| name | context-compaction-handoff |
| description | Guardrails for resuming work after context compaction or transcript handoff blocks; prioritize the latest real user request over stale summarized tasks and verify before answering. |
Context Compaction Handoff
Use this skill whenever the conversation includes a context-compaction summary, transcript handoff, or other generated summary block that says it is background/reference-only.
Core rule
A compaction block is background evidence, not the active instruction. Resume from the latest real user message after the compaction marker. If the latest user message conflicts with the summary's Active Task, the latest user message wins.
Procedure
-
Separate source types
- Treat explicit user messages as active instructions.
- Treat compaction summaries as historical context unless the latest user asks to resume them.
-
Locate the active request
- First scan below the compaction marker for the latest user-authored request.
- If no real user request exists after the marker, fall back to the latest visible real user request immediately before the compaction block, as long as it has not already been answered.
- If the compaction block says summary generation was unavailable or contains no reliable active-task section, do not treat old todos/final summaries as active; either continue the last visible unanswered user request or ask for clarification.
- Never invent a task from the generated summary itself.
-
Revalidate before acting
- If the active request asks for evidence, logs, state, or current facts, inspect the relevant sources before answering.
- Do not answer from the compaction summary when the user asked for live verification.
-
Avoid stale closeout leakage
- Do not use old todo lists, issue-closeout summaries, or prior final-response wording as the answer to a new post-compaction question.
- If a todo list was reconstructed from the compaction summary, keep it separate from the latest user request unless the request explicitly matches it.
Pitfalls
- Wrong: A compaction says
Active Task: None, but the pre-compaction visible user asked a new evidence-gathering question. The agent answers with a completed old closeout summary.
- Right: The agent identifies the latest real user request, loads relevant skills, inspects the requested evidence, and answers only that request.
References
references/compaction-latest-user-message-guard.md — session-specific failure mode and recovery checklist.