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
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

//! Uses the existing real-store overflow fixture to distinguish admission
//! evidence from primary replies and ground committed count recovery.
use super::*;
use crate::backend_probe::TokenCount;

fn measured(tokens: usize) -> TokenCount {
    TokenCount {
        tokens,
        response_bytes: format!("{{ \"input_tokens\" : {tokens} }}\n").into_bytes(),
        method: "llama_chat_input_tokens",
    }
}

fn prepare(harness: &SmartHarness, model: &str) -> (ContentId, Vec<u8>) {
    let body = serde_json::json!({"model":model,"messages":[
        {"role":"user","content":"retained operator prompt"}
    ]});
    harness
        .record_messages(body["messages"].as_array().unwrap())
        .unwrap();
    let bytes = harness.request(&body, "openai").unwrap();
    (harness.state().unwrap().request.unwrap(), bytes)
}

/// A measured rejection has no primary response. Its committed count and
/// recovery decision must remain reachable after the only writer exits.
#[test]
fn committed_token_count_allows_recovery_without_fabricating_a_model_reply() {
    let directory = tempfile::tempdir().unwrap();
    let harness = durable_harness(directory.path());
    let (request, bytes) = prepare(&harness, "counted-model");
    let count = measured(1500);
    harness
        .record_token_count(Ok(&count), 1000)
        .expect("the measured candidate must commit a request-bound count");
    let count_head = harness.head().unwrap();
    assert!(harness.state().unwrap().reply.is_none());
    assert!(harness
        .state()
        .unwrap()
        .session
        .pending_replies()
        .is_empty());
    harness
        .context_exceeded(&signal(Some(900)))
        .expect("an admitted oversized count is sufficient recovery evidence");
    let recovery_head = harness.head().unwrap();
    assert!(harness.state().unwrap().reply.is_none());
    assert!(harness
        .state()
        .unwrap()
        .session
        .pending_replies()
        .is_empty());
    assert!(
        harness.context_exceeded(&signal(None)).is_err(),
        "a consumed admission cannot authorize another recovery"
    );
    drop(harness);

    let restored = Session::restore(directory.path(), recovery_head, "local-session").unwrap();
    assert_eq!(restored.replay(request).unwrap(), bytes);
    assert!(restored.pending_replies().is_empty());
    assert_eq!(
        restored.restored_messages().unwrap(),
        vec![serde_json::json!({
            "role":"user","content":"retained operator prompt"
        })]
    );
    let store = FrameStore::open(directory.path()).unwrap();
    let count_journal = inspect(&store, count_head);
    assert_eq!(
        count_journal.record["request_intervention"]["request"],
        request.to_string()
    );
    let count_event: ContentId = count_journal.record["request_intervention"]["event"]
        .as_str()
        .unwrap()
        .parse()
        .unwrap();
    let event = inspect(&store, count_event);
    assert_eq!(event.record["origin"], "harness");
    assert_eq!(event.record["kind"], "intervention");
    let payload_id = event.record["payload"].as_str().unwrap().parse().unwrap();
    let payload: Value = serde_json::from_slice(&store.source(&payload_id).unwrap()).unwrap();
    assert_eq!(payload["tokens"], 1500);
    assert_eq!(payload["budget"], 1000);
    assert_eq!(payload["method"], count.method);
    assert_eq!(
        payload["response_body"].as_str().unwrap().as_bytes(),
        count.response_bytes
    );
    let recovery_journal = inspect(&store, recovery_head);
    assert_eq!(
        recovery_journal.parents,
        vec![count_head],
        "admission recovery must not insert a fabricated reply or model outcome"
    );
    let recovery_event: ContentId = recovery_journal.record["intervention"]["event"]
        .as_str()
        .unwrap()
        .parse()
        .unwrap();
    assert_eq!(inspect(&store, recovery_event).parents, vec![count_event]);
}

