💻 v0.1.6 — CodeMode Featured Release

CodeMode Featured Release

This release introduces CodeMode as a first-class, opt-in harness strategy in RLM Code.

Highlights

• Added strategy=codemode to harness runs, while keeping strategy=tool_call as the default baseline.
• Implemented CodeMode execution flow:
  • MCP discovery with search_tools
  • typed tool-surface prompt generation
  • single JS/TS program planning
  • guarded execution through call_tool_chain
• Added hard pre-execution guardrails for generated CodeMode programs.
• Added runtime caps for timeout, output size, code size, and tool-call count.
• Added benchmark support to compare tool_call vs codemode on the same preset/cases.
• Added CodeMode-specific telemetry in benchmark case payloads:
  • harness_strategy
  • codemode_chain_calls
  • codemode_search_calls
  • codemode_discovery_calls
  • codemode_guardrail_blocked

New CLI usage

/harness run "implement feature and add tests" steps=8 mcp=on strategy=codemode mcp_server=codemode
/rlm bench preset=dynamic_web_filtering mode=harness strategy=codemode mcp=on mcp_server=codemode

## Safety and rollout posture

- CodeMode remains opt-in.
- Default harness path stays strategy=tool_call.
- Strict MCP allowlist is preserved in harness benchmark path.
- Recommended rollout is side-by-side benchmarking with promotion gates.

## Documentation added

- Top-level CodeMode docs section.
- CodeMode quickstart.
- CodeMode architecture and provider/bridge separation model.
- CodeMode guardrails and release runbook.
- CodeMode evaluation and promotion-gate guidance.
- CodeMode sandbox responsibility and deployment matrix.

## Notes

- No change to default harness behavior unless strategy=codemode is explicitly selected.
- For production use, prefer explicit mcp_server=<name> and hardened bridge deployment.