# rlmx.yaml — Config for rlmx (RLM algorithm CLI)
# Docs: https://github.com/automagik-dev/rlmx
#
# Prompts live in separate .md files in this directory:
#   SYSTEM.md   — system prompt (LLM behavior & REPL instructions)
#   CRITERIA.md — output criteria (appended to system prompt for FINAL answers)
#   TOOLS.md    — custom Python tools (injected into REPL namespace)

# ─── Model ────────────────────────────────────────────────
# LLM provider and model selection.
# provider: any pi/ai provider (anthropic, openai, google, etc.)
# model: model ID for the main RLM loop
# sub-call-model: model for llm_query() sub-calls (cheaper/faster)
model:
  provider: google
  model: gemini-3.1-flash-lite-preview
  sub-call-model: gemini-3.1-flash-lite-preview

# ─── Context ──────────────────────────────────────────────
# Configure which files are loaded when --context points to a directory.
# extensions: file extensions to include (default: [.md])
# exclude: directories/patterns to skip
context:
  extensions:
    - .md
  exclude:
    - node_modules
    - .git
    - dist

# ─── Budget ───────────────────────────────────────────────
# Cost and resource limits. null = unlimited.
# max-cost: maximum USD spend per run
# max-tokens: maximum total tokens (input + output) per run
# max-depth: maximum recursive rlm_query depth
budget:
  max-cost: null
  max-tokens: null
  max-depth: null

# ─── Tools Level ──────────────────────────────────────────
# Controls which built-in functions are available in the REPL.
# core     — 6 paper functions only (default, paper-faithful)
# standard — core + batteries.py convenience functions + gemini batteries (if google)
# full     — standard + auto-detected package info in system prompt
tools-level: core

# ─── Storage ─────────────────────────────────────────────
# pgserve-backed large context handling.
# enabled: auto (activate when context exceeds model limit) | always | never
# mode: persistent (disk) | memory (in-memory only)
# data-dir: where pgserve stores data
# port: 0 = auto-assign
# chunk-size: null = auto-calculate from model context window
# chunk-utilization: fraction of context window to use per chunk (0-1)
# chars-per-token: character-to-token ratio for estimation
storage:
  enabled: auto
  mode: persistent
  data-dir: ~/.rlmx/data
  port: 0
  chunk-size: null
  chunk-utilization: 0.6
  chars-per-token: 4

# ─── Cache (CAG Mode) ───────────────────────────────────
# Enable cache-augmented generation. Full context is baked into the
# system prompt and cached at the provider for subsequent queries.
# cache:
#   enabled: false             # true to enable CAG mode (or use --cache flag)
#   retention: long            # short|long — maps to pi/ai cacheRetention
#   ttl: 3600                  # seconds — provider-specific TTL
#   expire-time: ""            # ISO 8601 — for Google explicit caching
#   session-prefix: ""         # prepended to content hash for sessionId

# ─── Gemini 3 Native ────────────────────────────────────
# Gemini-specific features. Silently ignored for non-Google providers.
# All features are opt-in and additive — no breaking changes.
# gemini:
#   thinking-level: null         # minimal|low|medium|high — controls thinking depth
#   google-search: false         # Enable Google Search grounding (web_search() battery)
#   url-context: false           # Enable URL Context (fetch_url() battery)
#   code-execution: false        # Enable server-side Python execution
#   media-resolution:            # Control media token costs per type
#     images: auto               # low|medium|high|auto
#     pdfs: auto                 # low|medium|high|auto
#     video: auto                # low|medium|high|auto
#   computer-use: false          # Planned for v0.5
#   maps-grounding: false        # Planned for v0.5
#   file-search: false           # Planned for v0.5

# ─── Output ─────────────────────────────────────────────
# Structured output configuration.
# output:
#   schema:                      # JSON Schema for structured output (Gemini only)
#     type: object               # When set, model output is guaranteed to match schema
#     properties:                # Falls back to FINAL() text parsing on non-Google
#       answer:
#         type: string