Documentation
¶
Overview ¶
Package review implements the MR-review pipeline: git diff → graph evidence → LLM review. Phase 3 of the optimization plan.
Two entry points:
- CLI: `codeiq review --base <ref> --head <ref>` (cli/review.go)
- MCP: `review_changes` tool (mcp/tools_review.go)
Both call ReviewService.Review which:
- Shells out to `git diff` for the file list + hunks.
- For each changed file, queries the graph for nodes-in-file, blast radius, and evidence packs.
- Renders a single prompt and calls the configured LLM (default: Ollama Cloud gpt-oss).
- Returns a structured review report (markdown for CLI, JSON for MCP).
Index ¶
Constants ¶
const SystemPrompt = `You are reviewing a pull request. Use the structured graph evidence to find correctness, security, and architectural issues. ` +
`Return strictly JSON in this shape: ` +
`{"summary": "<one-paragraph overview>", "findings": [{"file": "<path>", "line": <int>, "severity": "info|low|medium|high|critical", "comment": "<message>"}]}. ` +
`No prose before or after the JSON. ` +
`If the diff is trivial, return an empty findings array — do NOT invent issues.`
SystemPrompt is the single system message we use for every review. Plan §3.1 — "use the structured graph evidence to find correctness, security, and architectural issues".
Variables ¶
This section is empty.
Functions ¶
This section is empty.
Types ¶
type ChangedFile ¶
ChangedFile is one entry in the diff.
func GitDiff ¶
func GitDiff(cwd, base, head string) ([]ChangedFile, error)
GitDiff shells out to `git diff --unified=3 <base>..<head>` from cwd and returns the parsed result. Mirrors the Java DiffParser shellout.
func ParseDiff ¶
func ParseDiff(raw string) ([]ChangedFile, error)
ParseDiff parses raw unified `git diff` output into a per-file slice. File renames are reported under the new path. Binary diffs are recorded with empty hunks and zero line counts.
type Client ¶
Client wraps the OpenAI-compatible /chat/completions endpoint exposed by Ollama, Ollama Cloud, and proxies. The HTTPClient field is exported so tests can inject a stub.
type Config ¶
type Config struct {
// Endpoint is the OpenAI-compatible chat completions URL. Common
// values:
// - https://ollama.com/api/v1 (Ollama Cloud)
// - http://localhost:11434/v1 (local Ollama)
// - https://api.openai.com/v1 (OpenAI proxy / compatible)
Endpoint string
// APIKey is the bearer token. Empty for local Ollama.
APIKey string
// Model is the model name, e.g. "gpt-oss:20b".
Model string
// Timeout caps a single LLM HTTP call.
Timeout time.Duration
// MaxFiles caps the per-review change set. 0 = unlimited.
MaxFiles int
// MaxContextTokens is a soft per-call budget the prompt builder
// honors when assembling evidence (truncates oldest evidence first).
MaxContextTokens int
}
Config selects the LLM endpoint + model + timeouts for the review path. Defaults target local Ollama; setting OLLAMA_API_KEY flips to Ollama Cloud (mirrors the Java ReviewConfig contract).
func DefaultConfig ¶
func DefaultConfig() Config
DefaultConfig returns the standard Ollama defaults. When OLLAMA_API_KEY is set the endpoint flips to Ollama Cloud.
type Finding ¶
type Finding struct {
File string `json:"file"`
Line int `json:"line,omitempty"`
Severity string `json:"severity"` // info | low | medium | high | critical
Comment string `json:"comment"`
}
Finding is one structured review comment from the LLM.
type GraphContext ¶
type GraphContext interface {
// EvidenceForFile returns a short, deterministic textual summary of
// the graph context for a path: nodes-in-file, callers/dependents,
// frameworks detected. Free-form; whatever the LLM finds useful.
EvidenceForFile(path string) string
}
GraphContext is the small graph-side dep ReviewService needs. The concrete implementation lives in mcp/cli code; this interface keeps the review package decoupled from the graph.Store. nil is acceptable — review will just send the diff without graph evidence.
type KuzuGraphContext ¶
KuzuGraphContext implements GraphContext by querying an open Kuzu store. Wire from CLI: open store, NewKuzuGraphContext(store), pass to NewService.
EvidenceForFile returns a compact textual summary that the LLM finds useful: nodes-in-file with kind + layer, plus 1-hop blast radius node IDs. Strictly read-only.
func NewKuzuGraphContext ¶
func NewKuzuGraphContext(store *graph.Store) *KuzuGraphContext
NewKuzuGraphContext returns a context backed by store. nil Store yields empty evidence (the LLM degrades gracefully to diff-only review).
func (*KuzuGraphContext) EvidenceForFile ¶
func (k *KuzuGraphContext) EvidenceForFile(path string) string
EvidenceForFile satisfies GraphContext. Empty string when the store is missing or the file has no graph nodes.
type Report ¶
type Report struct {
Summary string `json:"summary"`
Findings []Finding `json:"findings"`
Model string `json:"model"`
RequestID string `json:"request_id,omitempty"`
}
Report is the structured LLM output. Both CLI (markdown) and MCP (JSON) paths consume this shape.
type Service ¶
type Service struct {
Client *Client
Graph GraphContext // may be nil
}
Service orchestrates diff → graph evidence → LLM call. Stateless; pass-by-value safe.
func NewService ¶
func NewService(client *Client, graph GraphContext) *Service
NewService wires the Service. Graph may be nil for non-enriched invocations (the LLM gets just the diff).
func (*Service) Review ¶
func (s *Service) Review(ctx context.Context, cwd, base, head string, focusFiles []string) (*Report, error)
Review runs the full pipeline against the given working tree and refs. Pass an empty base to compare against HEAD~1; empty head defaults to HEAD. FocusFiles, if non-empty, limits the review to those file paths.
Source Files