mirror of
https://github.com/langchain-ai/deepagents.git
synced 2026-08-28 05:00:04 -04:00
Mason Daugherty
457ac435e1
Related: #1117 - add a default-off `display.show_reasoning` preference with env, TOML, and `--show-reasoning` support - render canonical provider-visible reasoning separately in headless stderr and collapsible TUI rows - preserve reasoning/text ordering across streaming, virtualization, and checkpoint restore while suppressing nested and opaque reasoning ## Turning it on Off by default. Three ways in, all resolved at launch: | Surface | Value | | --- | --- | | CLI | `--show-reasoning` | | Env | `DEEPAGENTS_CODE_SHOW_REASONING=1` | | config.toml | `[ui] show_reasoning = true` | The CLI flag outranks the env var, which outranks the persisted config value. There is no runtime toggle, so a session decides once at startup. `--show-reasoning` is listed in `--help` and advertised by a startup tip. ## Headless Reasoning goes to stderr under a one-time `Reasoning:` heading per phase, so `-p` output stays pipeable: the answer alone reaches stdout, and reasoning is never accumulated into the final response. A phase is closed by the next text block, tool call, or round boundary, which is what keeps the heading from gluing onto the answer. ```console $ dcode -p "where does my config live?" --show-reasoning Your config lives at ~/.deepagents/config.toml. $ dcode -p "where does my config live?" --show-reasoning 2>&1 Reasoning: The user wants the config path. I should read it rather than guess. Your config lives at ~/.deepagents/config.toml. ``` Independent of `--no-stream`: reasoning is written as it arrives even when the answer is buffered. ## TUI Reasoning streams into its own muted row, expanded, then collapses to a `Reasoning` header once the phase ends so the answer is not pushed off screen. Click the header or press Ctrl+O to reopen it; that state survives transcript virtualization and thread restore. Provider text renders as plain content, never as Textual markup or markdown, so it cannot inject styling into the transcript. Subagent reasoning stays out of the main transcript, and opaque blocks (`non_standard`, redacted thinking) are skipped. Made by [Open SWE](https://openswe.vercel.app/agents/70049dde-4f1e-5c46-b2fd-d329629b11e2) ## References - Plan: https://openswe.vercel.app/agents/eba258eb-4008-5322-97df-c43b607fca26/plan --------- Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>