| name | anchor-context |
| description | Use when the user says '锚点上下文', 'anchor context', '查看锚点', '注入锚点', 'anchor inject', '上下文锚点', '之前的对话', '之前讨论的', '之前说的', mentions wanting to recall or recover information from earlier in a long conversation, asks about something discussed before, or needs context reconstruction after compaction — provides anchor-based context recovery from compacted or earlier conversation segments |
Anchor Context — 锚点上下文
Overview
Extract minimal structured anchors from conversations and reconstruct context on demand. Instead of reading full conversation summaries, query specific topics and get a targeted anchor window with the relevant conversation points.
Core principle: Anchor = entity (noun) + verb (label) + data values (exact numbers). ~97% compression rate, zero LLM cost for extraction.
When to Use
Activate when the user:
- Says trigger phrases:
锚点上下文, anchor context, 查看锚点, 注入锚点
- Asks about something discussed earlier: "之前讨论的...", "之前说的..."
- Wants to review what happened before: "之前那个方案是什么"
- After a
/compact — check if anchors were saved and offer to show them
How It Works
Compression (automatic, PreCompact hook):
Long conversation → context limit approaches
→ Hook fires → extracts anchors → saves to ~/.claude/anchors/
→ Original discarded, anchors preserved
Recovery (on demand, you trigger):
User says "锚点上下文" or asks about prior topic
→ Load saved anchors → format for display
→ Present anchor summary to user
Workflow
1. Check for Saved Anchors
Run the display command to see what's available:
python "<skill-path>/scripts/inject.py" --format
Replace <skill-path> with ~/.claude/skills/anchor-context or the actual path.
2. Present Anchors to User
If anchors exist, present them in a structured way:
- Show the session's anchor summary
- Highlight KEY anchors (DECISION / ANOMALY types)
- Ask which topic they want to explore
3. Reconstruct on Query
When the user asks about a specific topic, use:
python "<skill-path>/scripts/anchor/reconstructor.py" "<user-query>"
This returns a positional window of anchors around the most relevant match.
Do NOT fabricate or hallucinate conversation history. If no anchors are saved, tell the user: "No saved anchors — anchors are saved automatically during compaction. Keep working and they'll be generated when context fills up."
Red Flags
| Thought | Reality |
|---|
| "I'll just summarize from memory" | Memory summaries lose detail. Use saved anchors. |
| "The user didn't say the exact trigger phrase" | If they ask about past conversation, check for anchors. |
| "Anchors aren't relevant to this query" | Let the user decide. Show what's available first. |
Commands Quick Reference
python ~/.claude/skills/anchor-context/scripts/inject.py --format
python ~/.claude/skills/anchor-context/scripts/pre_compact.py save
python -m pytest ~/.claude/skills/anchor-context/scripts/../tests/
Technical Details
See REFERENCE.md for:
- Anchor format specification
- Entity class definitions and weights
- Position-based retrieval algorithm
- Hook integration details
- Troubleshooting
