Execution Tree

Execution Tree

The Execution Tree is a recursive visualization of everything that happened during a single turn — every sub-agent spawn, every tool call, every thinking block. It’s the main debugging and observability surface in Ptah.

Opening the tree

The tree lives in the right-hand pane of the chat. If it’s collapsed, click the tree icon in the chat header, or click any tool call card or sub-agent chip in the transcript — the tree opens with that node selected.

What a node represents

Each node is one unit of work:

Node type	Meaning
Message	An assistant text message or a user message.
Thinking	A reasoning block. Click to expand and read the model’s internal chain of thought (for providers that expose it).
Tool call	A single tool invocation — Read, Write, Bash, Grep, MCP tool, etc. Shows inputs, outputs, and timing.
Sub-agent	A spawned agent (via ptah_agent_spawn or the SDK agent tool). Expands to reveal that agent’s own tree.
Error	A failed tool call or aborted sub-agent.

What each node shows

Hover or click to see:

• Duration (ms).
• Tokens in, tokens out, cached tokens.
• USD cost for the node.
• Model and provider used.
• Full raw inputs and outputs (JSON view toggle).

Recursive by design

Sub-agents can spawn their own sub-agents. The tree renders that hierarchy naturally — collapse any branch to focus on the parts you care about. Depth is unlimited, but Ptah warns you if a branch goes deeper than four levels (usually a sign of runaway delegation).

Tip

The tree is the fastest way to answer “why did this turn cost so much?” or “which file did the agent actually read?” — every decision is visible.

Exporting a tree

From the tree header menu:

• Copy as JSON — the full tree as structured data for scripting.
• Copy as Markdown — a human-readable transcript with indentation.
• Save to file — writes .ptah/traces/<session>-<turn>.json in your workspace.

These exports are what the /orchestrate agents use to hand off work between specialists.