Labels

Log Book works without a model. Rules label what they can on every sync, on your computer; optional model labels add what a conversation was for, how it ended, and how you and the agent reacted to each other. A model run never starts on its own.

What works without a model

On every sync, rules label:

• what a shell call was for, such as running tests or installing packages;
• why a failed call failed, for the tools other than the shell;
• whether the agent recovered after a failed call;
• who started a session: you, or an agent.

So failures, retry loops, slow calls, timing, tokens and every conversation work from the first sync. The cards that need a model say so until you label:

The dashboard before any model labels: the cards that need them say so.

What labels add

A run has six tasks, each feeding its cards:

Task	Labels	Feeds
shell	What each shell call was for, and why a failed one failed	Tool problems, Unusually slow calls, Calls that went wrong
tool-failure	Why a failed call of another tool failed, where the rules could not tell	Tool problems, Calls that went wrong
session	What each conversation was for	Conversations by goal, Tokens per conversation, Time per turn, Time per conversation
outcome	How each conversation ended, with a note	Finished, Didn't finish
prompt	What each of your prompts did, and how you reacted to the agent	Reactions to the agent
reply	How the agent's last reply before your next prompt treated you	Reactions from the agent

A conversation's goal is one of the goals a model chooses from, such as build a feature, fix a bug, review and explore the codebase. Its outcome is done, partly done, handed off, blocked, failed, abandoned, no task or unclear.

Labelled, the reaction cards show how you and the agent dealt with each other, week by week:

Your corrections, pushback and praise, as shares of your prompts.

How often each model's replies asked permission, gave in, or pushed back.

Preview, then start a run

See what a run would send before it sends anything. The plan prints the model, the records per task and per agent, and an estimate of the tokens; a preview prints the exact text of a task's next batches:

logbook labels plan
logbook labels preview --task shell

Then start it, in a terminal or with Label in the top bar, which opens a page with the same plan and a Start labelling button:

logbook labels update

A run labels through your own Claude Code: it starts claude -p once per batch, signed in as you, so the requests count against your Claude plan like your own sessions. Before it starts, it says which claude it uses, how it is signed in and which model it runs on. When ANTHROPIC_API_KEY is set, it warns that Claude Code may bill the run to that key instead of your plan. Log Book does not support local models.

Privacy and security says exactly what each request holds.

Choose a model

One model labels a whole run: the one you pass with --model, or Claude Haiku (claude-haiku-4-5) when you pass none. Haiku is the cheapest, and its labels are less reliable:

Claude model id for every task of this run (default claude-haiku-4-5; pass a full id, not an alias). On a fixed 100-prompt sample Haiku agreed with Sonnet on 62% of prompt acts and of reaction sets and missed most redirects and praise.

So with Haiku, the praise line of Reactions to the agent reads lower than Sonnet's labels would make it. To choose another model, pass its full id, not an alias such as sonnet:

logbook labels update --model claude-sonnet-5-5

Log Book remembers no choice. A later logbook labels update with another model labels only the records no model has labelled yet.

What a run costs

Measured on one warehouse with Claude Code 2.1.286; treat the figures as examples, not estimates for your data. Claude Code adds about 440 input tokens of its own framing to every call. The calls count against your Claude plan like your own sessions.

Task	Measured with	Calls per 1,000 records	Input tokens per 1,000 records	Output tokens per 1,000 records	Time per 1,000 records (4 at once)
shell	claude-haiku-4-5	40	150,000 to 250,000	7,000	under a minute
tool-failure	claude-haiku-4-5	1 per 25 records the rules leave open (about 1 failure in 200)	about 330 per record	small	seconds
session	claude-sonnet-5-5	50	530,000	27,000	about a minute and a half
outcome	claude-sonnet-5-5	50	510,000	not measured	about a minute and a half
prompt	claude-sonnet-5-5	125	1,800,000	40,000	about 3.5 minutes
reply	claude-sonnet-5-5	125	1,700,000	not measured	about 2.5 minutes

The Sonnet rows have no Haiku figure, because a token count depends on the model's tokenizer.

Stop and resume

A run keeps every batch it finished, however it stops: Ctrl+C, Stop on the labelling page, stopping logbook, your Claude usage limit, or a killed process. Run the same command again to continue where it stopped.

A session's outcome and its last reply wait until the session has been quiet for an hour, so a run does not label a conversation that is still going.

Compare two models

To see how another model labels what you already have, label one task again with it, compare the two, and drop the labels you do not keep:

logbook labels run --task prompt --model claude-sonnet-5-5
logbook labels compare --task prompt --first claude-haiku-4-5 --second claude-sonnet-5-5
logbook labels drop --task prompt --labeller claude-haiku-4-5

logbook labels compare prints how often the two agree, field by field, and their most common disagreements. Reports read each record's newest model label, so drop the labeller you did not choose.