Seven of sixteen compacted agent sessions answered from the summary instead of re-reading the file

The hypothesis came out of repeated annoyance. After compacting, the author of the dev.to post reports, the model would state facts about code that were no longer true, and once corrected would go and read the file and say "ohh yes, I was wrong, the code states as you say" ^[18].

Compaction replaces the session with a summary of it, and a value the summary kept then exists twice: once in the file, once in prose about the file ^[6]. That prose copy is a second source of truth for something that can still change underneath it ^[1]. Re-reading costs a tool call. Quoting the summary costs nothing.

Two other paths put a stale value in front of the model. Compaction does not touch a file the agent itself wrote, so an agent that took notes reads its own notes instead of the code ^[7]. Then there is the handoff at the cut: "Claude Code returns your most recent files after compacting, and what you get is a photograph taken at the moment of the cut", the author wrote ^[8].

The sample is small. Eighteen sessions in total ^[4], sixteen of them compacted, which leaves a control arm of two ^[1]. The author states the limit before anyone else can: "This is mechanism validation, not a statistical study" ^[5]. Opus did better than Fable, and the behaviour turned up in both ^[9].

For that failure rate to transfer to another workload, re-reading has to be cheap and the agent has to skip it anyway. The trial built exactly that situation. Configuration files were chosen because code is the friendly case, sitting still in a file with re-reading one tool call away ^[13]. Where the source changes on its own and re-checking costs more than one file read, I would expect the skip to happen at least as often, and the post does not measure that case.

The savings side of the argument is asserted and untested. Compaction does clear the bloat of a long session, the exploratory scripts and the files read once and never needed again, and it points attention back at the current activity ^[17]. On tokens, the author wrote that "session pricing is dominated by cache reads in long sessions" and that "compacting often may result in x2 and even more savings", then added: "or at least that is the argument, I haven't measured it myself" ^[16].