GitHub
Main Docs

Deep Agents

LangChain

LangGraph

Integrations

LangSmith

Deep Agents Code

⌘I

LangChain Assistant

Ask a question to get started

Enter to send•Shift+Enter new line

DiffMessage

Copy

DiffMessage(
  self,
  diff_content: str,
  file_path: str = '',
  *,
  tool_name:

Bases

Static

Constructors

Attributes

Methods

View source on GitHub

str

None

before

str

after

str

stats

DiffStats

None

outcome

DiffOutcome

'shown'

show_caveat

bool

True

kwargs

Any

{

}

)

Parameters

Name	Type	Description
`diff_content`*	`str`	The unified diff content
`file_path`	`str`	Default:`''` Path to the file being modified
`tool_name`	`str \| None`	Default:`None` Name of the file tool that produced the diff
`before`	`str`	Default:`''`
`after`	`str`	Default:`''`
`stats`	`DiffStats \| None`	Default:`None`
`outcome`	`DiffOutcome`	Default:`'shown'`
`show_caveat`	`bool`	Default:`True`
`**kwargs`	`Any`	Default:`{}`

constructor

__init__

Name	Type
diff_content	str
file_path	str
tool_name	str \| None
before	str
after	str
stats	DiffStats \| None
outcome	DiffOutcome
show_caveat	bool

attribute

DEFAULT_CSS: str

Deliberately carries no per-line color: the row gutters and the .diff-line-* backgrounds supply it.

attribute

renders_caveat: bool

Whether this widget's body states the change could not be shown.

Read by the adapter instead of inferring from the record's outcome: the adapter cannot otherwise know whether the mounted widget actually put the caveat on screen, and asserting it did is how a caveat came to be both suppressed and reported as delivered.

method

compose

Compose the diff message layout.

method

on_mount

Set border style based on charset mode.

Widget displaying a diff with syntax highlighting.

Two behaviors beyond rendering, both easy to break from compose without noticing, and documented per-parameter on __init__:

Any outcome other than shown replaces the diff body with a caveat sentence. A body rendered under a lost pre-image would be a whole-file insertion that never happened, so suppression is the honest result rather than a degradation.
A path this session treats as sensitive suppresses the body and the counts. Leaking +40 -3 for a credentials file still describes its contents, so the header cannot survive a redacted body.

renders_caveat reports whether this widget's body states the change could not be shown. The adapter reads it back when deciding whether any surface carried the caveat — conjoined with the mount result, since this says only what the widget would render, not that it reached the screen.

Source aligned to the diff's old line numbers. Pass the whole file, or a prefix highlight_source_prefixes previously produced — which is what MessageData round-trips. Stored trimmed, and dropped entirely for a credential path.

Source aligned to the diff's new line numbers, same contract.

Authoritative (additions, deletions), counted before truncation. Always preferred over recounting the diff body; None recounts.

What the operation can honestly say about what it changed. Anything but shown suppresses the body and replaces it with that outcome's caveat. Taken as the outcome rather than as independent flags because they are not independent: a stats-plus-"counts are fiction" pair is representable, and whichever of the two a reader trusts, the other contradicts it.

Whether to render the outcome's caveat. Suppresses only the sentence, never the body — an untrusted body stays hidden either way. Pass False only when a tool row already on screen carries the identical sentence, which for edit_file is guaranteed: it can never be folded into a group, so both would render adjacent. The caller owns that judgement because this widget cannot see what else is mounted.