fix: handle array content, add model aliases, increase subprocess timeout

[Max-shipper]

Summary

Three bug fixes addressing real failures observed in production use with claude-max-proxy.

Fix 1 — extractContentText() in src/adapter/openai-to-cli.ts

OpenAI-compatible providers (including Claude Code's own proxy routing) send content as an array of content blocks rather than a plain string:

[{"type": "text", "text": "Hello world"}]

Without handling this, messagesToPrompt() was interpolating the raw array into the prompt string, producing [object Object] and sending garbage to the CLI.

Changes:

• Added extractContentText() helper that handles: string → passthrough, Array → filter type === "text" blocks and join with newline, null/undefined → empty string, anything else → String(content)
• Widened the OpenAIChatMessage.content type in src/types/openai.ts to string | OpenAIContentBlock[] | null
• Updated messagesToPrompt() to call extractContentText() for every message role

Fix 2 — MODEL_MAP additions in src/adapter/openai-to-cli.ts

The Claude 4-6 model generation IDs (claude-opus-4-6, claude-sonnet-4-6) were missing from MODEL_MAP. Any request specifying these IDs fell through to the default "opus", which worked for Opus but silently ignored Sonnet requests.

Added mappings:

"claude-opus-4-6"                    → "opus"
"claude-sonnet-4-6"                  → "claude-sonnet-4-6"
"claude-code-cli/claude-opus-4-6"    → "opus"
"claude-code-cli/claude-sonnet-4-6"  → "claude-sonnet-4-6"

Also extended ClaudeModel type to include "claude-sonnet-4-6" since the CLI accepts it as a full model ID (not just an alias).

Fix 3 — DEFAULT_TIMEOUT in src/subprocess/manager.ts

The 5-minute (300 000 ms) timeout was too short for agentic tasks that involve multiple tool calls — web searches, multi-step research, long code generation. These routinely run 8–12 minutes and were being killed mid-execution with a timeout error.

Increased to 15 minutes (900 000 ms) with an explanatory comment.

Test plan

• Send a request with array content blocks; verify prompt is correctly extracted
• Request with model: "claude-sonnet-4-6"; verify it routes to Sonnet, not Opus
• Trigger a long agentic task (web search chain); verify it completes without timeout

🤖 Generated with Claude Code