Wasabi

Wasabi is a Pi extension package with session cropping, LiteLLM session tracking, and Crunch file-change review. Each extension can be enabled independently with pi config.

Requirements

• Node.js 22.19.0 or newer
• npm
• Pi 0.84.4 or newer, installed from @earendil-works/pi-coding-agent

Development dependencies pin Pi's core packages to 0.84.4, the tested baseline. Runtime core packages are peers supplied by Pi, not separately bundled.

Install

From a fresh checkout, install the locked dependencies:

npm ci

Use npm install instead when intentionally updating dependencies and the lockfile.

Check and test

npm run check

check runs static TypeScript checking and the Jest test suite. Run only the behavioral tests with npm test. Tests use mocked model responses and Pi contexts; no API credentials or network calls are needed. Crunch diff tests exercise the actual built-in edit execution and renderers against temporary fixture files. Crop tests verify the copied suffix without creating real session files.

Install in Pi

Install all Wasabi extensions from GitHub:

pi install git:github.com/kc-ml2/Wasabi

After installation, use pi config to enable or disable individual extensions.

Run

Load the package root so Pi validates the extension declaration in package.json:

pi --offline --no-session --no-extensions -e .

At the Pi prompt, run:

/hello

Pi should display:

Hello! Wasabi is running.

This command does not call a model or persist session state.

Crop a session branch

/crop creates and switches to a new session containing the selected user turn and every entry after it on the current active branch. It is the suffix counterpart to /clone, so earlier turns are not copied into the new session. The original session remains unchanged.

/crop

Choose the user turn where the new session should start. The picker displays a number, entry ID, and preview for each user turn on the active branch. You can also use that number or full entry ID directly, which is useful without an interactive picker:

/crop 3
/crop a1b2c3d4

Cropping is intentionally restricted to user-turn boundaries. This prevents a new session from starting between an assistant tool call and its tool result. Model changes, thinking-level changes, extension messages, and branch summaries after the selected turn are copied. Labels, session names, and compaction checkpoints are not copied: they are source-session metadata, and a compaction checkpoint can point to entries that were intentionally removed by the crop.

Crunch focus mode

Crunch is a human approval step before Pi executes an edit or write tool call. Its main purpose is to explain what a proposed file change means and let the user approve it or provide feedback before it is applied.

Pi proposes an edit/write call
  -> Crunch explains the proposed change
  -> Approve: execute this call
  -> Reject: block this call
  -> Feedback: reconsider this change or stop the batch and revise the plan

AI-selected diff rows are a secondary review aid, not the main purpose. Crunch is not a rollback/checkpoint tool and does not split source code into new execution batches. A "batch" means the tool calls Pi already proposed together in one assistant message. One edit call may contain several replacement blocks; approving/rejecting it applies to the whole call, not selected blocks.

It is OFF by default. Run these commands before starting a task:

/crunch on
/crunch off
/crunch-model
/crunch-model current
/crunch-model <provider>/<model-id>

/crunch without an argument only shows usage; it does not toggle the mode. /crunch-model opens the summary-model picker. Mode and model choices are saved on the active session branch and restored on reload/resume.

Review and feedback

• A single-call batch offers Approve, Reject, and Reject this change with feedback.
• Multi-call batches additionally offer Stop all with feedback. The dialog lists sibling file changes and their review/completion states, plus the count of other tools. An approved call is not necessarily executed yet.
• Reject rejects only this call, without aborting the agent. Per-call feedback in a multi-call batch also lets siblings proceed.
• Stop all with feedback requests an abort even if the feedback editor is cancelled or left empty. The single-call feedback choice uses the same stop flow. Empty feedback stops without an automatic restart.
• With feedback, Crunch waits for the agent to settle, then requests a short review of the previous approach, feedback interpretation, and revised plan before work resumes. Files that finish while stopping are included in the verification instructions.

Stopping is cooperative; already-running tools may finish. Completed changes are not rolled back, and the review is not a sandbox: shell commands and other mutation tools are not subject to Crunch's approval gate. Headless modes without UI do not prompt or enforce the gate.

AI-selected edit diffs

In TUI mode, a collapsed local edit shows up to 12 real diff rows selected by the summary model, labeled Key diff · AI-selected, with the omitted-row count. The model returns row indices, never replacement code. Use Ctrl+O (or your configured expand-tools binding) to see the complete diff, including while the approval picker is open. write retains its existing display.

Malformed selections, oversized diffs, unavailable models, summary errors, and execution diffs that differ from the preview fall back to the built-in display. Summary requests time out after 30 seconds; failure does not bypass approval. Excerpts are a bounded, in-memory UI cache, not edits to the tool result or session history. Reloading restores the normal full-diff display for old calls.

Compatibility and model requests

• Initially OFF, Crunch registers no edit override. When enabled in TUI mode, it installs its renderer only if edit is the built-in local tool. Existing SSH, sandbox, SDK, or extension-provided edits are left intact; their calls can still be reviewed, but AI-selected diffs are disabled.
• Pi does not expose unregisterTool. After the renderer has been installed, /crunch off restores built-in rendering and execution behavior, but the wrapper remains registered until /reload. Disable and reload Crunch before switching to a different edit-tool extension. Pi may report that edit was overridden when the renderer is activated.
• RPC supports summaries and approval dialogs, but no custom diff rendering.
• Menus and notifications are English; generated summaries are Korean.
• Each reviewed change can make an additional model request, with extra latency and provider cost. File paths, proposed code, and possibly nearby diff context are sent to the selected summary provider. Pick an appropriate provider for sensitive source code. The normal Pi footer still shows the conversation model, not the separate summary model; Crunch displays Summarizing….
• Nested summaries do not pass through Pi's provider hooks. Crunch explicitly applies Wasabi's shared x-litellm-session-id policy to centinels and openai summary providers, while preserving other authentication headers. PR #1 added the conversation hook, and PR #2 added openai support. Crunch reuses that policy for its separate summary calls; the existing conversation behavior is unchanged. The shared helper is in src/litellm-session.ts. The header groups provider requests in LiteLLM; it does not add summaries to Pi's conversation history or account for their usage in Pi's session totals.

Local testing

To load only this package for local testing:

pi --offline --no-session --no-extensions -e .

Run /crunch on, then review a small edit call. A configured model is needed for AI summaries; without one, the approval gate uses a plain fallback summary.