Phase 2.1: CLI Core Agent Loop — Query → API → Tools → Repeat

[Delqhi]

Summary

• Core Agent Loop: Full implementation of the query → API call → tool execution → response cycle in packages/opensin-sdk/src/agent_loop/
• NDJSON Streaming: Every event (status, thinking, text_delta, tool_call_start/end, tool_result, error, retry, context_compressed, turn_complete, session_start/end) emitted as newline-delimited JSON to stdout
• Error Recovery: Exponential backoff with jitter for transient errors (rate limits, 429/502/503/504, timeouts, network errors); non-transient errors fail immediately without retry
• Tool Result Processing: Tool calls detected from LLM response, executed via ToolRegistry, results re-injected into context for next LLM turn — loop continues until no more tool calls
• Context Window Management: Auto-compression when utilization exceeds configurable threshold, system message preservation, priority-based message retention, conversation clearing
• Bug Fix: Fixed pre-existing bug in context_mgmt/compressor.ts where originalTokens was computed but never passed to sub-methods

Architecture

src/agent_loop/
├── types.ts          # All type definitions (AgentEvent, AgentLoopConfig, TurnState, LLMResponse, etc.)
├── agent_loop.ts     # Core AgentLoop class — the main loop engine
├── context.ts        # AgentLoopContext — context window management with auto-compression
├── ndjson.ts         # NDJSON serialization/parsing/streaming utilities
├── tool_registry.ts  # ToolRegistry — tool registration and execution dispatch
└── index.ts          # Public API exports

Acceptance Criteria

• Core loop runs continuously (until no tool calls or max iterations)
• Handles tool results correctly (re-injects into context, continues loop)
• Streams NDJSON output (all event types verified in tests)
• Error recovery works (exponential backoff retry, non-transient fail-fast)
• Tests pass (33 new tests, 334 total — all green)
• Code reviewed (self-review complete)
• Documentation updated (JSDoc on all public APIs)

Test Results

Test Files  10 passed (10)
Tests       334 passed (334)

33 new tests covering:

• NDJSON serialization/parsing/streaming (6 tests)
• Tool registry CRUD and execution (5 tests)
• Context window management (6 tests)
• Core agent loop (5 tests)
• Error recovery & retry logic (4 tests)
• NDJSON streaming integration (3 tests)
• Async NDJSON stream parsing (2 tests)
• Turn state tracking (2 tests)

[Delqhi]

✅ Already implemented via sin-claude integration