[Feature Request] ZCode: API key rotation, auto-retry, worktree orchestration & performance diagnostics

Feature Request: ZCode Improvements

Feedback from a new ZCode power user (Luke) who loves the platform but identified rough edges requiring custom workarounds.

---

Reporter Info

• User: Luke (Discord OP)
• Experience: New user, calls it "state of the art"
• Use Case: DeepSeek via API provider (not GLM)
• Timestamp: 8/8/26, 6:56 AM + 9:42 PM update

---

Feature #1: Native API Key Rotation

Problem:

User has multiple API keys on subscription plans and wants to distribute load evenly to maximize usage within rate limit windows.

Current Workaround:

"I created a proxy endpoint that does this key-rotation logic for me."

Requested Feature:

• Configure multiple API keys in settings
• Choose rotation strategy: Round-Robin / Least-Used / Random
• Built-in usage tracking per key
• Automatic failover when one key hits limits

Benefits:

• Maximize subscription value across keys
• No external proxy needed
• Works for users with multiple API subscriptions

---

Bug Fix #2: Auto-Retry for empty_message Errors

Problem:

~1 in 1000 API requests fail with empty_message response. This does NOT happen in Claude Code, OpenCode, PI, etc.

Impact:

• Background agents fail silently - particularly annoying
• Breaks automation workflows
• Requires manual retry or workaround

Current Workaround:

"Solution is endpoint proxy which can retry for errors like these."

Requested Fix:

ZCode should natively implement automatic retry with exponential backoff for:

• empty_message errors (the ~1/1000 case)
• Network timeouts (transient issues)
• Rate limit 429 (after appropriate wait)
• 502/503 server errors

Suggested: 3 retries with [500ms, 1000, 2000ms] backoff.

---

Feature #3: Git Worktree Orchestration Workflow

Problem:

User wants multiple features developed simultaneously using git worktrees, but ZCode lacks:

• Managing multiple worktree agents
• Automatic conflict resolution on merge
• Orchestrator pattern for parallel development

Current Workaround:

"I vibe-coded a plugin that uses an orchestrator - one per project - which implements features in background agents, then reports back."

Requested Architecture:

        ORCHESTRATOR (Main Branch)
        - Receives feature requests
        - Dispatches background agents
        - Has full context for conflict resolution
        - Stays idle for new user requests
              |
    +-----+-----+-----+
    |     |     |     |
Agent A   Agent B   Agent C
(Feature)  (Feature)  (Feature)
Worktree   Worktree    Worktree

Key Features Needed:

1. Parallel Development: Spawn agents per worktree with isolated context
2. Orchestrator Pattern: Main agent idle, background agents report progress
3. Auto Conflict Resolution: Orchestrator has ALL feature context, can intelligently merge
4. Status Dashboard: Show all features, progress, conflicts detected

---

Feature #4: Performance Diagnostics Dashboard

Problem:

ZCode becomes significantly slower as users accumulate memory, plugins, skills:

Configuration	Response Time Example
Clean install	~4 seconds for simple question
With memory + plugins + skills	~40 seconds same question

The Hidden Problem:

"Long term users keep adding skills, plugins, memory - making it slower without realizing why."

Requested Feature: Performance Breakdown

Show users WHAT is causing slowness:

• Time-to-First-Token (TTFT) total breakdown
• Per-component cost:
  • Memory/context: X seconds (Y%)
  • Each skill loaded: X seconds (Y%)
  • Each plugin active: X seconds (Y%)
  • System prompt overhead: X seconds (Y%)
• Payload size: Total KB vs context window
• Trend over time: Getting slower? By how much?

Proactive Alerts:

Warning: Response time increased 300% this week
Main cause: Memory grew from 50KB -> 180KB
Suggestion: Review/archive old entries

---

Summary

#	Type	Title	Priority	Effort
1	Feature	API Key Rotation	Medium	Low
2	Bug Fix	Auto-Retry empty_message	High	Low
3	Feature	Worktree Orchestration	Medium	High
4	Feature	Performance Diagnostics	High	Medium

---

Why This Matters

From reporter:

"Love everything about it, feels state of the art, but had to vibe-code fixes before I could use it productively."

These features bridge gap between casual use and productive/professional use.

---

Submitted on behalf of power user Luke who provided detailed technical feedback with existing workarounds.

No files, tests, or entry points are named. First split the request into separate proposals for key rotation, retry handling, worktree orchestration, and performance diagnostics, then locate the existing settings, API error, worktree, and timing components; each proposal is done only when its scope and acceptance criteria are independently testable.