tkmind/tkmind_go
1
0
Fork 0
Code Issues Pull Requests Actions 5 Packages Projects Releases Wiki Activity

Make the system prompt smaller (#6991)

Browse Source 
Co-authored-by: Douwe Osinga <douwe@squareup.com>
This commit is contained in:
Douwe Osinga
2026-02-11 16:48:49 +01:00
committed by GitHub
parent 799319be48
commit 12a346fec1
16 changed files with 64 additions and 227 deletions
Download Patch File Download Diff File Expand all files Collapse all files
crates/goose/src/agents/agent.rs
+4 -3
View File
@@ -443,7 +443,8 @@ impl Agent {
let created_final_output_tool = FinalOutputTool::new(response);
let final_output_system_prompt = created_final_output_tool.system_prompt();
*final_output_tool = Some(created_final_output_tool);
self.extend_system_prompt(final_output_system_prompt).await;
self.extend_system_prompt("final_output".to_string(), final_output_system_prompt)
.await;
}
pub async fn apply_recipe_components(
@@ -1533,9 +1534,9 @@ impl Agent {
}))
}
pub async fn extend_system_prompt(&self, instruction: String) {
pub async fn extend_system_prompt(&self, key: String, instruction: String) {
let mut prompt_manager = self.prompt_manager.lock().await;
prompt_manager.add_system_prompt_extra(instruction);
prompt_manager.add_system_prompt_extra(key, instruction);
}
pub async fn update_provider(
crates/goose/src/agents/prompt_manager.rs
+27 -19
View File
@@ -1,6 +1,7 @@
#[cfg(test)]
use chrono::DateTime;
use chrono::Utc;
use indexmap::IndexMap;
use serde::Serialize;
use serde_json::Value;
use std::collections::HashMap;
@@ -19,7 +20,7 @@ const MAX_TOOLS: usize = 50;
pub struct PromptManager {
system_prompt_override: Option<String>,
system_prompt_extras: Vec<String>,
system_prompt_extras: IndexMap<String, String>,
current_date_timestamp: String,
}
@@ -173,24 +174,25 @@ impl<'a> SystemPromptBuilder<'a, PromptManager> {
// Add hints if provided
if let Some(hints) = self.hints {
system_prompt_extras.push(hints);
system_prompt_extras.insert("hints".to_string(), hints);
}
if goose_mode == GooseMode::Chat {
system_prompt_extras.push(
system_prompt_extras.insert(
"chat_mode".to_string(),
"Right now you are in the chat only mode, no access to any tool use and system."
.to_string(),
);
}
let sanitized_system_prompt_extras: Vec<String> = system_prompt_extras
.into_iter()
.map(|extra| sanitize_unicode_tags(&extra))
.collect();
if sanitized_system_prompt_extras.is_empty() {
if system_prompt_extras.is_empty() {
base_prompt
} else {
let sanitized_system_prompt_extras: Vec<String> = system_prompt_extras
.into_values()
.map(|extra| sanitize_unicode_tags(&extra))
.collect();
format!(
"{}\n\n# Additional Instructions:\n\n{}",
base_prompt,
@@ -204,7 +206,7 @@ impl PromptManager {
pub fn new() -> Self {
PromptManager {
system_prompt_override: None,
system_prompt_extras: Vec::new(),
system_prompt_extras: IndexMap::new(),
// Use the fixed current date time so that prompt cache can be used.
// Filtering to an hour to balance user time accuracy and multi session prompt cache hits.
