Skip to main content

zeph_experiments/
error.rs

1// SPDX-FileCopyrightText: 2026 Andrei G <bug-ops>
2// SPDX-License-Identifier: MIT OR Apache-2.0
3
4//! Error types for the experiments module.
5
6/// Errors that can occur during experiment evaluation, benchmark loading, or persistence.
7///
8/// Most variants carry structured context (file paths, token counts, parameter names)
9/// so that callers can surface actionable diagnostics to the user.
10///
11/// # Examples
12///
13/// ```rust
14/// use zeph_experiments::EvalError;
15///
16/// let err = EvalError::BudgetExceeded { used: 1_500, budget: 1_000 };
17/// assert!(err.to_string().contains("1500"));
18///
19/// let err = EvalError::InvalidRadius { radius: -1.0 };
20/// assert!(err.to_string().contains("finite and positive"));
21/// ```
22#[derive(Debug, thiserror::Error)]
23pub enum EvalError {
24    /// The benchmark TOML file could not be opened or read.
25    #[error("failed to load benchmark file {0}: {1}")]
26    BenchmarkLoad(String, #[source] std::io::Error),
27
28    /// The benchmark TOML file could not be parsed.
29    #[error("failed to parse benchmark file {0}: {1}")]
30    BenchmarkParse(String, String),
31
32    /// [`BenchmarkSet::validate`] was called on an empty `cases` vec.
33    ///
34    /// [`BenchmarkSet::validate`]: crate::BenchmarkSet::validate
35    #[error("benchmark set is empty")]
36    EmptyBenchmarkSet,
37
38    /// The cumulative token budget for judge calls was exhausted.
39    ///
40    /// When this error is returned from [`Evaluator::evaluate`], the report will
41    /// have `is_partial = true` and only include cases scored before the budget was hit.
42    ///
43    /// [`Evaluator::evaluate`]: crate::Evaluator::evaluate
44    #[error("evaluation budget exceeded: used {used} of {budget} tokens")]
45    BudgetExceeded { used: u64, budget: u64 },
46
47    /// An LLM call failed (network, auth, timeout, or API error).
48    #[error("LLM error during evaluation: {0}")]
49    Llm(#[from] zeph_llm::LlmError),
50
51    /// The judge model returned a non-finite or structurally invalid score.
52    #[error("judge output parse failed for case {case_index}: {detail}")]
53    JudgeParse {
54        /// Zero-based index of the benchmark case that produced the invalid output.
55        case_index: usize,
56        /// Description of the parse failure (e.g., `"non-finite score: NaN"`).
57        detail: String,
58    },
59
60    /// The internal tokio semaphore used for concurrency control was closed.
61    ///
62    /// This is an internal invariant violation and should never occur in normal usage.
63    #[error("semaphore acquire failed: {0}")]
64    Semaphore(String),
65
66    /// The benchmark file exceeds the 10 MiB size limit.
67    #[error("benchmark file exceeds size limit ({size} bytes > {limit} bytes): {path}")]
68    BenchmarkTooLarge {
69        /// Canonicalized path of the file.
70        path: String,
71        /// Actual file size in bytes.
72        size: u64,
73        /// Maximum allowed size in bytes (currently 10 MiB).
74        limit: u64,
75    },
76
77    /// The benchmark file's canonical path escaped the expected parent directory.
78    ///
79    /// This indicates a symlink traversal attack and is rejected before any file I/O.
80    #[error("benchmark file path escapes allowed directory: {0}")]
81    PathTraversal(String),
82
83    /// A parameter value was outside its declared `[min, max]` range.
84    #[error("parameter out of range: {kind} value {value} not in [{min}, {max}]")]
85    OutOfRange {
86        /// Parameter name (e.g., `"temperature"`).
87        kind: String,
88        /// The value that was rejected.
89        value: f64,
90        /// Minimum allowed value (inclusive).
91        min: f64,
92        /// Maximum allowed value (inclusive).
93        max: f64,
94    },
95
96    /// All variations in the generator's search space have been visited.
97    #[error("search space exhausted: all variations in {strategy} have been visited")]
98    SearchSpaceExhausted {
99        /// Name of the strategy that exhausted (e.g., `"grid"`).
100        strategy: &'static str,
101    },
102
103    /// The [`Neighborhood`] radius was not finite and positive.
104    ///
105    /// [`Neighborhood`]: crate::Neighborhood
106    #[error("invalid neighborhood radius {radius}: must be finite and positive")]
107    InvalidRadius {
108        /// The invalid radius value that was rejected.
109        radius: f64,
110    },
111
112    /// An experiment result could not be persisted to SQLite.
113    #[error("experiment storage error: {0}")]
114    Storage(String),
115}