pi-ctx

Capture every LLM-bound prompt to disk, and optionally strip thinking/reasoning fields from the context sent to the provider. Designed for debugging context-window contents, replaying prompts offline, and shrinking wire payloads when chain-of-thought bloat is costing you tokens and context.

Features

On-disk prompt capture — When enabled (default off), on every turn the exact payload pi sends to the provider is serialised to <cwd>/ctx/<timestamp>.json. Writes are atomic (temp file + rename) so a concurrent reader never sees a half-written file. On session_start, when saving is on, the folder is wiped so its size is bounded by the current session's prompts. Off by default — the folder is never created until you opt in.

Thinking/reasoning stripping — Strips chain-of-thought out of the wire payload before pi sends it. Handles DeepSeek R1-style top-level reasoning_content strings, Anthropic {type:"thinking"} and {type:"redacted_thinking"} blocks, and OpenAI-style {type:"reasoning"} blocks. Retains the last N thinking-bearing messages intact (configurable, see Usage). Default off.

Footer status indicator — Live ctx(save|think) indicator in the pi footer:

• label is white when the toggle is on, dim when off
• | is dim if either neighbour is dim, white only if both are white
• ctx( / ) are dim when all labels are dim, white otherwise

Idempotent and safe — Saving and thinking are independent toggles. Disabling saving leaves the existing folder on disk for inspection. Stripping is a no-op when the payload carries no reasoning fields. Clone failures fall back to the un-stripped payload with a UI warning rather than failing the turn.

Installation

Install from npm:

pi install npm:pi-ctx

Install into the current project only:

pi install npm:pi-ctx -l

Or install from GitHub:

pi install git:github.com/ashLatham/pi-ctx

Try it without permanently installing:

pi -e npm:pi-ctx

Usage

All configuration is via the /ctx slash command. No config file.

/ctx                       show usage
/ctx save                  toggle saving on/off
/ctx think                 toggle thinking/reasoning stripping on/off
/ctx think N               keep last N thinking-bearing messages (0..3, default 1)

Examples:

• /ctx save then send a prompt — find the captured JSON in <cwd>/ctx/.
• /ctx think — strip all but the most recent reasoning block.
• /ctx think 0 — strip all reasoning. Note: this is a configure, not a toggle; it always enables stripping.
• /ctx think 3 — keep reasoning on the last three assistant turns.

Notification levels

• info — toggle state changes (ctx save: on (writing to <dir>), ctx think ON — keeping last 1, etc.)
• warning — non-fatal failures: wipe failed, clone failed (sends un-stripped), write failed
• error — unknown /ctx argument

What gets stripped

A message is "thinking-bearing" if it has any of:

• a top-level reasoning_content string field (DeepSeek R1 / some OpenAI-compatible providers)
• a content block with type of thinking, redacted_thinking, or reasoning

When a message is stripped, the reasoning fields are replaced with a single-space sentinel rather than dropped entirely — providers care about block indices and message shape, so removing fields would risk breaking the request.

Saved payload fidelity

The JSON written to <cwd>/ctx/ is exactly what pi sent to the provider:

• think off → raw wire payload
• think on → stripped payload (after pruning)

Links

• npm: https://www.npmjs.com/package/pi-ctx
• GitHub: https://github.com/ashLatham/pi-ctx
• Pi Agent: https://github.com/earendil-works/pi

License

MIT