Mandatory Debugging Workflow (ORCHESTRATION ONLY)

When tasked with debugging a crash or investigating unexpected behavior, the main agent MUST adhere to the following strict constraints:

1. Role Restriction: Orchestrator Only

• Delegate execution of heavy technical investigation tools to specialized subagents.
• Use subagents when running GDB (gdb-mcp).
• Delegate extensive searches or reading large C++ files to subagents.
• Your sole job is to Orchestrate, Delegate, and Synthesize.
• Continuous Delegation: Even after receiving reports, delegate the next investigation steps. Do not take over.
• Aggressive Parallelization: Do not wait for a subagent to finish or fail before spawning others for independent tasks. If a subagent reports a failure, a new hypothesis, or a new lead, immediately spawn a focused subagent to investigate that specific dimension in parallel.
• User Interface: Act strictly as the interface to the user. Keep them informed of high-level certainties. You may provide brief status updates on running agents.
• Naming: Do not refer to subagents by specific titles (e.g., "Spec Expert") in user communications. Use generic descriptions of the inquiry instead.

2. Mandatory First Actions (Three Parallel Tracks)

Upon receiving a debugging task, you MUST immediately initialize the following tracks in parallel:

Track A: Heavy Infrastructure

• Start building (gm.py) and preparing for runtime analysis (gdb-mcp or rr) immediately.

Track B: Conceptual Triage

• JS Source Reading: Analyze the reproducing JS script to identify language features (e.g., closures, eval, TDZ) and runtime flags/hints.
• Environment Preservation: Preserve the original reproducing script exactly as-is without modification. If the script fails due to missing flags or appears to have invalid calls (e.g., missing %PrepareFunctionForOptimization), keep it unaltered and ask the user to provide the correct flags or environment. Create scratch copies for minimization or testing if needed.
• Error Log Analysis: Analyze error logs and stack traces to identify the failing component or assertion.

Track C: Static Research & Eager Delegation

• V8/Spec Exploration: Use your internal knowledge to start reading relevant V8 C++ code (via code_search) and consulting ECMAScript spec requirements simultaneously.
• Work Branching: PROACTIVELY spawn subagents (e.g., via self) for every independent dimension of the task that can be parallelized.

3. Iterative Concept Discovery & Overlap Analysis

During investigation (especially via GDB or subagent reports):

• Continuous Discovery: If you encounter a new internal concept, spawn a subagent to research it immediately.
• Overlap & Re-evaluation: Actively check for overlaps between concepts. When an overlap is detected, spawn a subagent to analyze the interaction and re-evaluate previous findings.

4. Synthesis & Communication

• Proactive Updates: Keep the user informed of key milestones (e.g., triage results, parallel task status).
• Fix Presentation: When a fix is proposed, YOU MUST present it clearly to the user immediately, explaining the rationale and how it aligns with V8 best practices.
• Bug ID Association: Only attach bug IDs to a CL when there is proof that the CL fixes or affects the issue.

5. Subagent Coordination & Anti-Duplication

• Centralized Routing: The main agent MUST monitor subagent tasks to prevent duplicate investigations.
• Cross-Pollination: If Subagent A discovers information relevant to Subagent B, the main agent must immediately relay that information.
• Branching: If tasks overlap, instruct subagents to pivot to uninvestigated areas and share findings.
• GDB Resource Management: Avoid spawning multiple subagents running GDB concurrently unless they are explicitly investigating independent execution paths or orthogonal hypotheses.