channels.toml reference

~/.yachiyo/channels.toml holds channel credentials and tuning, kept separate from config.toml so bot tokens never mix with the rest of your settings.

Keys here are snake_case, unlike config.toml.

`[telegram]`

Key	Type	Description
`enabled`	`boolean`	Whether the Telegram bot runs
`bot_token`	`string`	Token from @BotFather
`model_provider`, `model_name`	`string`	Optional model override for Telegram threads
`[telegram.group]`	table	Group discussion settings — see below

`[qq]`

NapCatQQ over OneBot v11.

Key	Type	Description
`enabled`	`boolean`	Whether the QQ bot runs
`ws_url`	`string`	NapCatQQ forward WebSocket, e.g. `ws://localhost:3001`
`token`	`string`	Optional auth token for the WS connection
`model_provider`, `model_name`	`string`	Optional model override
`[qq.group]`	table	Group discussion settings

`[discord]`

Key	Type	Description
`enabled`	`boolean`	Whether the Discord bot runs
`bot_token`	`string`	Token from the Discord Developer Portal
`model_provider`, `model_name`	`string`	Optional model override
`[discord.group]`	table	Group discussion settings for server text channels

`[qqbot]`

QQ Official Bot API. Direct messages only — no group support.

Key	Type	Description
`enabled`	`boolean`	Whether the QQ Official Bot runs
`app_id`	`string`	App ID from the QQ Developer Portal
`client_secret`	`string`	Client secret
`model_provider`, `model_name`	`string`	Optional model override

DM effort

reasoning_effort is optional in [telegram], [qq], [discord], and [qqbot]. Set it with the Effort selector beside DM model. It applies to owner and guest DM replies, independently of group effort. Changes take effect on the next request without restarting the service or clearing history.

Default leaves the override unset and keeps the conversation’s existing effort behavior. Overrides do not overwrite saved conversation preferences. Owner DMs keep their own conversation model; an effort unsupported by that actual model is ignored rather than forcing a model change.

`[<platform>.group]`

Per-platform group discussion settings. Available for telegram, qq, and discord.

Key	Type	Default	Description
`enabled`	`boolean`	—	Whether group discussion runs on this platform
`mode`	`string`	`probe`	`probe` follows group activity; `mention` runs only on a direct mention of the bot account
`model_provider`, `model_name`	`string`	tool model	Model for the group probe
`vision`	`boolean`	`false`	Pass group images to the probe model
`active_check_interval_ms`	`number`	`60000`	Probe interval in the active phase
`engaged_check_interval_ms`	`number`	`30000`	Probe interval in the engaged phase
`wake_buffer_ms`	`number`	`60000`	Delay before waking on new activity
`dormancy_miss_count`	`number`	`3`	Quiet checks before dropping to dormant
`disengage_miss_count`	`number`	`3`	Quiet checks before leaving engaged
`probe_adapter`, `probe_adapter_provider`, `probe_adapter_model`	`string`	—	Headless probe adapter override

Both modes share the same conversation history and rolling summary. Saving a Probe/Mention mode or Effort change in Settings applies it without restarting the channel service or interrupting the current reply. Other configuration changes may still restart the service.

Set reasoning_effort in a platform’s group section to override effort for group replies only. The Effort selector next to Group model shows that model’s enabled options; Default leaves the override unset. DeepSeek V4 supports Off, Low, High, and Max. Off explicitly disables server-side thinking. Changing the group model clears the override. The headless Claude Code adapter does not expose this selector.

Group context keeps up to 100 recent buffered entries, without a short time cutoff. Each turn adds only the unseen entries; a new conversation starts with the buffered window. Mention mode buffers ordinary messages locally without calling the reply model, and describes their images only when selected for a turn. Images still follow the image-to-text setting. This is a local buffer, not a backfill of messages sent while the service was offline.

`[privacy]`

Key	Type	Description
`guest_instruction`	`string`	Custom context injected into the system prompt for guest conversations
`memory_filter_keywords`	`string[]`	Memory results containing these keywords are hidden from guests

`[image_to_text]`

Key	Type	Default	Description
`enabled`	`boolean`	`false`	Pre-describe images in group messages as alt text

Global group settings

Top-level keys applying across platforms, with per-platform values winning where both are set.

Key	Type	Default	Description
`check_interval_ms`	`number`	—	Global override for the active-phase probe interval
`dm_compact_token_threshold_k`	`number`	`64`	DM context budget, in thousands of tokens
`group_context_window_k`	`number`	`64`	Group probe sliding window, in thousands of tokens
`group_handoff_threshold_k`	`number`	`2 × group_context_window_k`	Probe thread size above which older transcript is summarized into a rolling handoff
`rewrite_model_provider`, `rewrite_model_name`	`string`	—	Model that rewrites outgoing group replies into the persona’s voice. Unset means replies go out as generated.

The handoff threshold’s floor is twice the context window — that gap is hysteresis, so a long-running group does not re-summarize on every turn.

channels.toml reference

[telegram]

[qq]

[discord]

[qqbot]

DM effort

[<platform>.group]

[privacy]

[image_to_text]