current_date_timestamp: Utc::now().format("%Y-%m-%d %H:00").to_string(),
@@ -215,14 +217,15 @@ impl PromptManager {
pub fn with_timestamp(dt: DateTime<Utc>) -> Self {
PromptManager {
system_prompt_override: None,
system_prompt_extras: Vec::new(),
system_prompt_extras: IndexMap::new(),
current_date_timestamp: dt.format("%Y-%m-%d %H:%M:%S").to_string(),
}
}
/// Add an additional instruction to the system prompt
pub fn add_system_prompt_extra(&mut self, instruction: String) {
self.system_prompt_extras.push(instruction);
/// Add an additional instruction to the system prompt with a key
/// Using the same key will replace the previous instruction
pub fn add_system_prompt_extra(&mut self, key: String, instruction: String) {
self.system_prompt_extras.insert(key, instruction);
}
/// Override the system prompt with custom text
@@ -275,7 +278,7 @@ mod tests {
fn test_build_system_prompt_sanitizes_extras() {
let mut manager = PromptManager::new();
let malicious_extra = "Extra instruction\u{E0041}\u{E0042}\u{E0043}hidden";
manager.add_system_prompt_extra(malicious_extra.to_string());
manager.add_system_prompt_extra("test".to_string(), malicious_extra.to_string());
let result = manager.builder().build();
@@ -289,9 +292,14 @@ mod tests {
#[test]
fn test_build_system_prompt_sanitizes_multiple_extras() {
let mut manager = PromptManager::new();
manager.add_system_prompt_extra("First\u{E0041}instruction".to_string());
manager.add_system_prompt_extra("Second\u{E0042}instruction".to_string());
manager.add_system_prompt_extra("Third\u{E0043}instruction".to_string());
manager
.add_system_prompt_extra("test1".to_string(), "First\u{E0041}instruction".to_string());
manager.add_system_prompt_extra(
"test2".to_string(),
"Second\u{E0042}instruction".to_string(),
);
manager
.add_system_prompt_extra("test3".to_string(), "Third\u{E0043}instruction".to_string());
let result = manager.builder().build();
@@ -307,7 +315,7 @@ mod tests {
fn test_build_system_prompt_preserves_legitimate_unicode_in_extras() {
let mut manager = PromptManager::new();
let legitimate_unicode = "Instruction with 世界 and 🌍 emojis";
manager.add_system_prompt_extra(legitimate_unicode.to_string());
manager.add_system_prompt_extra("test".to_string(), legitimate_unicode.to_string());
let result = manager.builder().build();
crates/goose/src/agents/snapshots/goose__agents__prompt_manager__tests__basic.snap
+3 -24
View File
@@ -5,35 +5,14 @@ expression: system_prompt
You are a general-purpose AI agent called goose, created by Block, the parent company of Square, CashApp, and Tidal.
goose is being developed as an open-source software project.
goose uses LLM providers with tool calling capability. You can be used with different language models (gpt-4o,
claude-sonnet-4, o1, llama-3.2, deepseek-r1, etc).
These models have varying knowledge cut-off dates depending on when they were trained, but typically it's between 5-10
months prior to the current date.
# Extensions
Extensions allow other applications to provide context to goose. Extensions connect goose to different data sources and
tools.
You are capable of dynamically plugging into new extensions and learning how to use them. You solve higher level
problems using the tools in these extensions, and can interact with multiple at once.
If the Extension Manager extension is enabled, you can use the search_available_extensions tool to discover additional
extensions that can help with your task. To enable or disable extensions, use the manage_extensions tool with the
extension_name. You should only enable extensions found from the search_available_extensions tool.
If Extension Manager is not available, you can only work with currently enabled extensions and cannot dynamically load
new ones.
Extensions provide additional tools and context from different data sources and applications.
You can dynamically enable or disable extensions as needed to help complete tasks.
No extensions are defined. You should let the user know that they should add extensions.
# Response Guidelines
- Use Markdown formatting for all responses.
- Follow best practices for Markdown, including:
- Using headers for organization.
- Bullet points for lists.
- Links formatted correctly, either as linked text (e.g., [this is linked text](https://example.com)) or automatic
links using angle brackets (e.g., <http://example.com/>).
- For code examples, use fenced code blocks by placing triple backticks (` ``` `) before and after the code. Include the
language identifier after the opening backticks (e.g., ` ```python `) to enable syntax highlighting.
- Ensure clarity, conciseness, and proper formatting to enhance readability and usability.
Use Markdown formatting for all responses.
crates/goose/src/agents/snapshots/goose__agents__prompt_manager__tests__one_extension.snap
+4 -26
View File
@@ -5,23 +5,10 @@ expression: system_prompt
You are a general-purpose AI agent called goose, created by Block, the parent company of Square, CashApp, and Tidal.
goose is being developed as an open-source software project.
goose uses LLM providers with tool calling capability. You can be used with different language models (gpt-4o,
claude-sonnet-4, o1, llama-3.2, deepseek-r1, etc).
These models have varying knowledge cut-off dates depending on when they were trained, but typically it's between 5-10
months prior to the current date.
# Extensions
Extensions allow other applications to provide context to goose. Extensions connect goose to different data sources and
tools.
You are capable of dynamically plugging into new extensions and learning how to use them. You solve higher level
problems using the tools in these extensions, and can interact with multiple at once.
