puq v1.0.1

@puq-code/ai

Fixed

• Tool calls whose argument JSON is incomplete are refused instead of executed with partial arguments.
• Replayed history drops tool names that were really invocation text, avoiding provider rejections.
• Auth-gateway bearer tokens are sent only in the Authorization header.

@puq-code/coding-agent

Fixed

• Launching in your home directory no longer treats a single git repo under it as the active project; the session stays in home.
• Resuming a session that was interrupted mid-tool-call recovers the pending tool calls instead of breaking the conversation.
• A failure in post-turn session maintenance no longer leaves the run hanging or the history pruned.
• Closing a session (including subagents) releases its persistent shells.
• Awaiting Python eval handles keeps the surrounding run context.
• Signed thinking stays valid when memory rewrites the system prompt mid-session.
• Edits keep their read provenance across line-neutral hunks, so follow-up edits are not rejected as unread.
• Shell pipelines whose stages exit before the shell waits on them no longer hang.
• Terminal output no longer drops writes when the TTY is non-blocking.

puq v1.0.0

The first release of puq code: a coding agent that lives in your terminal, run as puq.

Install

curl -fsSL https://github.com/puq-ai/code/releases/download/channels/install.sh | sh

Then run puq in any project and sign in with /login puq (a puq.ai token). Runs on macOS
(arm64, x64) and Linux (x64, arm64, glibc).

What's in 1.0.0

• Works in your codebase: reads, searches and edits files, runs shell commands, uses language
  servers (LSP) for definitions, references and diagnostics, connects MCP servers, and runs
  subagents in parallel.
• Spends less per turn: the optimizer ranks code_search hits and trims tool output;
  /optimizer shows what was pruned and saved, and the turn_cost status-bar segment shows each
  turn's cost.
• Modes on Shift+Tab: Manual → Auto-edit → Plan → Yolo; the thinking level cycles on Ctrl+N.
  tools.approvalMode: "auto" lets the optimizer approve low-risk commands and asks you for the rest.
• Safe to experiment: /undo and /redo rewind code and conversation together from a snapshot
  taken before every prompt, and shell commands can run in an OS-level sandbox (sandbox-exec on
  macOS, bubblewrap on Linux; bash.sandbox).
• Claude Code-compatible hooks: hooks in .claude/settings.json and executables in
  .claude/hooks/ work as-is; project hooks ask for approval the first time they are seen.
• Formats on write with the project's own formatter (prettier, biome, gofmt, rustfmt, ruff, black).
• Web search through puq with your /login puq key.
• puq media: image, music, speech and video generation and transcription (/image, /voice,
  /music, /video, /transcribe), and /live voice conversations.
• Scriptable: puq -p --output-schema returns JSON that matches a schema, and
  puq github install adds a GitHub Actions workflow triggered by a /puq comment.
• /init writes or updates AGENTS.md for your project.

Updates

puq checks its channel on launch and every 6 hours, verifies the SHA-256 and switches over on the
next start; puq update does it by hand. stable (the default) gets versions after about a week
on latest; switch with puq config set update.channel latest.

License

MIT. puq code is derived from oh-my-pi (MIT);
puq --license shows the copyright lines and third-party notices.

puq update channels

Rolling pointers consumed by puq update and the install script — not a versioned release.