Conversation context

The context module (context.py) triages which previous messages to include in the prompt. This keeps token usage reasonable while preserving relevant history. One step of the algorithm below lives elsewhere: the recency window (step 2) is applied by _apply_recency_window_talk / _apply_recency_window_db in executor.py, before context.py sees the list.

Selection algorithm

1. Fetch last lookback_count (25) messages for the conversation
2. Apply recency window: if context_recency_hours > 0, exclude messages older than the cutoff while always keeping at least context_min_messages (10)
3. If total <= skip_selection_threshold (3): include all, skip LLM selection
4. Most recent always_include_recent (5) messages are always included
5. Older messages are triaged by a fast model that returns which message IDs are relevant to the current request. Triage runs through the task's own brain, not a hardcoded claude CLI call: a claude_code task triages with Haiku via the CLI subprocess, a native task triages through its own provider's fast completer. If a completer can't be built (e.g. a native task with no API key), triage fails open — all older messages are kept rather than dropped
6. Selected older messages + guaranteed recent messages are combined in chronological order
7. On any triage error: fail open — keep every older message rather than dropping it. Losing context silently is the worse failure, so an unavailable triage costs tokens instead of history

Reply-to messages are force-included regardless of selection. Actions taken (tool use descriptions) are appended after bot responses so Claude can see what it did previously.

Talk context

Talk tasks use a poller-fed local cache (talk_messages table) populated by the talk poller. This avoids redundant API calls. Bot responses are captured in the cache via :result reference IDs.

Cache size is bounded per conversation (talk_cache_max_per_conversation, default 200). The Talk context limit (talk_context_limit, default 100) controls how many messages are fetched from the Talk API.

Email context

Email tasks use DB-based context from completed tasks matching the conversation token.

Configuration

All context settings live in the [conversation] section:

Setting	Default	Purpose
enabled	true	Enable conversation context
lookback_count	25	Messages to consider
skip_selection_threshold	3	Include all if history is this small
selection_model	"fast"	Role alias for relevance matching, resolves to Haiku by default (resolved through the active brain; under native the brain's own fast completer is used)
selection_timeout	30.0	Timeout for LLM selection
use_selection	true	Use LLM selection (false = include all)
always_include_recent	5	Always include this many recent messages
context_truncation	0	Max chars per bot response (0 = no limit)
context_recency_hours	0	Exclude messages older than this (0 = disabled)
context_min_messages	10	Minimum messages to keep when recency filtering
previous_tasks_count	3	Recent unfiltered tasks to inject
talk_context_limit	100	Messages to fetch from Talk API

Background task exclusion

Scheduled and briefing task results are excluded from interactive conversation context. This prevents cron job output from cluttering a user's chat history. The previous_tasks_count setting ensures scheduled/briefing messages aren't completely orphaned when a user replies to them.