Fixed the Agenda Planner and ERI Server assistants losing parts of your input (#1142)

This commit is contained in:
Thorsten Sommer authored and GitHub committed 2026-10-31 15:24:33 +01:00
1 parent 8a4443ab4b
commit 652d3a509f
19 files changed
+1161 -339

No files matched your search

@@ -0,0 +1,82 @@
using AIStudio.Assistants.ERI;
namespace AIStudio.Tests.Assistants.ERI;
/// <summary>
/// Checks how the ERI Server assistant reads which files the model wants to create.
/// </summary>
/// <remarks>
/// Up to v26.10.1, the assistant asked for invalid JSON and took every string of the answer for a file,
/// apart from the word "files". A path with a space was lost that way, and any other key the model came
/// up with became a file of its own.
/// </remarks>
[TestFixture]
public sealed class ERIServerFileListTests
{
private static readonly string[] FILES = ["docker-compose.yml", "src/my-server/main.py", "docs/read me.md"];
private static readonly string[] PATH_WITH_BACKSLASH = [@"src\main.py"];
private static readonly string[] MAIN_PY = ["main.py"];
[Test]
public void AListBetweenProseAndAFenceIsRead()
{
const string ANSWER = """
Here are the files:
```json
{ "files": [ "docker-compose.yml", "src/my-server/main.py", "docs/read me.md" ] }
```
""";
var isRead = ERIServerFileList.TryParse(ANSWER, out var files);
Assert.Multiple(() =>
{
Assert.That(isRead, Is.True);
Assert.That(files, Is.EqualTo(FILES));
});
}
[Test]
public void EscapedCharactersAreDecoded()
{
var isRead = ERIServerFileList.TryParse("""{ "files": [ "src\\main.py" ] }""", out var files);
Assert.Multiple(() =>
{
Assert.That(isRead, Is.True);
Assert.That(files, Is.EqualTo(PATH_WITH_BACKSLASH));
});
}
[Test]
public void EmptyPathsAreLeftOut()
{
var isRead = ERIServerFileList.TryParse("""{ "files": [ "main.py", "", " " ] }""", out var files);
Assert.Multiple(() =>
{
Assert.That(isRead, Is.True);
Assert.That(files, Is.EqualTo(MAIN_PY));
});
}
[TestCase("""[ "main.py" ]""", TestName = "A list without its object")]
[TestCase("""{ "paths": [ "main.py" ] }""", TestName = "A key other than files")]
[TestCase("""{ "files": [] }""", TestName = "An empty list")]
[TestCase("""{ "files": [ "main.py", }""", TestName = "Broken JSON")]
[TestCase("""{ "files": "main.py" }""", TestName = "A single path instead of a list")]
[TestCase("The files are main.py and README.md.", TestName = "No JSON at all")]
public void AnAnswerWithoutAListOfFilesIsRefused(string answer)
{
var isRead = ERIServerFileList.TryParse(answer, out var files);
Assert.Multiple(() =>
{
Assert.That(isRead, Is.False);
Assert.That(files, Is.Empty);
});
}
}
@@ -0,0 +1,90 @@
using AIStudio.Assistants.ERI;
using Microsoft.Extensions.Logging.Abstractions;
namespace AIStudio.Tests.Assistants.ERI;
/// <summary>
/// Checks what the ERI Server assistant reports for each file it is asked to write below the base directory.
/// </summary>
/// <remarks>
/// Up to v26.10.1, a file which was not written only showed up in the log, and an error while writing
/// ended the whole run. Now every file gets an outcome, which the assistant tells the user about.
/// </remarks>
[TestFixture]
public sealed class ERIServerFileWritingTests
{
private const string MAIN_PY = "print('ERI server')\n";
private const string ANSWER = $"## main.py\n\n```python\n{MAIN_PY}```";
private string baseDirectory = string.Empty;
[SetUp]
public void CreateBaseDirectory()
{
this.baseDirectory = Path.Combine(Path.GetTempPath(), $"ai-studio-eri-server-{Guid.NewGuid():N}");
Directory.CreateDirectory(this.baseDirectory);
}
[TearDown]
public void RemoveBaseDirectory()
{
try
{
Directory.Delete(this.baseDirectory, true);
}
catch (IOException)
{
// A temporary directory we could not remove says nothing about the code under test.
