Context and information control
Context is the material Ghast can see for one request. It affects whether the model can remember the goal, understand files, and use Tools correctly; it is more than the visible chat text.
What can be included
A request can combine:
- system and safety rules;
- Personal Instructions;
- the current message and attachments;
- recent conversation messages;
- summaries of older conversation history;
- Workspace content and Tool results;
- loaded Skill instructions, scripts, and templates;
- descriptions of currently available Tools;
- MCP Tools and Resources enabled for the conversation;
- automatically retrieved Memory;
- web results and relevant Team state;
- intermediate output produced during the run.
The exact set depends on mode, selector state, Model, and runtime. Do not assume every request includes every item.
Three sources of growth
History and results
Long history, complete logs, large diffs, full web pages, and many images consume context quickly. Collapsing a card changes presentation and does not necessarily remove the material already sent.
Capability descriptions
All Tools, multiple MCP Servers, and multiple Skills give the model more descriptions to read and more options to choose between. Disable capabilities that the current task does not need.
Long-term information
Personal Instructions, Memory, automatically loaded Skills, and Team Memory can carry older information into a new task. Keep long-term content short, stable, scoped, and confirmed.
More context can improve grounding but can also add stale instructions, send sensitive data to a Provider, and reduce room for the answer.
Four controls
| Control | Use |
|---|---|
| Current message | Goal, scope, and constraints |
| Composer | Tool, Skill, and MCP selection |
| Chat Settings | Defaults, preferences, compaction |
| Memory / Skill | Reusable facts or methods |
Reduce context deliberately
- Read entry points before broad directories.
- Limit Grep and WebSearch results.
- Ask Ghast to summarize confirmed constraints while preserving paths, IDs, commands, and versions.
- Turn off irrelevant MCP Servers and Skills.
- Split long tasks into stages.
- Record paths, IDs, commands, versions, and constraints before compaction.
Read context status
When the current conversation, model, and usage data are available, the composer shows context usage independently of Show token usage. That Chat Settings switch controls Provider usage details on completed messages. Compacting context can appear while the operation runs. After success, the summary and compaction marker are hidden from the ordinary timeline; do not expect a separate Context compacted, Earlier messages summarized, or Tokens saved receipt. Reopen context usage and compare the numbers. Provider token accounting may not be directly comparable.
When to start a new conversation
Start fresh when:
- the goal has changed;
- the project, branch, account, or external page has changed;
- older logs and errors are no longer relevant;
- a read-only review is becoming an execution task;
- the Model or capability set needs to change completely.
A new conversation does not migrate file state, external state, or an unsaved conclusion. Leave a short, verified handoff first.