deepseekcode

Reasoning Tape (/tape)

The Reasoning Tape is deepseekcode’s headline feature. It turns DeepSeek’s reasoning_content channel into a first-class, browsable timeline of the model’s thinking.

Inline view (default)

Every reasoning span the model produces is rendered into the scrollback as a folded line:

▸ thinking (3.1s · ~450 tok)

Keys (when the input is empty):

• r — toggle the most recent reasoning block
• R — toggle every reasoning block to the same state

Expanded blocks render the raw reasoning_content in dimmed italic so they don’t compete visually with the assistant’s final answer.

Fullscreen view (/tape)

Type /tape (or hit the configured hotkey) to open a scrubbable timeline of the entire session:

/tape  23 entries · cursor 5/23
▶ ◇ flash  reasoning  3.1s  ~450 tok
  ◇ flash  tool_call  read_file
  ◆ pro    validation  1.2s  approved
  ◇ flash  tool_call  edit_file
  ◇ flash  reasoning  collapsed
  ...
  j/k move · ⏎ select · esc back · q quit

The glyph attributes each entry to its model:

• ◇ flash
• ◆ pro

Keys:

• j / k — step forward / back
• ⏎ — expand the cursor entry (v0.1 returns to chat; deeper inline expansion lands in v0.2)
• esc / q — back to chat

Why this exists

Other agents throw reasoning_content away or dump it raw. The Tape treats it as a structured signal you can rewind, fold, and learn from. For DeepSeek specifically, where reasoning_content is a separate API field, this is essentially free differentiation — every interaction becomes inspectable without the model paying extra tokens for the “explain your reasoning” prompt.

Coming in v0.2

• Branching (b from the Tape) — fork a child session at any prior step via SQLite parent_id + branch_point. Architecture is already in place (internal/session/branch.go); only the keybinding wiring remains.
• Sharing — dsc tape export <sessionID> > tape.html for embedding reasoning timelines in PRs, blog posts, or bug reports.

This site is open source. Improve this page.