harn-vm 0.10.127

Async bytecode virtual machine for the Harn programming language
Documentation 
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

use super::cost_context::{
    estimate_json_tokens, project_llm_call_context_breakdown, project_llm_call_tokens,
    LlmContextTokenBreakdown,
};

fn segment_tokens(breakdown: &LlmContextTokenBreakdown, id: &'static str) -> Option<i64> {
    breakdown
        .segments
        .iter()
        .find(|segment| segment.id == id)
        .map(|segment| segment.tokens)
}

#[test]
fn context_breakdown_reports_request_segments_and_matches_projection() {
    let mut opts = crate::llm::api::options::base_opts("openai");
    opts.system = Some("System policy".to_string());
    opts.messages = vec![
        serde_json::json!({"role": "user", "content": "fix the bug"}),
        serde_json::json!({"role": "assistant", "content": "I will inspect"}),
        serde_json::json!({"role": "tool", "content": "test failed"}),
        serde_json::json!({"role": "developer", "content": "keep it small"}),
    ];
    opts.native_tools = Some(vec![serde_json::json!({
        "type": "function",
        "function": {"name": "read_file", "parameters": {"type": "object"}}
    })]);
    opts.provider_tools = vec![serde_json::json!({
        "type": "web_search_preview",
        "search_context_size": "low"
    })];
    opts.max_tokens = 128;

    let breakdown = project_llm_call_context_breakdown(&opts);
    let (input_tokens, output_tokens) = project_llm_call_tokens(&opts);

    assert_eq!(breakdown.schema, "harn.llm.context_token_breakdown.v1");
    assert_eq!(breakdown.message_count, 4);
    assert_eq!(breakdown.native_tool_count, 1);
    assert_eq!(breakdown.provider_tool_count, 1);
    assert_eq!(breakdown.input_tokens, input_tokens);
    assert_eq!(breakdown.output_budget_tokens, output_tokens);
    assert_eq!(
        breakdown.context_tokens,
        input_tokens.saturating_add(output_tokens)
    );
    assert!(segment_tokens(&breakdown, "system_prompt").unwrap_or(0) > 0);
    assert!(segment_tokens(&breakdown, "user_messages").unwrap_or(0) > 0);
    assert!(segment_tokens(&breakdown, "assistant_messages").unwrap_or(0) > 0);
    assert!(segment_tokens(&breakdown, "tool_results").unwrap_or(0) > 0);
    assert!(segment_tokens(&breakdown, "other_messages").unwrap_or(0) > 0);
    assert!(segment_tokens(&breakdown, "native_tool_schemas").unwrap_or(0) > 0);
    assert!(segment_tokens(&breakdown, "provider_tools").unwrap_or(0) > 0);
    assert_eq!(segment_tokens(&breakdown, "output_budget"), Some(128));
}

#[test]
fn context_breakdown_counts_user_role_tool_result_content_as_tool_results() {
    let mut opts = crate::llm::api::options::base_opts("openai");
    opts.messages = vec![
        serde_json::json!({
            "role": "user",
            "content": "ordinary user request that mentions [result of run] without a closing envelope",
        }),
        serde_json::json!({
            "role": "user",
            "content": "[result of run]\nCommand failed\n[end of run result]\n",
        }),
        serde_json::json!({
            "role": "user",
            "content": [
                {
                    "type": "tool_result",
                    "tool_use_id": "toolu_001",
                    "content": "anthropic-shaped result",
                }
            ],
        }),
        serde_json::json!({
            "role": "user",
            "content": [
                {
                    "toolResult": {
                        "toolUseId": "toolu_002",
                        "content": [{"text": "bedrock-shaped result"}],
                    }
                }
            ],
        }),
        serde_json::json!({
            "role": "assistant",
            "content": "next step",
        }),
    ];
    let ordinary_user_tokens = estimate_json_tokens(&opts.messages[0], &opts.model);
    let tool_result_tokens: i64 = opts.messages[1..4]
        .iter()
        .map(|message| estimate_json_tokens(message, &opts.model))
        .sum();

    let breakdown = project_llm_call_context_breakdown(&opts);

    assert_eq!(breakdown.message_count, 5);
    assert_eq!(
        segment_tokens(&breakdown, "user_messages"),
        Some(ordinary_user_tokens)
    );
    assert!(segment_tokens(&breakdown, "assistant_messages").unwrap_or(0) > 0);
    assert_eq!(
        segment_tokens(&breakdown, "tool_results"),
        Some(tool_result_tokens)
    );
    assert_eq!(segment_tokens(&breakdown, "other_messages"), Some(0));
}

#[test]
fn deferred_tools_leave_the_resident_segment_without_changing_the_total() {
    // #7768: a single tool segment made "this tool is deferred, so it costs
    // nothing yet" and "the measurement never fired" the same number. The
    // split is only worth anything if a deferred tool actually moves, so both
    // directions are asserted against the same two-tool request.
    let resident = serde_json::json!({
        "name": "read_file",
        "description": "Read a slice of a file",
        "input_schema": {"type": "object"},
    });
    let deferred = serde_json::json!({
        "name": "run_migration",
        "description": "Run a database migration",
        "input_schema": {"type": "object"},
        "defer_loading": true,
    });

    let mut both_resident = crate::llm::api::options::base_opts("anthropic");
    both_resident.native_tools = Some(vec![resident.clone(), {
        let mut undeferred = deferred.clone();
        undeferred
            .as_object_mut()
            .expect("tool object")
            .remove("defer_loading");
        undeferred
    }]);

    let mut one_deferred = crate::llm::api::options::base_opts("anthropic");
    one_deferred.native_tools = Some(vec![resident, deferred]);

    let all = project_llm_call_context_breakdown(&both_resident);
    let split = project_llm_call_context_breakdown(&one_deferred);

    // The measured zero, and the non-null read that makes it mean something.
    assert_eq!(segment_tokens(&all, "deferred_tool_schemas"), Some(0));
    assert!(
        segment_tokens(&split, "deferred_tool_schemas").unwrap_or(0) > 0,
        "deferring a tool must put tokens in the deferred segment"
    );
    assert!(
        segment_tokens(&split, "native_tool_schemas").unwrap_or(0)
            < segment_tokens(&all, "native_tool_schemas").unwrap_or(0),
        "the resident segment must shrink by what the deferred segment gained"
    );

    // Counts: every tool is still sent, only residency changed.
    assert_eq!(split.native_tool_count, 2);
    assert_eq!(split.resident_tool_count, 1);
    assert_eq!(all.native_tool_count, all.resident_tool_count);

    // Splitting a segment must not change any budget taken from it: the two
    // parts still add up to what one combined segment reported, which is what
    // `input_tokens` and every ceiling downstream are computed from.
    let combined: i64 = one_deferred
        .native_tools
        .as_ref()
        .expect("native tools set")
        .iter()
        .map(|tool| estimate_json_tokens(&serde_json::json!(tool.to_string()), &one_deferred.model))
        .sum();
    assert_eq!(
        segment_tokens(&split, "native_tool_schemas").unwrap_or(0)
            + segment_tokens(&split, "deferred_tool_schemas").unwrap_or(0),
        combined,
    );
}