# Error Handling
Two error paths in codescout. The model's behavior differs.
## `RecoverableError` → `isError: false`
Input-driven, expected failure. The MCP server serialises this as
`isError: false` with a JSON body containing `"error"`, optional
severity-tagged guidance (one of `hint` / `warning` / `must_follow`),
and any structured `extra` fields spliced at the top level.
**Sibling tool calls in the same turn survive** — Claude Code aborts
parallel siblings only when it sees `isError: true`.
Examples of recoverable conditions:
- Unknown symbol name passed to `symbols(name=...)`.
- Topic not found in `get_guide(topic)`.
- Path not found, unsupported file type, empty glob match.
- No index built yet for the queried project.
- User declined an elicitation prompt.
The response body carries a guidance field describing how to recover.
Read it, adjust the call, retry — do not bail the whole sequence.
## `anyhow::bail!` → `isError: true`
Genuine tool failure: LSP server crashed, filesystem error, panic in
the tool, security violation. Fatal — the rest of the turn's parallel
tool calls are at risk.
The model should not blind-retry without changing inputs. The user may
need to intervene.
## How to tell them apart
The MCP response has an `isError` field.
- `isError: true` → fatal; stop and surface to the user.
- `isError: false` → recoverable; read the message + guidance field
(`hint` / `warning` / `must_follow`) and adapt the next call.
For tool authors: return `RecoverableError::new(message)` (chainable
with `.with_hint(...)`, `.with_warning(...)`, `.with_must_follow(...)`,
`.with_extra(key, value)`) when the failure is the model's fault — bad
input it can self-correct. Return `anyhow::bail!` when it's not.
See `src/tools/core/types.rs` for the type definition and
`src/tools/mod.rs::route_tool_error` for the serialisation path.