tga 10.2.0

Developer productivity analytics — git commit collection, classification, and reporting
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
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233

//! #131 / #111: prompt restriction, fail-closed reply validation, and token
//! usage recorded from mocked provider replies. No network.

use crate::classify::rules::CategoryDef;
use crate::classify::tiers::llm::LlmClassifier;
use crate::classify::tiers::llm_prompt::{
    resolve, restricted_system_prompt, LlmOutcome, LlmUsage, ABSTAIN_LABEL,
};
use crate::core::config::LlmEffort;
use wiremock::matchers::{method, path};
use wiremock::{Mock, MockServer, ResponseTemplate};

fn allowed() -> Vec<CategoryDef> {
    vec![
        CategoryDef {
            name: "bug_fix".into(),
            description: Some("Corrects wrong\n  behaviour.".into()),
        },
        CategoryDef {
            name: "enablement".into(),
            description: None,
        },
    ]
}

fn verdict_json(category: &str) -> String {
    format!(
        "{{\"category\":\"{category}\",\"subcategory\":\"x\",\"confidence\":0.8,\"complexity\":2}}"
    )
}

const USAGE: LlmUsage = LlmUsage {
    input_tokens: 120,
    output_tokens: 9,
};

/// Why (#131): the LLM must be offered only the configured categories plus
/// the abstain label, never the built-in list.
/// What: builds the restricted prompt and checks each line.
/// Test: this test.
#[test]
fn restricted_prompt_lists_only_configured_categories() {
    let p = restricted_system_prompt(&allowed());
    assert!(p.contains("- bug_fix: Corrects wrong behaviour.\n"), "{p}");
    assert!(p.contains("- enablement\n"), "{p}");
    assert!(p.contains(&format!("- {ABSTAIN_LABEL}: ")), "{p}");
    assert!(p.contains("<one of: bug_fix|enablement|unclear>"), "{p}");
    let offered: Vec<&str> = p.lines().filter(|l| l.starts_with("- ")).collect();
    assert_eq!(offered.len(), 3, "exactly two categories plus abstain: {p}");
    for builtin in ["feature|", "bugfix|", "chore", "documentation", "refactor|"] {
        assert!(!p.contains(builtin), "built-in `{builtin}` leaked: {p}");
    }
}

/// Why (#131): fail-closed — a category outside the configured set is an
/// abstention and must never be stored as given.
/// What: resolves a `chore` reply against the configured set.
/// Test: this test.
#[test]
fn out_of_set_reply_is_an_abstention() {
    let cats = allowed();
    let call = resolve(Some(&verdict_json("chore")), Some(&cats), Some(USAGE));
    assert_eq!(call.outcome, LlmOutcome::OutOfSet);
    assert!(call.verdict.is_none());
    assert_eq!(call.usage, Some(USAGE), "tokens are billed either way");
}

/// Why (#131): the abstain label is an abstention, whatever its case; a
/// configured name is kept under its configured spelling, without the
/// model's invented subcategory.
/// What: resolves `UNCLEAR` and `Bug_Fix` replies.
/// Test: this test.
#[test]
fn abstain_label_is_an_abstention() {
    let cats = allowed();
    let call = resolve(Some(&verdict_json("UNCLEAR")), Some(&cats), None);
    assert_eq!(call.outcome, LlmOutcome::Abstained);
    assert!(call.verdict.is_none());

    let call = resolve(Some(&verdict_json("Bug_Fix")), Some(&cats), None);
    assert_eq!(call.outcome, LlmOutcome::Answered);
    let v = call.verdict.expect("verdict");
    assert_eq!(v.category, "bug_fix");
    assert_eq!(v.subcategory, None);
}

/// Why: configs without `extend_defaults: false` keep the pre-#131
/// behaviour exactly — a bare object parses, a fenced one does not; only
/// restricted mode extracts the object from a fence.
/// What: resolves plain, fenced and unparseable replies in both modes.
/// Test: this test.
#[test]
fn unrestricted_reply_is_kept() {
    let call = resolve(Some(&format!(" {} ", verdict_json("chore"))), None, None);
    assert_eq!(call.outcome, LlmOutcome::Answered);
    let v = call.verdict.expect("verdict");
    assert_eq!(v.category, "chore");
    assert_eq!(v.subcategory.as_deref(), Some("x"));

    let fenced = format!("```json\n{}\n```", verdict_json("bug_fix"));
    assert_eq!(
        resolve(Some(&fenced), None, None).outcome,
        LlmOutcome::Failed
    );
    let cats = allowed();
    assert_eq!(
        resolve(Some(&fenced), Some(&cats), None).outcome,
        LlmOutcome::Answered
    );

    let call = resolve(Some("not json"), None, Some(USAGE));
    assert_eq!(call.outcome, LlmOutcome::Failed);
    assert_eq!(call.usage, Some(USAGE));
}

async fn anthropic_server(category: &str) -> MockServer {
    let server = MockServer::start().await;
    let body = serde_json::json!({
        "content": [
            {"type": "thinking", "thinking": ""},
            {"type": "text", "text": verdict_json(category)}
        ],
        "stop_reason": "end_turn",
        "usage": {"input_tokens": 120, "output_tokens": 9}
    });
    Mock::given(method("POST"))
        .and(path("/v1/messages"))
        .respond_with(ResponseTemplate::new(200).set_body_json(body))
        .mount(&server)
        .await;
    server
}

fn anthropic_llm(server: &MockServer) -> LlmClassifier {
    LlmClassifier::build_anthropic("claude-sonnet-5", Some("sk-ant-test".to_string())) // pragma: allowlist secret
        .with_endpoint(format!("{}/v1/messages", server.uri()))
}

