diff --git a/website/src/content/docs/learning-hub/automating-with-hooks.md b/website/src/content/docs/learning-hub/automating-with-hooks.md index 6e1b7c7ae..31b2b15ae 100644 --- a/website/src/content/docs/learning-hub/automating-with-hooks.md +++ b/website/src/content/docs/learning-hub/automating-with-hooks.md @@ -3,7 +3,7 @@ title: 'Automating with Hooks' description: 'Learn how to use hooks to automate lifecycle events like formatting, linting, and governance checks during Copilot agent sessions.' authors: - GitHub Copilot Learning Hub Team -lastUpdated: 2026-04-01 +lastUpdated: 2026-04-04 estimatedReadingTime: '8 minutes' tags: - hooks @@ -98,6 +98,8 @@ Hooks can trigger on several lifecycle events: | `subagentStart` | A subagent is spawned by the main agent | Inject additional context into the subagent's prompt, log subagent launches | | `subagentStop` | A subagent completes before returning results | Audit subagent outputs, log subagent activity | | `errorOccurred` | An error occurs during agent execution | Log errors for debugging, send notifications, track error patterns | +| `PermissionRequest` | When the agent is about to show a permission approval dialog | Programmatically approve or deny permission requests without user interaction | +| `notification` | **Asynchronously** after shell completion, permission prompts, elicitation dialogs, or agent completion | Send async system alerts and log events without blocking the agent | > **Key insight**: The `preToolUse` hook is the most powerful — it can **approve or deny** individual tool executions. This enables fine-grained security policies like blocking specific shell commands or requiring approval for sensitive file operations. @@ -125,6 +127,80 @@ When multiple IDE extensions (or a mix of extensions and a `hooks.json` file) ea Hook event names can be written in **camelCase** (e.g., `preToolUse`) or **PascalCase** (e.g., `PreToolUse`). Both are accepted, making hook configuration files compatible across GitHub Copilot CLI, VS Code, and Claude Code without modification. Hooks also support Claude Code's nested `matcher`/`hooks` structure alongside the standard flat format. +### PermissionRequest Hook + +The `PermissionRequest` hook fires when the agent is about to display a **permission approval dialog** — a prompt asking the user to allow or deny a specific tool use or action. This lets you write scripts that approve trusted actions automatically without requiring manual user interaction. + +```json +{ + "version": 1, + "hooks": { + "PermissionRequest": [ + { + "type": "command", + "bash": "./scripts/auto-approve.sh", + "cwd": ".", + "timeoutSec": 5 + } + ] + } +} +``` + +Unlike `preToolUse` (which runs before *every* tool call), `PermissionRequest` fires specifically for the interactive permission dialogs — making it suitable for CI/CD environments or unattended sessions where you want fully automated approvals for known-safe operations. + +### Enhanced preToolUse: permissionDecision Output + +In addition to controlling tool execution via exit codes, `preToolUse` hooks can now output a JSON object with a `permissionDecision` field. When the hook script writes `{"permissionDecision": "allow"}` to stdout, the tool approval prompt is **suppressed** and the tool proceeds without user confirmation: + +```bash +#!/usr/bin/env bash +# Auto-approve read-only tools, block destructive ones +INPUT=$(cat) +TOOL=$(echo "$INPUT" | jq -r '.tool_name // empty') + +case "$TOOL" in + view|glob|grep|read) + echo '{"permissionDecision": "allow"}' + exit 0 + ;; + *) + exit 0 # Default: show the normal approval prompt + ;; +esac +``` + +This is useful for workflows where you want to pre-approve low-risk tools while still prompting for higher-risk operations. + +### notification Hook + +The `notification` hook is unique: it fires **asynchronously**, meaning the agent does not wait for it to finish before continuing work. It fires at key points in the session lifecycle: + +- After a shell command completes +- When a permission prompt appears +- When an elicitation dialog is shown +- When the agent completes its response + +Use `notification` for logging and alerting integrations that should not block the agent: + +```json +{ + "version": 1, + "hooks": { + "notification": [ + { + "type": "command", + "bash": "./scripts/notify.sh", + "cwd": ".", + "timeoutSec": 10 + } + ] + } +} +``` + +> **Important**: `notification` hooks run in the background. Their exit code does **not** affect agent behavior — they cannot approve, deny, or block any action. Use them purely