#[test]
fn only_an_admitted_current_request_rejection_can_authorize_count_recovery() {
    let directory = tempfile::tempdir().unwrap();
    let harness = durable_harness(directory.path());
    let count = measured(1500);
    let head = harness.head().unwrap();
    assert!(harness.record_token_count(Ok(&count), 1000).is_err());
    assert_eq!(harness.head().unwrap(), head);
    prepare(&harness, "first");
    harness
        .record_token_count(Ok(&measured(500)), 1000)
        .unwrap();
    let accepted_head = harness.head().unwrap();
    assert!(harness.context_exceeded(&signal(None)).is_err());
    assert_eq!(harness.head().unwrap(), accepted_head);

    let error = anyhow::anyhow!("token-count endpoint returned 401: unauthorized");
    harness.record_token_count(Err(&error), 1000).unwrap();
    let failure_head = harness.head().unwrap();
    assert!(harness.context_exceeded(&signal(None)).is_err());
    assert_eq!(harness.head().unwrap(), failure_head);

    harness.record_token_count(Ok(&count), 1000).unwrap();
    prepare(&harness, "different-template");
    let new_request_head = harness.head().unwrap();
    assert!(harness.context_exceeded(&signal(None)).is_err());
    assert_eq!(harness.head().unwrap(), new_request_head);
}

#[test]
fn committed_tokenizer_capacity_error_allows_recovery_without_a_model_outcome() {
    let directory = tempfile::tempdir().unwrap();
    let harness = durable_harness(directory.path());
    prepare(&harness, "test");
    let error =
        anyhow::anyhow!("token-count endpoint returned 400: Context size has been exceeded");
    harness.record_token_count(Err(&error), 1000).unwrap();
    let count_head = harness.head().unwrap();
    harness.context_exceeded(&signal(None)).unwrap();
    let store = FrameStore::open(directory.path()).unwrap();
    assert_eq!(
        inspect(&store, harness.head().unwrap()).parents,
        vec![count_head]
    );
    assert!(harness
        .state()
        .unwrap()
        .session
        .pending_replies()
        .is_empty());
    assert!(harness.state().unwrap().reply.is_none());
}

/// Grounds recording failure in a blocked real checkpoint locator. The error
/// reports observed count/error evidence as well as the failed persistence.
#[test]
fn token_count_persistence_failure_surfaces_evidence_and_stops_further_requests() {
    for probe_failed in [false, true] {
        let directory = tempfile::tempdir().unwrap();
        let harness = durable_harness(directory.path());
        prepare(&harness, "test");
        harness
            .record_token_count(Ok(&measured(1600)), 1000)
            .unwrap();
        let head = harness.head().unwrap();
        std::fs::rename(
            directory.path().join("heads"),
            directory.path().join("retained-heads"),
        )
        .unwrap();
        std::fs::write(
            directory.path().join("heads"),
            b"blocked checkpoint directory",
        )
        .unwrap();
        let count = measured(1500);
        let observed =
            anyhow::anyhow!("token-count endpoint returned 400: Context size has been exceeded");
        let observation = if probe_failed {
            Err(&observed)
        } else {
            Ok(&count)
        };
        let error = harness.record_token_count(observation, 1000).unwrap_err();
        let diagnostic = format!("{error:#}");
        assert!(diagnostic.contains("token-count evidence"), "{diagnostic}");
        assert!(
            diagnostic.contains("frame storage") || diagnostic.contains("run writer conflict"),
            "retain the failed persistence cause: {diagnostic}"
        );
        assert!(
            diagnostic.contains(if probe_failed {
                "Context size has been exceeded"
            } else {
                "1500"
            }),
            "{diagnostic}"
        );
        assert_eq!(harness.head().unwrap(), head);
        let recovery_error = harness.context_exceeded(&signal(None)).unwrap_err();
        assert!(
            recovery_error.to_string().contains("observed rejection"),
            "a failed replacement count must clear stale admission authority: {recovery_error}"
        );
        assert!(harness
            .request(
                &serde_json::json!({"messages":[{"role":"user","content":"task"}]}),
                "openai"
            )
            .is_err());
    }
}