/// Request body the mock server received first.
async fn sent_body(server: &MockServer) -> serde_json::Value {
    let reqs = server.received_requests().await.expect("recording on");
    serde_json::from_slice(&reqs[0].body).expect("json body")
}

/// Why (#131): the restricted prompt must actually be sent, and an
/// out-of-set reply dropped end to end.
/// What: a classifier restricted to two categories calls a mock Anthropic
/// endpoint that answers `chore`.
/// Test: this test.
#[tokio::test]
async fn classifier_sends_restricted_prompt_and_drops_out_of_set() {
    let server = anthropic_server("chore").await;
    let llm = anthropic_llm(&server).with_allowed_categories(allowed());
    let call = llm.classify_detailed("tidy up").await;
    assert_eq!(call.outcome, LlmOutcome::OutOfSet);
    assert!(call.verdict.is_none());

    let body = sent_body(&server).await;
    let system = body["system"].as_str().expect("system");
    assert!(system.contains("- enablement"), "{system}");
    assert!(!system.contains("chore"), "built-in list sent: {system}");
    assert!(body.get("output_config").is_none(), "no effort unless set");
}

/// Why (#111): every call's tokens come from the provider reply's `usage`.
/// What: mock Anthropic reply with `usage`; also checks `effort` is sent.
/// Test: this test.
#[tokio::test]
async fn anthropic_usage_is_recorded() {
    let server = anthropic_server("feature").await;
    let llm = anthropic_llm(&server).with_effort(Some(LlmEffort::Low));
    let call = llm.classify_detailed("feat: add login").await;
    assert_eq!(call.outcome, LlmOutcome::Answered);
    assert_eq!(call.usage, Some(USAGE));
    assert_eq!(sent_body(&server).await["output_config"]["effort"], "low");
    assert_eq!(llm.provider_label(), "anthropic-api");
    assert_eq!(llm.model(), "claude-sonnet-5");
}

/// Why (#111, owner default): only the Jev source pseudonymizes; every
/// other provider receives the commit message as written.
/// What: a message with an e-mail, a ticket key and a path reaches the
/// mock Anthropic endpoint byte for byte.
/// Test: this test.
#[tokio::test]
async fn non_jev_providers_receive_the_real_message() {
    let server = anthropic_server("feature").await;
    let message = "fix ACME-42 for jane.doe@corp.example.com in src/billing/ledger.rs";
    anthropic_llm(&server).classify_detailed(message).await;
    let sent = sent_body(&server).await["messages"].to_string();
    assert!(sent.contains(message), "{sent}");
    assert!(
        !sent.contains("EMAIL_") && !sent.contains("TICKET_"),
        "{sent}"
    );
}

/// Why (#111): the OpenAI-compatible path reports `prompt_tokens` /
/// `completion_tokens`.
/// What: mock chat-completions reply with `usage`.
/// Test: this test.
#[tokio::test]
async fn openai_usage_is_recorded() {
    let server = MockServer::start().await;
    let body = serde_json::json!({
        "choices": [{"message": {"content": verdict_json("feature")}}],
        "usage": {"prompt_tokens": 120, "completion_tokens": 9, "total_tokens": 129}
    });
    Mock::given(method("POST"))
        .and(path("/v1/chat/completions"))
        .respond_with(ResponseTemplate::new(200).set_body_json(body))
        .mount(&server)
        .await;
    let llm = LlmClassifier::new("gpt-4o-mini", Some("sk-test".to_string()))
        .with_endpoint(format!("{}/v1/chat/completions", server.uri()));
    let call = llm.classify_detailed("feat: add login").await;
    assert_eq!(call.outcome, LlmOutcome::Answered);
    assert_eq!(call.usage, Some(USAGE));
}

/// Why (#131): the non-2xx log line carries the provider's reason, bounded,
/// without splitting a multi-byte character.
/// What: truncates a long body and leaves a short one alone.
/// Test: this test.
#[test]
fn error_body_is_truncated() {
    use crate::classify::tiers::llm::truncate_chars;
    assert_eq!(truncate_chars("short", 500), "short");
    let long = "é".repeat(600);
    let cut = truncate_chars(&long, 500);
    assert_eq!(cut.chars().count(), 501);
    assert!(cut.ends_with('…'));
}