Context Scout

Use pre-send workspace discovery to attach relevant files, maps, and task context

Context Scout is a pre-send workspace discovery flow. Before a prompt is sent, Scout scans the active workspace, finds likely relevant files, builds a task brief, and adds selected context to the model input.

Use it when answer quality depends on repository or workspace context and you want Claw to start from better evidence.

When to use Context Scout

Use Scout for:

  • Bug fixes where Claw needs to find the likely fault path
  • Implementation planning before code changes
  • Code review or regression checks
  • Large repositories where manual file selection is slow
  • Prompts that mention behavior but not exact files

Skip Scout for:

  • Simple questions that do not need workspace context
  • Chats with no workspace folder
  • Agent chats
  • Slash commands
  • Repeated follow-ups where the needed files are already in context

Scout is meant to improve the first turn of workspace-aware work, not to run on every message.

Availability

Context Scout is available when:

  • The active chat has a workspace folder
  • The chat is not an agent chat
  • The prompt is not a slash command
  • A provider and model are available

If no workspace is selected, the Scout control prompts you to choose a folder.

Core flow

  1. Select a workspace chat
  2. Open the Context Scout control near the chat input
  3. Turn Scout before send on
  4. Choose mode and depth
  5. Optionally enable Review before send
  6. Send the prompt

After a successful scouted send, Claw turns Scout off for the next turn. Re-enable it when a later prompt needs fresh discovery.

Modes

Choose the mode based on intent.

  • Fix: finds files likely to explain, reproduce, fix, or validate a bug
  • Plan: finds entry points, contracts, and patterns before implementation
  • Review: finds changed or risky paths, tests, shared contracts, and regression risks

The mode changes how Scout prioritizes files and frames the generated task brief.

Depth

Depth controls how wide the discovery pass should go.

  • Lean: narrow pass for the strongest starting points
  • Balanced: default pass with a steadier spread of likely files
  • Deep: wider pass that includes more call sites, contracts, tests, and maps

Practical guidance:

  • Use Lean when you know roughly where the work is
  • Use Balanced for most workspace tasks
  • Use Deep when the request spans multiple modules or the fault path is unclear

What Scout adds

Scout can add:

  • A task brief
  • Suggested starting files
  • File reasons
  • Code maps and symbol summaries
  • Project map entries
  • Model-generated findings, approach notes, risks, or open questions when model preflight is used

Treat Scout output as a strong starting point, not a hard boundary. Claw should still verify facts by reading files before editing.

Model assignment

Context Scout can use a dedicated model through Settings > Model Assignments > Context Scout.

If assigned, Scout uses that model for the preflight investigation. If the assigned provider is unavailable, Claw falls back to the active chat model where possible.

Use a dedicated Scout model when:

  • You want a cheaper or faster model for pre-send discovery
  • Your main chat model should stay focused on execution
  • You want consistent discovery behavior across chats

Native scan and model preflight

Scout has two layers:

  • Native scan: scans the workspace for relevant files and builds file/code-map context
  • Model preflight: uses the selected model to refine the brief and suggest better file hints

If model preflight is unavailable, fails, or stalls, Scout falls back to the native scan so the send can still proceed with useful context.

Review before send

Enable Review before send when you want to inspect or adjust Scout's result before the prompt goes out.

The review drawer lets you:

  • Inspect the generated task brief
  • Review selected files
  • Include or exclude file context
  • Adjust the prompt before sending
  • Cancel the scouted send

Use review mode for higher-risk changes, broad scans, or work where context selection matters.

Language coverage

Scout currently recognizes these language families:

  • TypeScript
  • JavaScript
  • Python
  • Go
  • Rust

Other files can still appear through project maps or adjacent context, but these languages have first-class scanning support.

Interaction with Data Protection

When Data Protection preview is active, Scout prepares discovery from the protected prompt path so sensitive content handling stays aligned with the send flow.

If a workflow handles sensitive data, validate the Data Protection and Scout combination in a low-risk prompt before using it for production work.

Troubleshooting

If Scout is unavailable:

  • Confirm the chat has a workspace folder
  • Confirm you are not inside an agent chat
  • Confirm the prompt is not a slash command
  • Confirm provider/model setup works
  • Confirm the selected Scout model provider is visible and available

If Scout finds weak context:

  • Switch from Lean to Balanced or Deep
  • Rewrite the prompt with more domain terms, file names, or expected behavior
  • Use Review before send and remove weak hits
  • Ask Claw to inspect additional files in the main chat after send

Related detail pages

Workspace Profile Files
Use SOUL.md, IDENTITY.md, and USER.md from workspace roots as local context