Conversation Context

Conversations can carry durable context through file references, attached docs, and binary uploads. The agent sees this context alongside the conversation history.

File References

Type @ in the composer to fuzzy-search workspace files. Matching files appear in a dropdown. Select one to insert a reference.

The referenced file content is injected into the prompt when the message is sent. The agent sees the full file contents, not just the path.

User: @src/auth.ts Can you review this file?
      @src/auth.test.ts And its tests?

Agent sees: contents of auth.ts + auth.test.ts + the prompt

Attached Context Docs

Drag a markdown file into the composer or use the attach button to pin it as conversation-scoped durable context.

Attached docs:

• Persist across all turns in the conversation
• Appear in the Knowledge tree when the workbench is open
• Are loaded by the agent on every turn
• Do not modify the knowledge base — they are scoped to the conversation

Multiple docs can be attached to a single conversation. Remove an attached doc from the Knowledge tree.

Binary Attachments

Images, PDFs, and other binary files can be attached to individual messages:

Images are sent to image-capable models as image content. The composer downscales large images before upload, and text-only image probing caps each prompt at 8 images / 8 MiB per image. For text-only models, Neon Pilot saves the images locally, assigns stable img_<hash> IDs, and injects the probe_image tool only when a preferred vision model is configured in Settings. The agent calls probe_image with one or more image IDs and a focused question; multi-image calls handle comparisons. Other binary formats are stored as conversation attachments but may not be visible to all models.

Existing conversations can also reopen saved drawings through the Open Drawing Picker conversation command, then attach the latest revision or a specific saved revision back into the composer.

Context Loading Order

When the agent builds context for a turn, it merges inputs in this order:

1. Instruction files — standing behavior and policy from the knowledge base, ~/.config/agents, config, or project discovery
2. Attached context docs — durable docs pinned to this conversation
3. Inline @ file references — files referenced in the current message
4. Binary attachments — images and files attached to the current message
5. Conversation history — previous turns in this thread

Later entries have higher priority for conflicting instructions.

Use Cases

• Project setup — attach a CONTEXT.md doc that describes the project
• Code review — @ reference specific files for the agent to analyze
• Design feedback — paste a screenshot and ask for visual feedback
• Multi-file changes — reference several files and ask for coordinated edits

Edit this page on GitHub →