Cline's automatic context compaction is supposed to summarize prior conversation history when a long session runs low on context. The code that decides whether to install that summary only checks whether it came back completely empty — not whether it got cut off partway through.
What the source says
cline/cline#14732, filed October 1, 2026 by a user (feiiiiii5), traces the gap to runAgenticCompaction in core/src/extensions/context/agentic-compaction.ts. generateSummary already records a SummaryGenerationResult.incompleteReason when generation is cut short by the output budget, but the installing code only guards against if (!rawSummary) — a fully empty result. A truncated-but-non-empty summary passes that check and replaces the full conversation history exactly as if it were complete, with nothing in the application logs distinguishing the two cases. The reporter traces this becoming more likely after PR #13137, which started applying the user's own max-output-tokens setting to the summarization request itself, so a setting lowered for unrelated reasons can now silently truncate compaction.
A second commenter, chenhz01, confirmed the problem the same day and proposed warning the user and retrying with a smaller input. A third, aldi949, named the downstream risk on October 9: because the agent has no signal that its own context is incomplete, "mandatory requirements can disappear" from the task state after compaction.
The reporter opened three follow-ups — PR #14741 (logging the truncation and adding one retry), issue #15005 (a fidelity signal for whether retained context is provably complete), and discussion #15006 (a user-visible marker) — but said directly that the retry PR "does not claim" to recover requirements a truncated summary already dropped. As of this writing the issue is open with no maintainer decision on which fix to take.
What it doesn't establish
No maintainer has confirmed the root cause or committed to a fix direction — this is the reporter's own trace through the code, corroborated by two commenters, not a maintainer-acknowledged bug. There's no report yet of a specific task that failed because of a dropped requirement; the "requirements can disappear" risk is aldi949's inference about what the gap permits, not a documented instance of it happening. No data loss or security impact is claimed.
Why it's still worth logging
Agentic coding tools lean hard on compaction to survive long sessions, and this is a case where the mechanism meant to preserve continuity can silently drop information with zero signal — not because summarization failed outright, but because "cut short" and "complete" produce the same code path. The reporter and two commenters agree on the gap; what's missing is a maintainer saying which of the three proposed fixes actually ships.
Full record, severity rationale, and frontmatter: https://www.stupidllm.com/incident/STUPID-2026-0143/
Top comments (0)