Claude Agent View

A VS Code sidebar that shows what Claude Code's workflows and subagents are doing right now — phases, per-agent state, tokens and the tool each agent is holding.

Not affiliated with or endorsed by Anthropic. It reads Claude Code's on-disk session files, which are an internal layout rather than a published interface, so a Claude Code release can change or break what this shows. It only ever reads: nothing here writes to those files, and the worst case is an empty view.

The panel

The layout follows Claude Code on the web's "Background tasks" panel, measured from the live page rather than eyeballed:

Every surface is color-mixed from --vscode-foreground, so the panel follows whatever VS Code theme is active instead of pinning a light one.

Three values were measured, then changed, because Claude's own are tuned for its own backgrounds and did not survive the move:

• the 20% done cell measured 1.40:1 on a light ground and 1.64:1 on a dark one, well under the 3:1 a meaningful non-text mark needs. It is 62% / 52% now, reaching 3.34 and 3.56.
• the amber #fab219 reads at 1.46:1 on light, so light uses #9c6606 (3.88).
• the web panel tints a running phase's title with its warning colour. Lightened for a dark theme that lands on the stalled amber, and it put one state in two colours a row apart, so a running phase is blue here — the same blue as its cells.

Every state mark now clears 3:1 in both themes. Muted label text reaches 4.06–4.42 on light rather than 4.5: --vscode-foreground is #3b3b3b in Light Modern, so no alpha of it gets there. Body text is 5.2+.

Why

Claude Code's /workflows browser is a terminal-only view: its slash command is registered as a terminal-rendered command with no equivalent for the VS Code extension, so there is nothing to draw there. The extension's own progress UI tracks Task-tool subagents but not workflow runs — a running workflow reaches it only as the "session is busy" spinner, with no detail behind it.

So this reads the same run data from disk instead.

Where the data comes from

Everything is read from ~/.claude/projects/<project>/<session>/. Nothing is written, and Claude Code's own configuration is left untouched.

Which session a run or a subagent came from is read from ~/.claude/sessions/<pid>.json, keyed by sessionId. Its name is the label Claude Code's Rename edits (nameSource says whether it was renamed or generated), and its pid plus procStart say whether that session is still running — matching the process start time as well as the pid, because pids are recycled. Sessions with no record, such as a headless -p run, fall back to a short session id.

A live run records neither the model nor a per-agent duration in its metadata — both are only in the end-of-run snapshot. They are recovered from the transcript instead: assistant messages carry model, and the first and last timestamps bracket the agent's runtime.

The snapshot is the richest source but it is written once, when the run reaches a terminal state — polling confirmed a single write across a whole run. So:

• Finished runs are rendered from the snapshot. One small JSON per run instead of every transcript, which took a cold scan of 49 runs from 33s to 0.6s.
• Running runs are rendered from the journal plus incremental reads of each agent transcript.

Reading a transcript

Clicking an agent opens a reading view rather than its raw .jsonl: the prompt, the assistant's replies, and each tool call as a row showing the tool and what it was actually asked to do — the command for Bash, the path for Read, the pattern for Grep. A row opens to its input and result; string fields are shown as themselves, so a shell command reads as a shell command instead of one line of escaped JSON. Failed calls are outlined and marked. "Open raw JSONL" is still there for when the raw record is what you want.

Long transcripts reach tens of megabytes, so the view caps at 500 steps and clips very long inputs and results, saying so where it does.

Agent state

A run with no snapshot whose files have been quiet for ten minutes is reported as abandoned rather than running — that is a session that died mid-run.

A run whose agents are all done still reports as running until its snapshot lands. That is not a lag in this view: the script itself is still finishing, and the run is not over until Claude Code says it is.

Performance

Transcripts reach several megabytes, so they are never re-read whole. Each file keeps a byte offset and only newly appended bytes are parsed. A transcript that was already cold when the extension first saw it is primed from its last 64 KB, which is enough to decide the state; its token and tool totals are reported as unknown rather than guessed. Steady-state scans measured 1–17 ms.

That leaves blanks in a run that was already going when the view opened, so those agents are queued and re-read in full, four per scan. Measured on a live 29-agent run: first paint 181 ms, every total filled by the fifth scan.

Settings

Install

npx @vscode/vsce package --allow-missing-repository --skip-license
code --install-extension claude-agent-view-0.1.0.vsix

Then reload the window. The view lives in the activity bar; a status bar item appears while any agent is running.

Caveats

Every path and file shape above is Claude Code's internal on-disk layout, not a published interface. It was read from version 2.1.266 and can change in any release. Nothing here writes to those files, so the worst case is an empty view.

The extension has to run wherever Claude Code runs. If you use Claude Code inside WSL or a container, install it into that remote, not the local UI side — otherwise it has no ~/.claude to read.