}
}
[Test]
public async Task AFileIsWrittenBelowTheBaseDirectory()
{
var outcome = await ERIServerFiles.WriteAsync(this.baseDirectory, "src/my-server/main.py", ANSWER, NullLogger.Instance);
Assert.Multiple(() =>
{
Assert.That(outcome, Is.EqualTo(ERIServerFileOutcome.WRITTEN));
Assert.That(File.ReadAllText(Path.Join(this.baseDirectory, "src", "my-server", "main.py")), Is.EqualTo(MAIN_PY));
});
}
[Test]
public async Task APathLeadingOutsideTheBaseDirectoryIsNotWritten()
{
var file = $"../outside-{Guid.NewGuid():N}.py";
var outcome = await ERIServerFiles.WriteAsync(this.baseDirectory, file, ANSWER, NullLogger.Instance);
Assert.Multiple(() =>
{
Assert.That(outcome, Is.EqualTo(ERIServerFileOutcome.OUTSIDE_BASE_DIRECTORY));
Assert.That(File.Exists(Path.GetFullPath(Path.Join(this.baseDirectory, file))), Is.False);
});
}
[Test]
public async Task AnAnswerWhichBreaksOffWritesNothing()
{
var outcome = await ERIServerFiles.WriteAsync(this.baseDirectory, "main.py", "## main.py\n\n```python\nprint(", NullLogger.Instance);
Assert.Multiple(() =>
{
Assert.That(outcome, Is.EqualTo(ERIServerFileOutcome.NO_COMPLETE_CODE_BLOCK));
Assert.That(File.Exists(Path.Join(this.baseDirectory, "main.py")), Is.False);
});
}
[Test]
public async Task AFileWhichCannotBeWrittenIsReportedInsteadOfEndingTheRun()
{
// A directory of the same name stands where the file would go:
Directory.CreateDirectory(Path.Join(this.baseDirectory, "main.py"));
var outcome = await ERIServerFiles.WriteAsync(this.baseDirectory, "main.py", ANSWER, NullLogger.Instance);
Assert.That(outcome, Is.EqualTo(ERIServerFileOutcome.WRITE_FAILED));
}
}
@@ -0,0 +1,88 @@
using AIStudio.Assistants.ERI;
namespace AIStudio.Tests.Assistants.ERI;
/// <summary>
/// Checks how the ERI Server assistant reads the files of a server out of the answers of the model.
/// </summary>
/// <remarks>
/// Up to v26.10.1, a file such as docker-compose.yml never reached the disk: the regular expression which
/// read its content expected a heading without a hyphen, and the file only showed up in the log. The same
/// expression read up to the last fence of an answer, so a usage example after the file ended up in it.
/// </remarks>
[TestFixture]
public sealed class ERIServerFilesTests
{
private const string COMPOSE_FILE = "services:\n eri:\n build: .\n ports:\n - \"8080:8080\"\n";
[TestCase("## docker-compose.yml", TestName = "A heading with a hyphen")]
[TestCase("## `docker-compose.yml`", TestName = "A heading with the path in backticks")]
[TestCase("### File: deploy/my-server/docker-compose.yml", TestName = "A heading with a label in front of the path")]
[TestCase("Here is the file:", TestName = "No heading at all")]
public void TheContentIsReadWhateverStandsAboveIt(string introduction)
{
var answer = Answer(introduction, string.Empty, "```yaml", COMPOSE_FILE.TrimEnd('\n'), "```");
Assert.That(ERIServerFiles.ReadFileContent(answer), Is.EqualTo(COMPOSE_FILE));
}
[Test]
public void ACodeBlockAfterTheFileStaysOutOfIt()
{
var answer = Answer(
"## main.py",
string.Empty,
"```python",
"print('ERI server')",
"```",
string.Empty,
"Start the server with:",
string.Empty,
"```bash",
"python main.py",
"```");
Assert.That(ERIServerFiles.ReadFileContent(answer), Is.EqualTo("print('ERI server')\n"));
}
[Test]
public void AFileWithCodeBlocksOfItsOwnStaysWholeInALongerFence()
{
var answer = Answer(
"## README.md",
string.Empty,
"````markdown",
"# ERI Server",
string.Empty,
"```bash",
"python main.py",
"```",
"````");
