Download free

The Steps view

See every step the agent took, search them and check the chat's performance.

Updated · version 2.0.240 · Markdown

«Steps» shows everything the agent did in the chat, step by step: what it thought, which tool it used and what it got.

Steps

  1. Above the conversation, press «Steps». To go back, press «Chat».
  2. At the top you see how many events there are, such as «24 events».
  3. Type in «Search the steps» to find one.
  4. Use the «Timeline overview» to jump to a moment.
  5. Press a row to open that step whole. Each row carries only what you need to recognise it: its kind, its title, how it ended, when and how long it took, and a short line of what it got. Pressing it opens a window with the step's title and, below, «Input», «Output» and «Raw event»; while the app reads the whole step, each part says «Reading…». «Copy raw» copies the event and «Close» goes back to the list.

Because only the step you open is read whole, a long history draws and moves without stalling, and earlier pages also arrive with short rows. «Search the steps» looks in what the row shows (kind, title, state, moment and that short line), ignoring capitals and accents, and every word you type must be in the row; it does not look inside the full input or output of a step you did not open.

The task record

Beside the number of events, the «Task record» button opens what the app keeps of each of the chat's tasks: what you asked, how it ended, the files, the commands and tests, the errors and what fixed them. It is what the model carries when a long conversation is compacted. See Context, compaction and the task record.

The line while it works

From the moment you press Enter, a line under the chat says in plain words what the agent is doing right now and, when something takes a while, how long it has been. It is not made-up text: it comes from what the app is doing.

  • Before the model starts. The app prepares the turn and the line says so: «Reading your message», «Checking what you have open», «Running the commands you set up», «Choosing how to work on this», «Shortening the conversation» or «Getting your tools ready». They usually pass in under a second; if one takes longer, the line shows how long.
  • While the model works. It says «Thinking» with moving dots and then what it does: the command it runs, the page it opens, the file it reads. A step that usually takes a minute or more on your Mac says so while it runs («This usually takes about 4 min»).
  • When the wait is on you. The line says so at once, with who is waiting and what it needs: an approval, a question, a plan to review or a browser page it handed to you. «Review» and «Show the page» take you to where you act. For a handed page the line also says how much time is left (ten minutes) and what to do: do the step there and press «Ready», or «I can’t» if you can’t.

How a turn went

When a turn ends, the line above its answer says how long it worked, how many steps it took and what it did, counted by kind, such as «Worked for 3 min 55 s · 9 steps · 3 edits · 5 commands · 2 lookups». (A turn whose record does not say the kinds shows the total instead, such as «23 actions».) A stopped turn says «Stopped after …» and a failed one «Failed after …». «Steps» has each one of those steps.

Open what a turn did

Press that line («Show what it did») and it opens in place, under the line, into what the turn did, in groups: «Read files», «Searched the files», «Edited files», «Ran commands», «Searched the web», «Used the browser», «Worked with other agents», «Asked for approval», «Thought it through» and so on; the calls of a connected program are grouped under its name. Press a group to see its steps. A group with a failed step says «1 failed» on its closed line, and a turn that ended in an error shows «Stopped before it finished» with the error. Press the line again («Hide what it did») to close it.

  • How much to show («Groups», «Steps», «Everything»): only the kinds of work; every step, one line each; or every step with what it returned (the output of a command included). Your choice is kept on this Mac.
  • «Hide the thinking»: the eye hides the agent's reasoning from this list.
  • A turn without tools says «It did not use any tools in this turn.» and a turn older than the kept record says «The details of this turn are no longer kept.»

The changed files of a turn

When a turn changed files, a card ends it with «Changed 3 files», each file with its added and removed lines and a «new» tag on the ones it created. Press a file to see that turn's change; the icon «Open in the changes» opens it in the Changes panel, and the undo icon («Undo this turn’s changes») puts back just that turn's files. See Go back and undo files.

When the service fails and the agent goes on

If the model's service fails in the middle of an answer, the Steps view shows a line such as «Retry 2 · went on from what was already written»: the agent continued from where the answer stopped. See Stop, steer or queue.

When a step makes no progress

If a tool gives no sign of life for longer than that step usually takes, the app cuts it cleanly and the turn goes on from its last good point: the model knows the step was cut and that how it ended is unknown, and decides from there. How long a step usually takes is learned by the app from its last runs on your Mac; a step that has not yet run twice is not watched, and neither is one that waits for your answer or another agent's turn. The app never repeats the step on its own, because a click or a send could happen twice.

  • The line under the chat says it in plain words: «I restarted … at …: its step» and, at the end, «… had made no progress for … min».
  • In «Show what it did», the group «Restarted» holds the step with the time and the minutes; in «Steps» it is a row of the kind «Watchdog».
  • This happens only once per turn. If another step makes no progress in the same turn, the app stops the turn and says so: the line starts with «I stopped … at …» and the chat says «I stopped this turn at …: a second step (…) had made no progress for … min after I had already restarted it once. What it was doing is unfinished; tell me if I should go on.»

One review pass at the end

When a turn that codes ends, the app gathers everything it has to tell the agent (errors the code left, what it said it checked, open items on its list) into a single notice, and the agent fixes it all at once. Before, there could be several notices in a row, and each one went over the whole conversation again: fewer rounds, so the turn finishes sooner and costs less. What it did stays in view in «Show what it did».

The reasoning

In the chat, the agent's reasoning appears collapsed. Open it if you want to see how it thought.

The chat's performance

Under the message box, the token bar shows the chat's tokens and the «Conversation space used» (Your balance, your tokens and the status strip). Press the tokens to open «Chat performance», which details «Total tokens», «Estimated cost», «Model time», «Tool time», «Turns» and «Cache hit». They are exact measurements: missing values are not estimated. The context percentage opens the «Context» window.