Add AI-powered issue triage system with correction tracking

[MackinnonBuck]

Summary

Adds an end-to-end AI-powered issue triage pipeline using GitHub Agentic Workflows. When an issue is created, an AI classifier routes it to a type-specific handler agent that investigates the issue, applies a label, and leaves a comment. Humans can submit corrections via a slash command, which are collected into tracking issues and eventually fed back to improve the agents — creating a continuous improvement loop.

How the flow works

1. Automated triage (on issue create)

A classifier agent reads newly opened issues and routes them to one of four handler workflows based on the issue type: handle-bug, handle-enhancement, handle-question, or handle-documentation.

Each handler is an independent agentic workflow with its own instructions and scope. Some are simple (e.g., the enhancement handler just adds a label), while others do deeper investigation. For example, the bug handler reads the issue, analyzes the relevant source code, attempts to determine whether the behavior is actually a bug or working as designed, and may reclassify the issue if appropriate. Some handlers leave a comment explaining their reasoning.

Handlers run in per-issue concurrency groups so multiple issues can be triaged in parallel without blocking each other.

2. Human correction (on demand)

If the agent's classification was wrong, a human can run the /triage_feedback slash command on the issue. This opens a modal dialog where the user can describe what the agent got wrong and what the correct action should have been. Submitting the form dispatches a repository_dispatch event that triggers the correction collection pipeline.

3. Correction collection (automated)

The collect-corrections workflow:

1. Parses the slash command payload
2. Finds (or creates) an open tracking issue labeled triage-agent-tracking
3. Appends a row to the tracking issue's markdown table with the issue link, feedback, submitter, and date
4. When the correction count reaches a configurable threshold, auto-assigns Copilot to the tracking issue

4. Addressing corrections (Copilot)

When Copilot is assigned to the tracking issue, it:

1. Reads each linked correction and the original issues for context
2. Identifies patterns in misclassifications
3. Opens a PR with targeted improvements to the agentic workflow instructions
4. The tracking issue is closed when the PR merges, and a new one is created for future corrections

This creates a flywheel: triage → corrections → improvements → better triage.

What changed

Agentic Workflows

• Re-initialized gh-aw to the latest version (new agent file, updated lock format)
• Added 5 agentic workflows: issue-classification.md (classifier) + 4 handlers (handle-bug, handle-enhancement, handle-question, handle-documentation)
• Fixed 2 existing workflows (issue-triage.md, sdk-consistency-review.md) for new gh-aw syntax (roles: moved under on:)
• Each handler uses job-discriminator: ${{ inputs.issue_number }} for per-issue concurrency

Correction pipeline

• Added /triage_feedback slash command (.github/commands/triage_feedback.yml)
• Added collect-corrections workflow triggered by repository_dispatch
• Added scripts/corrections/collect-corrections.js — the core collection logic with full JSDoc type annotations (@actions/github types, zero tsc errors)
• 28 vitest tests covering pure functions and mocked Octokit integration

CI & tooling

• Added corrections-tests.yml — runs tests on push/PR to scripts/corrections/
• Added verify-compiled.yml — ensures gh aw compile is run when agentic workflow frontmatter changes
• Added install-corrections and test-corrections recipes to the justfile, wired into top-level install and test

[Copilot]

Pull request overview

Adds an end-to-end AI-driven GitHub Issue triage + human feedback/correction loop using GitHub Agentic Workflows, plus a small Node-based correction-collector script with tests and CI to keep the pipeline reliable.

Changes:

• Added an issue classifier workflow and per-type handler workflows (bug/enhancement/question/documentation) using gh-aw compiled lockfiles.
• Added a /triage_feedback slash command and a repository_dispatch-triggered correction collection workflow backed by scripts/corrections/collect-corrections.js (with Vitest tests).
• Updated repo automation/tooling: just recipes for installing/testing corrections, and a workflow to verify .lock.yml files are up-to-date.

Reviewed changes

Copilot reviewed 31 out of 35 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
scripts/corrections/collect-corrections.js	Implements correction tracking issue discovery/creation, table row appending, and Copilot auto-assignment threshold logic.
scripts/corrections/test/collect-corrections.test.ts	Adds unit tests for correction collector helpers and Octokit interactions.
scripts/corrections/package.json / scripts/corrections/tsconfig.json / .gitignore	Sets up the corrections script package/tooling for tests and typechecking.
.github/workflows/collect-corrections.yml	Runs the correction collector on repository_dispatch events from the slash command.
.github/commands/triage_feedback.yml	Defines the /triage_feedback command (modal form) that dispatches correction events.
.github/workflows/issue-classification.md + handler workflows/lockfiles	Adds the classifier + handler workflow set for automated issue triage.
.github/workflows/verify-compiled.yml	Adds a CI check to ensure gh-aw compilation outputs are committed.
.github/workflows/corrections-tests.yml	Adds CI coverage for the corrections script tests.
justfile	Integrates corrections install/test into top-level just install and just test.
.github/aw/actions-lock.json and .github/agents/*	Updates gh-aw configuration/agent files to the newer gh-aw structure.

Files not reviewed (1)

• scripts/corrections/package-lock.json: Language not supported