Assert.That(ERIServerFiles.ReadFileContent(answer), Is.EqualTo("# ERI Server\n\n```bash\npython main.py\n```\n"));
}
[Test]
public void AFenceOfTildesIsReadAsWell()
{
var answer = Answer("## main.py", "~~~python", "print('ERI server')", "~~~");
Assert.That(ERIServerFiles.ReadFileContent(answer), Is.EqualTo("print('ERI server')\n"));
}
[Test]
public void ACodeBlockWhichBreaksOffIsNoFile()
{
var answer = Answer("## main.py", "```python", "print('ERI server')");
Assert.That(ERIServerFiles.ReadFileContent(answer), Is.Empty);
}
[Test]
public void AnAnswerWithoutACodeBlockIsNoFile()
{
Assert.That(ERIServerFiles.ReadFileContent("I cannot create this file."), Is.Empty);
}
private static string Answer(params string[] lines) => string.Join('\n', lines);
}
@@ -0,0 +1,81 @@
using AIStudio.Assistants.ERI;
using OperatingSystem = AIStudio.Assistants.ERI.OperatingSystem;
namespace AIStudio.Tests.Assistants.ERI;
/// <summary>
/// Checks that the system prompt of the ERI Server assistant carries every input of the form, in a shape
/// the model can read.
/// </summary>
/// <remarks>
/// Up to v26.10.1, the description of the planned authentication never reached the model. Besides, the
/// heading of the data source ran into the text after it, and all authentication methods stood on a
/// single line.
/// </remarks>
[TestFixture]
public sealed class ERIServerPromptTests
{
private const string AUTH_DESCRIPTION = "Editors log in with their username and password; the public catalog needs no login.";
private static readonly string[] HEADINGS =
[
"# Introduction",
"# Data Source",
"# Authentication and Authorization",
"# Security",
"# Retrieval Processes",
"# Remarks"
];
private static readonly string[] TOKEN_AND_PASSWORD_LINES = ["- Login by static token per user.", "- Login by username and password."];
[Test]
public void EveryHeadingStandsOnItsOwnLine([Values] DataSources dataSource)
{
var lines = Lines(new ERIServerPrompt { DataSource = dataSource, AuthenticationMethods = [Auth.TOKEN] });
Assert.That(lines, Is.SupersetOf(HEADINGS));
}
[Test]
public void EachAuthenticationMethodStandsOnItsOwnLine()
{
var lines = Lines(new ERIServerPrompt { AuthenticationMethods = [Auth.TOKEN, Auth.USERNAME_PASSWORD] });
Assert.That(lines, Is.SupersetOf(TOKEN_AND_PASSWORD_LINES));
}
[Test]
public void TheDescriptionOfTheAuthenticationReachesTheModel()
{
var lines = Lines(new ERIServerPrompt { AuthenticationMethods = [Auth.NONE, Auth.USERNAME_PASSWORD], AuthDescription = AUTH_DESCRIPTION });
Assert.Multiple(() =>
{
Assert.That(lines, Does.Contain("The authentication process is planned as follows:"));
Assert.That(lines, Does.Contain(AUTH_DESCRIPTION));
});
}
[Test]
public void WithoutADescriptionOfTheAuthenticationItsSectionIsLeftOut()
{
var prompt = new ERIServerPrompt { AuthenticationMethods = [Auth.TOKEN] }.ToSystemPrompt();
Assert.That(prompt, Does.Not.Contain("The authentication process is planned as follows:"));
}
[Test]
public void KerberosStillNamesTheOperatingSystemAfterTheDescription()
{
var lines = Lines(new ERIServerPrompt { AuthenticationMethods = [Auth.KERBEROS, Auth.TOKEN], AuthDescription = AUTH_DESCRIPTION, OperatingSystem = OperatingSystem.LINUX });
Assert.That(lines, Does.Contain("The server will run on Linux operating systems. Keep"));
}
/// <summary>
/// The lines of the system prompt, whatever line endings the source of the prompt was checked out with.
/// </summary>
private static string[] Lines(ERIServerPrompt prompt) => prompt.ToSystemPrompt().ReplaceLineEndings("\n").Split('\n');
}