newt-core 0.8.0

Newt-Agent core types, errors, and the NeMoCode-style tier router
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

use super::test_support::{run, tool_heavy, EST};
use super::*;
use crate::agentic::trim::PromptTracker;
use crate::agentic::{calibrate_down, calibrate_up};

#[test]
fn reported_one_and_a_half_ratio_shrinks_the_next_budget() {
    let mut state = CompressState::new();
    let prior_budget = calibrate_down(9_000, state.calibration.ratio(None));
    state.calibration.observe(Some(6_000), 4_000);
    let next_budget = calibrate_down(9_000, state.calibration.ratio(None));
    assert_eq!(prior_budget, 9_000);
    assert_eq!(
        next_budget, 6_000,
        "server usage must tighten the next request"
    );
    assert_eq!(
        calibrate_up(next_budget, state.calibration.ratio(None)),
        9_000
    );
}

#[test]
fn missing_usage_leaves_the_prior_and_learned_ratio_untouched() {
    let mut state = CompressState::new();
    state.calibration.observe(None, 4_000);
    assert_eq!(state.calibration.ratio(None), 1.0);
    assert_eq!(state.calibration.ratio(Some(1.25)), 1.25);
    state.calibration.observe(Some(6_000), 4_000);
    state.calibration.observe(None, 7_000);
    assert_eq!(state.calibration.ratio(None), 1.5);
}

#[test]
fn measured_ratios_above_the_old_cache_clamp_remain_authoritative() {
    let mut state = CompressState::new();
    state.calibration.observe(Some(30_000), 4_000);
    assert_eq!(state.calibration.ratio(None), 7.5);
    assert_eq!(calibrate_down(30_000, state.calibration.ratio(None)), 4_000);
    state.calibration.observe(Some(100), 4_000);
    state.calibration.observe(Some(0), 4_000);
    state.calibration.observe(Some(30_000), 0);
    assert_eq!(
        state.calibration.ratio(None),
        7.5,
        "partial or unusable counts cannot loosen the gate"
    );
}

#[test]
fn overflow_uses_stronger_evidence_or_tightens_the_failed_ratio() {
    let mut state = CompressState::new();
    assert_eq!(state.calibration.overflow(1.0), 1.5);
    assert_eq!(state.calibration.overflow(1.5), 2.25);
    state.calibration.observe(Some(4_000), 1_000);
    assert_eq!(state.calibration.overflow(1.0), 4.0);
    assert_eq!(state.calibration.ratio(None), 4.0);
}

#[test]
fn calibration_uses_measured_usage_before_guessing_an_overflow_ratio() {
    let mut state = CompressState::new();
    state.calibration.observe(Some(6_000), 4_000);
    assert_eq!(
        state.calibration.overflow(1.5),
        1.5,
        "an observed token ratio must not be replaced by a guessed multiplier"
    );
    assert_eq!(state.calibration.ratio(None), 1.5);
}

#[test]
fn repeated_usage_free_overflows_cannot_exceed_the_heuristic_ceiling() {
    let mut state = CompressState::new();
    // Six turns, each with an initial rejection and two failed shrinks.
    for _ in 0..18 {
        state.calibration.overflow(state.calibration.ratio(None));
    }
    let ratio = state.calibration.ratio(None);
    assert!(
        ratio <= 3.0,
        "usage-free overflow guesses must stay at or below 3.0, got {ratio}"
    );
    assert_eq!(
        ratio, 3.0,
        "repeated rejection must still tighten the prior"
    );
    assert_eq!(calibrate_down(9_000, ratio), 3_000);
}

#[test]
fn overflow_keeps_authoritative_measurements_above_the_guess_ceiling() {
    let mut state = CompressState::new();
    state.calibration.observe(Some(30_000), 4_000);
    for _ in 0..18 {
        assert_eq!(state.calibration.overflow(7.5), 7.5);
    }
    assert_eq!(calibrate_down(30_000, state.calibration.ratio(None)), 4_000);
}

#[test]
fn discarded_cache_hit_usage_retains_the_observed_sample_policy_on_overflow() {
    for inferred in [false, true] {
        let mut state = CompressState::new();
        if inferred {
            state.calibration.overflow(1.0);
            state.calibration.overflow(1.5);
        }
        let retained = state.calibration.ratio(None);
        // A reported suffix count is an observation, but cannot loosen the
        // retained prior. It still selects usage over an invented multiplier.
        for tokens in [1, 100, 500] {
            state.calibration.observe(Some(tokens), 4_000);
        }
        for _ in 0..18 {
            assert_eq!(state.calibration.overflow(retained), retained);
        }
    }
}

#[tokio::test]
async fn calibration_survives_compaction_anchor_invalidation_and_turn_reuse() {
    let messages = tool_heavy("preserve the operator task", 12, 2_000);
    let mut state = CompressState::new();
    let estimate = estimate_tokens(&messages, EST);
    state
        .calibration
        .observe(Some((estimate * 3) as u32), estimate * 2);
    let mut tracker = PromptTracker::new();
    tracker.record((estimate * 3 / 2) as u32, messages.len());
    let out = run(&messages, estimate / 2, None, None, &mut state).await;
    assert!(out.fired);
    tracker.invalidate();
    let next_turn = state.clone();
    assert_eq!(next_turn.calibration.ratio(None), 1.5);
    assert_eq!(
        tracker.current(&out.messages, None, next_turn.calibration.ratio(None), EST),
        calibrate_up(estimate_tokens(&out.messages, EST), 1.5)
    );
    state.reset();
    assert_eq!(
        state.calibration.ratio(None),
        1.0,
        "a new conversation starts with the cold prior"
    );
}