feat: reduce agent token usage by 84% per deployment

[raghavyuva]

Summary

• Token reduction: Simple deploy dropped from 724K → 113K input tokens (84%), complex deploy from 881K → 275K (69%)
• Cost reduction: $0.89 → $0.13 per simple deploy, $0.93 → $0.29 per complex deploy
• Agent architecture consolidation: Single delegateTool, instruction skills loaded on demand, trimmed workspace tools

Changes

Context pruning (biggest win)

• Added TokenLimiterProcessor (128K cap) to prevent unbounded context growth within a run
• Built ToolResultPruner processor to strip transient search_tools/load_tool results from conversation history

Tool routing fixes

• Introduced queryKeys in tool-factory.ts to correctly route id as a query param instead of body, fixing add_application_domain 400 errors that caused retry spirals
• Moved createProject and generateRandomSubdomain to core tools so they're always available without search/load overhead

Skill loading fix

• Fixed SKILLS_SOURCE path in workspace-factory.ts using import.meta.url instead of process.cwd() — resolves correctly in both dev (tsx) and production (Docker bundled .mjs)

Response compaction

• Enhanced compact-output.ts with recursive nested object flattening (e.g. GitHub owner objects)
• Embedded critical domain and monitoring rules directly in system prompt as fallback

Production logging

• Default log level to info in production, debug in development

Test plan

• All 468 existing tests pass
• Build succeeds (mastra build)
• Verified simple deploy (vanilla-js): 113K tokens, $0.13
• Verified complex deploy (django-cms with Dockerfile gen): 275K tokens, $0.29
• Verified import.meta.url resolves correctly in Docker layout (/app/.mastra/output/ → ../../skills → /app/skills/)
• Skills load successfully (hasSkills: true in workspace logs)
• Domain attached at project creation time (no post-deploy add_application_domain needed)