Custom primary agents instantly hit "Free usage exceeded" on opencode/big-pickle while built-in build/plan work (A-B-A reproduced)

Description

Summary

On the Zen free tier (opencode/big-pickle), any custom primary agent defined in
.opencode/agents/*.md that has a body (custom system prompt) fails on its very
first request with Free usage exceeded, subscribe to Go [retrying...]
(server log: AI_APICallError: Error from provider (Console): Rate limit exceeded. Please try again later.).

The built-in build and plan agents work fine in the same session, same account,
same model, seconds apart. Context size is not the cause — the failing request has
0 tokens of conversation context.

Environment

• opencode 1.18.16 (TUI and opencode run, Windows)
• Provider/model on every request (verified in logs): providerID=opencode modelID=big-pickle
• No model/provider pin anywhere: agent frontmatter only has description + mode: primary;
  no model field in project or global config
• Reproduced identically on two independent machines/accounts

Reproduction (A-B-A)

1. Create .opencode/agents/ztest.md:

   ---
   description: throttle test
   

   ---

   description: throttle test
   mode: primary

   Reply with OK only.

2. Run the following back-to-back (~30s total):
   • opencode run with agent build → works
   • agent ztest (with body) → fails instantly (Rate limit exceeded, retries forever)
   • agent ztest with the body removed (frontmatter only, default system prompt applies) → works
   • body added back → fails again (~0.6s)

The only variable changed between pass and fail is the agent body / system prompt content.
Agent name, mode: primary, tools, temperature, model are identical.

Additional observation

The internal title-generation request (agent=title, small=true — also a non-default
prompt) was throttled with the same error in the same second in which a build
main request succeeded. So the limiter appears to differentiate by request content
(default vs. non-default system prompt), not by actual usage volume.

Log excerpt (timestamps UTC):

11:18:20 stream providerID=opencode modelID=big-pickle small=false agent=build mode=primary   -> success
11:18:45 stream providerID=opencode modelID=big-pickle small=false agent=ztest mode=primary   -> stream error: Rate limit exceeded (0.35s)
11:22:09 stream ... agent=build   -> success
11:22:13 AI_RetryError: Failed after 3 attempts ... Rate limit exceeded (agent=title, same account, same window)
11:23:41 stream ... agent=ztest (no body, default prompt) -> success
11:24:27 stream ... agent=ztest (body restored)           -> stream error: Rate limit exceeded (0.6s)

Expected

Either custom primary agents are subject to the same free-tier limits as built-in
agents, or the limitation is documented (docs currently say primary agents without a
model pin use the globally configured model; there is no mention of free-tier
restrictions for custom agents/prompts).

Question

Is this intended behavior of the free tier (custom system prompts excluded), or a bug
in the gateway's rate limiting? If intended, a clear error message ("custom agents
require a subscription") instead of a generic rate-limit error would save users a lot
of debugging.

Plugins

No response

OpenCode version

No response

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response