[Feature]: Allow to append system prompt

[quan-ngo-ts]

What would you like to be added?

I’d like auggie to support an --append-system-prompt flag, similar to the Claude CLI (--append-system-prompt as documented here: https://code.claude.com/docs/en/cli-reference). This flag would allow users to append additional system-level instructions that persist across long-running tasks.

Why is this needed?

Currently, auggie only supports the --instruction flag for providing an initial instruction. In long-running tasks—such as automated PR reviews—the model may eventually compress or lose parts of the original instruction due to context window limits. This can cause the assistant to deviate from the expected behavior or fail to follow the intended rules toward the end of the run.

Supporting an --append-system-prompt flag would allow users to inject persistent system-level guidance that is less likely to be compacted or deprioritized, resulting in more reliable output for extended tasks.

Possible solution or alternatives

Add a new flag: --append-system-prompt, matching Claude CLI behavior.

Ensure appended system prompts are added at a system or meta-instruction level rather than as user instructions, so they are less prone to compression.

Alternatively, support a configuration file where persistent system rules can be defined and automatically included in each request.

Another option is exposing a dedicated “persistent context” option in the CLI that remains unchanged across chunked or streaming interactions.

Additional context

I frequently use auggie for long-running tasks like PR review automation. Over time, the initial instruction is compressed or deprioritized, leading to outputs that no longer follow the defined rules. Being able to append a system prompt—rather than rely solely on --instruction—would help preserve important constraints and improve reliability in these workflows

[justinxu421]

thanks for the feedback, this is on our backlog, but we do support a --rule flag which is similar here that allows you to enter a path to a file which will add the contents to the system prompt through our rule system

[quan-ngo-ts]

I tried using --rule, but in longer contexts the AI seems to stop respecting the rule—possibly because it gets compressed.
@justinxu421 could you help confirm whether the rule is appended to the system prompt and remains uncompressed when the LLM is using the full context window?

[justinxu421]

I tried using --rule, but in longer contexts the AI seems to stop respecting the rule—possibly because it gets compressed. @justinxu421 could you help confirm whether the rule is appended to the system prompt and remains uncompressed when the LLM is using the full context window?

Rules will get truncated when it is too long, documentation here https://docs.augmentcode.com/setup-augment/guidelines#what-are-the-current-limitations. But similar limits would likely still apply if we applied a --append-system-prompt flag to allow us to maintain a high quality agent

[quan-ngo-ts]

@justinxu421 Thanks! To clarify, I’m asking whether the system prompt/rule itself gets compressed when the AI runs out of context — not about the size limit of the rules.

[justinxu421]

@justinxu421 Thanks! To clarify, I’m asking whether the system prompt/rule itself gets compressed when the AI runs out of context — not about the size limit of the rules.

ah nope, we should only be compressing chat history when we run out of context, never the system prompt