my-pi

A collection of pi extensions for an enhanced coding experience.

pi install https://github.com/Loongphy/my-pi

(git:github.com/Loongphy/my-pi is equivalent.) pi clones the repo into ~/.pi/agent/git/github.com/Loongphy/my-pi and records the source in your settings.json (packages). The extensions listed in the package manifest (package.json → pi.extensions) are loaded on the next startup — no manual copying, and no filename conflicts: the package lives in its own directory, fully separate from ~/.pi/agent/extensions/.

Then reload pi:

/reload

Update later with pi update (all packages) or pi update https://github.com/Loongphy/my-pi.

How it shows up in pi:

• pi list → the source string you installed (https://github.com/Loongphy/my-pi or git:github.com/Loongphy/my-pi) with install path ~/.pi/agent/git/github.com/Loongphy/my-pi
• loaded-resources panel (compact labels) → Loongphy/my-pi:editor.ts, Loongphy/my-pi:status

A comprehensive status bar suite with multiple modules:

Files: status/index.ts, status/header.ts, status/git.ts, status/tps.ts, status/title.ts, status/theme.ts, status/statusline.ts

• Composer — codex-style input area with a bold ❯ prompt (highlighted in !bash mode, ！ alias included)
• Skill mentions — $skill mentions render bold in the theme accent; typing $ opens the mention picker with all indexed skills (agents, codex, claude, pi); unknown $tokens are left untouched. Chinese IMEs work too: Shift+4 gives ￥, which opens the same picker and is rewritten to $ in the composer the moment the token names a skill (￥coss → $coss under the caret), so highlighting and the submitted prompt keep the plain $ rules — and a non-resolving ￥100 stays a price
• Bash alias — Chinese IMEs: Shift+1 gives ！ (U+FF01), rewritten to ! in the composer when leading (！ls → !ls, ！！/！!/!！ → !!); a ！ elsewhere stays punctuation (你好！ never runs)

File: editor.ts

Logs every provider request to ~/.pi/agent/requests/<session>.request.log — HTTP status, headers, token counts, model info, and the complete decoded request body (including gzip-compressed provider payloads) — with sensitive query parameters and auth headers sanitized.

File: request-logger.ts

Alt+C copies the current editor content to the system clipboard. (Formerly Ctrl+Shift+C, which Windows Terminal intercepts for its own Copy action whenever any text is selected in the terminal — the key never reached pi in that state.)

File: shortcuts.ts

Beyond what this extension registers, the local config remaps two pi built-in actions to match codex-cli's input habits:

{
  "app.message.followUp": "ctrl+enter",
  "app.message.dequeue": "alt+up"
}

Two caveats, both from pi's own docs/terminal-setup.md:

• Ctrl+Enter needs a terminal that reports modified Enter distinctly (Kitty keyboard protocol: Kitty, Ghostty, WezTerm, iTerm2, Windows Terminal, VS Code ≥1.109.5). xfce4-terminal and terminator cannot tell it apart from plain Enter — there the chord arrives as Enter, i.e. it steers instead of queueing.
• After editing keybindings.json, run /reload in pi to apply it.

File: ~/.pi/agent/keybindings.json (not an extension)

Retries transient HTTP 429 responses automatically (any provider). The wait follows an incremental sequence — 5s, 10s, 20s, 30s, 60s, 90s, ... (+30s per retry after 30s), up to 15 attempts, or the server's Retry-After / body reset time when provided — with a live status-bar countdown. Hard limits fail fast: a wait longer than 10 min, or a provider's permanent-limit signature (e.g. workbuddy quota exhausted, opencode usage-limit errors), surfaces the response immediately with the reset time instead of retrying.

OpenCode-specific: before an opencode.ai usage-limit 429 is handed back to the SDK, the response body is annotated with the server-provided reset time — resets_at in the codex TUI's format (14:30 / 14:30 5 Mar) plus the remaining duration resets_in, both top-level and inside error, with a (Free usage limit resets at 14:30) suffix appended to error.message — so the final error message the TUI displays carries the reset time (e.g. {"type":"FreeUsageLimitError","message":"Rate limit exceeded. Please try again later. (Free usage limit resets at 14:30)","resets_at":"14:30","resets_in":"8h 7m 17s"}). Other providers keep the historical top-level resets_in only.

Command: /429-retry toggles on/off · /429-retry <seconds> sets a fixed wait time for every retry

File: 429-retry.ts

Remembers the reasoning level per model and defaults each model to its highest supported level.

• Switching back restores its remembered level; a model without one is raised to its maximum — never inheriting a lower level (a high-only model followed by a max-capable one lands on max). Fallback: scoped --models model:level.
• /model lists the levels each model supports, current one highlighted. No commands, no switches.

Storage: ~/.pi/agent/thinking-level-memory.json · File: thinking-level.ts