choplin/my-agent-skills/code-review-session-import-ai

Import AI review findings (ingestion)

Add review items from an AI code review. This is one of the ways items enter the record (alongside code-review-session-import-pr, code-review-session-import-ci, code-review-session-run-checks, and direct feedback). It ingests findings — it does not resolve or fix them; that is code-review-session-resolve's job.

Select this skill when the requested deliverable is a persistent review record. A one-off "review this diff and tell me what's wrong" wants findings in the conversation and no record: that is quick-code-review on its own.

Who performs the review

This skill contains no review procedure. It calls one:

The caller may also pass findings it already has, in which case this skill only records them.

Input

• review_dir — where the record lives (default: standalone; see

code-review-session-base skill (references/procedures/review-init-guide.md)).

• scope — the change to review (default: current branch diff).
• reviewer — optional; which reviewer to invoke (default

quick-code-review).

• findings — optional; findings already produced, to record without reviewing.
• mode — optional; force full review. Other modes are derived from recorded

rounds and the current target.

Process

1. Resolve the record

Resolve review_dir and ensure review.md exists (create via code-review-session-base skill (references/procedures/review-init-guide.md) if not). Apply code-review-session-base skill (references/procedures/ai-review-rounds.md), conforming to its ledger (references/contracts/ai-ledger.md), and load or initialize {review_dir}/sources/ai.json.

2. Plan the round

Resolve the current target revision and stable content fingerprint when possible. Compare them with the latest completed round for the same scope, then select full, incremental, provided-findings, or skipped by the loaded AI-round procedure. Allocate the next round ID and record the attempt before invoking a reviewer.

If the mode is skipped, record why, report that no review-relevant input changed, and stop without changing review.md.

3. Obtain findings

Unless findings was supplied, invoke the selected reviewer over scope. Require each finding to carry a one-line summary, the location, and the detail; a severity and confidence if the reviewer offers them. Build iterative context as specified by the loaded AI-round procedure. In particular, let prior items guide an ordinary incremental review without exposing prior reviewer conclusions inside artifact-review's blind Lens passes.

If the reviewer fails or its output cannot be normalized, record the round as failed with the reason and stop without adding items.

4. Record findings and finish the round

For each finding, append an item to review.md under ## Items (see code-review-session-base skill (references/contracts/review-state.md)):

• Source: ai:round/{round}
• Status: open
• Detail: the finding text, including its location
• Summary: the one-line heading

Determine the next item number from the highest existing item. Dedupe: skip a finding that matches an existing item (same location + gist) so re-running the ingestion does not double-add.

In sources/ai.json, associate each new item with the round and store available location, severity, and confidence metadata. Record duplicates under matched_items. Mark the round completed only after both the generic items and the companion ledger agree.

Update the Resolved: X / Y denominator to the new total.

5. Report

AI review round {round} completed ({reviewer used}, {mode}, {base..head or scope}).
{count} items added; {duplicate count} matched existing items.
 
Open items: {open count}. Run `code-review-session-resolve` to work through them.

This skill does not fix anything or change the review Phase. It leaves the items open for code-review-session-resolve.

Success Criteria

• [ ] The review itself is delegated to a reviewer; no review procedure is

re-implemented here

• [ ] Every attempted review has a round in sources/ai.json with its target

revision and terminal outcome

• [ ] Each new finding is an open item with Source: ai:round/{round} and

source-specific metadata stays in the AI ledger

• [ ] Duplicates against existing items are skipped
• [ ] Iterative context prevents stale repetition without weakening

artifact-review reviewer independence

• [ ] Does not run project checks (that is code-review-session-run-checks)
• [ ] Does not fix, gate, or change the review Phase