fix(code): record shell output as user context

[Mason Daugherty (mdrxy)]

Local ! commands are user-initiated environment context, not assistant-authored output. Before this change, dcode represented a non-incognito shell command as a synthetic HumanMessage/AIMessage pair:

Human: !git pull
AI: ```text
Already up to date.
```

That made shell output look like something the model said, which is misleading in traces and can confuse downstream history consumers. This switches the model-visible record to a single structured user-context message, matching the pattern used by tools like Codex for user-entered shell commands:

Human: <user_shell_command>
    <command>
        git pull
    </command>
    <result>
        Exit code: 0
        Output:
        Already up to date.
    </result>
</user_shell_command>

Agent-invoked shell tools can still use normal tool-call/tool-result semantics. This change is only for user-entered ! commands that run outside the agent graph and are later flushed into the thread state before the next user turn.

Changes

• Change DeepAgentsApp._buffer_shell_for_model_context to append one structured HumanMessage per ! command instead of a HumanMessage command plus synthetic AIMessage output.
• Include the command, exit status, and combined stdout/stderr in explicit XML-like sections so the model can distinguish user intent from environment output.
• Record Exit code: unknown for interrupted/timeout paths where the subprocess has no return code.
• Preserve the existing privacy and timing behavior: !! incognito commands and --startup-cmd output stay local-only, and non-incognito ! output is still deferred until the next user send.
• Update shell-context tests for successful output, nonzero exits, empty output, interrupted commands, timeout/OSError paths, flush ordering, and retained buffers.

[open-swe]

✅ Open SWE Review: No issues found

Open SWE reviewed this PR and found no potential bugs to report.

Open in Web • View Open SWE trace