v2.1.3

Security fixes (GO-with-conditions)

External GPT Pro security re-audit passed. All High/Blocker findings resolved.

Round 1-2 (prior commits on main)

• OpenAI API-key URL double /v1 fix
• Azure Responses v1 endpoint handling
• WebSocket Origin allowlist
• openUrl() command injection → spawn() + shell:false
• config.json atomic writes with 0o600
• parallel_tool_calls catalog guard

Round 3 — this release

• auth.json permission migration: atomicWriteFile() + hardenExistingSecret() at load time ensures existing 0644 files are fixed to 0600 on first startup
• Directory hardening: hardenConfigDir() fixes existing 0755 → 0700 at all entry points
• Non-streaming tool output: buildResponseJSON() now handles tool_call_start/delta/end, preserves output order, distinguishes function_call/custom_tool_call/tool_search_call, and propagates error events
• Reasoning summary suppression: hideThinkingSummary now consumed in both SSE and non-streaming paths, including web-search sidecar
• MCP namespace round-trip: Anthropic/Google adapters flatten namespace via namespacedToolName(), Zod schema preserves namespace field, findToolById() propagates namespace to tool results
• Upstream parallel_tool_calls:false: openai-chat adapter explicitly sends the field when tools are present
• Stall detection fix: upstream activity tracked on event receipt, not emission — prevents false response.incomplete when thinking_delta is suppressed

Remaining conditions (v2.1.4 candidates)

• Pre-header connect timeout (Medium)
• Native catalog parallel_tool_calls preservation (Low)
• reasoning.summary undefined handling (Medium)
• Regression test coverage for new paths (Medium)