fix: merge goal continuation into primary system block (closes #1)

[fank]

What changed

experimental.chat.system.transform now merges the goal continuation block into output.system[0] instead of output.system.push(...)-ing a second array entry. When the system array is empty, it falls back to pushing — preserving current behavior for callers that rely on goal-plugin owning the entire system prompt.

Why

The previous push pattern produced two consecutive role: "system" messages on the wire. Backends that enforce the OpenAI convention of exactly one system message at index 0 reject this with HTTP 400 "System message must be at the beginning.". Affected backends include:

• Qwen family on vLLM (the Qwen3.5/3.6 chat templates raise_exception on a second system message)
• Several Llama.cpp templates
• Some Mistral templates

OpenAI's and Anthropic's hosted APIs silently tolerate multiple system messages, which is why this bug is invisible to users on those providers.

The opencode community has documented that plugins using experimental.chat.system.transform should merge rather than push (see anomalyco/opencode#23660 and anomalyco/opencode#15059). This change brings goal-plugin in line with that guidance.

How

// before
output.system.push(
  [buildGoalBlock(goal), ...].filter(Boolean).join("\n"),
)

// after
const goalBlock = [buildGoalBlock(goal), ...].filter(Boolean).join("\n")
if (output.system.length === 0) {
  output.system.push(goalBlock)
} else {
  output.system[0] = `${output.system[0]}\n\n${goalBlock}`
}

Verification

Checks

$ npm run check
... 35 pass, 0 fail
$ npm run pack:check
... opencode-goal-plugin-0.1.8.tgz (12.0 kB)

Manual OpenCode smoke test

Against opencode 1.15.11 with this plugin, provider @ai-sdk/openai-compatible pointing at vLLM 0.19.x serving Qwen3.5-122B-A10B (same setup as issue #1):

Plugin code	Result
Original (push)	HTTP 400 from vLLM: {"message":"System message must be at the beginning.","type":"BadRequestError","code":400}
This PR (merge)	/goal draw a ball succeeds; model executes write tool, creates the artifact, step finishes cleanly with reason: "tool-calls", goal auto-continues to next iteration

Token usage in the passing case: 9120 input / 283 output — confirming both the base system prompt and the goal continuation block reach the model, just collapsed into a single role: "system" message instead of two.

Tests

• Existing "system transform is idempotent" — unchanged, still passes (idempotence guard preserved).
• New "merges into existing system block instead of adding a second one" — pins the bug fix; asserts output.system.length === 1 and the base prompt is preserved as a prefix.
• New "pushes a new block when system array is empty" — pins the empty-array fallback.

Related

• Issue Goal continuation block sent as a second role: system message breaks strict-template backends (Qwen vLLM and similar) #1 in this repo (filed alongside this fix)
• anomalyco/opencode#15059 — canonical upstream bug
• anomalyco/opencode#23660 — guidance request for plugin authors

Closes #1.