feat(context): per-tool output caps + stale tool-result pruning

[arniesaha]

Summary

Two independent levers for in-session token hygiene:

1. Tool-level output caps

• read_file gains offset / limit params (line-based), defaults to 500 lines capped at ~20K chars, includes pagination hint in the truncated response so the model knows how to page
• run_shell and claude-subagent switch to head+tail truncation (70/30) via a shared helper src/tools/truncate.ts#headAndTail — so neither the initial plan nor the final error/answer is silently lost

2. pruneStaleToolResults in transformContext

• Runs every turn (before the compaction threshold check)
• Preserves the last FRESH_TURNS user turns (default 4, override via MAX_FRESH_TURNS)
• Older toolResult bodies replaced with a short stub — tool name, length, head prefix
• Structure-preserving: toolCallId / role / isError untouched so tool_use ↔ tool_result pairing remains valid for Anthropic's API
• Idempotent (second pass returns the same reference)

Why

Before: a single heavy gh pr diff / browser-scrape early in a session re-bills itself at full size on every subsequent agent iteration. After: stale bodies collapse to a ~200-char breadcrumb after 4 turns.

This is the biggest in-session lever other than caching: it operates on the active messages array on every turn, not just at compaction time.

Test plan

• 7 new unit tests in tests/context.test.ts (no-op, prune-stale, keep-fresh, skip-small, idempotence, structural preservation)
• Existing claude-subagent truncation test updated to verify head+tail
• Full npm test — 78/78 pass
• npm run build clean
• Manual: heavy session, check context_info shows token drop after turn 4

🤖 Generated with Claude Code