v0.3.0 — turnkey install, entry wizard, health panel, jump-to-source

Self-hosted LangGraph debugger for VS Code. Install from the Marketplace (search
"GraphLoupe") or the attached .vsix (Extensions ▸ ⋯ ▸ Install from VSIX…), then point it
at a compiled graph. Full docs: https://bygao.github.io/GraphLoupe/

Highlights since v0.2.0

• Turnkey Python setup — resolves your interpreter (Python-extension selection → PATH), runs the sidecar in its own managed environment and your graph in your interpreter, and a preflight doctor gives actionable errors. Fixes Windows / venv install.
• Manual entry wizard — Select Graph ▸ Specify manually… connects a graph AST discovery can't find (including one exposed as a compiled module variable), generating a .graphloupe/entry.py adapter — no hand-editing settings.json.
• Health panel — a new Health tab shows what's wired up (checkpointer, input schema, LLM nodes, node→source) and what isn't, with the GraphLoupe / langgraph versions and the interpreter running your graph.
• Jump to source — click a node's ↗ to open its function in your editor, beside the graph — even for langgraph-wrapped nodes.
• Reliability — generous graph-load timeout for cold langchain/langgraph imports, clearer doctor / error messages, dismissable error banner.

Notes

• Step debugging, state diff, and time travel need compile(checkpointer=…).
• No account, no license, no cloud — the IDE never imports your graph (isolated worker + frozen protocol).

v0.2.0 — time travel, token economy, clean abort

Self-hosted LangGraph debugger for VS Code. Install the attached .vsix
(Extensions ▸ ⋯ ▸ Install from VSIX…), then point it at a compiled graph.

Highlights since v0.1.0

• ⏱ Time travel — every pause shows the run's checkpoint timeline (the live lineage). Click any earlier checkpoint to rewind and re-run from it.
• Token economy — per-node prompt/completion counts and a run total; expand a node to see the actual prompt/response.
• Clean abort — ■ Stop cancels the run mid-stream without reloading the graph.
• Manual inference — pause a node, copy the rendered prompt into any chat (Copilot/ChatGPT/Claude), paste the answer back, resume.
• Node kinds — ⚡ llm / ✋ manual / script badges, ELK orthogonal layout, branch + loop edge labels.
• Inspector layout — Input / State / Tokens / Manual tabs; read-only locked canvas.

Notes

• Step debugging, state diff, and time travel need compile(checkpointer=…).
• No account, no license, no cloud — the IDE never imports your graph (isolated worker + frozen protocol).

v0.1.0 — early preview

GraphLoupe is a self-hosted LangGraph debugger for VS Code: step through a graph
at the node level, inspect & diff state between nodes, time-travel from a checkpoint,
count tokens per LLM node, and pause a node to paste an answer from any chat
(Copilot/ChatGPT/Claude) back in as the model's output.

Early preview — runs from source. This is a working tool, but not a one-click
install yet: the VS Code extension spawns a Python sidecar, so you build it and run
it from an Extension Development Host (see the README Quick start). No Marketplace
.vsix in this release.

What works today

• ✅ Graph visualization — ELK orthogonal auto-layout, nodes coloured by kind, conditional branches & loops drawn distinctly, active node highlights.
• ✅ Step debugging — breakpoints, state snapshot + diff, step, time-travel fork (needs compile(checkpointer=…)).
• ✅ Manual inference — interrupt → copy prompt → paste into any chat → paste answer back → resume (text + tool_call).
• ✅ Token economy panel — per-node prompt/completion + run total (exact when the model reports usage, else a flagged estimate).

Not yet (roadmap)

• Copilot auto-path via vscode.lm (today: manual paste-back).
• A sandbox for untrusted graph code — see security note below.
• A Marketplace .vsix / bundled Python bootstrap.

Run it

Clone, npm install, npm run build, open in VS Code, press F5 — full steps and a feature-showcase graph in the README.

Security & privacy

100% local, no telemetry, no phone-home; discovery is AST-only (never runs your code); execution is in an isolated sidecar subprocess. Process isolation guards against buggy/runaway graphs, not deliberately malicious code — point it at graphs you'd run yourself. Full threat model: SECURITY.md.

Verified: CI green (quality gate L0 + L1 — lint, types, security scan, PIN tests, TS↔Py contract round-trip).