Skip to content

configuration.md

Latest commit

 

History

History
151 lines (124 loc) · 4.39 KB

configuration.md

File metadata and controls

151 lines (124 loc) · 4.39 KB
Raw

Configuration

re-agent is configured primarily through re-agent.yaml, plus the supported environment variables and runtime CLI flags documented below.

Priority Order

Supported CLI overrides > supported environment variables > YAML config > defaults

Environment Variables

Variable Maps to
RE_AGENT_LLM_PROVIDER llm.provider
RE_AGENT_LLM_API_KEY llm.api_key
RE_AGENT_LLM_MODEL llm.model
RE_AGENT_LLM_BASE_URL llm.base_url
RE_AGENT_BACKEND_CLI_PATH backend.cli_path
RE_AGENT_BACKEND_TIMEOUT backend.timeout_s

LLM Config

llm:
  provider: "claude"        # claude | claude-cli | openai | openai-compat | codex
  model: "claude-sonnet-4-5-20250929"
  api_key: null
  base_url: null
  max_tokens: 4096
  temperature: 0.0
  timeout_s: 1800
  input_cost_per_million: 0.0
  output_cost_per_million: 0.0

Notes:

  • claude uses the Anthropic SDK and typically reads ANTHROPIC_API_KEY
  • openai and openai-compat use the OpenAI-compatible chat completions provider and typically read OPENAI_API_KEY
  • codex uses the local codex CLI and ChatGPT login credentials instead of an API key
  • claude-cli uses the local Claude Code CLI login. cli_path, max_budget_usd, and effort are optional.

Independent role overrides inherit the top-level llm block only when the role is omitted:

agents:
  reverser:
    provider: claude-cli
    model: sonnet
    max_budget_usd: 1.0
    effort: high
  checker:
    provider: codex
    model: gpt-5.4

A present role block is a complete LLMConfig; its individual fields are not merged with the top-level block.

Backend Config

backend:
  type: ghidra-bridge
  cli_path: ghidra-bridge
  timeout_s: 45

The ghidra-ai-bridge package installs the ghidra-bridge executable. Prepare its exports separately before running reversal commands.

Project Profile

The project_profile section makes re-agent work across different RE projects. This example is specifically for GTA-reversed-style source:

project_profile:
  hook_patterns:
    - 'RH_ScopedInstall\s*\(\s*(\w+)\s*,\s*(0x[0-9A-Fa-f]+)'
  stub_markers: ["NOTSA_UNREACHABLE"]
  stub_call_prefix: "plugin::Call"
  source_root: "./source/game_sa"
  source_extensions: [".cpp", ".h", ".hpp"]

Parity Config

parity:
  enabled: true
  call_count_warn_diff: 3
  inline_wrapper_autoskip: false

Orchestrator Config

orchestrator:
  max_review_rounds: 4
  max_functions_per_class: 10
  objective_verifier_enabled: true
  objective_call_count_tolerance: 3
  objective_control_flow_tolerance: 2
  investigation_enabled: true
  max_investigations: 8
  selection_strategy: dependency-order # dependency-order | easiest-first | high-impact
  max_attempts_per_function: 3

Candidate Validation

Generated code is written to a safe overlay. Commands can use both format placeholders and environment variables.

validation:
  enabled: true
  # true copies the project to a temporary isolated directory before commands
  copy_project: false
  project_root: .
  build_commands:
    - 'clang++ -fsyntax-only "{candidate_file}"'
  test_commands: []
  runtime_commands: [] # optional differential/record-replay harness
  require_build: true
  require_tests: false
  require_runtime: false
  require_verified: true # UNKNOWN validation results block acceptance
  trust_configured_commands: false # explicit attestation for project-owned shell gates
  parity_fail_on_red: true
  parity_fail_on_yellow: false
  command_timeout_s: 900
  working_directory: .
  keep_project_copy: false # delete isolated full-project copies after validation

With copy_project: true, the default working directory becomes the isolated project copy and the source candidate replaces the real relative source file. Configure the project inside that copy (for example cmake -S . -B build) before building because generated build/ directories are intentionally not copied. With it disabled, build commands must consume {candidate_file} or RE_AGENT_CANDIDATE_FILE explicitly.

Shell commands are user-defined and cannot be proven meaningful merely by inspecting their text. They therefore produce UNKNOWN until trust_configured_commands: true explicitly attests that the configured project commands compile/test the candidate. The non-isolated placeholder check is an additional mistake detector, not a semantic proof.