Unroll 0.9.1

Unroll 0.9.1

A documentation release; the extension's behavior is unchanged from 0.9.0.

• The Marketplace listing, README and website now describe Unroll for reading eval runs and trajectory datasets: selecting runs by outcome or any field, checking each step against its raw record, opening large files, reading files on remote machines, and data handling.
• The listing is now named Unroll: Agent Trace and Eval Run Viewer.
• The guide adds notes on remote machines and unencrypted HTTP endpoints, and explains how investigation estimates use your own labels.
• The light-theme dataset screenshot now shows the resolved is false condition.

Install from the Visual Studio Marketplace (mhermon.unroll). The attached VSIX is the same package, for offline installs.

Unroll 0.9.0

Unroll 0.9.0

Unroll is now on the Visual Studio Marketplace as mhermon.unroll. This VSIX is the same package, for offline installs.

Highlights:

• Select research cases: narrow a dataset with conditions on any field, including nested metadata. Fields list what they hold, value suggestions count the whole file, and only conditions that fit a field are offered.
• Smoother reading: resize the run list, its conditions and the investigation panel; scroll continuously across loaded messages in both directions; continue and resume long conversation searches.
• Investigations (experimental, off by default): ask a language model you choose about a trace or dataset. The model reads the file through read-only tools, cites the steps it used and may ask you to label examples. Turn them on with Unroll: Enable Investigations (Experimental)…. They run only in trusted workspaces, their settings can be set only in your user settings, and until they're on Unroll doesn't load any model-provider code.
• Refreshed site, screenshots and walkthrough.

Reading traces stays fully offline, with no account, API key or telemetry.

Validation included 578 Node tests, 72 browser tests, TypeScript checks, VSIX content and notice validation, and the native extension suite on Linux, macOS and Windows for VS Code 1.96.4 and stable.

Unroll 0.8.0

Unroll 0.8.0

Unroll now opens saved JSONL and NDJSON directly from disk without a configured file-size cutoff, while keeping windows, indexes, searches, and message reads bounded and cancellable.

Highlights:

• Stackable case filters for Boolean, numeric, text, regex, presence, and typed value-list conditions.
• Read filtered messages in context and return to the same result and reading position.
• Large messages retain their role and decoded preview, then open in a focused full-text reader with section navigation, section search, copy, pause/resume, and exact Original JSON mode.
• Preview-aware content totals, improved source provenance, tool-call linking, keyboard navigation, lifecycle handling, and recovery from stale or malformed reads.
• Updated documentation, examples, screenshots, and walkthrough.

The VSIX is fully offline and requires no account or API key. Install it from the attached unroll-0.8.0.vsix file or from the VS Code Marketplace when the listing updates.

Validation included 305 Node tests, 46 browser tests, TypeScript checks, VSIX content validation, a Python build with 12 tests, bounded full-message benchmarks, and the native extension suite on Linux, macOS, and Windows for VS Code 1.96.4 and stable. The local macOS Electron runner in this development environment could not start its host process and was not used as release evidence.

Unroll 0.7.0

Unroll now reads files of many traces: one conversation or agent trace per JSONL row or JSON array item, as in agent-trajectory datasets from Hugging Face.

New

• Conversation list. Rows are listed with their id, message and tool-call counts, and their outcome when the row has one (resolved, success, reward, exit_status, label…). Filter by id, row field or message text; rows that matched only in their messages are marked, and opening one highlights the match. Press C to show or hide the list, and Shift+J / Shift+K to move between conversations.
• Row fields. Each row's other fields, such as model, outcome, reward or a generated patch, appear above its conversation. Long values fold away.
• More row shapes. SWE-agent trajectory, OpenHands messages, τ-bench traj, AgentInstruct and Hermes conversations, xLAM query / answers, and agent turns written as thought + action with an environment reply (Agent-SafetyBench, ReAct-style logs). Rows can be bare message arrays, hold a conversation in a field (including content: [[...]]), or be wrapped in one key of a JSON object.
• Better source links. Items in a JSON array link to the line where each starts, instead of the top of the file.

Fixed

• Plain messages were dropped from conversations that mixed them with Claude-style content blocks.
• A model field could turn a dataset row into a single message.

Install

Download unroll-0.7.0.vsix below and choose Extensions → … → Install from VSIX… in VS Code, or run code --install-extension unroll-0.7.0.vsix. Requires desktop VS Code 1.96 or newer. Upgrading from 0.6.0 keeps your settings and keybindings.

Verify the download with SHA256SUMS. Guide: https://mhermon.github.io/unroll/guide.html#datasets

Unroll 0.6.0

RunTranscript is now Unroll: a VS Code extension for reading agent traces and chat logs.

Open Claude Code, Codex and other JSONL conversation files, or chat datasets you download, as a readable conversation. Search messages and tool calls, see each tool call's arguments and result together, open the original JSON record, and jump to the source line.

Install

1. Download unroll-0.6.0.vsix below.
2. In VS Code, open the Extensions view, then … → Install from VSIX…, or run code --install-extension unroll-0.6.0.vsix.
3. Right-click a .jsonl, .ndjson or .json file and choose Unroll: Open as Agent Trace.

Requires desktop VS Code 1.96 or newer.

Upgrading from RunTranscript 0.5.0

The extension ID changed to mhermon.unroll. Uninstall RunTranscript (mhermon.runtranscript) first, because both register the same commands. Command and setting IDs keep the chatview.* prefix, so keybindings and settings carry over.

Changes

• New name and extension ID; the description now covers chat logs and datasets as well as agent traces.
• New website at https://mhermon.github.io/unroll/, with a walkthrough video, guide and two synthetic example files.

Verify the download with SHA256SUMS. Preview release; see the guide for limits.

RunTranscript 0.5.0 — first public preview

Read AI agent traces as conversations inside VS Code. This preview supports common Claude Code, Codex, OpenAI, Anthropic, Gemini and conversation-dataset record shapes.

• Search messages, tool names, arguments and results.
• Follow a tool call to its recorded result and inspect original JSON.
• Jump to a source line beside the preview.
• Follow updates to a growing JSONL trace.
• Use light/dark themes and keyboard navigation.

Homepage and demo · Guide · Report a problem

Download runtranscript-0.5.0.vsix below. In VS Code, choose Extensions → … → Install from VSIX…. Requires VS Code 1.96+. If you installed the earlier local ChatView extension, uninstall it first to avoid duplicate command registrations. The normal text editor remains the default; right-click a trace and choose RunTranscript: Open as Agent Trace.

Marketplace publication is pending publisher account setup. This release can be installed directly from the VSIX.

This is a preview, with a 100 MB default file limit and in-memory parsing. Estimated token totals are not billed usage or request context sizes. The extension has no telemetry or provider connection; processing occurs in the extension host and webview, including a remote host when applicable. Implementation source is private; this repository contains public documentation and compiled binary releases. The distributed extension is MIT licensed.

The SHA256SUMS file verifies the attached VSIX. The video and screenshots use synthetic data from the actual installed extension. Idle gaps in the recording were shortened.