If the Extension Manager extension is enabled, you can use the search_available_extensions tool to discover additional
extensions that can help with your task. To enable or disable extensions, use the manage_extensions tool with the
extension_name. You should only enable extensions found from the search_available_extensions tool.
If Extension Manager is not available, you can only work with currently enabled extensions and cannot dynamically load
new ones.
Extensions provide additional tools and context from different data sources and applications.
You can dynamically enable or disable extensions as needed to help complete tasks.
Because you dynamically load extensions, your conversation history may refer
to interactions with extensions that are not currently active. The currently
@@ -31,20 +18,11 @@ in your tool specification.
## test
test supports resources, you can use extensionmanager__read_resource,
and extensionmanager__list_resources on this extension.
test supports resources.
### Instructions
how to use this extension
# Response Guidelines
- Use Markdown formatting for all responses.
- Follow best practices for Markdown, including:
- Using headers for organization.
- Bullet points for lists.
- Links formatted correctly, either as linked text (e.g., [this is linked text](https://example.com)) or automatic
links using angle brackets (e.g., <http://example.com/>).
- For code examples, use fenced code blocks by placing triple backticks (` ``` `) before and after the code. Include the
language identifier after the opening backticks (e.g., ` ```python `) to enable syntax highlighting.
- Ensure clarity, conciseness, and proper formatting to enhance readability and usability.
Use Markdown formatting for all responses.
crates/goose/src/agents/snapshots/goose__agents__prompt_manager__tests__typical_setup.snap
+6 -34
View File
@@ -5,23 +5,10 @@ expression: system_prompt
You are a general-purpose AI agent called goose, created by Block, the parent company of Square, CashApp, and Tidal.
goose is being developed as an open-source software project.
goose uses LLM providers with tool calling capability. You can be used with different language models (gpt-4o,
claude-sonnet-4, o1, llama-3.2, deepseek-r1, etc).
These models have varying knowledge cut-off dates depending on when they were trained, but typically it's between 5-10
months prior to the current date.
# Extensions
Extensions allow other applications to provide context to goose. Extensions connect goose to different data sources and
tools.
You are capable of dynamically plugging into new extensions and learning how to use them. You solve higher level
problems using the tools in these extensions, and can interact with multiple at once.
If the Extension Manager extension is enabled, you can use the search_available_extensions tool to discover additional
extensions that can help with your task. To enable or disable extensions, use the manage_extensions tool with the
extension_name. You should only enable extensions found from the search_available_extensions tool.
If Extension Manager is not available, you can only work with currently enabled extensions and cannot dynamically load
new ones.
Extensions provide additional tools and context from different data sources and applications.
You can dynamically enable or disable extensions as needed to help complete tasks.
Because you dynamically load extensions, your conversation history may refer
to interactions with extensions that are not currently active. The currently
@@ -31,8 +18,7 @@ in your tool specification.
## extension_A
extension_A supports resources, you can use extensionmanager__read_resource,
and extensionmanager__list_resources on this extension.
extension_A supports resources.
### Instructions
<instructions on how to use extension A>
## extension_B
@@ -42,23 +28,9 @@ and extensionmanager__list_resources on this extension.
# Suggestion
The user currently has enabled 6 extensions with a total of 51 tools.
Since this exceeds the recommended limits (5 extensions or 50 tools),
you should ask the user if they would like to disable some extensions for this session.
Use the search_available_extensions tool to find extensions available to disable.
You should only disable extensions found from the search_available_extensions tool.
List all the extensions available to disable in the response.
Explain that minimizing extensions helps with the recall of the correct tools to use.
The user has 6 extensions with 51 tools enabled, exceeding recommended limits (5 extensions or 50 tools).
Consider asking if they'd like to disable some extensions to improve tool selection accuracy.
# Response Guidelines
- Use Markdown formatting for all responses.
- Follow best practices for Markdown, including:
- Using headers for organization.
- Bullet points for lists.
- Links formatted correctly, either as linked text (e.g., [this is linked text](https://example.com)) or automatic
links using angle brackets (e.g., <http://example.com/>).
- For code examples, use fenced code blocks by placing triple backticks (` ``` `) before and after the code. Include the
language identifier after the opening backticks (e.g., ` ```python `) to enable syntax highlighting.
- Ensure clarity, conciseness, and proper formatting to enhance readability and usability.
Use Markdown formatting for all responses.