Review inbox: 'kai review analyze' headless Finding emitter#26
Open
jschatz1 wants to merge 9 commits into
Open
Review inbox: 'kai review analyze' headless Finding emitter#26jschatz1 wants to merge 9 commits into
jschatz1 wants to merge 9 commits into
Conversation
New command: kai review analyze <review-id> --format json — assembles the structured Finding for an agent change and prints it as JSON. This is the headless surface both the TUI and the server web UI render. Serializes existing engine data (review.State -> verdict, changeset intent + optional contract enrichment -> intent-vs-code, semanticdiff -> per-file patch with +/- counts) and drives the new evidence layer: transitive blast radius (safetygate.Reachback), grounded callers() claims, and should-touch (graph candidates from shouldtouch.Detect + an LLM judge gated on ai.IsConfigured, so it emits nothing rather than assert an unverified candidate).
…aw ai.Client The should-touch LLM judge was using the legacy ai.Client (hardcoded ANTHROPIC_API_KEY + api.anthropic.com). Switch it to the same provider path the rest of the engine uses — provider.FromEnv + provider.New via buildGateProvider — so it goes through kailab -> OpenRouter when logged in, with ANTHROPIC_API_KEY only as an offline fallback. Degrades to no should-touch findings when no provider is available, unchanged.
CI review pods had no image with the kai CLI (registry has only kailab-* images; ubuntu:22.04 default has no kai/git/jq). Add a kai-ci image: builds the kai CLI from the local workspace (kai-cli+kai-core+kai-engine via go.work, because kai review analyze needs kai-engine packages not in the published tag), plus git/jq/curl/bash. CGO on for tree-sitter, amd64 builder. build.sh logs in via kailab-registry-credentials and pushes registry.kaicontext.com/kai-ci.
Make the image actually build (verified: 208MB, kai 0.34.3, git/jq/curl/bash): - Resolve local modules via replace directives, not go.work (the go.work + dockerignore whitelist stripped kai-engine's contents in-container, so the build tried to fetch the private kai-engine v0.1.3 and failed). - Cross-compile the CGO binary (tree-sitter) on a native BUILDPLATFORM builder instead of compiling under qemu (qemu emulation crashes the Go toolchain with 'illegal instruction'). Needs the target cross-gcc AND libc6-dev-<arch>-cross headers (--no-install-recommends omits the latter). - Retry the runtime apt layer (transient dpkg/perl flake under emulation). - buildx --load/--push in build.sh.
hydrateSnapshot() materializes a snapshot's full graph on fetch-by-ref — the snapshot node plus every file node, content blob, and HAS_FILE edge — so a clean store can diff/analyze a snapshot resolved from a git.<sha> ref. Previously fetch-by-ref inserted only the bare root node, leaving the snapshot undiffable (kai diff saw no files; a review analyzed to files:0). Push now propagates the git.<sha> bridge refs: the push allowlist includes the git. prefix, and the 'Already up to date' gate is ref-aware (compares each ref to its remote/<remote>/<name> tracking ref) instead of keying on snap.latest alone — which had both misfired across remotes and never published a ref whose target snapshot already existed. Together these make the data plane SHA-addressable, so a server-side review can resolve a PR's base/head SHA to snapshots and diff them without re-capture.
runReviewAnalyze is now a thin wrapper: open the store, build the should-touch judge from client config, call reviewanalyze.Analyze, print JSON. The analysis logic moved to github.com/kaicontext/kai-engine/reviewanalyze.
Thin wrapper over reviewanalyze.AnalyzeSnapshots — emits the Finding JSON for an arbitrary base->head snapshot pair with no persisted review. Parity with the server-side entry point and useful for testing.
kai review-commit built its diff from `git --numstat` (line counts only), so the finding's diff.files carried Added/Removed but no Patch — the findings inbox Code tab rendered file headers with an empty body. Add rcCommitPatches(): run the real `git diff`/`git show`, split per file, strip file headers, keep hunk headers + +/-/context body, and attach as DiffFile.Patch keyed by path. Blast radius is still computed elsewhere and left as a separate follow-up.
Walk the captured graph from the changed files so the finding shows what the change reaches, not just the diff. Non-fatal panel. Durably commits the wiring the shipped kai-ci image already builds from the working tree.
🔍 Kai reviewKai is reviewing this pull request… the verification result will appear here shortly. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
CLI half of the agent-review inbox. New command
kai review analyze <review-id> --format jsonassembles the structured Finding for an agent change and prints it as JSON — the headless surface both the TUI and the server web UI render.Serializes existing engine data (review state → verdict, changeset intent + optional contract enrichment → intent-vs-code, semanticdiff → per-file patch with +/- counts) and drives the new evidence layer: transitive blast (safetygate.Reachback), grounded callers() claims, and should-touch (shouldtouch.Detect candidates + an LLM judge gated on ai.IsConfigured, so it emits nothing rather than assert an unverified candidate).
Depends on kai-engine#4 (finding/reachback/shouldtouch packages).