Tour the product

Log Book's page goes from a pattern to its evidence: the dashboard shows where time and failures pile up, every number opens the steps or conversations behind it, and a conversation shows the exact turn where it went wrong.

Spot a pattern on the dashboard

The dashboard over 30 days of the demo's data.

The top bar sets the range every page shows: Today, 7 days (where Log Book starts), 30 days, All, or a Custom span of days. The four figures at the top compare the range with the one before it.

The cards fall into three groups:

• Failures: Tool problems groups failed calls by cause and says when a cause is rising. Unusually slow calls lists calls that took over 10 times their tool's usual time and over 30 seconds, and Retry loops the same call tried again after failing. Calls that went wrong leaves out real results, such as failing tests or a search that found nothing, since the tool did its job.
• Time and context: Time per turn and Time per conversation leave out the time an agent waited on you, so the gap between active and total time is your own. Tokens by tool shows what each tool puts in the context.
• Labelled patterns: Didn't finish, Conversations by goal, Tokens per conversation and the two reaction cards read labels a model writes. Until you label, they say so and the rest of the dashboard works as usual.

Open the evidence behind a number

A number that counts calls opens the steps behind it: a cause in Tool problems its failed steps, a row of Unusually slow calls that tool's slow calls, and Calls that went wrong every failed step.

Following the most common cause from the Tool problems card: each failed step, what it ran, and how long the tool usually takes.

A step opens its conversation at that turn, with the step open in the Turn pane. Numbers that count conversations open the conversation list with its filter filled in, so you can change it from there.

Find a conversation

Conversations lists every conversation in the range, latest message first. Type words to search what was said and the titles, and add fields to narrow it:

Field	Matches	Example
goal:	The goal a model labelled	goal:"fix a bug"
outcome:	The outcome a model labelled	outcome:done
by:	Who started it: me, agent or script	by:me
harness:	The agent tool, claude or opencode	harness:opencode
agent:	The agent's name	agent:build
model:	Part of a model's name	model:opus
project:	The project's directory, or its last part	project:shop
tool:	A tool it called	tool:Edit
cause:	The cause of a failed call	cause:"Edit didn't match the file"
has:	failures, retry, correction or spawned	has:retry

A value with a space takes quotes, commas give alternatives (outcome:blocked,failed,abandoned), and different fields must all match. To find the bug fixes in one project that hit failures: project:billing goal:"fix a bug" has:failures. The buttons under the filter add the common ones with one click.

Every conversation in the range, with its goal and outcome once a model has labelled it.

Labels lists every goal and outcome a model chooses from.

Read what happened in one conversation

A conversation from its top: the session map, the thread and the Turn pane.

The header separates how long the conversation took from how long the agent was active, and counts its failed steps, the agents it spawned and its compactions. Plugins lists the tools each MCP server or plugin offered, which were used, and what their definitions added to the context.

The Session map draws the whole conversation as one strip, with marks for failed steps, your reactions, compactions and spawned agents, so you can see where to look in a long run. Switch it between by time and by turn, and click or drag on it to jump. Under each prompt, the reactions a model labelled mark how you answered the agent: a correction, pushback, a clarification, a redirect, teaching or praise.

Inspect a turn

The Turn pane splits the turn's time into Model, Tools and Waiting, shows what filled its Context, and lists its steps as a tree. Open a step to see its input and output, how long the tool usually takes, and how many steps like it the range holds.

A turn whose test run failed three times before it passed.

As you scroll, the map and the Turn pane follow the turn you are reading, and the address bar keeps it, so a link you copy reopens that turn.

Key	Moves to
j, k	The next or the previous turn
J, K	The next or the previous turn with a failed step
n, N	The next or the previous failed step
gg, G	The first or the last turn
[, ]	Collapse or expand the spawned agents
Esc	Close the open step

See what your tools cost in context

Tokens by tool shows which tools and skills put the most text into your agents' context. The numbers are estimates, at about four characters a token, from the text of each call.

Tool context, grouped by plugin, then every tool and skill, largest first.

By plugin sums the tools by where they come from: each plugin's server, the agent's own tools under Built in, then Skills. The list below separates a tool's Definition, which its name, description and input schema add to every request that carries the tool whether or not the agent calls it, from what its calls put in the context. Definition shows the largest one conversation in the range recorded loading, or none when no session recorded it.