Make tkn CLI agent-friendly for AI/LLM consumption

Summary

AI agents are increasingly using CLIs as their primary interface to external systems. The tkn CLI should be optimized for both human and agent consumption, following principles from You Need to Rewrite Your CLI for AI Agents.

Motivation

AI coding agents (Copilot, Claude Code, Cursor, etc.) use CLIs directly. They need:

• Deterministic, machine-readable output — not tables designed for human eyes
• Self-describing conventions — agents need to know how to use the CLI efficiently
• Structured input — agents generate JSON more reliably than bespoke flag combinations
• Reliable exit codes — agents depend on exit codes, not output parsing

Currently, tkn has inconsistent --output support, no non-TTY auto-detection, undocumented exit codes, and no structured JSON input path.

Sub-issues

Ordered by priority / effort-to-value ratio:

Phase 1 — Foundations

1. #2856 — Ship a CONTEXT.md for AI agents (zero code, highest ROI)
2. #2850 — Consistent --output json on all commands + non-TTY auto-detection
3. #2855 — Consistent, documented exit codes

Phase 2 — Enhanced agent support

4. #2851 — Structured JSON input for start commands
5. #2853 — Field selection (--fields) — client-side filtering to reduce agent token consumption
6. #2854 — Agent-friendly logs output (structured NDJSON logs)

Phase 3 — Advanced capabilities

7. #2859 — Adversarial input validation for agent-generated commands
8. #2852 — Runtime schema introspection for agent discovery
9. #2860 — Dry-run and validation for mutating operations
10. #2861 — MCP surface for direct agent integration

References

• You Need to Rewrite Your CLI for AI Agents — Justin Poehnelt (Google)
• The MCP Abstraction Tax — followup on protocol layers

The issue names no files, tests, or entry points. Start by reviewing the linked Phase 1–3 sub-issues and choose one narrowly scoped task; done means implementing and validating that selected sub-issue rather than the entire agent-friendly CLI initiative.