# Example coral config. Copy to config.toml and tweak.
# Secrets belong in secrets.toml.env, not here.
#
# Launch with:
#   coral --config /path/to/config.toml --secrets /path/to/secrets.toml.env

# ── Model presets ────────────────────────────────────────────────────────────
# Define named model configurations once, then reference them from [model],
# [fallback], [summary], and [embeddings] via `preset = "name"`.
# Secrets (api_key) are deep-merged from the secrets file into the same
# [models.<name>] table, so they don't need to be repeated here.

[models.umans]
provider = "openai"
model = "umans-glm-5.2"
base_url = "https://api.code.umans.ai/v1"
tool_choice = "required"
enable_thinking = true
thinking_effort = "low"

[models.glm_openrouter]
provider = "openrouter"
model = "zai-org/glm-4.7-flash"
base_url = ""
tool_choice = "required"
referer = ""
title = ""
n_ctx = 4096
context_margin = 256
hard_overflow_tokens = 1024
# enforce_context_limit = false

[models.gemini_embedding]
provider = "openai"
model = "google/gemini-embedding-2-preview"
base_url = "https://openrouter.ai/api/v1"
dimensions = 3072
referer = ""
title = ""

# ── LLM bridge (our own OpenRouter) ──────────────────────────────────────────
# Optional unified inference proxy. Accepts OpenAI-format requests and routes to
# the best provider (pi-ai registry + custom providers below). Claude models are
# discovered dynamically from Anthropic's /v1/models — nothing is hardcoded.
# [llm_bridge]
# enabled = true
# host = "127.0.0.1"
# port = 4040
#
# [[llm_bridge.providers]]
# name = "umans"
# api_key = "sk-..."                 # or deep-merged from secrets.toml.env
# base_url = "https://api.code.umans.ai/v1"
#
# [[llm_bridge.providers]]
# name = "anthropic-oauth"
# # pi-ai auth.json: { "anthropic": { "type": "oauth", "refresh", "access", "expires" } }
# # Tokens refresh + rotate back in place, just like codex's ~/.codex/auth.json.
# auth_path = "~/.pi/auth.json"

[server]
port = 3000

[agent]
# Display name used in summarizer prompt and grounding context.
name = "niri"
# Set to "local" to skip the primary model and route everything through the fallback.
env = "default"
# Boredom wake: if > 0, wake the agent after this many minutes of inactivity.
# Any real wake (DM, heartbeat, etc.) resets the timer. 0 disables.
# boredom_wake_min = 0
# Git upstream check: when true, runs `git fetch` on every file read/write
# and nags the agent if the remote has changed. Throttled to 10 min per repo.
# git_upstream_nag = false

# Optional: route shell tools through a Docker container instead of the local shell.
# Set both to enable. Mutually exclusive with [ssh].
[container]
name = ""
user = ""

# Optional: route shell tools through SSH instead. Mutually exclusive with [container].
[ssh]
target = ""
identity = ""

[model]
# Reference a preset defined under [models.<name>], or set fields directly.
preset = "umans"
# Any field here overrides the preset.

[fallback]
# Used when the primary endpoint is unreachable or rate-limited (429/5xx).
preset = "glm_openrouter"

[summary]
# Optional dedicated summarizer model. Leave empty to reuse primary/fallback.
preset = "glm_openrouter"

[embeddings]
# Reference a preset, or set fields directly.
preset = "gemini_embedding"

[context]
compact_target_tokens = 65000
compact_trigger_tokens = 90000
compact_recent_messages = 80
compact_chunk_messages = 32
compact_summary_max_chars = 16000
# 0 disables idle compaction.
idle_compaction_minutes = 45
# When true, no summaries are produced — the conversation stays raw to
# preserve prompt cache. When false, old large tool results are proactively
# summarized using the [summary] model (or primary model if no [summary] is set).
respect_cache = true
# Minimum number of turns (assistant messages) that must pass after a tool
# result before it's eligible for proactive compaction. Only applies when
# respect_cache = false.
compaction_turn_age = 5
# Minimum character length for a tool result to be considered "large" enough
# to summarize. Only applies when respect_cache = false.
compaction_min_chars = 2000

[runner]
# Hard cap on assistant turns in a single wake cycle.
max_turns = 120
# Stop if the same assistant tool call + result pattern repeats.
max_identical_tool_turns = 6

[image_tool]
# Max bytes accepted before rejecting an image attachment.
max_bytes = 150000
# Absolute image root exposed to the agent. Defaults based on container/ssh mode.
root = ""

[discord]
gateway_enabled = true
gateway_trace = false
gateway_raw_fallback = true
gateway_raw_fallback_all = false
# Channel ids scanned by discord_scan when no channel_ids arg is passed.
scan_channel_ids = []
wake_on_event = false
wake_on_dm = true
batch_interval_ms = 60000
batch_only_configured = true
pending_auto_seen_minutes = 10
batch_scan = true
batch_max_messages = 40
rest_max_attempts = 3
rest_retry_base_ms = 1000
# Auto-detected from the gateway READY payload if left blank.
bot_user_id = ""

[metrics]
retention_days = 3

# ── Tool filtering (main agent) ───────────────────────────────────────────────
# Control which tools the main agent can use.
# - enabled_tools: if empty (default), ALL tools are enabled.
#   If non-empty, only the listed tools are available.
# - blacklisted_tools: always removed from the available set, regardless of
#   enabled_tools.
#
# [tools]
# enabled_tools = ["shell", "read_file", "memory_search"]
# blacklisted_tools = ["rest"]

# ── Subagents ─────────────────────────────────────────────────────────────────
# Define named subagents the main agent can spawn via the `subagent` tool.
# Each [[subagents]] entry creates a new subagent with its own system prompt,
# tool set, and turn limit. Subagents run a simplified loop — no event waiting,
# no compaction, no Discord access, no nesting.
#
# [[subagents]]
# name = "researcher"
# model = ""                    # blank = inherit main agent's model
# system_prompt = "You are a research assistant. Be thorough and cite sources."
# enabled_tools = ["shell", "read_file", "memory_search"]   # blank = all (minus auto-blacklisted)
# blacklisted_tools = ["edit_file"]                            # additional exclusions
# max_turns = 30
#
# [[subagents]]
# name = "writer"
# model = ""
# system_prompt = "You are a technical writer. Produce clean, concise prose."
# enabled_tools = []
# blacklisted_tools = []
# max_turns = 20