Skip to content

CodeMode Integration

This page documents the implemented CodeMode strategy in HarnessRunner.

Scope

CodeMode is a harness strategy (strategy=codemode) for MCP-backed tool chaining.

It is not a replacement for default harness planning. The baseline remains strategy=tool_call.

Layer Separation

CodeMode in RLM is intentionally split into three layers:

Layer Owned by Responsibility Requirement in RLM
Harness strategy RLM Code (HarnessRunner) Prompting, guardrails, orchestration, telemetry Always present
MCP bridge contract External MCP server Expose bridge tools consumed by harness Must expose search_tools and call_tool_chain
Bridge implementation Provider specific (UTCP, Cloudflare-based, custom) How bridge is implemented/deployed Any implementation is acceptable if contract matches

This means RLM does not bind to Cloudflare-only or UTCP-only internals. RLM binds to MCP tool names and schemas.

Strategy Matrix

Strategy Planner shape Execution shape Default
tool_call Iterative JSON action loop (tool / final) Multiple tool calls across steps Yes
codemode Single generated JS/TS program One guarded chain execution via MCP No

End-to-End Lifecycle

sequenceDiagram
    participant User as User (/harness run)
    participant HR as HarnessRunner
    participant MCP as MCP Server
    participant LLM as Model

    User->>HR: strategy=codemode, mcp=on
    HR->>MCP: search_tools(task_description, limit=10)
    MCP-->>HR: discovered tool interfaces
    HR->>LLM: task + typed tool surface prompt
    LLM-->>HR: {"code": "..."}
    HR->>HR: guardrail validation
    HR->>MCP: call_tool_chain(code, timeout, max_output_size)
    MCP-->>HR: execution payload
    HR-->>User: final response

Implementation path is in rlm_code/harness/runner.py:

  • HarnessRunner.run(..., strategy="codemode")
  • _run_codemode_strategy(...)
  • _validate_codemode_code(...)

Required MCP Tools

CodeMode requires both tools on the selected MCP server:

  • search_tools
  • call_tool_chain

In strict MCP mode, harness defaults to this allowlist:

  • search_tools
  • list_tools
  • tools_info
  • get_required_keys_for_tool
  • call_tool_chain

This list is defined as HarnessRunner.STRICT_MCP_TOOL_ALLOWLIST.

Provider note:

  • Cloudflare CodeMode packages and UTCP bridges are implementation choices.
  • RLM harness only requires MCP tools with compatible names/inputs.
  • If a provider stack does not expose this bridge contract directly, add an adapter MCP server.

Typed Tool Surface Generation

CodeMode creates a prompt surface from two sources:

  1. search_tools discovery payload (tools[].typescript_interface when available)
  2. MCP tool schemas already visible to harness (input_schema)

The planner then receives:

  • task
  • target server
  • typed declarations for MCP tools on that server
  • constraints blocking import/network/process/fs/eval APIs

Guardrail Validation

Before execution, generated code is validated for:

  • non-empty snippet
  • max code length
  • blocked API patterns
  • max inferred tool-call count

Blocked patterns include:

  • module loading: import, require(...)
  • network APIs: fetch, XMLHttpRequest, WebSocket, http, https, net, dns, tls
  • process APIs: process., child_process, spawn, exec
  • filesystem APIs: fs, path, readFile, writeFile
  • dynamic eval: eval, new Function

If validation fails, harness returns a guardrail block and does not execute call_tool_chain.

Runtime Controls

HarnessRunner.run() supports these CodeMode controls:

Parameter Default Meaning
codemode_timeout_ms 30000 call_tool_chain timeout
codemode_max_output_chars 200000 max output bytes/chars returned from MCP chain
codemode_max_code_chars 12000 max generated code length allowed
codemode_max_tool_calls 30 max inferred tool-call expressions in code

Current slash command surface does not expose these knobs directly; it uses defaults.

Output Semantics

After chain execution:

  • if payload contains success=false + error, harness returns a failure message.
  • if payload contains nonMcpContentResults, harness returns that content.
  • otherwise harness returns raw chain output text.

CLI Usage

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

Strategy coercions in slash commands:

  • strategy=codemode with mcp=off auto-enables MCP with warning.
  • tools=... allowlist is ignored for strategy=codemode.
  • strategy only applies to /rlm bench when mode=harness.

Programmatic API Example

from rlm_code.harness.runner import HarnessRunner

runner = HarnessRunner(llm_connector=llm_connector, mcp_manager=mcp_manager, workdir=workdir)

result = runner.run(
    task="implement task and add tests",
    max_steps=8,
    include_mcp=True,
    strategy="codemode",
    mcp_server="codemode",
    codemode_timeout_ms=30000,
    codemode_max_output_chars=200000,
    codemode_max_code_chars=12000,
    codemode_max_tool_calls=30,
)