Skip to main content

Guides

How to Use the Agent Flow Graph in Power Claude

Flow Explorer Path on map with SessionStart and Stop nodes and the Detail inspector
Recorded on the live try-it demo: Design map, What ran, Path on map, then Simple Flow — same chrome as the extension.

JSONL is storage. The graph is how you read a run. Power Claude’s Flow Explorer has three modes — Design map, What ran, Path on map — plus Simple Flow for humans and Graph for the DAG. This is the click-by-click walkthrough, using the same UI that ships in the extension and the free browser demo.

Start here (no install)

Open the harness-aware demo. Synthetic data only. Real sessions stay on your machine.

If the demo fails to load CSS or the map, that is a product bug, not “how graphs work.” The walkthrough below matches the live Harness Flow chrome: mode pills on the top row, view pills on the second row, Detail on the right.

The three modes (do not skip this)

The top segment is the whole product:

PillMode idQuestion it answers
1 · Design mapblueprintWhat can the harness run? Full pipeline of gates and hooks.
2 · What ransessionWhat did this session do? Prompts, tools, tokens, evidence.
3 · Path on mapcombinedWhat ran, drawn on the design pipeline. Dimmed = never reached.
A common mistake: staying on Design map and wondering where your prompts are. They live in **What ran**. The join is **Path on map**.

View row: Simple Flow vs Graph

Under the modes:

ViewWhen
Simple FlowFirst read. 7–12 human stages (“Start the agent”, “Prepare the workspace”, “Check harness health”).
GraphInteractive DAG. Click a node, pan, zoom.
ListSame steps as a list when the canvas is noisy.
DiagramMermaid of the current zoom/filters.
FilesRead / write / modified paths per step.
TimingDuration waterfall from known start/end only.
**More charts…** hides Sankey, Treemap, Heatmap, Sequence, Gantt, Network, Stats, Hooks, Quality. Those need session data. They stay empty on a pure design map. That is honest, not broken.

Click-by-click: learn a run in five minutes

1. Path on map + Simple Flow

  1. Click 3 · Path on map.
  2. Click Simple Flow.
  3. Read the stage list. Stages marked ACTIVE have evidence in this run.
  4. Open Show details on a stage, then Open technical graph when you need the underlying nodes.

This is the operator view. Use it in a postmortem before you drown in tools.

2. Path on map + Graph

  1. Stay on 3 · Path on map.
  2. Click Graph.
  3. Press ▶ Start (or Go to start on the how-to card).
  4. Click SessionStart (or the first prompt node, depending on zoom).
  5. Read Detail on the right: WHAT, WHY THIS PATH, PROCESS CHAIN, DECISIONS & GATES.

On a Combined overlay you should see executed events (for example SessionStart) and the Stop end. Dimmed design nodes were not taken.

3. What ran + Graph

  1. Click 2 · What ran.
  2. Stay on Graph.
  3. You should see session-scoped nodes: prompts, tools, sub-agents.
  4. Click a Prompt diamond.
  5. Switch the inspector to Evidence for tool I/O when the node is a tool turn.
  6. Open Taken path and press Play to walk chronological legs.

If badges say snapshots missing or tokens missing, believe them. The UI does not invent token counts.

4. Design map (pipeline only)

  1. Click 1 · Design map.
  2. Raise Semantic zoom (Level 3–5) to see every gate.
  3. Filter Show kinds: Gates, Agents, Matchers.
  4. This is the same picture as the Harness Flow Graph article — interactive.

5. Export

  • Copy export — Mermaid or JSON of the current view.
  • Copy link — deep link with mode, view, zoom, selection.
  • CLI parity: pc tree --format mermaid for the session tree.

Paste Mermaid into GitHub, Notion, or mermaid.live. Do not redraw the run by hand.

Controls that matter

ControlDoes
▶ StartJump to the flow entry.
FitFit the graph in the viewport (f).
ExpandReveal nested tools and sub-agents (e).
Focus contextIsolate the selected branch (Enter).
HeatColor by token weight when tokens exist.
ErrorsShow only error / blocked nodes.
TourGuided Start → path → Evidence. Safe to re-run.
Harness guideIn-product help for Blueprint vs Session vs Combined.
Semantic zoom: lower = big picture, higher = tools and files. Only levels with data for the current mode are populated.

Browser demo vs the extension

Try-it demoVS Code / code-server
DataSynthetic scenarios (harness-aware, simple edit, failed tool, fan-out)Your local transcripts
Badge“Mock demo”Live session source
PrivacyNothing of yours is uploadedGraphs stay on the machine
Pick **Harness-aware run** in the demo dropdown. It is the scenario that shows Combined (session path on the pipeline).

What not to do

  • Do not treat the old marketing still of four rounded boxes as the product. That was a 7-second HTML card. The real UI is Design / What ran / Path on map.
  • Do not caption “tokens: 12k” unless the inspector shows it.
  • Do not call Path on map “the harness graph.” The harness graph is Design map. Path on map is the overlay.

Frequently asked questions

The graph looks empty.

You are probably on Design map at zoom 1 with filters that hide everything, or What ran on a session with no indexable structure. Press Fit, set zoom to Level 2, turn all kinds on, then ▶ Start.

Simple Flow says Loading…

Give the payload a second, or switch to Graph and back. If it stays empty, the fixture or session has no simple-flow projection. That is a data problem, not a missing feature.

Is this Claude Code’s session history?

No. Claude Code stores transcripts. Flow Explorer projects structure (agents, tools, edges, gates) and exports Mermaid.

Does export send data to Neural-LLM?

No.

Can I deep-link a graph?

Yes. Copy link encodes mode, view, zoom, and selection. The public demo also accepts ?scenario=harness-aware.

Why do I see “Trace: degraded”?

Integrity is shown honestly. Degraded or compromised is never upgraded to ok. Fix capture; do not hide the pill.

Relation to messaging?

Messaging coordinates live sessions. Flow Explorer explains what one session did. Use both.

Related

Independent third-party software. Not affiliated with Anthropic.