feat(claude): forward --system-prompt / --append-system-prompt via ACP _meta [AI-assisted]

[Vercantez]

Summary

Adds two global CLI flags that forward through to claude-agent-acp via the ACP _meta.systemPrompt field that landed in agentclientprotocol/claude-agent-acp#91:

• --system-prompt <text> — replace the Claude Code system prompt entirely
• --append-system-prompt <text> — append to the default Claude Code system prompt

Both are mutually exclusive and validated at flag-resolve time. The option is persisted in session_options.system_prompt so acpx claude sessions ensure / reuse flows keep the override across turns, and is emitted alongside the existing claudeCode.options in _meta on session/new:

// --system-prompt "..."
{ "_meta": { "systemPrompt": "you are an obsidian assistant" } }

// --append-system-prompt "..." --model sonnet
{
  "_meta": {
    "claudeCode": { "options": { "model": "sonnet" } },
    "systemPrompt": { "append": "always speak in spanish" }
  }
}

Codex and other adapters ignore the field, so this is safe to keep as a global flag rather than a claude-specific one, matching how --model / --allowed-tools are already handled.

Why

Today there is no way to pass a system prompt override through acpx into Claude Code: acpx drives claude-agent-acp over ACP rather than forwarding unknown args to the claude binary, so --system-prompt / --append-system-prompt on the underlying CLI are unreachable. claude-agent-acp has supported _meta.systemPrompt (string = replace, {append} = append) since PR #91, so this is purely an acpx-side plumbing gap.

Changes

• SessionAgentOptions / AcpClientOptions.sessionOptions gain a systemPrompt?: string | { append: string } field; the persisted shape is session_options.system_prompt (snake_case, passes lint:persisted-key-casing).
• buildClaudeCodeOptionsMeta lifts systemPrompt to the top level of _meta (not nested under claudeCode.options), matching the claude-agent-acp parser.
• New resolveSystemPromptFlag helper enforces mutual exclusion and returns the correct union shape.
• Persistence round-trips the field through conversation-model, persistence/parse, session-management, and mode-preference.
• All three sessionOptions construction sites in command-handlers.ts (runOnce, sessions new, sessions ensure) now forward the flag.

Tests

• pnpm check passes locally (format, typecheck, oxlint, build, viewer typecheck/build, test:coverage): 528 / 528 tests pass.
• New unit tests:
  • test/client.test.ts — AcpClient createSession emits _meta.systemPrompt as a string, and emits { append } alongside claudeCode.options.
  • test/cli-flags.test.ts — resolveSystemPromptFlag string/append/none/mutual-exclusion cases.
  • test/session-persistence.test.ts — listSessions round-trips both the string and the { append } shapes through session_options.system_prompt.
• Live end-to-end against @anthropic-ai/claude-code@2.1.92 + @agentclientprotocol/claude-agent-acp:
  • acpx --system-prompt "...respond with EXACTLY SP_OVERRIDE_TOKEN_42..." claude exec "say hi" → returns exactly SP_OVERRIDE_TOKEN_42, confirming the replace path.
  • acpx --append-system-prompt "...always include APPEND_TOKEN_99..." claude exec "What is 2+2?" → returns 4 followed by APPEND_TOKEN_99, confirming the append path.
  • Passing both flags errors with Use only one of --system-prompt or --append-system-prompt.

Test plan

• pnpm check
• Live verification against claude-agent-acp for both replace and append
• Mutual exclusion guard exercised
• Reviewer sanity check that top-level _meta.systemPrompt is the shape they want vs. nesting under claudeCode

AI-assisted PR

Per CONTRIBUTING.md:

• Marked AI-assisted in the title
• Fully tested (unit tests + live e2e against real claude-agent-acp)
• I understand what the code does and will own any review follow-through