An Aider user gets a well-reasoned, correct code change back from the model — and Aider throws it away without a word, because the response didn't match the edit format Aider guessed it should be in.
What happened
Filed as Aider-AI/aider#5486 by the affected user (noelsaw1) on 2026-07-23. Working in a large repo (~1,100 files, ~179k tokens of context) with qwen3.8-max-preview via an OpenAI-compatible endpoint — a model with no entry in Aider's model-settings.yml — Aider had no format to go on and auto-selected whole (full-file rewrites). The model didn't cooperate: it replied with a standard unified diff instead.
Aider's parser tried to match that diff against the whole-format path, failed, concluded there were "no tracked changes," and dropped the entire response. No warning, no error — just an empty, no-op-looking turn. Setting --edit-format diff explicitly avoided the bug, which is how the reporter isolated the cause to the format guess rather than the model's output.
Why it matters more than it looks
The model did the work correctly. The failure is entirely on Aider's side, in the gap between "format I guessed" and "format the response actually arrived in" — and the tool resolves that mismatch by silently deleting the response rather than telling the user anything went wrong. A user in that position has no signal to go on: they might re-prompt, assume the model failed, or just lose the change.
It's also not a one-off. It's the third distinct silent-failure report in Aider from roughly the same window: a partial-hunk-application warning that fails to fire (#5573), headless mode exiting 0 after a fatal API failure (#5552), and now this. Three different code paths, same shape — work is lost or a failure occurs, and nothing downstream is told.
What the source establishes — and what it doesn't
The issue is filed by the person who hit it, with a specific reproduction path (unlisted model → auto-selected whole format → model returns a diff instead), and the reporter's own fix (--edit-format diff) confirms where the fault lies. That's why this is reproducible.
What it doesn't establish: no maintainer response had landed as of publication, so it's unconfirmed and unverified — this is one user's report, not something StupidLLM's process has independently reproduced. No specific production incident (lost hours, a missed deadline) was reported alongside the filing — the damage here is the mechanism itself, not a documented downstream consequence.
Status
Open, unconfirmed by maintainers, as of this writing. Any Aider user on a model without a model-settings.yml entry who doesn't pin --edit-format explicitly is exposed to this today.
Full incident record and severity scoring (4.6 / medium): STUPID-2026-0082
This is one of 82+ verified, severity-scored AI agent incidents documented at StupidLLM, an open incident database for AI coding agent failures.
Top comments (0)