Why?

If you've ever opened an AI agent's chat-history .jsonl file in a plain editor, you've seen this:

{"id":"chatcmpl-abc","parentId":"...","timestamp":1782133268740,"type":"function_call",
"providerData":{"messageId":"…","model":"…","argumentsDisplayText":"subagent_type: …"},
"callId":"tooluse_x9fHWvn496nS0JPI4QYdiB","name":"Read","arguments":"{\"file_path\": …}",
"sessionId":"…","cwd":"…"}

…one giant unreadable line per message, sub-agent calls scattered across sibling files, no way to tell which assistant turn produced which tool call.

Chat History Viewer parses these files, normalizes them into a uniform conversation model, and renders them as messages — with markdown, tool arguments, sub-agent links, and a timeline view for multi-session debugging.

Features

🧵 Single-session view

• Per-message inspection: type · timestamp · id · role · tool name · tool arguments — all visible at a glance, click an id to copy.
• Markdown rendering for assistant text and tool results, with auto-collapse of long blocks and an inline expand/collapse toggle.
• Bulk selection & hide: tick messages, hide them to declutter, then Show hidden / Restore all to bring them back.
• Type filter chips (USER, ASSISTANT, TOOL CALL, TOOL RESULT, SNAPSHOT, OTHER) and a free-text search across id / text / tool name / arguments.
• Search with context: when you find a match, ±N surrounding messages are kept visible so you can read the up- and down-stream conversation without losing the filter. Adjustable from "match only" up to "all".
• Raw JSON drawer on every message for when you really need to dig in.
• Full IME support — Chinese / Japanese / Korean input works correctly in the filter box.

🔗 Sub-agent linking

When an Agent tool invokes a sub-agent, the viewer detects the link via three independent channels and resolves the sub-agent's .jsonl file. Click ▶ Open sub-agent on the tool call and the sub-session opens as a new tab, with breadcrumbs back to the parent.

📊 Side-by-side compare mode

Open N sessions as columns sharing a single time axis. Messages from every session interleave in real chronological order, with idle gaps collapsed into compact ↕ gap 12m separators so you can see who was doing what while another session waited.

   time      col1: main        col2: planner    col3: vision-reader
   10:01:23  user: 帮我...
   10:01:25  assistant: ok…
   10:01:30  ▶ Agent ─────────▶  …                ·
   10:01:42                       user: …            ·
   10:01:55                       assistant: …       ·
   ↕ gap 4m
   10:05:31  tool_result        ·                 user: 截图描述
   10:05:48  assistant: …       ·                 assistant: …

Two entry points:

• Command palette → Chat History: Compare Multiple Sessions…
• Explorer: select several .jsonl files, right-click → Chat History: Compare as Side-by-Side

Once you're in compare mode, clicking a ▶ Open sub-agent link appends the sub-agent as a new column to the right so you can watch parent and child agents collaborate on the same timeline.

� Supported formats

Out of the box this extension parses CodeBuddy / Claude-Code style JSONL transcripts. Support for additional formats (Cursor sessions, OpenAI Assistants logs, …) is on the roadmap.

Install

1. From a .vsix (recommended for now):

   code --install-extension chat-history-viewer-<version>.vsix
   

2. Or, search "Chat History Viewer" in the VS Code Extensions view (publisher: devoszhang).

Getting started

Open a single session

Three ways, pick whichever is closest to where you are:

Compare multiple sessions

• Command palette: Chat History: Compare Multiple Sessions…, then multi-select the files in the dialog.
• Or, multi-select .jsonl files in the Explorer (Ctrl/Cmd + click), right-click, Chat History: Compare as Side-by-Side.

File layout convention

The default parser expects sub-agent files to live next to the parent:

<dir>/<session>.jsonl
<dir>/<session>/subagents/agent-XXXX.jsonl

Sub-agent references are detected via three independent channels (most specific wins):

1. providerData.toolResult.subAgent.sessionId on a function_call_result
2. Trailing [Agent ID: agent-xxxx] text inside the tool output
3. Agent tool calls with arguments.subagent_type (provides the label; id is resolved from the matching result)

Configuration

FAQ

Q. My .jsonl file is huge (10MB+). Will this hang VS Code? No — the parser is streaming-friendly and the webview keeps the original records in memory but only renders visible messages. Files in the 100k-line range render in well under a second.

Q. Can I edit messages in this view? No, this is a read-only viewer. The original .jsonl is never written to. Use open source in the toolbar to jump to the raw file if you need to edit.

Q. The Chinese / Japanese / Korean filename in the tab is showing %E9%9C%80...? Fixed in 0.1.0 — file basenames are URL-decoded before being displayed.

Q. Does compare mode work with sub-agents? Yes, and it's where compare mode shines. Open a parent session, switch to compare view via the tab bar's ⇆ button, then click any ▶ Open sub-agent to append it as a new column.

Roadmap

• Proportional timeline mode (row height ∝ real time elapsed)
• Column-width drag handles in compare mode
• Export filtered conversation to Markdown
• Built-in parsers for Cursor and OpenAI Assistants logs

License

Released under the MIT License. See the LICENSE file shipped with the extension for the full text.

© devoszhang