Repo guidelines & linked-issue context are computed but never injected into the review prompt

[mk7luke]

Summary

Repo guidelines and linked-issue context are computed in Reviewer but never injected into the review prompt. Both features appear wired up but are silent no-ops for the actual code review.

• Guidelines: dropped entirely — never reach the review or the walkthrough.
• Linked issues: reach the walkthrough but not the review.

Evidence (src/reviewer.ts)

The review prompt receives repo knowledge only through the learnings array (knowledgeLearnings), which is passed to this.ai.review(context, repoConfig, knowledgeLearnings):

// Build enhanced prompt context by injecting knowledge into the AI prompt
// (Guidelines and issues are injected via the prompt builder's learnings param)
const knowledgeLearnings = [...relevantLearnings];
// Add guideline content as synthetic learnings
const guidelinesPrompt = formatGuidelinesForPrompt(relevantGuidelines); // reviewer.ts:756 — assigned, never used
const issuesPrompt   = formatIssuesForPrompt(linkedIssues);             // reviewer.ts:757 — assigned, never used

knowledgeLearnings is only ever [...relevantLearnings] (754), then unshift-ed with the language preference (761); it is never given guidelinesPrompt or issuesPrompt before being passed to this.ai.review(...) (813). Despite the comment ("injected via the prompt builder's learnings param" / "Add guideline content as synthetic learnings"), that injection never happens.

Corroborating facts:

• formatGuidelinesForPrompt (src/guidelines.ts:147) and formatIssuesForPrompt (src/issues.ts:88) are pure string formatters — no side effects — so lines 756/757 are pure dead computation.
• Only call site of each is the dead assignment above (formatGuidelinesForPrompt → reviewer.ts:756; formatIssuesForPrompt → reviewer.ts:757).
• relevantGuidelines has no other consumer (defined 744, used only at the dead 756).
• linkedIssues is used for the walkthrough (reviewer.ts:1217, formatIssuesForWalkthrough), so issue context reaches the walkthrough but not the review.
• buildReviewPrompt (src/ai/prompt.ts) takes no guidelines/issues parameters — the learnings array is the only channel into the review prompt.

Impact

• Repo guidelines configured by users have no effect on inline review findings.
• Linked-issue context does not inform the review (only the walkthrough).
• Both fail silently — no error, no log — so the gap is invisible to operators.

Suspected intended behavior

Push the formatted strings into knowledgeLearnings as synthetic Learning entries before the review call, e.g.:

if (guidelinesPrompt) knowledgeLearnings.push({ id: "__guidelines__", repo: repoFullName, content: guidelinesPrompt, createdAt: "" });
if (issuesPrompt)     knowledgeLearnings.push({ id: "__issues__",     repo: repoFullName, content: issuesPrompt,     createdAt: "" });

(Exact shape/placement TBD — verify against how learnings are rendered in buildReviewPrompt and add a regression test asserting guideline/issue content reaches the built prompt.)

Provenance / notes

• Introduced by befbb49 ("Add CodeRabbit-parity features: walkthrough, chat, config, knowledge base, finishing touches"). Pre-existing on main.
• Surfaced via the advisory-lint warning 'guidelinesPrompt' is assigned a value but never used while reviewing PR Add backup AI provider / failover (off by default) #78 (backup AI failover). Unrelated to that PR and intentionally left out of its scope to keep the diff focused.