STUPID-2026-0143

Cline silently installs a truncated post-compaction summary with no warning when the summary generation hits the output budget

3.3low
October 1, 2026

  1. Instruction given

    No specific instruction triggers this — it occurs automatically whenever Cline's agentic context compaction runs during a long session, generating a summary of prior conversation history to free up context space.

  2. Expected behavior

    If summary generation is cut off before completing (an "incomplete" summary, as opposed to a fully empty one), Cline should warn the user and/or retry with a smaller input before installing it, or abort compaction and keep the full history, rather than silently treating a partial summary as equivalent to a complete one.

  3. Actual behavior

    In `core/src/extensions/context/agentic-compaction.ts`, `generateSummary` records a `SummaryGenerationResult.incompleteReason` when the summary is cut short, but `runAgenticCompaction` only checks whether the summary is completely empty (`if (!rawSummary)`) before installing it. A truncated-but-non-empty summary silently replaces the full conversation history with no warning anywhere, including the logs: as the reporter put it, "nothing in the logs distinguishes 'summary complete' from 'summary hit the output budget.'" A prior change (PR #13137) made this more likely to occur by applying the user's own max output tokens setting to summarizer requests, so a low setting can now truncate compaction summaries that previously would not have been affected. Another commenter (aldi949) noted the downstream risk: because the agent proceeds believing the installed summary is complete, "mandatory requirements can disappear" from the task state after compaction with nothing to signal that they were dropped.

  4. Damage

    No data loss or security impact is reported. The risk is silent context loss: after compaction, the agent's continuation note can "stop mid-thought" and the task can proceed on an incomplete record of prior requirements with no error surfaced to the user. As of this writing the issue remains open and unresolved — a second commenter (chenhz01) confirmed the problem, and the reporter opened three follow-up items (PR #14741, making truncation at least observable and adding one retry; issue #15005, on where a fidelity signal should live; and discussion #15006, requesting a user-visible marker) — but no maintainer has yet chosen a fix direction, and the reporter's own follow-up PR stops short of claiming the underlying requirements-loss risk is resolved.

A Cline user (feiiiiii5) filed an issue on October 1, 2026 describing a gap in Cline's agentic context compaction: when a long session triggers automatic summarization of prior history, the code path that decides whether to install the resulting summary only checks for a completely empty result. It does not check `SummaryGenerationResult.incompleteReason`, the field that is set when generation was cut off partway through. A summary that is truncated but non-empty is installed exactly as if it were complete, with nothing in the application logs distinguishing the two cases. The reporter traced this to have become more likely after PR #13137, which began applying the user's own max-output-tokens setting to the summarization request itself — so a user who had lowered that setting for unrelated reasons can now unknowingly cause their own compaction summaries to be truncated. A second commenter, chenhz01, confirmed the problem the same day and proposed warning the user and retrying with a smaller input before falling back to installing a flagged-incomplete summary. A third commenter, aldi949, flagged the downstream consequence on October 9: because the agent has no way to know its own context is incomplete, "mandatory requirements can disappear" from the task state after compaction without any signal that this happened. The reporter opened three follow-up items in response — PR #14741 (making truncation observable in logs and adding one retry), issue #15005 (proposing a fidelity signal for whether retained context is provably complete), and discussion #15006 (requesting a user-visible marker on truncated summaries) — but said explicitly that the retry PR "does not claim" to preserve requirements that a truncated summary already dropped. As of this writing the original issue remains open with no maintainer decision on which fix to adopt, and the risk aldi949 described — a requirement silently vanishing from a task's working context mid-session — is unresolved.

Classification

Agent
Cline
Failure mode
Logic Error
Root cause
Logic Error
Domain
Other
Source
Github Issue

Related incidents

10.0
Clinejection: an AI issue-triage workflow enabled arbitrary code execution on the CI runner
8.4
Amazon AI coding agent mistake blamed on human employees
7.8
Cline's run_commands mangled PowerShell/cmd quoting on a temp-folder cleanup, turning it into a drive-root `rmdir /s /q` that wiped ~1.5TB of a user's photos, videos, and documents (GitHub #14864)
5.8
Cline's hub daemon silently reverted workspace files to stale state, repeatedly erasing a gitignored .env.local and collapsing a node_modules folder from 588 entries to 2 (GitHub #14819)
Discuss on Hacker News

Get told when an agent breaks something

We document AI agent failures daily, severity-scored against a published scale. When one lands at 7.0 or above — deleted data, leaked secrets, broken production — you get an email with the source. When nothing does, you get nothing.

This database is callable over MCP — query it from inside your agent.