omcustomcodex:auto-improve

/omcustomcodex:auto-improve — Automated Improvement Workflow

Purpose

Reads improvement suggestions from eval-core analysis, lets the user select which to apply, applies changes in an isolated worktree with sauron verification, and creates a PR for review.

Usage

/omcustomcodex:auto-improve              # Interactive selection from pending suggestions

Prerequisites

• eval-core analysis data exists (run /omcustomcodex:improve-report first if empty)
• Pending improvement suggestions in proposed status

Workflow

Step 1: Read Suggestions

1. Run bun run packages/eval-core/src/cli/index.ts analyze --format json --save via Bash
2. Parse JSON output for improvement suggestions
3. If no suggestions: display "No improvement suggestions available" and exit

Step 2: Display & Select

Display numbered list:

[Auto-Improve] Available suggestions:
  1. [HIGH] agent:lang-golang-expert — Escalate model frontier/medium→frontier/high (3 failures in 5 uses)
  2. [MED]  routing:dev-lead-routing — Add Flutter keyword mapping (2 routing misses)
  3. [LOW]  skill:systematic-debugging — Add timeout guard (1 timeout in 10 uses)

Select items: [1,2,3] / "all" / "cancel"

Self-reference filter: Exclude items where targetName matches:

• omcodex:auto-improve, auto-improve
• pipeline-guards, evaluator-optimizer
• Any item targeting this skill itself

Step 3: Approve (State Transition)

For each selected item:

1. Call eval-core API: transition proposed → approved
2. Display: [Approved] {N} items selected for application

Step 4: Worktree Isolation

• Use EnterWorktree tool with name auto-improve-{YYYYMMDD}
• Creates isolated branch from HEAD

Step 5: Apply Changes

Map each approved item to the appropriate subagent by targetType:

Spawn agents in parallel (max 4 per R009). Each agent receives:

• Action description and evidence data
• Target file path
• Specific modification instructions

Step 6: Verification

1. Delegate to mgr-sauron: full R017 verification
2. If PASS: proceed to Step 7
3. If FAIL: display failures, offer options:
   • fix → re-apply with sauron feedback (max 2 cycles)
   • reject → transition all to rejected, ExitWorktree(remove)
   • manual → keep worktree for user inspection

Step 7: PR & Finalize

1. Delegate to mgr-gitnerd: commit + create PR
   • Title: chore(auto-improve): apply {N} improvement suggestions
   • Body: table of applied items with evidence
2. Transition all items to applied with appliedAt timestamp and PR URL
3. ExitWorktree(action: "keep") — keep branch for PR
4. Display PR URL to user

Safety Guards

Error Handling

Display Format

[Auto-Improve] Starting improvement workflow
├── Suggestions: {N} available ({high}H/{medium}M/{low}L confidence)
├── Self-reference filtered: {count} items excluded
└── Select items to apply: [1,2,3] or "all" or "cancel"

[Auto-Improve] Applying {N} improvements in worktree
├── Worktree: auto-improve-{date}
├── Agents: {count} parallel
└── Pipeline guards: max 20 items, 2 retry cycles

[Auto-Improve] Verification
├── Sauron: {PASS|FAIL}
├── PR: #{number} created
└── Status: {N} items → applied