review

package
v0.2.1 Latest Latest
Warning

This package is not in the latest version of its module.

 Go to latest Published: May 26, 2026 License: MIT Imports: 8 Imported by: 0

Documentation

Overview

Package review is the Stage-1 namespace for self-review / critique / quality scoring types in package engine. See ../REFACTOR_PLAN.md.

Index

Constants

This section is empty.

Variables

This section is empty.

Functions

func CompareApproaches 

func CompareApproaches(solutions []Solution) string

CompareApproaches analyzes how different attempts approached the problem and returns a human-readable comparison.

func DefaultScoreFn 

func DefaultScoreFn(solution string) float64

DefaultScoreFn combines length and completeness scoring.

func FormatConsensus 

func FormatConsensus(result *ConsensusResult) string

FormatConsensus produces a human-readable summary of the consensus result.

func FormatInline 

func FormatInline(comments []ReviewComment) string

FormatInline produces GitHub-style inline review comments.

func FormatReport 

func FormatReport(report *ReviewReport) string

FormatReport produces a human-readable summary of a review report.

func FormatReview 

func FormatReview(result *ReviewResult) string

FormatReview produces a formatted summary of the review result.

func FormatSelfAssessment 

func FormatSelfAssessment(a *Assessment) string

FormatSelfAssessment produces a human-readable summary of an assessment.

func PairwiseSimilarity 

func PairwiseSimilarity(a, b string) float64

PairwiseSimilarity computes the Jaccard similarity between two strings based on their word sets.

func ScoreByCompleteness 

func ScoreByCompleteness(content string) float64

ScoreByCompleteness scores based on structural indicators: code blocks, file mentions, numbered steps.

func ScoreByLength 

func ScoreByLength(content string) float64

ScoreByLength scores content based on reasonable length. Too short or too long content gets penalized.

func ShouldRetry 

func ShouldRetry(solutions []Solution) bool

ShouldRetry determines whether additional attempts should be made. Returns true if the best score so far is below 0.7.

Types

type Assessment 

type Assessment struct {
	Score        float64            // overall score 0.0 - 1.0
	Dimensions   map[string]float64 // per-dimension scores
	Strengths    []string           // things that went well
	Weaknesses   []string           // things that went poorly
	Improvements []string           // actionable suggestions for next time
	TaskType     string             // classification of the task
	Timestamp    time.Time          // when the assessment was made
}

Assessment captures the result of a self-evaluation after completing a task.

type Bot 

type Bot = ReviewBot

Bot is the rule-driven review bot for diffs.

func NewBot 

func NewBot() *Bot

NewBot returns a fresh review bot with the default rule set.

type Comment 

type Comment = ReviewComment

Comment is one finding emitted by a Bot.

type ConsensusResult 

type ConsensusResult struct {
	Winner     *Sample
	AllSamples []Sample
	Agreement  float64
	Method     string
}

ConsensusResult holds the outcome of multi-sample consensus.

type ConsensusSampler 

type ConsensusSampler struct {
	NumSamples int
	Strategy   string // "majority", "best_score", "synthesize"
	ScoreFn    func(solution string) float64
	// contains filtered or unexported fields
}

ConsensusSampler implements the multi-sample consensus pattern inspired by SWE-agent's "Ask Colleagues" approach: generate N solutions in parallel, then select the best one using a configurable strategy.

func NewConsensusSampler 

func NewConsensusSampler(numSamples int) *ConsensusSampler

NewConsensusSampler creates a ConsensusSampler with the given number of samples. If numSamples is <= 0, it defaults to 3.

func (*ConsensusSampler) SampleSolutions 

func (cs *ConsensusSampler) SampleSolutions(ctx context.Context, prompt string, generateFn func(context.Context, string) (string, error)) (*ConsensusResult, error)

SampleSolutions generates N solutions in parallel, scores each, and selects a winner based on the configured strategy.

type Critic 

type Critic struct {
	// contains filtered or unexported fields
}

Critic provides fast pre-validation of patches using a cheap model before expensive execution. It generates a prompt for the cheap model and parses the response into a structured verdict.

func NewCritic 

func NewCritic(model string) *Critic

NewCritic creates a new critic that will use the given model for screening.

func (*Critic) BuildPrompt 

func (c *Critic) BuildPrompt(original, patched, intent string) string

BuildPrompt constructs a prompt for the cheap model to evaluate a patch.

func (*Critic) Model 

func (c *Critic) Model() string

Model returns the model name used for pre-screening.

func (*Critic) ParseVerdict 

func (c *Critic) ParseVerdict(response string) *PatchVerdict

ParseVerdict parses a model response into a structured PatchVerdict.

func (*Critic) PreScreenPatch 

func (c *Critic) PreScreenPatch(originalContent, patchedContent, intent string) *PatchVerdict

