Phase 8: Context Management — token counting, compaction, CLAUDE.md, memory

[gHashTag]

Problem

tri-api accumulates messages with no size management. After ~5-10 tool-heavy turns, context exceeds Claude's 200K token limit → API returns 400 → session dies.

Solution: 3 new files + 2 modified (~300 LOC)

P0 — Context Survival

• context.zig (~150 LOC): Token estimation (bytes/4), two-phase compaction:
  1. Truncate old tool outputs (keep last 3 turns)
  2. Summarize via API call (if still over 80%)

P1 — Instructions Layer

• claude_md.zig (~80 LOC): Load CLAUDE.md hierarchy → system prompt
• memory.zig (~60 LOC): Persistent learnings in ~/.tri-api/MEMORY.md

Wiring

• main.zig (+30 LOC): Compaction before each turn, system prompt in API request
• tool_protocol.zig (+5 LOC): Support for system field

Acceptance Criteria

• Token estimation works
• Compaction triggers at 80% of 180K tokens
• Tool output truncation preserves last 3 turns
• CLAUDE.md loaded from hierarchy
• System prompt injected into API request
• Memory loads from ~/.tri-api/MEMORY.md
• TUI shows context usage per turn
• zig build clean, all tests pass

[gHashTag]

Implemented in 3b64d64. Phase 8 complete:

• context.zig (276 LOC): Token estimation + two-phase compaction (truncate → summarize)
• claude_md.zig (83 LOC): CLAUDE.md hierarchy loader → system prompt
• memory.zig (100 LOC): Persistent learnings in ~/.tri-api/MEMORY.md
• main.zig (+65 LOC): Compaction before each turn, system prompt in API body

17/17 tests pass. tri-api now 2,555 LOC across 11 files.