SPIKE: Evaluate Claude GitHub Actions agent for auto-triggering /close-issue on new issues

## Overview
Evaluate the feasibility and value of setting up Claude GitHub Actions agent to automatically start working on new issues as they're created, using the `/close-issue` workflow.

## Core Concept
When a new issue is created in this repo, a GitHub Action would:
1. Detect the new issue event
2. Trigger Claude agent with the `/close-issue {issue_number}` command
3. Agent autonomously works through the issue to PR completion

## Motivation
- **Instant response**: No delay between issue creation and work starting
- **True automation**: Issues flow directly to completion without manual intervention
- **Compound benefits**: Each resolved issue adds to knowledge base automatically

## Technical Exploration

### GitHub Actions Integration with Claude Pro/Max
```yaml
name: Auto-resolve issues with Claude
on:
issues:
types: [opened, labeled]

jobs:
claude-agent:
if: contains(github.event.label.name, 'auto-resolve')
runs-on: ubuntu-latest
steps:
- name: Trigger Claude Agent
run: |
# NOTE: Claude Pro/Max uses session-based auth, not API keys
# Need to explore GitHub's integration options
```

### Key Questions

1. **Authentication & Sessions**
- How can GitHub Actions authenticate with Claude Pro/Max (no API keys)?
- Can we use GitHub's built-in AI integrations?
- Is there a Claude GitHub App or OAuth flow available?

2. **Context & Knowledge Base**
- Can the agent access the full `knowledge/` directory?
- How to ensure it has the right CLAUDE.md context?
- Will it respect our principles and procedures?

3. **Safety & Control**
- Should this be opt-in via label (e.g., `auto-resolve`)?
- How to prevent runaway agents or infinite loops?
- Emergency stop mechanism?

4. **Observability**
- How to track what the agent is doing?
- Can we preserve the full conversation for retros?
- Integration with our `_meta` field ideas?

## Implementation Considerations

### Pros
- **Ultimate automation**: Issues → PRs without human intervention
- **24/7 availability**: Work happens even when you're away
- **Consistent quality**: Agent follows procedures exactly
- **Learning opportunity**: See how autonomous agents handle real work

### Cons
- **Authentication complexity**: Pro/Max doesn't use API keys
- **Context limitations**: Agent might miss nuanced requirements
- **Session management**: GitHub Actions can't maintain Claude sessions
- **Debugging complexity**: Harder to troubleshoot failed runs

## Selective Optimization Check
Does this align with our throughput goals?
- ✅ Increases agent management capability
- ✅ Reduces cognitive load (issues handle themselves)
- ❓ Authentication barriers might make this impractical
- ❓ Could reduce learning opportunities from manual work

## Research Tasks

- [ ] Investigate GitHub's AI assistant integrations
- [ ] Check if Anthropic offers GitHub Apps for Pro/Max users
- [ ] Explore alternative approaches (webhooks to local Claude?)
- [ ] Research GitHub Copilot's issue-to-PR capabilities
- [ ] Define safety constraints and guardrails

## Alternative Approaches

Since Claude Pro/Max doesn't use API keys, consider:

1. **GitHub Copilot Integration**: Use GitHub's native AI for issue triage
2. **Webhook to Local**: GitHub webhook triggers local Claude Code instance
3. **Semi-automated**: Create draft issues/PRs for manual Claude processing
4. **Scheduled Local Runs**: Cron job that checks for new issues and runs Claude locally

## Success Criteria

- Find viable authentication method for Claude Pro/Max in GitHub Actions
- Agent can successfully complete simple issues autonomously
- Maintains code quality standards (tests, linting)
- Follows our git workflow and conventions
- Provides visibility into its decision-making

## Next Steps

- [ ] Research GitHub's AI integration options for Pro users
- [ ] Investigate if local Claude Code can be triggered remotely
- [ ] Explore GitHub Apps marketplace for Claude integrations
- [ ] Consider building a bridge service if needed
- [ ] Evaluate alternative AI solutions with better GitHub integration

**Principles**: `ose`, `throughput-definition`, `selective-optimization`

Start by reviewing the knowledge/ directory and CLAUDE.md, then investigate GitHub's AI integrations, Anthropic GitHub Apps, Copilot issue-to-PR support, and webhook or scheduled local alternatives. Document a viable authentication path, safety and observability constraints, and whether a simple issue can reach a tested PR while following the repository workflow.

Written by the indexing model from the issue text.