PreScreenPatch asks the cheap model whether a patch looks correct given the stated intent. It builds a prompt, and returns a verdict. In this implementation, the caller is expected to send the prompt to the model and pass the response to ParseVerdict. This method constructs a PatchVerdict based on a simple heuristic comparison when no model call is available.

func (*Critic) ShouldBlock 

func (c *Critic) ShouldBlock(verdict *PatchVerdict) bool

ShouldBlock returns true if the verdict indicates the patch should be blocked (verdict is "incorrect" with confidence > 0.8).

type PatchVerdict 

type PatchVerdict struct {
	Likely     string   // "correct", "incorrect", "uncertain"
	Issues     []string // specific issues found
	Confidence float64  // 0-1
}

PatchVerdict is the result of a critic's pre-screening of a patch.

type QualityScorer 

type QualityScorer struct {
	Weights ScoreWeights
	History []ScoredResponse
	// contains filtered or unexported fields
}

QualityScorer evaluates LLM response quality across multiple dimensions and provides feedback for the self-improvement loop.

func NewQualityScorer 

func NewQualityScorer() *QualityScorer

NewQualityScorer creates a QualityScorer with default weights.

func (*QualityScorer) AverageScore 

func (qs *QualityScorer) AverageScore(n int) float64

AverageScore computes the average composite score over the last n responses.

func (*QualityScorer) FormatReport 

func (qs *QualityScorer) FormatReport(last int) string

FormatReport generates a formatted quality report for the last n responses.

func (*QualityScorer) GenerateFeedback 

func (qs *QualityScorer) GenerateFeedback(scored *ScoredResponse) []string

GenerateFeedback produces human-readable suggestions based on the scored response.

func (*QualityScorer) Score 

func (qs *QualityScorer) Score(ctx ResponseContext) *ScoredResponse

Score evaluates a response across all quality dimensions and returns a composite result.

func (*QualityScorer) TrendAnalysis 

func (qs *QualityScorer) TrendAnalysis() string

TrendAnalysis returns a human-readable description of quality trends.

type Report 

type Report = ReviewReport

Report aggregates Comments for a single review run.

type ResponseContext 

type ResponseContext struct {
	UserPrompt        string
	AssistantResponse string
	ToolCallCount     int
	ToolErrors        int
	FilesModified     []string
	TestsPassed       bool
	LintPassed        bool
	TokensUsed        int
	Duration          time.Duration
}

ResponseContext provides the context needed to evaluate a response's quality.

type ReviewBot 

type ReviewBot struct {
	Rules    []ReviewRule
	Severity string // minimum severity to report: "error", "warning", "info"
	// contains filtered or unexported fields
}

ReviewBot is a rule-based code review engine that produces structured feedback without requiring an LLM call.

func NewReviewBot 

func NewReviewBot() *ReviewBot

NewReviewBot creates a ReviewBot pre-loaded with 20+ built-in rules.

func (*ReviewBot) ReviewDiff 

func (rb *ReviewBot) ReviewDiff(diffInput string) (*ReviewReport, error)

ReviewDiff parses a unified diff and reviews only changed lines.

func (*ReviewBot) ReviewFile 

func (rb *ReviewBot) ReviewFile(path, content string) (*ReviewReport, error)

ReviewFile reviews a full file's content.

type ReviewComment 

type ReviewComment struct {
	File       string
	Line       int
	Severity   string // "error", "warning", "info"
	Category   string
	Message    string
	Suggestion string
	RuleID     string
}

ReviewComment represents a single piece of review feedback.

func FilterBySeverity 

func FilterBySeverity(comments []ReviewComment, minSeverity string) []ReviewComment

FilterBySeverity returns only comments at or above the specified minimum severity.

type ReviewReport 

type ReviewReport struct {
	Comments      []ReviewComment
	FilesReviewed int
	IssuesFound   int
	BySeverity    map[string]int
	Duration      time.Duration
}

ReviewReport summarizes the results of a code review.

type ReviewResult 

type ReviewResult struct {
	Best          *Solution
	All           []Solution
	Attempts      int
	TotalDuration time.Duration
	TotalTokens   int
	Agreement     float64
}

ReviewResult holds the outcome of the multi-attempt review process.

type ReviewRule 

type ReviewRule struct {
	ID       string
	Name     string
	Category string // "security", "performance", "correctness", "style", "testing"
	Language string
	Check    func(file string, lines []string, diffLines []diff.DiffLine) []ReviewComment
}

ReviewRule defines a single review check that can be applied to code.

type Rule 

type Rule = ReviewRule

Rule is a single check in a Bot's rule set.

type Sample 

type Sample struct {
	ID         int
	Content    string
	Score      float64
	Duration   time.Duration
	TokensUsed int
}

Sample represents a single generated solution with metadata.

func BestScore 

func BestScore(samples []Sample) *Sample

BestScore picks the highest-scored solution.

