Context Management for Long-Running Sessions

The Core Problem

LLMs degrade gradually starting around 30K tokens. Information in the middle of context is recalled 10-40% worse than at the edges. After ~100K tokens, you're operating at reduced capacity.

Prevention: Keep Context Lean

File Reading

• Read only files you're about to edit
• Use Grep to find specific patterns instead of reading whole files
• Spawn subagents for exploratory reading — their context is separate

Subagent Isolation

• Each subagent gets its own fresh context window
• Send them specific instructions, get back structured results
• The main agent's context only grows by the summary, not the raw exploration

Progressive Disclosure

• Start with file names (Glob)
• Then search for patterns (Grep)
• Then read specific lines (Read with offset/limit)
• Only read the full file when you're ready to edit it

Detection: Know When Context Is Full

Warning signs that you're approaching limits:

• You've been working for 20+ tool calls without spawning subagents
• You've read more than 10 files in the current session
• You're re-reading files you read earlier (sign of context loss)
• Your responses are getting less specific or accurate

Recovery: The Handoff Protocol

When you detect context pressure, write a handoff file:

.autopilot/handoff.md Structure

# Handoff — [Timestamp]

## Current Task
[What I was working on]

## What's Done
- [Completed item 1]
- [Completed item 2]

## What's Left
- [Remaining item 1 with specific file:line references]

## Key Context
- [Critical fact 1 that the next session MUST know]
- [Critical fact 2]

## Failed Approaches
- Tried [X], failed because [Y] — DO NOT retry

## Files Modified This Session
- `path/to/file.ts` — [what changed]

## Environment State
- Branch: [current branch]
- Tests: [passing/failing]
- Uncommitted changes: [yes/no, what]

What Makes a Good Handoff

1. Concrete, not narrative: File paths and line numbers, not stories
2. Failed approaches: The most valuable section — prevents re-doing dead ends
3. Actual code snippets: For complex logic, include the key 5-10 lines
4. Decision rationale: Not just "chose X" but "chose X because Y"
5. Verification state: What tests pass, what's broken

What to Omit

• General knowledge about the codebase (it's in CLAUDE.md)
• Full file contents (they're on disk)
• Explanations of how tools work
• Narrative of the session ("First I tried...")

The Manus Pattern: Todo Recitation

For very long sessions, periodically rewrite your current task state to keep it fresh at the context boundary:

• Every 10 tool calls, mentally review: "What am I doing? What's left?"
• Write current state to handoff.md even if you're not switching sessions
• This fights the "lost in the middle" effect