If your agent tooling shows a session as "failed," you'd assume it stopped calling the model. This report says that assumption doesn't always hold.
What the source says
GitHub issue #14256, filed against cline/cline on September 18, 2026 by user romemozu: of their 10 non-Anthropic Cline Desktop App sessions in the local session store, 9 were marked failed — and several kept making API calls for 20 to 23 hours after the point of failure. One session alone made 416 API calls and consumed 112,057,985 input tokens, averaging roughly 183,696 input tokens per call, because Cline was resending the full, growing conversation on every call rather than the failed session actually halting. Across the reporter's full session set: 1,156 calls, 212,352,674 input tokens, 834,634 output tokens, and 165,124,705 cache-read tokens (77.8% of input).
A separate defect compounded it: the reporter's Cline Credits balance went negative (-$0.26) while the sessions in question were running models explicitly labeled FREE — cline-free/deepseek-v4.1-flash and cline-free/muse-spark-1.3-contributor.
Neither problem is new. The report links cline/cline#13367, describing the same runaway-after-failure behavior, and cline/cline#8074, describing the same FREE-model billing bug — both filed earlier and still open. As of this incident's publication, #14256 had no visible maintainer response.
What it doesn't establish
This is a single first-hand report, and it isn't marked reproducible in our record. There's no maintainer confirmation of root cause, no word on scope (how many users are affected), and no dollar figure attached to the paid-model portion of the token burn — only the raw counts above, which come from one user's local session store.
Why it's still worth logging
A session lifecycle bug and a billing-integrity bug are two different failure classes, and this report has both, with hard numbers instead of a vague complaint. "Marked failed" doing nothing to stop further API calls is the kind of defect that turns a single bad run into a runaway one — and a FREE-labeled model still deducting credits means the field meant to tell a user "this costs you nothing" is, at least here, wrong. Both were reported before this issue and remain open.
Full incident record, including frontmatter fields and severity scoring: https://www.stupidllm.com/incident/STUPID-2026-0119/
Top comments (0)