omcustom:auto-improve

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

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

Prerequisites

• eval-core analysis data exists (run /omcustom-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 sonnet→opus (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:

• omcustom-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:

targetType	Agent	Action
agent	mgr-creator	Modify agent frontmatter/body
skill	mgr-creator	Revise skill SKILL.md (R010 Protected Paths)
routing	mgr-creator	Update routing patterns (R010 Protected Paths)
model-escalation	general-purpose	Update model field in agent frontmatter

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

Guard	Implementation
Self-reference prevention	Blocklist filter in Step 2
User approval gate	Step 2 interactive selection
Worktree isolation	Step 4 EnterWorktree
Sauron verification	Step 6 mandatory pass
PR-based merge	Step 7 — no direct push to develop
Max items per run	20 default, 50 hard cap
Max fix cycles	2 retries before rejection
Rollback	git revert via mgr-gitnerd post-merge

Error Handling

Scenario	Action
No suggestions available	Display message, exit
User cancels selection	Exit, no state changes
Sauron verification fails 2x	Reject all, cleanup worktree
Agent application error	Mark individual item as rejected, continue others
EnterWorktree fails	Report error, exit

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