# agnostic-ai v0.96.0: let a hook answer routine prompts

> Source: <https://agnostic-ai.org/updates/2026-11-20-v0.96.0/>
> Published: 2026-11-19 23:00:00+00:00

**Key takeaway:** Setting `approvals` to `yolo` installs a hook via `sync --global` that auto-approves one safe call at a time while still prompting on spec and tool-folder edits, Git config reads, credential-file links and user `ask` rules.

## Highlights

- Set `approvals` to keep prompts, use a tool's own reviewer, or let a trusted hook answer routine prompts in Claude Code, Codex, Kiro, Qwen Code and Qoder on macOS and Linux. See[approvals](https://agnostic-ai.org/docs/approvals/) .
- The companion lists cached Claude app chats and Cowork tasks next to your coding sessions, plus cached ChatGPT chats when `sqlite3` is available. See[companion list](https://agnostic-ai.org/docs/cli-reference/companion-list/) .
- On macOS, Ctrl+Alt+V opens the send picker, or pastes into the only eligible Claude Code or Codex chat shown in Terminal or Ghostty.
- On macOS 13 and later, `companion install` can download the matching app without Go or Xcode tools when run from a release CLI without`--source` . See[companion](https://agnostic-ai.org/docs/cli-reference/companion/) .
- `sync` stops when a locked pack is missing or a configured output path points into a source folder. See[sync](https://agnostic-ai.org/docs/cli-reference/sync/) .

## What changed

### Approvals

Set `approvals` in your home config to one of three modes. `ask` keeps every prompt. `review` hands routine prompts to the tool's own reviewer, such as Claude Code auto mode or Codex Auto-review. `yolo` installs a hook with `sync --global` that approves one safe call at a time; edits to specs and tool folders, reads of Git config, and links to credential files still ask, and so does any call that matches one of your own `ask` rules.

`approvals status` shows how each tool handles prompts. `approvals log` prints the hook's decisions, with `--since` and `--json`. `approvals ask`, or **Ask me** in the companion, turns hook answers off on the next request. On Windows, `yolo` asks in every tool. Each tool's limits are on the [approvals page](https://agnostic-ai.org/docs/approvals/).

### More chats in the companion

- Claude app chats, Cowork tasks and ChatGPT chats appear in the list. Unread replies land under **Needs you** , and they open in their own app or on chatgpt.com.
- Expand a chat to see its working Claude Code and Codex subagents. A chat that waits for your permission, an answer or a forgotten background shell shows a lock.
- The chat menu marks a chat complete, undoes that, hides an archived chat, and on macOS stops a live Claude Code terminal session after asking, keeping its transcript.

### Send copied text to a chat

On macOS, Ctrl+Alt+V opens a picker of chats that can take text. Pick one and the companion fills its input in Terminal or Ghostty and leaves Enter to you. Copied web text can paste as Markdown. Pasting into Terminal needs macOS Accessibility access. See [companion list](https://agnostic-ai.org/docs/cli-reference/companion-list/).

### Sync protects your sources

- `sync` ,`status` and`validate` refuse an`outputs` path that resolves into`.agnostic-ai/` or a spec folder.
- `sync` stops when`agnostic.packs.lock` is invalid or names a missing pack, before it removes that pack's generated files.
- `doctor --fix` and`sync` back up edits to generated files, including edits through links into the project.

See the [release notes](https://github.com/Chemaclass/agnostic-ai/releases/tag/v0.96.0) for the remaining fixes.

## What changed in the targets

- **Goose 1.60.0** removed`goose review` ([release](https://github.com/aaif-goose/goose/releases/tag/v1.60.0) ). This release stops writing Goose review files;`import goose` still reads an old root`.agents/REVIEW.md` . See[Goose reviews](https://agnostic-ai.org/docs/targets/goose-reference/#reviews) .
- **Kiro CLI 3.4.0** runs`PermissionRequest` command hooks in V3 sessions ([hooks](https://kiro.dev/docs/hooks/types.md) ). This release validates and syncs them, and the`approvals` hook uses them. Kiro CLI 3.5.0 and IDE 1.4.236 add remote MCP`oauth.mode` ([MCP](https://kiro.dev/docs/mcp/configuration.md#organization-sign-in) ), which import and sync now keep.
- **Qwen Code**`PermissionRequest` hooks ([hooks](https://qwenlm.github.io/qwen-code-docs/en/users/features/hooks/) ) carry the`approvals` hook. Qwen Code 0.28.0`model.reasoningEffort` ([settings](https://qwenlm.github.io/qwen-code-docs/en/users/configuration/settings/) ) now round-trips portable`effort` .
- **Qoder**`PermissionRequest` hooks ([hooks](https://docs.qoder.com/cli/hooks) ) carry the`approvals` hook. Qoder CLI 1.1.93 reads an explicit MCP`protocolVersion` ([MCP](https://docs.qoder.com/cli/mcp-reference) ), which import and sync now keep.
- **Mistral Vibe 3.1.3** adds`pre_llm` and`post_llm` hooks ([release](https://github.com/mistralai/mistral-vibe/releases/tag/v3.1.3) ).`sync` and`sync --global` now write them.
- **Grok Build 1.0.88** adds`headers_helper` for dynamic MCP headers ([changelog](https://x.ai/changelog/build) ). Import and sync keep`x-grok.headers_helper` .

Checked on November 20, 2026 against the linked vendor docs and releases. Native behavior was not run, except the Qwen Code hook, which ran with Qwen Code 0.28.0 and a local test model.
