harn-stdlib 0.10.150

Embedded Harn standard library source catalog
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

import {
  ConfidenceKind,
  EvaluationPolicy,
  EvaluationQuestion,
  choice,
  evaluation_windows,
} from "std/predicate"

pub type CompactionClassifyItem = {index: int, role: string, content: string}

pub type CompactionClassifyChoice = "keep" | "reword" | "drop"

pub type CompactionClassifyDecision = {
  index: int,
  question_id: string,
  choice: CompactionClassifyChoice,
  confidence: float,
  confidence_kind: ConfidenceKind,
  evaluation_receipt: string,
}

pub type CompactionFixtureInput = {items: list<CompactionClassifyItem>, anchor: string, round: int}

pub type CompactionClassifyDecisions = list<CompactionClassifyDecision>

pub type CompactionClassifyFixture = fn(CompactionFixtureInput) -> CompactionClassifyDecisions

pub type CompactionClassifyRequest = {
  items: list<CompactionClassifyItem>,
  anchor: string,
  round: int,
  budget_tokens: int,
  max_questions: int,
  policy: EvaluationPolicy,
  fixture?: CompactionClassifyFixture,
}

pub type CompactionClassifyRound = {
  kind: "answered" | "unavailable" | "window_refused",
  decisions: list<CompactionClassifyDecision>,
  reason: string,
}

/**
 * Classify one round through the shared evaluator. Compaction owns confidence
 * floors and message application; this export only batches questions and
 * preserves the evaluator's reported answers.
 *
 * @effects: [llm.call]
 * @errors: ["invalid window budget", "invalid classifier answer", "cancelled"]
 */
pub fn classify_compaction_round(
  harness: {fs: HarnessFs, llm: HarnessLlm},
  request: CompactionClassifyRequest,
) -> CompactionClassifyRound {
  const planning = try {
    evaluation_windows(
      harness.llm,
      request.items,
      {
        anchor: request.anchor,
        budget_tokens: request.budget_tokens,
        max_items: request.max_questions,
      },
    )
  }
  if is_err(planning) {
    return {kind: "window_refused", decisions: [], reason: to_string(unwrap_err(planning))}
  }
  const planned = unwrap(planning)
  if request?.fixture != nil {
    return {
      kind: "answered",
      decisions: request.fixture(
        {items: request.items, anchor: request.anchor, round: request.round},
      ),
      reason: "",
    }
  }
  let decisions: list<CompactionClassifyDecision> = []
  for window in planned.windows {
    const items = request.items.slice(window.first_index, window.last_index + 1)
    let questions: dict<string, EvaluationQuestion> = {}
    for item in items {
      questions["message_" + to_string(item.index)] = choice(
        harness.fs.render_prompt(
          "std/agent/prompts/compaction_classify.harn.prompt",
          {index: item.index},
        ),
        {
          keep: "Keep the complete original message because its details remain necessary.",
          reword: "Keep the information in a faithful short rewrite; full detail is unnecessary.",
          drop: "Remove the message because it is irrelevant, redundant, or superseded.",
        },
      )
    }
    const state: {anchor: string, items: list<CompactionClassifyItem>} = {
      anchor: request.anchor,
      items: items,
    }
    const outcome = harness.llm.evaluate_request(
      "agent.compaction.classify.v1",
      state,
      questions,
      request.policy + {threshold: 0.0},
    )
    match outcome.kind {
      "cancelled" -> {
        throw {category: "cancelled", message: "compaction classification cancelled"}
      }
      "answered" -> {
        for item in items {
          const question_id = "message_" + to_string(item.index)
          const answer = outcome.value[question_id]
          if answer.kind != "choice"
            || (answer.choice != "keep" && answer.choice != "reword" && answer.choice != "drop") {
            throw "invalid classifier answer"
          }
          decisions = decisions
            + [
              {
                index: item.index,
                question_id: question_id,
                choice: answer.choice,
                confidence: answer.confidence,
                confidence_kind: answer.confidence_kind,
                evaluation_receipt: outcome.receipt,
              },
            ]
        }
      }
      _ -> { return {
        kind: "unavailable",
        decisions: [],
        reason: outcome.kind + ": " + outcome.receipt,
      } }
    }
  }
  return {kind: "answered", decisions: decisions, reason: ""}
}