Add max-tokens and max-iterations execution bounds to engine configuration

[Copilot]

Extends execution bounding configuration beyond max-turns to support max-tokens (context window limit) and max-iterations (alternative turn limit nomenclature) based on Claude and Copilot CLI capabilities.

Changes

• EngineConfig extensions: Added MaxTokens and MaxIterations string fields for token budget and iteration limits
• Capability detection: Extended CapabilityProvider interface with SupportsMaxTokens() and SupportsMaxIterations() methods
• Engine support matrix:
  • Claude: max-turns ✓, max-tokens ✓ (documented in CLI)
  • Custom: All flags ✓ (maximum flexibility)
  • Copilot/Codex: None ✓ (pending CLI verification)
• Validation: Added validateMaxTokensSupport() and validateMaxIterationsSupport() with strict/non-strict mode handling (error vs warn)
• Schema: Extended engine_config definition with field examples and descriptions
• Tests: 12 integration tests covering supported/unsupported engine scenarios and capability detection

Usage

engine:
  id: claude
  max-turns: 15      # existing
  max-tokens: 4096   # new: context window limit

engine:
  id: custom
  max-iterations: 3  # new: alternative to max-turns

Unsupported flags produce errors in strict mode, warnings otherwise, maintaining compilation safety while allowing engine capability evolution.

Original prompt

Review the Claude cli documentation and add every flag that allows to bound the execution of the agent to the engine front matter element.

Typically a "max-"...

Look for compatibility for other engines. Centralize in a single interface . Adapt the agentic engine interface to review the flags and refuse unsupported flags in strict mode, warning in non strict.

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.

[pelikhan]

@copilot remove max-iterations, we already have max turns