# Conversation context management

The LangGraph checkpoint retains original messages and artifacts. Each model
call constructs a bounded view; compaction never rewrites the checkpoint.

The prompt budget includes system instructions, saved memories, tool schemas
(including admin tools), message content and tool arguments. The default target
is 16,000 estimated tokens. A 4,096-token generation limit and 2,048-token safety
reserve also constrain the budget against the configured 32,768-token window.

Estimates use UTF-8 bytes divided by three, message overhead, and a fixed image
allowance. These are estimates rather than model-specific tokenizer counts.
Compare `prompt_budget.estimated_prompt_tokens` with actual `stream_request`
model-call usage and adjust the target/window for the deployed backend.

The current user request is always retained. Oversized tool results become
marked excerpts preserving their beginning and end; all tool IDs and arguments
remain. If necessary, intermediate assistant text is also shortened. If the
mandatory request and exchange structure cannot fit, the request fails with an
explicit error rather than silently dropping the question.

Recent history is selected as complete turns. Older user messages, assistant
answers and tool results are capped; earlier images become placeholders.
Incomplete tool exchanges from cancelled calls and orphan results are removed
from the prompt. A small budget is reserved for bounded question/answer excerpts
from older turns. These excerpts are not an LLM-generated rolling summary and do
not guarantee retention of every historical decision. No extra model call is
added to the request path. Saved memories use a separate 1,000-token estimated
budget, preferring recent entries without deleting omitted memories.

Obsidian reads accept `start_line` (one-based content line, excluding frontmatter)
and `max_lines`, allowing focused follow-up reads. Full artifact content remains
available to the interface and checkpoint.

Configuration:

- `CONTEXT_WINDOW_TOKENS`: smallest supported context window across backends.
- `PROMPT_BUDGET_TOKENS`: prompt target, including fixed instructions and tools.
- `RESPONSE_RESERVE_TOKENS`: maximum generated tokens and reserved answer space.

`context_budget` logs history estimates, message counts, dropped turns and excerpt
inclusion. `prompt_budget` logs fixed overhead and the complete prompt estimate.
Neither log includes message content.