Documentation
¶
Overview ¶
Package gates ports the two parallel `.ai()` post-processing passes from the Python PR-AF pipeline: the merge-blocker gate (merge_gate.py) and the comment polish pass (polish.py). Both fan out one AI call per item, preserve input order, and degrade gracefully on per-item failure.
Merge gate — a separate lens from severity. Severity asks "how bad is this issue?"; the merge gate asks "must this be fixed BEFORE the PR ships, or can it ship and be addressed in a follow-up?". Failure mode: default to blocking=false on every error path, because a false negative (a real blocker slips through as advisory) is recoverable by human review, whereas a false positive (advisory issue flagged blocking) erodes trust and blocks merge.
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func ClassifyFindings ¶
func ClassifyFindings(ctx context.Context, app AICaller, findings []schemas.ScoredFinding) []schemas.ScoredFinding
ClassifyFindings ports merge_gate.classify_findings: one AI call per finding, fired in parallel, original order preserved. Returns a NEW slice with Blocking and BlockingReason populated; the input slice and its elements are not mutated (pure, mirroring pydantic model_copy).
Concurrency: Python uses asyncio.gather WITHOUT return_exceptions, but _gate_one catches every exception internally and never raises, so no error ever propagates through gather. The faithful Go mirror is a WaitGroup writing into a pre-indexed results slice — order is positional, not completion-order.
func PolishComments ¶
func PolishComments(ctx context.Context, app AICaller, comments []schemas.GitHubComment) []schemas.GitHubComment
PolishComments ports polish.polish_comments: rewrite each comment body in parallel, original order preserved. Returns a NEW slice; each comment's Body is replaced with the rewritten text while every other field (Path, Line, Side) is copied unchanged — path preservation is guaranteed structurally here, and callout/markdown preservation is enforced by the polish system prompt. Pure — the input slice and its elements are not mutated.
Concurrency mirrors classify_findings: Python's asyncio.gather never sees an exception because _polish_one catches its own AI errors, so a WaitGroup over a pre-indexed results slice is the faithful order-preserving fan-out.
Types ¶
type AICaller ¶
type AICaller interface {
AI(ctx context.Context, prompt string, opts ...ai.Option) (*ai.Response, error)
}
AICaller is the minimal one-method seam the two `.ai()` gates drive. It is appx.AIer's one method, so the review's App satisfies it unchanged and tests supply a fake. This is the Go analogue of Python's `app.ai(prompt, system=..., response_format="json")`: the system prompt is passed as ai.WithSystem and JSON mode as ai.WithJSONMode.
Source Files