### Success Criteria
- [x] Create skill directory structure
- [x] Write main SKILL.md
- [ ] Write session-state module
- [ ] Update context_warning.py hook
- [ ] Create continuation agent definition
- [ ] Test the full workflow

---

## Progress Summary

### Completed
- [x] Created /plugins/conserve/skills/clear-context/
- [x] Created /plugins/conserve/skills/clear-context/modules/
- [x] Wrote SKILL.md with full documentation

### In Progress
- [ ] Writing session-state.md module (this file)

### Remaining
- [ ] Update context_warning.py with 80% threshold
- [ ] Create continuation-agent.md
- [ ] Integration testing

---

## Key Decisions Made

| Decision | Rationale | Alternatives |
|----------|-----------|--------------|
| Use subagent delegation | Can't programmatically /clear | Manual clear workflow |
| 80% default threshold | Leaves 20% buffer | 75%, 85% |
| .claude/session-state.md | Project-scoped | /tmp, ~/.claude |

---

## Continuation Instructions

### Immediate Next Step
Complete the session-state.md module, then update context_warning.py

### Context to Re-read
1. `/plugins/conserve/skills/clear-context/SKILL.md` - Main skill reference
2. `/plugins/conserve/hooks/context_warning.py` - Hook to update

### Warnings/Gotchas
- CLAUDE_CONTEXT_USAGE env var may not be set
- Design must work with estimation fallback

Continuation Agent Response

I've read the session state checkpoint. Let me verify my understanding:

**Objective**: Implement clear-context skill for conserve plugin
**Progress**: SKILL.md is complete, session-state module in progress
**Next Step**: Complete session-state.md, then update context_warning.py

Reading the referenced files now to continue...