DC-CTX-013: Conversation summarization engine [v2.0]

[radoxtech]

DC-CTX-013 — Conversation summarization engine

Epic: #187 (Context Management v2)
Sprint: 3
Depends on: DC-CTX-010

---

Goal

Replace older conversation messages with intelligent summaries to free context space (Plandex ConvoSummary pattern).

Scope

• Summarizer agent (from MVP roster) generates summaries
• Summary replaces N oldest messages, keeping:
  • First user message (original task context)
  • Last K messages (recent working context)
  • All messages after the most recent summary
• Summary format: structured markdown with key decisions, file changes, current state
• Incremental: new summary builds on previous summary (not full re-summarization)
• Token target: summary should be 10-20% of original messages' token count

Acceptance Criteria

• Summarization preserves key decisions and file change records
• First user message always preserved
• Recent K messages always preserved
• Summary token count is 10-20% of replaced messages
• Incremental summarization works (builds on previous summary)
• Agent can still function correctly after summarization

Technical Notes

• Uses LOW tier model for summarization (cost-efficient)
• Incremental summary: append new summary to previous, then re-summarize if too long
• Key extraction: file paths mentioned, decisions made, errors encountered, current approach

Implementation Research (Mandatory)

Before implementing, research:

• Plandex ConvoSummary — exact implementation of conversation summarization
• mem0 memory system — how mem0 extracts and compresses conversation memories
• LangChain ConversationSummaryMemory — summary-based memory implementation
• GPT-4 summarization benchmarks — what compression ratios are achievable
• Anthropic context caching — can summaries leverage prompt caching?
• Review analiza-context-management.md Section 2.3

References

• analiza-context-management.md Section 2.3 (Plandex ConvoSummary)
• Summarizer agent (MVP roster)