selfware 0.6.2

Your personal AI workshop — software you own, software that lasts
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174

//! Prompt profiles for benchmark and agent modes.
//!
//! Profiles encapsulate system-prompt and task-prompt generation so that
//! different evaluation pipelines (SWE-bench Pro, local ad-hoc fixes, etc.)
//! can use tailored prompts without scattering prompt strings across the
//! codebase.

#[cfg(feature = "bench-harness")]
use crate::bench_harness::swebench_pro::dataset::{coerce_string_list, Instance};

#[cfg(feature = "bench-harness")]
pub type SwebenchProInstance = Instance;

/// Named prompt generation strategy.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum PromptProfile {
    /// The generic agent prompt (broad, conversational).
    Default,
    /// Tight, tool-oriented prompt optimized for Qwen3.6 local models on
    /// SWE-bench Pro.
    SwebenchPro,
}

impl PromptProfile {
    /// System prompt for the profile.
    pub fn system_prompt(&self) -> String {
        match self {
            PromptProfile::Default => {
                // The default system prompt is injected by the agent layer;
                // returning empty here lets the caller fall back to its own
                // default.
                String::new()
            }
            PromptProfile::SwebenchPro => {
                "You are a software repair agent. Your job is to produce the \
                 simplest correct patch for the issue described.\n\n\
                 RULES:\n\
                 1. Output a valid tool call ONLY, or a concise final answer.\n\
                 2. NEVER write prose, reasoning, or explanations before a tool XML tag.\n\
                 3. Make the smallest code change that resolves the issue.\n\
                 4. Verify honestly; do not gold-plate or refactor unrelated code.\n\
                 5. Do NOT modify test files.\n\
                 6. Run relevant tests and confirm they pass before finishing.\n"
                    .to_string()
            }
        }
    }

    /// Task prompt for a SWE-bench Pro instance.
    ///
    /// * `mode` – `"diagnostic"` includes fail-to-pass tests and selected test
    ///   files; `"official"` omits them.
    #[cfg(feature = "bench-harness")]
    pub fn task_prompt(&self, inst: &SwebenchProInstance, mode: &str) -> String {
        match self {
            PromptProfile::Default => build_default_prompt(inst, mode),
            PromptProfile::SwebenchPro => build_swebench_pro_prompt(inst, mode),
        }
    }
}

#[cfg(feature = "bench-harness")]
fn build_default_prompt(inst: &SwebenchProInstance, mode: &str) -> String {
    let trimmed = inst.problem_statement.trim();
    // Unwrap at most ONE balanced surrounding pair of quotes (a wrapping
    // artifact) instead of greedily stripping every leading/trailing quote,
    // which mangled statements that legitimately begin or end with a quote.
    let problem = trimmed
        .strip_prefix('"')
        .and_then(|s| s.strip_suffix('"'))
        .unwrap_or(trimmed)
        .replace("\\n", "\n");

    if mode == "official" {
        format!(
            "[mode: official]\n\n\
             You are working on a real codebase in the current directory. \
             Resolve this issue:\n\n{problem}\n\n\
             Steps:\n\
             1. Read the codebase to understand the expected behavior.\n\
             2. Make the smallest code change that resolves the issue.\n\
             3. Do NOT modify the test files themselves.\n\
             4. Run tests if possible to verify your fix.\n\
             5. When done, summarize what you changed."
        )
    } else {
        let tests = coerce_string_list(&inst.selected_test_files_to_run);
        let fail = coerce_string_list(&inst.fail_to_pass);

        let fail_str = fail
            .iter()
            .map(|t| format!("  - {}", t))
            .collect::<Vec<_>>()
            .join("\n");
        let test_str = tests.join(", ");

        format!(
            "[mode: diagnostic]\n\n\
             You are working on a real codebase in the current directory. \
             Resolve this issue:\n\n{problem}\n\n\
             The fix needs to make these tests pass:\n{fail_str}\n\n\
             Relevant test files: {test_str}\n\n\
             Steps:\n\
             1. Read the failing test files to understand the expected behavior.\n\
             2. Read the implementation files mentioned in the tests.\n\
             3. Make the smallest code change that resolves the issue.\n\
             4. Do NOT modify the test files themselves.\n\
             5. Run the failing tests if possible to verify your fix.\n\
             6. When done, summarize what you changed."
        )
    }
}

#[cfg(feature = "bench-harness")]
fn build_swebench_pro_prompt(inst: &SwebenchProInstance, mode: &str) -> String {
    let trimmed = inst.problem_statement.trim();
    // Unwrap at most ONE balanced surrounding pair of quotes (a wrapping
    // artifact) instead of greedily stripping every leading/trailing quote,
    // which mangled statements that legitimately begin or end with a quote.
    let problem = trimmed
        .strip_prefix('"')
        .and_then(|s| s.strip_suffix('"'))
        .unwrap_or(trimmed)
        .replace("\\n", "\n");

    if mode == "official" {
        format!(
            "[mode: official]\n\n\
             Resolve the issue in the current directory.\n\n\
             ISSUE:\n{problem}\n\n\
             CONTRACT:\n\
             - Valid tool call ONLY, or concise final summary.\n\
             - NO prose before tool XML.\n\
             - Simplest correct patch; verify honestly; do not gold-plate.\n\
             - Do NOT modify test files.\n\
             - Run relevant tests and confirm they pass before finishing."
        )
    } else {
        let tests = coerce_string_list(&inst.selected_test_files_to_run);
        let fail = coerce_string_list(&inst.fail_to_pass);

        let fail_str = if fail.is_empty() {
            "  (none specified)".to_string()
        } else {
            fail.iter()
                .map(|t| format!("  - {}", t))
                .collect::<Vec<_>>()
                .join("\n")
        };
        let test_str = if tests.is_empty() {
            "(none specified)".to_string()
        } else {
            tests.join(", ")
        };

        format!(
            "[mode: diagnostic]\n\n\
             Resolve the issue in the current directory.\n\n\
             ISSUE:\n{problem}\n\n\
             FAIL-TO-PASS TESTS:\n{fail_str}\n\n\
             RELEVANT TEST FILES: {test_str}\n\n\
             CONTRACT:\n\
             - Valid tool call ONLY, or concise final summary.\n\
             - NO prose before tool XML.\n\
             - Simplest correct patch; verify honestly; do not gold-plate.\n\
             - Do NOT modify test files.\n\
             - Run the failing tests and confirm they pass before finishing."
        )
    }
}

#[cfg(all(test, feature = "bench-harness"))]
#[path = "../../tests/unit/config/prompt_profiles/prompt_profiles_test.rs"]
mod tests;