test: reduce /create-bug root cause analysis cost with two-phase approach

[chrisleewilcox]

Description

The root cause analysis in /create-bug (Step 6) currently runs entirely on Opus — all file reads, grep searches, git log/blame calls, and reasoning happen in one expensive context. This refactors Step 6 into a two-phase approach:

• Phase 1 (Gather Evidence): A Haiku subagent (model: "haiku") performs all the I/O-heavy work — grep/glob/read for code tracing, git log for regression PRs, and pattern searching for scope analysis. Returns a structured summary.
• Phase 2 (Analyze & Post): The main Opus context receives the distilled summary and does the reasoning — determines root cause, identifies regression PRs, assesses scope, files separate bugs, and posts the comment.

This reduces cost by ~5-8x per analysis while maintaining output quality since reasoning stays on Opus.

	Opus (current)	Haiku gathering + Opus reasoning
Gathering phase tokens	~100-150k on Opus	~100-150k on Haiku
Reasoning phase tokens	(included above)	~10-20k on Opus
Estimated cost	$1-3	~$0.15-0.40

Related issues

MMQA-1530

Manual testing steps

Feature: Two-phase root cause analysis in /create-bug

  Scenario: user opts in to root cause analysis
    Given a bug report has been created via /create-bug

    When user opts in to root cause investigation at Step 5
    Then a Haiku subagent is launched to gather evidence
    And the subagent returns a structured summary with relevant files, code flow trace, git history findings, and pattern matches
    And the main context analyzes the summary and posts a comment with all required sections

Screenshots/Recordings

Before

Root cause analysis ran entirely on Opus — all file reads, greps, git logs, and reasoning in one context.

After

Run /create-bug in Claude Code terminal, opt in to root cause analysis, and verify the Agent tool is called with model: "haiku" for evidence gathering while the final comment still contains all required sections.

Changelog

CHANGELOG entry: null

Pre-merge author checklist

• I've followed MetaMask Contributor Docs
• I've completed the PR template to the best of my ability
• I've included tests if applicable
• I've documented my code using JSDoc format if applicable
• I've applied the right labels on the PR (see Labeling Guideline)

Pre-merge reviewer checklist

• I've manually tested the PR (e.g. pull and round trip)
• I confirm that this PR addresses all acceptance criteria described in the ticket

---

Note

Low Risk
Documentation-only change to the /create-bug Claude command; no production code paths or data handling are modified. Risk is limited to potential workflow/prompt regressions in how root-cause analysis is executed and formatted.

Overview
Refactors /create-bug Step 6 root-cause investigation guidance to a two-phase flow: a Haiku subagent performs I/O-heavy evidence gathering (code search/trace, git log regression hunting, and scope/pattern matching) and returns a structured summary, then the main context performs the reasoning and posts a single consolidated issue comment.

Adds explicit instructions for the Haiku prompt (required sections and constraints on output) and updates the main-context checklist to consume that summary to determine root cause, identify regression PRs, assess blast radius, and file/link follow-up bugs as needed.

^{Written by Cursor Bugbot for commit 932e944. This will update automatically on new commits. Configure here.}

[github-actions]

CLA Signature Action: All authors have signed the CLA. You may need to manually re-run the blocking PR check if it doesn't pass in a few minutes.

[github-actions]

🔍 Smart E2E Test Selection

• Selected E2E tags: None (no tests recommended)
• Selected Performance tags: None (no tests recommended)
• Risk Level: low
• AI Confidence: 98%

click to see 🤖 AI reasoning details

E2E Test Selection:
The only changed file is .claude/commands/create-bug.md, which is a Claude AI assistant command configuration file. This file defines how Claude should help create bug reports for MetaMask Mobile - it's internal tooling documentation, not application code.

The changes restructure the "Root Cause Analysis" workflow from three sequential phases to a two-phase approach using a Haiku subagent for evidence gathering and the main context for analysis. This is purely a workflow optimization for the AI assistant.

This file:

• Is NOT part of the MetaMask Mobile application code
• Is NOT part of the E2E test infrastructure (tests/, page-objects, fixtures)
• Is NOT part of CI/CD workflows (.github/workflows/)
• Does NOT affect any user-facing functionality
• Does NOT affect any test execution or test framework

No E2E tests are needed because this change has zero impact on the application or testing infrastructure.

Performance Test Selection:
This change is to a Claude AI assistant command configuration file (.claude/commands/create-bug.md). It has no impact on app performance, UI rendering, data loading, or any application code. No performance tests are needed.

View GitHub Actions results