albert.toml

Rama main · réplica de lectura del repositorio original.

# Albert runtime config. Its path is set by ALBERT_CONFIG in .env (default: next
# to the binary, else this crate dir). Secrets stay in the environment, named here
# by their env-var name (never the value) — so this file is safe to commit.

# LLM backend — two auth modes (details in docs/configuration.md):
#   auth = "api_key"      -> base_url + openai_key_env (OpenRouter / OpenAI-compatible). The default.
#   auth = "subscription" -> a ChatGPT subscription via the OpenAI Codex backend (OAuth). An option.
auth  = "api_key"
# Optional live model pool; see docs/model-pool.md and config/models.toml.example.
# model_pool = "config/models.toml"
model = "deepseek/deepseek-v4-flash"
base_url = "https://openrouter.ai/api/v1"
openai_key_env = "ALBERT_OPENAI_KEY"

# Vision: can the model see photos sent in chat? Omit for the default heuristic
# (subscription -> always yes; api_key -> guessed from the model name).
# multimodal = true

# The subscription ORGANS — hearing and transcribing (voice messages in chat go to it too),
# speaking, drawing — are connectors, each on while its manifest is present
# (config/connectors/{transcribe,speak,imagegen}), with its settings there. They ride on the
# [subscription].auth_json token, whatever `auth` the brain uses. No transcribe organ ->
# Albert says, with regret, that he can't hear voice messages.

# Live progress in chat while a turn runs (typing indicator is always on; this
# also streams tool calls / thoughts as an in-place edited status message).
# stream_status = true

# --- Subscription mode (option) — set auth = "subscription" above to switch ---
# model = "gpt-5.4"     # Codex slug: gpt-5.6-sol|terra|luna, gpt-5.5, gpt-5.4, gpt-5.4-mini
# [subscription]
# # Reuse the OAuth tokens the `codex` CLI already manages (run `codex login` once);
# # Albert reads auth.json each turn. base_url defaults to the Codex backend.
# auth_json = "~/.codex/auth.json"
# base_url = "https://chatgpt.com/backend-api/codex"

# Owner's timezone (IANA name). The agent's "current time" — and thus its sense
# of "today" and any reminder times it computes — is rendered in this zone rather
# than UTC. Defaults to UTC if omitted. Set the calendar connector's own
# `timezone` (config/connectors/calendar) to match, so listed events read local.
timezone = "Europe/Moscow"

[prompt]
# Persona and operating instructions — hot-reloaded (edit and save; no restart).
# Paths resolve relative to this file.
soul   = "soul.md"
system = "system.md"

[reflection]
period_secs = 21600   # base memory-reflection routine cadence (6h)

# Connectors (Telegram channel, calendar) are config-driven separately, under
# config/connectors/*/*.toml (Octo's from_config_file). See those manifests.

[history]
# Retained session journals. Relative paths follow this config file's directory.
# Session startup requires SQLite; legacy memory/file stores are not auto-converted.
backend = "sqlite:state/history.db"
# Consistent snapshots before internal history/control schema upgrades (WAL included).
backup_dir = "state/history-backups"

[sessions.limits]
temporary_max_sessions = 32
temporary_max_turns = 10000
temporary_max_bytes = 33554432
max_page_size = 100                 # ordinary chat/history pages still default to 10
telegram_tracking_max_scopes = 256
telegram_tracking_max_ids = 10000
telegram_cleanup_batches = 4         # at most 100 exact message IDs per API call
telegram_cleanup_timeout_ms = 3000

[scheduler]
state_path = "state/scheduler.json"

[agent]
max_tool_turns = 30

[skills]
# Declarative skills: a folder of `<name>/SKILL.md` (frontmatter name/description +
# instructions). Albert sees the catalog each turn, lists (skill_list) and applies
# (skill_apply); an applied skill's instructions stay hot in an LRU cache of this
# many. Path resolves relative to this file; a missing folder just means no skills.
dir   = "skills"
cache = 5
# Page size for skill_list only; names and descriptions stay visible every turn.
page  = 10

[code]
# Ephemeral workspace root the file tools (read/write/edit/list/glob/grep) are
# jailed to. Resolved relative to this file; octo-code creates it on first use. A
# future storage/telegram workspace bridge should point at the SAME directory.
workspace = "state/workspace"

# --- Memory backend: embedded kaeru (default), or a dedicated kaeru MCP server.
[memory]
backend = "embedded"
# To use external memory, replace the section above with:
# [memory]
# backend = "mcp"
# transport = "http"
# url = "http://127.0.0.1:9876/mcp"
# timeout_secs = 30
# token_env = "ALBERT_MEMORY_TOKEN"   # optional bearer token, value in .env
#
# Or launch a local MCP stdio process:
# [memory]
# backend = "mcp"
# transport = "stdio"
# command = "kaeru-mcp"
# args = ["--stdio"]
# env = { KAERU_MCP_AUTH_TOKEN = "ALBERT_MEMORY_TOKEN" } # child var -> parent var NAME
#
# At startup, an absent albert initiative is created by migration v1. Existing
# initiatives are untouched. Switching backends does not copy existing memories.

# --- Cloud memory (option) — kaeru-cloud endpoints Albert can share to / pull from.
# Absent = local-only (the default): the share/pull/cloud_recall tools are not even
# shown to the model. Add one [clouds.<name>] per cloud (its URL + the env-var NAME
# of its bearer token; the value stays in .env). Optional [clouds] default picks the
# cloud a tool hits when its `cloud` arg is omitted.
# [clouds]
# default = "family"
#
# [clouds.family]
# url       = "https://cloud.family.example"
# token_env = "ALBERT_CLOUD_FAMILY_TOKEN"
#
# [clouds.work]
# url       = "https://kaeru.work.example"
# token_env = "ALBERT_CLOUD_WORK_TOKEN"

# Bounded one-level delegation. Omit a child's connectors/tools to grant none.
# See docs/subagents.md; model IDs come from the configured model pool.
[subagents]
enabled = true
max_concurrent = 4
max_retained = 64
max_runs_per_turn = 8
max_tool_turns = 17
default_tool_turns = 11
timeout_secs = 600

# Compact journal inspection; full records remain in history storage.
[subagents.inspection]
page_size = 20
preview_chars = 300
artifact_bytes = 8192
read_chars = 8192

# Retained SQLite history: compact + fresh messages, never deletion by turn count.
# Set window_tokens to the model's real limit (pool members can override downward).
[context]
enabled = true
window_tokens = 1000000
reserve_tokens = 32768
response_tokens = 4096
new_messages_percent = 70
compact_output_tokens = 16384
# Optional additional instructions, used for automatic and manual compaction:
# compact_prompt = "Preserve decisions, unresolved tasks, and source links."
tokenizer = "o200k"
image_tokens = 4096
timeout_secs = 300
request_timeout_ms = 180000
continuation_retries = 1
continuation_retry_delay_ms = 1000
max_compaction_passes = 16

# Large tool results are saved before entering a model request.
[context.artifacts]
enabled = true
threshold_bytes = 16384
preview_chars = 2048
read_chars = 8192
max_file_bytes = 67108864
max_scope_bytes = 536870912
retention_secs = 604800

# Optional bounds for the RAM working scratchpad (no SQL persistence or TTL).
# [scratchpad]
# pads = 128
# stages = 128
# notes = 128
# attempts = 32
# text_bytes = 16384
# pad_bytes = 1048576
# render_bytes = 16384
# page_bytes = 65536
# page_size = 32