SWE-chat is a dataset of real coding-agent sessions introduced by Baumann et al. in SWE-chat: Coding Agent Interactions From Real Users in the Wild, which describes it as the first large-scale dataset of real coding-agent sessions from open-source developers in the wild. It is a living dataset: its collection pipeline automatically and continually discovers and processes new sessions from public repositories. It has also been used as one of the two datasets behind How Coding Agents Fail Their Users: A Large-Scale Analysis of Developer-Agent Misalignment in 20,574 Real-World Sessions, where it covers CLI workflows specifically and complements the IDE-weighted SpecStory exports that study pairs it with.
Contents
A record is a single coding-agent session: interleaved user prompts, agent responses, and tool-call traces such as file edits, shell commands and code searches, together with token usage. What sets the dataset apart, according to its authors, is that sessions are linked to commits with line-level attribution of which committed lines a human wrote and which the agent wrote; among the agent datasets they compare it with, it is the only one combining human prompts, agent tool-use trajectories, code diffs and code attribution. As of April 2026 it held almost 6,000 sessions across more than 200 repositories, with over 13,000 checkpoints, 63,000 user prompts and 355,000 agent tool calls; the full logs contain 2.7 million events, including streamed progress events, tool-call return values and a small set of reasoning traces. About 85% of the data comes from Claude Code, the first agent the logging tool supported; sessions from OpenCode, Gemini CLI, Cursor and Factory AI Droid are also recorded. The introducing paper also annotates sessions and prompts with LLM judges for session success, user persona, prompt intent and user pushback.
In How Coding Agents Fail Their Users: A Large-Scale Analysis of Developer-Agent Misalignment in 20,574 Real-World Sessions the corpus contributes 5,785 sessions across 198 repositories. CLI sessions run longer than IDE ones in that study: it reports a median of five user-authored messages per session across its CLI group against three across its IDE group, though the per-agent medians in its combined dataset vary widely (eight for OpenCode, five for Claude Code, one for OpenAI Codex).
Agent identity is recorded for these sessions, unlike the early SpecStory exports that study pairs them with: within that study’s CLI group, sessions are attributed to Claude Code (6,648), OpenCode (624), OpenAI Codex (517), Gemini CLI (39) and Cursor CLI (32), with 483 of unknown agent. Those counts describe the study’s combined CLI group rather than SWE-chat alone. The study does not analyse results by model identity, giving two reasons: SpecStory exports do not record it, and within SWE-chat, Claude-family models account for 94.9% of annotated responses, leaving insufficient variation for meaningful comparison.
Provenance
The corpus is collected via Entire CLI, an open-source tool that logs coding-agent sessions and records them, with checkpoint metadata, on a dedicated branch of the repository. The introducing paper’s pipeline finds public GitHub repositories that use it through the GitHub code search API, downloads their checkpoint directories and parses the raw transcripts into structured tables; it filters out data that appears to be generated by automated bots. Only repositories whose licenses allow redistribution are included, images attached to prompts are not collected, and before release personally identifiable information is removed from every user prompt and assistant response with Microsoft Presidio and credentials such as API keys and OAuth tokens with TruffleHog. The study procedure was deemed exempt by Stanford’s Institutional Review Board.
Its authors caution that because it captures only developers who opt into public checkpoint logging, the dataset reflects an early-adopter population, excludes proprietary enterprise codebases, and may not generalize to all coding-agent users; a large fraction of early data came from the logging tool’s own repository, a share they report falling below 20% as adoption grew. The later study that reuses it describes its sessions as public checkpoint logs from developers on GitHub who opted in between January and April 2026, characterises that public availability as reflecting deliberate developer action rather than incidental exposure, and reports redacting personally identifiable information — names, emails, phone numbers and credentials such as API keys and OAuth tokens — from extracted records before analysis.
Use
SWE-chat: Coding Agent Interactions From Real Users in the Wild uses the dataset to characterize real-world coding-agent usage — including a bimodal split between sessions where the agent writes virtually all committed code and sessions where the human writes all of it — and failure modes such as low code survival and frequent user pushback. Its authors also propose it as material for benchmarks grounded in real workflows and for training user simulators for offline evaluation.
How Coding Agents Fail Their Users: A Large-Scale Analysis of Developer-Agent Misalignment in 20,574 Real-World Sessions combines SWE-chat with a re-crawl of SpecStory exports to assemble 20,574 sessions from 1,639 repositories, having first verified that the two datasets contain no overlapping repositories. That study’s CLI-versus-IDE contrasts — notably that CLI sessions are more prone to constraint violations and to damage reaching project and external state — draw on SWE-chat for the CLI side. Because the CLI group in that analysis combines sessions from both datasets, the authors re-ran the comparison within the SpecStory data alone and report that all directions held, though the absolute values differed.