func MajorityVote 

func MajorityVote(samples []Sample) *Sample

MajorityVote finds the most similar/common solution using pairwise similarity. The sample with the highest average similarity to all others wins.

func Synthesize 

func Synthesize(samples []Sample) *Sample

Synthesize combines elements from all solutions, weighted by score. It selects unique paragraphs from higher-scoring samples first.

type ScoreWeights 

type ScoreWeights struct {
	Completeness float64 // did it address the full request?
	Correctness  float64 // is the code syntactically valid?
	Conciseness  float64 // not overly verbose?
	ToolUsage    float64 // efficient use of tools?
	Safety       float64 // no dangerous operations?
}

ScoreWeights defines the relative importance of each quality dimension. All values should be in [0,1] and sum to 1.

func DefaultWeights 

func DefaultWeights() ScoreWeights

DefaultWeights returns a balanced set of scoring weights.

type ScoredResponse 

type ScoredResponse struct {
	Score     float64            // 0-1 overall composite score
	Breakdown map[string]float64 // per-dimension scores
	Feedback  []string           // human-readable improvement suggestions
	Timestamp time.Time
	Model     string
	TaskType  string
}

ScoredResponse holds the quality evaluation of a single LLM response.

type SelfAssessor 

type SelfAssessor struct {
	History []Assessment
	// contains filtered or unexported fields
}

SelfAssessor evaluates agent performance after each task and tracks trends.

func NewSelfAssessor 

func NewSelfAssessor() *SelfAssessor

NewSelfAssessor creates a new SelfAssessor with an empty history.

func (*SelfAssessor) Assess 

func (sa *SelfAssessor) Assess(ctx TaskContext) *Assessment

Assess evaluates the agent's performance on a completed task across multiple dimensions and records the assessment in history.

func (*SelfAssessor) AverageScore 

func (sa *SelfAssessor) AverageScore(n int) float64

AverageScore computes the average overall score of the last n assessments. If n is 0 or exceeds history length, all assessments are averaged.

func (*SelfAssessor) GetTrend 

func (sa *SelfAssessor) GetTrend(dimension string) string

GetTrend analyzes the trend of a given dimension over the last 10 assessments. Returns "improving", "stable", or "declining".

func (*SelfAssessor) IdentifyStrengths 

func (sa *SelfAssessor) IdentifyStrengths(ctx TaskContext) []string

IdentifyStrengths returns a list of things that went well.

func (*SelfAssessor) IdentifyWeaknesses 

func (sa *SelfAssessor) IdentifyWeaknesses(ctx TaskContext) []string

IdentifyWeaknesses returns a list of things that went poorly.

func (*SelfAssessor) SuggestImprovements 

func (sa *SelfAssessor) SuggestImprovements(ctx TaskContext) []string

SuggestImprovements returns actionable suggestions for future tasks.

type Solution 

type Solution struct {
	ID            int
	Content       string
	Score         float64
	Duration      time.Duration
	TokensUsed    int
	Errors        []string
	FilesModified []string
}

Solution represents a single attempted solution with metadata.

type SolutionReviewer 

type SolutionReviewer struct {
	MaxAttempts int
	ScoreFn     func(solution string) float64
	// contains filtered or unexported fields
}

SolutionReviewer implements a multi-attempt solution review pattern inspired by SWE-agent's reviewer: run the agent N times, score each solution, and select the best one. This improves reliability by sampling multiple approaches.

func NewSolutionReviewer 

func NewSolutionReviewer(maxAttempts int) *SolutionReviewer

NewSolutionReviewer creates a SolutionReviewer with the given max attempts. If maxAttempts is <= 0, it defaults to 3.

func (*SolutionReviewer) ReviewAndSelect 

func (sr *SolutionReviewer) ReviewAndSelect(ctx context.Context, task string, solveFn func(context.Context, string) (*Solution, error)) (*ReviewResult, error)

ReviewAndSelect runs solveFn up to MaxAttempts times, scores each solution, selects the best one, and calculates agreement across attempts.

func (*SolutionReviewer) ScoreSolution 

func (sr *SolutionReviewer) ScoreSolution(solution *Solution) float64

ScoreSolution evaluates a solution using default scoring criteria:

  • Has code changes (+0.3)
  • No errors (+0.3)
  • Reasonable length (+0.2)
  • Files modified (+0.2)

type TaskContext 

type TaskContext struct {
	Goal          string
	ToolCalls     int
	Errors        int
	Retries       int
	FilesModified int
	TestsPassed   bool
	Duration      time.Duration
	TokensUsed    int
	UserFeedback  string
}

TaskContext captures all relevant metrics from a completed task for assessment.

Source Files

View all Source files

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL
go.dev uses cookies from Google to deliver and enhance the quality of its services and to analyze traffic. Learn more.