Module jit 

Available on crate feature vcs-git only.

Just-in-time (commit-level) defect-induction risk scoring (issue #331).

Where score ranks files at a ref, this module scores a single commit at check-in time — the unit a CI gate actually reviews. It is the static, rule-based counterpart to the machine-learning just-in-time (JIT) defect-prediction models in the literature; no model is trained or persisted, so there is nothing to re-fit as a project ages.

Why static rules rather than a model

The JIT defect-prediction literature (Kamei et al., A Large-Scale Empirical Study of Just-in-Time Quality Assurance, IEEE TSE 39(6), 2013; the systematic survey by Zhao et al. in ACM Computing Surveys 55(4), 2022) is mature and consistently finds commit-level prediction high-value at check-in. The survey’s key tooling caveat is that trained JIT models lose predictive power within about a year and must be re-trained on recent data, so a rule-based scorer with no model to drift is the most maintainable starting point. The signed direction of every term below is taken from that literature, not fitted, so the score needs no retraining.

Features (the Kamei change measures)

Grouped exactly as Kamei et al. group them, with the open-source replications [Commit Guru (Rosen, Grawi & Shihab, FSE 2015 tool demo)] and [McIntosh & Kamei, Are Fix-Inducing Changes a Moving Target?, IEEE TSE 44(5), 2018] confirming the directions on independent corpora:

• Size — lines added/deleted, files touched, diff hunks. Larger changes are more defect-prone (Kamei LA/LD/NF).
• Diffusion — distinct subsystems and directories touched, plus the within-commit change entropy. Scattered changes are riskier (Kamei NS/ND/Entropy).
• History — the touched files’ priors: prior change count, distinct prior authors, prior bug- and security-fix counts, and the composite file-level risk_score. Files with turbulent history induce more defects (Kamei NDEV/NUC; the file priors fold in the #328 composite).
• Experience — the author’s prior commit count, long and recent. This term is negatively signed: experienced authors induce fewer defects (Kamei EXP/REXP, the one robustly protective signal in their models).
• Purpose — whether the commit is a fix (itself defect-prone in Kamei’s FIX), a security fix (weighted higher here), or a revert (corrective, so dampened).

The score is ordinal

score returns a non-negative composite plus its per-group JitContributions (so a consumer sees why a commit scored as it did). Like the file-level risk score it is ordinal: rank commits by it, compare a commit against a project’s own distribution, but do not read the absolute magnitude as a probability. Any change to the term set or weights must bump JIT_SCORE_VERSION.

Scope

score / JitReport cover a real commit (all five groups). Scoring an arbitrary unprovenanced diff (bca vcs jit --diff <file>) is supported as a deliberately partial path (issue #580): a bare diff carries no author, parent, or file history, so only the size and diffusion groups are computable. That path produces a distinct JitDiffReport whose unavailable groups are absent from the type (not present as zero), and whose partial_risk_score is not comparable to a commit score — see JitDiffReport. ML-based JIT and server-side hook integration remain out of scope per issues #331 / #580.

Structs

JitCommit

Structural facts about the scored commit.

JitContributions

Per-group contributions to the composite score. They sum to the score before the non-negative floor (so experience is typically negative); surfaced so a consumer can see which group drove the result.

JitDiffContributions

The contributions available from a bare diff: size and diffusion only. History, experience, and purpose are omitted (no input), so — unlike JitContributions — there is no zero-valued field a consumer could misread as “this group is low risk”.

JitDiffReport

The result of scoring an arbitrary unified diff (issue #580).

JitDiffusion

How widely the change is spread (Kamei NS/ND/Entropy).

JitExperience

The author’s prior activity (Kamei EXP/REXP). Higher means more experience, which lowers the score.

JitFeatures

Every numeric feature of one commit, grouped as Kamei groups them. The score’s purpose term is supplied separately (JitPurpose) so this struct is the pure numeric feature vector.

JitHistory

Priors of the touched files, measured from history before the scored commit (Kamei NDEV/NUC plus the #328 file composite).

JitPurpose

Keyword classification of the commit message (Kamei FIX, plus the security and revert refinements this crate already detects).

JitReport

The full result of scoring one commit: the resolved commit, its features, the per-group contributions, and the composite score.

JitSize

Size of the change (Kamei LA/LD/NF, plus diff hunks).

Enums

JitSource

Which input a JIT report was scored from. Serializes to a lowercase string ("commit" / "diff") so consumers can branch on it.

Constants

JIT_SCHEMA_VERSION

Output-shape version for a JitReport. Bump on any change to the serialized field set.

JIT_SCORE_VERSION

Version of the composite JIT formula. Increment on any change to the term set, weights, or bumps in score. Separate from the file-level RISK_SCORE_VERSION so the two scores version independently.

Functions

score

Compute the composite JIT risk score and its per-group breakdown.

score_diff_features

Compute the partial (size + diffusion only) JIT score for an arbitrary diff, reusing the same score math as a commit so the two terms are computed by one code path. The history, experience, and purpose terms are left at their zero defaults (no input), and their zero contributions are discarded — only size and diffusion survive into the returned JitDiffContributions.