Add mix loop: autonomous AI-driven issue resolution

[KTSCode]

Summary

Introduces mix loop, a new Mix task that autonomously resolves GitHub issues by driving a self-correcting AI loop. The task picks an open issue, classifies it as a feature or bug, repeatedly invokes the claude agent to implement/fix it, verifies each attempt with mix quality, and feeds failures back into subsequent attempts until the suite passes or the iteration budget is exhausted.

Key Changes

• New task module (lib/mix/tasks/loop.ex): 619-line implementation with:

  • Issue fetching and classification (feature vs. bug based on labels)
  • Iterative prompt generation (initial + retry prompts with failure context)
  • Self-correcting loop: runs claude headless → mix quality → decision logic
  • Git workflow: branch creation, commit, push, PR opening, issue commenting
  • Comprehensive error handling with clear user feedback

• Pure functions (unit-testable):

  • parse_issues/1, parse_issue/1: JSON decoding from gh CLI
  • classify_issue/1: label-based feature/bug detection
  • slugify/1, build_branch_name/2: branch naming conventions
  • pr_title/2, commit_message/2: conventional commit formatting
  • build_initial_prompt/3, build_retry_prompt/5: context-aware prompts
  • should_continue?/3: loop termination logic (done/retry/exhausted)
  • build_summary/4, extract_pr_url/1: result reporting

• Test suite (test/mix/tasks/loop_test.exs): 302 lines covering all pure functions with 40+ test cases

• Documentation:

  • Design spec (docs/specs/2026-06-08-feature-debug-loop-design.md)
  • README section under Development workflow
  • CHANGELOG entry

Notable Implementation Details

• Non-destructive on failure: exhausted loops leave the branch in place for manual review rather than force-pushing or reverting
• No auto-merge: opens a PR for human review (unlike mix todo)
• Bounded execution: --max flag (default 5) prevents infinite loops
• Tool validation: checks gh and claude are installed upfront
• Failure context: retry prompts include the tail of mix quality output (last 40 lines by default) to guide the agent's fixes
• Clean CLI: status messages to stderr, clean summary to stdout
• Flexible issue selection: --issue N for specific issues, --label L to filter, or auto-pick the next open issue
• Quick mode: --quick (default) for faster iteration; --full for comprehensive quality checks

Usage

mix loop                 # work the next open issue
mix loop --issue 42      # work a specific issue
mix loop --label bug     # only consider issues with this label
mix loop --max 8         # cap the loop at 8 attempts (default 5)
mix loop --full          # run the full `mix quality` each iteration

https://claude.ai/code/session_01FJ7GGnsNFWoyqr8KzDEyyp

[KTSCode]

Closing this. Severance isn't complex enough to justify an autonomous feature/debug loop — the manual mix todo flow covers our needs and keeps a human in the loop. Branch left in place for reference.