for observability and notifications. + ### Plugin Hooks Environment Variables When hooks are defined inside a **plugin**, the hook scripts receive two additional environment variables automatically: diff --git a/website/src/content/docs/learning-hub/creating-effective-skills.md b/website/src/content/docs/learning-hub/creating-effective-skills.md index 1cf911019..30c7a45e3 100644 --- a/website/src/content/docs/learning-hub/creating-effective-skills.md +++ b/website/src/content/docs/learning-hub/creating-effective-skills.md @@ -3,7 +3,7 @@ title: 'Creating Effective Skills' description: 'Master the art of writing reusable, shareable skill folders that deliver consistent results across your team.' authors: - GitHub Copilot Learning Hub Team -lastUpdated: 2026-02-26 +lastUpdated: 2026-04-04 estimatedReadingTime: '9 minutes' tags: - skills @@ -35,6 +35,12 @@ Skills are folders containing a `SKILL.md` file and optional bundled assets. The - Skills are **more normalised across coding agent systems** via the open [Agent Skills specification](https://agentskills.io/home) - Skills still support **slash-command invocation** just like prompts did +### Built-in CLI Skills + +Starting with v1.0.17, GitHub Copilot CLI ships with **built-in skills** that are available without any installation. These provide ready-made guidance for common tasks directly in the CLI. The first built-in skill covers [customizing the Copilot cloud agent's environment](../cli-for-beginners/) — helping you configure `.github/copilot-setup-steps.yml` and related setup. + +You can browse and invoke built-in skills just like your own project skills. They serve as both useful capabilities and reference examples for writing your own. + ### How Skills Differ from Other Customizations **Skills vs Instructions**: diff --git a/website/src/content/docs/learning-hub/using-copilot-coding-agent.md b/website/src/content/docs/learning-hub/using-copilot-coding-agent.md index dcd48b51a..40cb35cde 100644 --- a/website/src/content/docs/learning-hub/using-copilot-coding-agent.md +++ b/website/src/content/docs/learning-hub/using-copilot-coding-agent.md @@ -3,7 +3,7 @@ title: 'Using the Copilot Coding Agent' description: 'Learn how to use GitHub Copilot coding agent to autonomously work on issues, generate pull requests, and automate development tasks.' authors: - GitHub Copilot Learning Hub Team -lastUpdated: 2026-03-25 +lastUpdated: 2026-04-04 estimatedReadingTime: '12 minutes' tags: - coding-agent @@ -334,6 +334,34 @@ This repository provides a curated collection of agents, skills, and hooks desig > **Example workflow**: Combine a `test-specialist` agent with a `database-migrations` skill and a linting hook. Assign an issue to the coding agent using the test-specialist agent — it will automatically pick up the migrations skill when relevant, and the hook ensures all code is formatted before completion. +## Critic Agent (Experimental) + +GitHub Copilot CLI includes an experimental **Critic agent** that automatically reviews plans and complex implementations using a complementary AI model. Rather than relying on a single model to both generate and evaluate code, the Critic uses a separate model to independently critique the work — catching logical errors, missing edge cases, and implementation gaps earlier. + +**How it works**: + +``` +1. Coding agent develops a plan or implementation + ↓ +2. Critic agent reviews the work using a complementary model + ↓ +3. Critique is fed back to the coding agent + ↓ +4. Agent refines the plan or implementation + ↓ +5. Final result is higher quality before you see the PR +``` + +**Key characteristics**: +- Available in **experimental mode** for Claude models only +- Activates automatically during complex plans and implementations — no manual invocation needed +- Runs transparently in the background as part of the agent's normal workflow +- Helps catch a category of errors that single-model systems miss (the "blind spot" of self-evaluation) + +**Why it matters**: When the same model generates *and* evaluates its own code, it tends to miss the same classes of errors in both phases. The Critic agent's use of a complementary model introduces genuine independence — a design similar to code review by a second developer. + +> **Note**: The Critic agent is an experimental feature. Check the [CLI release notes](https://github.com/github/copilot-cli/releases) for the latest status. + ## Hooks and the Coding Agent Hooks are especially valuable with the coding agent because they provide deterministic guardrails for autonomous work: