feat: CC-aligned token optimization with multi-provider cache support

[Starlitnightly]

Summary

Re-integrates the 5-stage CC-aligned token optimization pipeline (from PR #54 and #57) on top of main's litellm-free architecture, with new multi-provider cache support.

What's included

1. Token optimization pipeline restored and adapted — all 5 stages working without litellm:

   • Stage 1: Tool Result Budget (externalize large outputs to disk)
   • Stage 2: History Snip (drop oldest messages over token budget)
   • Stage 3: Time-Based Microcompact (clear old compactable tool results)
   • Stage 4: Context Collapse (fold consecutive read/search groups)
   • Stage 5: Autocompact (LLM-based summarization, adapted to use acompletion with model_params)

2. Cache control extended to Qwen — new supports_explicit_cache_control() function enables cache_control: {"type": "ephemeral"} injection for both Anthropic and Qwen models (same format).

3. Multi-provider benchmark — new scripts/benchmark_multi_provider.py tests token reduction + cache behavior across 6 providers.

Benchmark Results

Local Token Optimization (provider-independent)

Optimization	Medium (128K tok)	Large (512K tok)
Micro-Compact	79.9%	89.9%
HISTORY_SNIP	90.0%	95.0%
contextCollapse	100.0%	100.0%
Full Pipeline	50.0%	85.0%
ALL COMBINED	79.9%	90.0%

Live API Token Reduction (10x50KB scenario)

Provider	Model	Raw	Optimized	Saved
OpenAI	gpt-4.1-mini	170,827	34,287	79.9%
Gemini	gemini-2.5-flash	248,316	49,757	80.0%
Kimi	kimi-k2.5	170,849	34,301	79.9%

Cache Hit Comparison

Provider	Unstable c2	Stable c2	Cache Type
DeepSeek	448	448	Auto disk (64-token granularity)
Kimi	0	588	Auto prefix (stable ordering required!)
OpenAI	0	0	Auto prefix (needs longer warmup)
Gemini	0	0	Implicit auto (no guarantee)

Key finding: stabilize_tool_definitions() is critical for Kimi — unstable tool ordering = 0 cache hits, stable ordering = 588 cached tokens.

Cache Strategy Per Provider

Provider	Explicit Markers	Strategy
OpenAI	No	Auto prefix — stable tool definitions help
Anthropic	Yes	cache_control markers injected
Gemini	No	Implicit auto — stable prefixes help
DeepSeek	No	Auto disk — 64-token granularity
Qwen	Yes (NEW)	Same cache_control format as Anthropic
Kimi	No	Auto prefix — stabilize_tool_definitions verified

litellm Adaptation

• acompletion_litellm → acompletion (unified adapter-based routing)
• max_tokens/temperature → model_params dict (new acompletion signature)
• Removed litellm version check in inject_cache_control_markers (native SDK handles it)
• is_anthropic_model check → supports_explicit_cache_control (Anthropic + Qwen)

Files Changed

• pantheon/agent.py — AgentRunContext cache fields, build_llm_view_async pipeline, cache marker injection
• pantheon/utils/token_optimization.py — all 5 stages + litellm-free adaptations + Qwen support
• pantheon/team/pantheon.py — delegation cache helpers and structured fork context
• pantheon/utils/llm.py — per-tool limits in process_tool_result
• pantheon/utils/truncate.py — persisted-output format
• pantheon/settings.py — max_tool_content_length (50K), max_file_read_chars (500K)
• pantheon/toolsets/python/python_interpreter.py — shared_executor support
• scripts/benchmark_multi_provider.py — NEW: 6-provider benchmark
• scripts/benchmark_token_optimization_live.py — live API optimization test
• tests/test_token_optimization.py — 65 unit tests (all mocked, no API keys needed)

Test plan

• pytest tests/test_token_optimization.py — all 65 tests pass (no API keys needed)
• pytest tests/test_truncate.py — truncation tests pass
• Run scripts/benchmark_multi_provider.py with API keys to verify live token reduction
• Run scripts/benchmark_token_optimizations.py for full per-optimization breakdown