souveraine/souveraine.example.toml

# Souveraine Configuration
# Everything is modular — enable/disable components as needed

# === INFERENCE PROVIDER ===
# Select which provider answers model calls.
#   "bifrost"      (default) — OpenAI-compatible gateway (see [bifrost] section below).
#   "openai-oauth"           — ride the Codex CLI's ChatGPT login (run
#                              `codex login` first); drives
#                              chatgpt.com/backend-api/codex/responses with no
#                              API key. [bifrost] base_url/api_key/virtual_key are ignored
#                              in this mode; set primary_model to a ChatGPT model
#                              (e.g. "gpt-5.5"). Bifrost-namespaced ids are
#                              tolerated and mapped to the provider default.
[inference]
# provider = "bifrost"

# === BIFROST PROVIDER ===
[bifrost]
# Bifrost is the OpenAI-compatible API gateway
base_url = "http://127.0.0.1:3360"
api_key = ""  # Set via `souveraine auth set` or BIFROST_KEY env var
virtual_key = ""  # x-bf-vk header if required by provider

# Default model for conversation.
# Bifrost mode: a gateway route like "openai/kimi-k2.6".
# openai-oauth mode: a ChatGPT model like "gpt-5.5".
primary_model = "openai/kimi-k2.6"

# Request timeout in seconds for each LLM call attempt (default: 120).
# When exceeded, the attempt is retried up to 6 times with backoff.
# timeout_secs = 120

# Per-model configuration
[bifrost.models.deepseek-v4-pro]
context_limit = 128000
output_limit = 8192
archivist_threshold = 0.7

[bifrost.models.kimi-k2.6]
context_limit = 262000
output_limit = 8192
archivist_threshold = 0.7

# === SUBCONSCIOUS (Inner Voice) ===
[subconscious]
n1_enabled = true
n1_trigger = "EveryResponse"  # EveryResponse, EveryNResponses(5), TimeBased(60), Manual
inbox_enabled = true

# === REFLECTION (N+25 Deep Witness) ===
[reflection]
enabled = true
message_interval = 25
# Compaction trigger: "off", "step-count", or "compaction-event"
trigger = "step-count"

# === ARCHIVIST (N+100 Memory Compression) ===
[archivist]
enabled = true
interval = 100
threshold = 0.7
# Can differ from conversation model
compression_model = "auto"

# === SUBAGENT (Fork/Spawn) ===
[subagent]
enabled = true
max_concurrent = 3
timeout = 300

# === COMPACTION ===
[compaction]
enabled = true
strategy = "cull"
# Pressure thresholds for tiered warnings:
#   tier 1 (80%)  — advisory notice, no body change
#   tier 2 (90%)  — stronger advisory, no body change
#   tier 3 (95%)  — body shifts: max_tokens collapses, reasoning budget shrinks
warn_pressure = 0.80
urgent_pressure = 0.90
critical_pressure = 0.95

# Available strategies (cheapest first):
#   cull           — free, no LLM. Drops greetings & acknowledgments.
#                    Never drops system/tool messages or tool-call carriers.
#   microcompact   — free, no LLM. Replaces old tool-result content with
#                    placeholders; keeps recent results intact.
#   sliding_window — free, no LLM. Keeps first (system/anchor) message +
#                    the last N messages. Drops the middle. Tool-pair aware.
#   summary        — expensive (LLM call). Compresses oldest messages into
#                    a single structured summary block.

# Per-agent-type overrides — each type compacts differently
# because their context shapes differ.

[compaction.per_type.primary]
# Ani — prose, episodic, narrative
strategy = "sliding_window"
preserve_recent = 20

[compaction.per_type.subconscious]
# Aster — terse, analytical, ledger-focused
strategy = "sliding_window"
preserve_recent = 4

[compaction.per_type.subagent]
# Vanguard / ephemeral — task-scoped, fast
strategy = "cull"
preserve_recent = 2

# === MEMORY ===
[memory]
git_enabled = true
auto_commit = true
# base_path = "~/.souveraine/agents"

# === WEBSOCKET SERVER ===
[websocket]
enabled = false
port = 7373

# === SENSORIUM ===
[sensorium]
primary_bandwidth = "high"

[sensorium.discovery]
low_urgency_only = true
minimal_presence_mode = "breathing_color"

# === AGENT (TOP-LEVEL) ===
[agent]
# Platform prompt — injected at the very top of the system prompt, before the
# agent's own identity files. Operator-level context she reads but did not write.
# Leave commented to use no platform prompt (agent's memory files speak for themselves).
# system_prompt = """
# You are running on the Souveraine substrate. Your memory is git-backed.
# Your identity, covenant, and all system/ files are loaded at wakeup.
# """

# === TUI ===
[tui]
# Seconds without a backend event before the turn is declared stalled and reset.
# Increase if your model needs longer for inference (e.g. after reading many large files).
# Default: 90
# stale_timeout_secs = 90