> For the complete documentation index, see [llms.txt](https://docs.publicai.io/publicai-documentation/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.publicai.io/publicai-documentation/publicai-trajector-cli/reasoning-summaries.md). # Reasoning summaries From **Oct 17, 2026 (00:00 UTC)**, a session recorded without reasoning summaries is not rewarded. It is still recorded and uploaded, but it earns nothing. This page is the rule. It says what counts as a session with reasoning summaries, and how to turn them on for each way you run Claude Code. The dashboard, the upload service's messages and `trajector enable` link here. The short version: > **Turn reasoning summaries on before you start a session — in a terminal with `showThinkingSummaries`, in the desktop app with the Thinking transcript view, with `--thinking-display summarized` for `claude -p`, and with `thinking.display: "summarized"` in the Agent SDK.** *** ### What reasoning summaries are Before it answers, the model reasons. Since v2.1.89, Claude Code no longer returns that reasoning as readable text by default: what comes back is empty, apart from an occasional one-line progress note. Turning summaries on makes Claude Code return a summary of the reasoning instead, and that text is what makes a session worth something to the buyer of the data. Turning them on does not change your token usage. ### The rule A session has reasoning summaries when, in its main conversation, **at least half of the model's responses that contain reasoning carry reasoning text**. * **Main conversation only.** Subagents are not counted: whether a subagent's model writes summaries is not something you control. * **At least three responses with reasoning.** A session with fewer is too short to tell, and this rule does not decide it; the other checks on the [Rewards](/publicai-documentation/publicai-trajector-cli/rewards.md) page still apply. * **Only the model's reasoning text counts.** Claude Code's occasional progress notes are not summaries. * **The whole session is judged, and judged again as it grows.** A session that is not rewarded under this rule stays that way, even if later responses carry summaries. * **It applies to sessions first uploaded on or after Oct 17, 2026.** A session that had started uploading before then is not affected by this rule, however long it runs. Two consequences follow from the rule: 1. **Turn it on before you start a session.** Whether a session has summaries is decided over the whole session, so switching them on halfway doesn't rescue the session you are in — the next one counts. 2. **It applies to every way of running Claude Code.** If your setup can't produce reasoning summaries, its sessions aren't rewarded. There is no exemption by how you run Claude Code. *** ### Turn them on #### Claude Code in a terminal The switch is Claude Code's `showThinkingSummaries` setting. Either: * **Run `trajector enable` in the project and answer yes** when it asks about `showThinkingSummaries`. Add `--no-proxy` and `--no-earlier` if the project uses them. This turns the setting on for that project only, in the project's local Claude Code settings, and `trajector disable` puts it back. * **Or set it yourself for every project**, in your user settings file `~/.claude/settings.json`: ```json { "showThinkingSummaries": true } ``` trajector never writes this file. It is your change to make and to undo. Then start a new session. #### Claude desktop app The desktop app does not use `showThinkingSummaries`. Its switch is the **transcript view**: summaries are produced while a session's view is **Thinking** or **Verbose**. * **For every new session:** Settings → **Default transcript view** → **Thinking**. * **For one session:** press **Ctrl+O**, or pick **Transcript view** from the session's menu, and choose **Thinking** before your first message. The app confirms with *"Thinking summaries will appear from the next response."* What it costs you, honestly: * The Thinking view shows the summaries in the transcript, so the transcript gets longer. * Switching a session back to **Normal** stops the summaries about three minutes later. Responses from then on have none, and they count against the session. **If you also use Claude Code in a terminal**, keep `showThinkingSummaries` on there. Whether it is set for a project or in `~/.claude/settings.json`, it does not get in the way of the desktop app's Thinking view. **If the Thinking view shows no summaries**, the desktop app has not made them available to your account, and its sessions cannot have them. Record the sessions you contribute with Claude Code in a terminal instead. #### `claude -p` and the Agent SDK Neither reads `showThinkingSummaries`. Ask for summaries on each run: ```sh claude -p "your prompt" --thinking-display summarized ``` In the Agent SDK, set `display` in the `thinking` option: ```ts query({ prompt: "your prompt", options: { thinking: { type: "adaptive", display: "summarized" } }, }); ``` #### Anywhere else We haven't checked how other ways of running Claude Code, such as IDE extensions or remote sessions, turn summaries on. Run one session, then look at My data: if it shows no summaries, record with Claude Code in a terminal instead. *** ### Check that it worked * **Your dashboard's My data page is the check that counts.** It shows how many of your sessions in the last seven days had reasoning summaries, grouped by how you ran Claude Code, and marks each session. It counts the same way this rule does; sessions too short to tell are listed apart. * **`trajector status`** in the project gives a rough local count, subagents included: ``` 12 of 40 assistant lines carry no reasoning. ``` * **In the session itself**, when most of its reasoning carries no text, trajector tells you at most once a day per project, with the next step for the way you ran it. For example: ``` trajector: no reasoning summaries in this session; run trajector enable, then start a new session ``` It never blocks you. ### When this takes effect * **From Oct 14, 2026 to Oct 17, 2026:** nothing changes for your rewards. Sessions without reasoning summaries are marked on your dashboard, and the upload service's message in `trajector status` says how many of each upload had none, so you can fix your setup in time. * **From Oct 17, 2026:** sessions first uploaded from then on without reasoning summaries are not rewarded. They show as rejected in My data — see [Rewards](/publicai-documentation/publicai-trajector-cli/rewards.md). * **Nothing is retroactive.** Sessions uploaded before Oct 17, 2026, and rewards already credited, are not touched. trajector 0.5.0 and later ask for `showThinkingSummaries` as a required setting, and from Oct 17, 2026 uploads from older versions are refused with a message asking you to upgrade. Its data agreement changes with it, so after upgrading, recording pauses until you run `trajector enable` once (with `--no-proxy` and `--no-earlier` if you use them) to confirm the updated agreement. It asks about this setting at the same time.