feat: support ZAI token metadata, trigger compaction on idle sessions, add GLM system prompt

[processtrader]

Summary

Fix token extraction for ZAI/Anthropic API, improve auto-compaction timing for idle sessions, and add dedicated system prompt for GLM-4.6 model.

Changes

Token Extraction Fix (session/index.ts)

• Problem: ZAI (Anthropic endpoint) returns token counts in metadata.anthropic.usage instead of the top-level usage object, causing the sidebar to show 0 tokens.
• Solution: Extract tokens from metadata.anthropic.usage when top-level values are missing/zero.
• Supports both camelCase (inputTokens) and snake_case (input_tokens) field names.
• Also supports Bedrock metadata format for consistency.

Compaction Timing Fix (session/prompt.ts)

• Problem: Auto-compaction check happened after the exit condition in the main loop, so idle sessions at high context usage would exit before compaction could trigger.
• Solution: Move compaction check before the exit condition.
• Added hasPendingCompaction guard to prevent infinite loops when a compaction task is already queued.

GLM System Prompt (session/prompt/glm.txt)

Added a dedicated system prompt optimized for GLM-4.6's capabilities:

• Why? GLM-4.6 has distinct strengths (200K context, advanced reasoning, 30% better token efficiency, stronger agentic performance) that benefit from tailored instructions.
• Concise structure - Leaner prompt (~57 lines) leverages GLM's token efficiency, leaving more context for actual work.
• Reasoning-first approach - Added "Reasoning Approach" section to encourage systematic problem breakdown.
• Tool-agnostic language - Avoids explicit tool names to prevent agents (like Plan) from attempting unauthorized operations.
• Consistent style - Follows similar structure to anthropic.txt for maintainability.

Testing

• Verified token counts display correctly in sidebar with ZAI endpoint
• Confirmed auto-compaction triggers at ~84% context usage
• Tested compaction works on both active and idle sessions