pi-claude-cli · Pi Extensions Directory

Safety notes

Verified with notes
Routes model calls through Claude Code CLI; useful for subscription-backed experiments.

Static scan findings

This is a first-pass static screen, not a formal audit. It flags patterns worth reading before install.

MEDIUM · token_access

index.ts

maxTokens: model.maxTokens,
apiKey: "unused",

MEDIUM · token_access

README.md

- Streams text, thinking, and tool call tokens in real-time

MEDIUM · token_access

tests/stream-parser.test.ts

message: { usage: { input_tokens: 10 } },

MEDIUM · token_access

tests/event-bridge.test.ts

maxTokens: 8192,
input_tokens: 100,
output_tokens: 0,

MEDIUM · spawn_shell

tests/provider.test.ts

// Mock child_process.execSync for validateCliPresence/validateCliAuth
vi.mock("node:child_process", () => ({
execSync: vi.fn(() => Buffer.from("1.0.0")),

MEDIUM · token_access

tests/provider.test.ts

maxTokens: 8192,
maxTokens: 16384,
expect(config.apiKey).toBe("unused");

MEDIUM · spawn_shell

tests/process-manager.test.ts

import type { ChildProcess } from "node:child_process";
// Mock child_process.execSync for validation tests
vi.mock("node:child_process", () => ({

MEDIUM · token_access

.planning/ROADMAP.md

- [x] **Phase 3: Extended Thinking and Usage** - Bridge thinking token events and usage metrics with configurable thinking budgets (completed 2026-03-14)
**Goal**: Pi receives thinking token streams and usage metrics, with configurable thinking effort per model
2. After each response, pi receives accurate usage metrics (input tokens, output tokens, cache tokens)

MEDIUM · token_access

.planning/STATE.md

- [Phase 03]: Used --effort levels instead of --thinking-budget tokens (CLI does not support --thinking-budget) (03-01)

MEDIUM · token_access

.planning/REQUIREMENTS.md

- [x] **PROV-02**: Provider exposes all current Claude models derived from `getModels("anthropic")` with correct context windows, max tokens, and cost info
- [x] **STRM-05**: Extension tracks and reports usage metrics (input tokens, output tokens, cache tokens) from `message_start` and `message_delta` events
- **PERF-01**: Persistent subprocess sessions to avoid ~12s startup overhead per request and O(n^2) token growth from replaying full conversation history each turn

Package scripts captured

package.json

{
  "test": "vitest run --reporter=verbose",
  "test:coverage": "vitest run --coverage",
  "typecheck": "tsc --noEmit",
  "lint": "eslint .",
  "format:check": "prettier --check .",
  "prepare": "husky"
}