context-compaction

Compaction is not summarization for its own sake. It is energy recovery.

The goal is to preserve decision-useful signal while discarding stale context mass.

1) Use When

• a task has shifted phases
• earlier exploration is no longer needed in full
• repeated tool results are crowding out active work
• a worker or route is carrying too much irrelevant history
• token burn is rising faster than useful progress

2) Keep vs Drop

Keep:

• chosen approach
• file paths that still matter
• constraints and assumptions
• exact failures still being debugged
• verification commands and outcomes

Drop:

• dead-end explorations
• duplicate file summaries
• verbose tool output once the conclusion is known
• alternative approaches already rejected
• long prose that can be collapsed into one line

3) Compaction Pattern

Before:
- 8 files explored
- 4 rejected theories
- 3 repeated test outputs
- 1 active fix path

After:
- active fix path
- 2 relevant files
- exact failing test
- 1 sentence on rejected theory only if it prevents rework

4) Phase Boundary Rule

Compact aggressively at:

• research -> implementation
• implementation -> verification
• pre-release handoff
• after a large failed attempt

Do not carry research noise into implementation unless it is still causally relevant.

5) High-Value Compaction Formats

• decision log
• current target files
• active hypothesis
• exact next command
• known no-go paths

6) Anti-Patterns

• compressing away the one error string that matters
• keeping giant excerpts "just in case"
• summarizing every explored file equally
• treating compaction as archival completeness instead of operational focus

7) Checklist

• [ ] Removed stale exploration from the active context
• [ ] Preserved the chosen path and why
• [ ] Kept exact errors and commands still needed for verification
• [ ] Dropped duplicate summaries and obsolete alternatives
• [ ] Reduced token load without reducing decision quality