context-degradation（环境退化）

Context Degradation Monitor

Detects context window degradation and prescribes corrective actions before accuracy drops.

Severity Zones

Reference: Models advertise 200K but reliability drops past 32K. "Lost in the middle" effect: middle tokens have 20-40% lower recall past 100K.

Early Warning Indicators

Symptoms that context degradation is affecting output quality (regardless of token count):

1. Repeated tool calls — Agent re-reads files already read in same session
2. Contradictory reasoning — Later steps contradict earlier decisions
3. Missing prior context — Agent "forgets" task scope or constraints stated at session start
4. Over-explanation — Agent re-explains concepts already established
5. Stale references — Agent references file paths or task IDs that were resolved earlier

If 2+ indicators are present, treat as one zone higher than token count suggests.

Corrective Routing by Zone

Yellow (32-64K):

• Invoke Skill({skill: 'context-compressor'}) at the current phase boundary
• Remove completed phase content from active context
• Keep: current task spec, key decisions, in-progress file list

Orange (64-100K):

• Invoke Skill({skill: 'context-compressor'}) — aggressive summarization
• Write phase summary to .claude/context/tmp/phase-summary-{date}.md
• Prune: all resolved task details, intermediate research, superseded plans

Red (100-140K):

• Do NOT continue complex multi-step tasks in current agent
• Spawn a fresh subagent with only the compressed summary as context
• Current agent: write handoff doc → call TaskUpdate(completed) with handoff path in metadata

Critical (>140K):

• Halt immediately
• Write emergency summary: what was done, what remains, key decisions
• Route to session-handoff skill
• Invoke Skill({skill: 'session-handoff'}) before context window forces truncation

Detection Checklist (Run at Each Phase Boundary)

[ ] Token count below 32K? → Green, no action
[ ] Token count 32-64K? → Yellow, begin compression
[ ] Token count 64-100K? → Orange, compress aggressively
[ ] Token count 100K+? → Red/Critical, spawn fresh agent
[ ] 2+ early warning indicators? → Upgrade one severity zone
[ ] Compression reminder file exists? → Invoke context-compressor immediately

Integration

• Pairs with: context-compressor, context-compressor, session-handoff
• Called by: planner (at plan start), developer (after each phase), router (before large spawns)
• Trigger: Check at every phase boundary, not just when problems appear

Iron Laws

1. ALWAYS check token count at every phase boundary — not just when problems appear or after completing a large task.
2. NEVER continue complex multi-step tasks past 100K tokens in the same agent context — spawn a fresh subagent with a compressed handoff instead.
3. ALWAYS treat 2+ early warning indicators as one severity zone higher than the raw token count suggests.
4. ALWAYS invoke context-compressor at Yellow zone (32–64K) before context bloat becomes severe — prevention is cheaper than recovery.
5. NEVER claim a task complete without writing a context summary when operating in Red or Critical zone — if it's not written down, the next agent won't know it happened.

Anti-Patterns

Memory Protocol (MANDATORY)

Before starting: Read .claude/context/memory/learnings.md

After completing:

• New pattern -> .claude/context/memory/learnings.md
• Issue found -> .claude/context/memory/issues.md
• Decision made -> .claude/context/memory/decisions.md

ASSUME INTERRUPTION: If it's not in memory, it didn't happen.