Optimize token usage across provider pipeline

[apriendeau]

Summary

Reduces token consumption by ~65% on multi-turn sessions through five optimizations:

• Tiered tool result compression — Prior-turn tool results are replaced with one-line summaries ([read /src/main.rs: 458 lines]) before sending to the provider. Current-turn results stay full. Error results and non-zero bash exits are always kept in full to preserve diagnostic context. Full results remain on disk in SessionStore.
• Compact line numbering — Read tool output changed from padded format ( 1\t) to compact (1:), saving ~15-20% overhead per file read.
• Compact skill catalog — Replaced verbose XML skill catalog in system prompt with plain-text one-liner format. Cross-LLM compatible and saves 500-1000 tokens/turn.
• Cached tool schemas — Tool schemas are built once at session start instead of regenerated every loop iteration. Prevents accidental Anthropic cache invalidation.
• Split Anthropic cache breakpoints — system_prompt changed from Option<String> to Vec<String>. Anthropic provider gives each block its own cache_control marker so the stable base prompt stays cached when the context suffix changes. Other providers join the blocks.

Changed files

File	Change
session/compress.rs	New module — compress_for_provider() with per-tool summarizers
tools/read.rs	Compact line number format
session/skills.rs	Plain-text skill catalog
providers/traits.rs	SendOptions.system_prompt: Option<String> → Vec<String>
providers/anthropic.rs	Multi-block system prompt with per-block cache control
providers/{openai,google,ollama}.rs	Join system prompt blocks
session/{conversation,agent_loop}.rs, tui/conversation.rs	Wire compression + cached schemas + Vec system prompt

Test plan

• All 418 tests pass
• Clippy clean, fmt clean
• Manual test: multi-turn session, verify prior-turn results show as summaries in provider logs
• Manual test: bash with non-zero exit in prior turn stays full
• Manual test: session resume loads full history, compression applies on next send

🤖 Generated with Claude Code