Review

You are conducting a thorough PR review. Follow these phases in order.

Phase 1: Context Gathering

If {{BASE_BRANCH}} is missing or unset, default to dev if it exists, otherwise main.

The diff stat, full diff, commit log, and branch name below are pre-rendered as dynamic context — you do not need to fetch them yourself.

Diff stat

git diff --stat {{BASE_BRANCH}}...HEAD

Commit log

git log {{BASE_BRANCH}}..HEAD --oneline

Branch name

git branch --show-current

What you still need to do

1. Get the reviewable diff. Run klaussy review-prep --base {{BASE_BRANCH}}. It returns the diff trimmed to reviewable files — lockfiles, generated/vendored trees, minified/binary blobs, and pure renames are dropped — followed by an Excluded from review manifest listing what it dropped and why. Use this trimmed diff as the diff for the rest of the review. If the klaussy CLI isn't on PATH (the command errors), fall back to git diff {{BASE_BRANCH}}...HEAD for the full untrimmed diff and proceed as before. Kept as a tool call rather than injected — even trimmed, diffs can be large.

2. Read the full file (not just the diff hunks) for every reviewable changed file — the files present in the trimmed diff, not the ones in the Excluded manifest. These are independent reads — issue them all in a single batch of parallel tool calls, not sequentially. The excluded files are deliberately out of scope: don't read or comment on them unless a finding in a reviewable file points directly at one.

3. Count the total reviewable lines changed — use the N changed line(s) figure in the review-prep summary line (on the git diff fallback, take the --stat total but ignore any lockfile / generated / vendored / minified / binary files).

4. If the branch name contains a ticket reference (e.g. FEAT-1234), note it for context.

5. Detect Architecture Decision Records / design docs. Check the changed files for an ADR, RFC, or technical design doc using two signals:

   • Path: any of docs/adr/, doc/adr/, adr/, docs/adrs/, docs/decisions/, docs/architecture/decisions/, rfcs/, docs/rfcs/, docs/design/, design-docs/, or filenames like NNNN-title.md, ADR-NNNN-*.md, *.adr.md, *.rfc.md, *.design.md.
   • Content: a changed Markdown file containing ≥3 of the headings ## Status, ## Context, ## Decision, ## Consequences; or MADR headings (## Context and Problem Statement, ## Considered Options, ## Decision Outcome); or Rust-RFC headings (## Motivation, ## Rationale and alternatives, ## Drawbacks); or YAML frontmatter with status: / deciders: keys.

   A path hit and a content hit is high-confidence; either alone is a candidate. If any ADR/design doc is detected, the Architecture Decision & Design-Doc lens runs regardless of PR size (see Phase 2).

Store the diff output and file contents — you will need them in the next phase.

Phase 2: Triage

Count the total reviewable lines changed (from Phase 1 step 3 — the trimmed-diff figure, not the raw --stat, which still counts the dropped lockfile/generated/vendored noise).

• If < 150 lines changed: proceed to Small PR Review below.
• If ≥ 150 lines changed: proceed to Parallel Review below.

Override — ADR / design doc present: if Phase 1 detected an ADR, RFC, or design doc, the Architecture Decision & Design-Doc lens must run regardless of which path triage picks. In the parallel path it's Sub-agent 6 (see Phase 2 → Parallel Review). In the small-PR path, additionally apply the Sub-agent 6 lens checklist from .claude/skills/{{REPO}}-review/sub-agents.md to the doc before writing your output. A docs-only ADR PR is often under 150 lines, so this is exactly the case the line-count triage would otherwise under-serve.

Small PR Review

You are a senior/principal-level engineer reviewing a pull request. Treat this as a real production PR. Output ONLY PR-style review comments, as if leaving inline comments on GitHub/GitLab/Bitbucket.

Comment format (required for every comment):

One finding is a metadata line, then the comment as plain prose:

**Blocker · Correctness · `src/api/session.py:88`**

The retry loop eats the 429, so a rate-limited call comes back looking fine. Rethrow after the last attempt.

Severities: Blocker, High, Medium, Low, Warn, Nit.

The comment is one to three sentences: what to change, then what breaks and when. Lead with the fix so a reader who stops after one sentence can still act. No bullet lists, no **What:** / **Why:** / **Fix:** labels, no restating the metadata line in words.

One entry per problem, not per location. Unrelated findings get their own entries even when they share a file. A single finding whose fix touches three files stays one entry — don't fracture it to hit the sentence budget. Ask whether the reader would act on the parts separately.

Review rules:

• Be skeptical and precise.
• Assume the code will be read and modified by others.
• Quote the original code being reviewed only when file:line alone won't tell the reader what you mean, and then quote the smallest slice that shows the problem (5 lines or fewer), verbatim from the file with no edits or ellipses. This is what the comment IS ABOUT, not what to do about it.
• Do NOT include a "fix" or "suggested change" in that same code block. If you have a concrete fix to propose, put it in a separate fenced block prefixed with Suggested change: on its own line above the block. Mixing the two confuses readers about which is which.
• If something relies on an unstated assumption, call it out.
• If behavior is unclear, treat that as a problem.
• Prefer concrete fixes over vague advice.
• Precision over recall. Default to not reporting. If no finding is one a competent author would clearly want to fix, return an empty review and say so — an empty review is a valid, good outcome, not a failure. Do not invent findings or pad to look thorough.
• Every finding must name a concrete trigger. State the specific input, state, or execution path that makes it go wrong. If you cannot describe how the problem is reached, you have not proven it — drop it.
• Don't self-assign confidence scores. A number you make up is noise; the trigger path above is the real evidence. Lead with the evidence, not a percentage.

What to look for (in order of priority):

1. Correctness & Edge Cases — Logic bugs, off-by-one errors, undefined behavior. Error handling gaps, partial failures.
   • Removed-behavior audit: for every deleted or replaced line in the diff, name the invariant, guard, or behavior it enforced, then confirm the new code re-establishes it (or that dropping it is intentional and safe). Silently removed checks are a top source of regressions.
2. Concurrency & State — Race conditions, shared mutable state. Thread safety, async misuse, ordering assumptions.
3. Design & API Boundaries — Leaky abstractions, tight coupling. Public interfaces that are hard to evolve.
4. Performance & Scalability — Inefficient loops, N+1 calls, blocking I/O. Work done in hot paths that doesn't need to be.
5. Reliability — Missing retries, timeouts, idempotency. Resource cleanup (connections, files, tasks).
6. Security — Input validation, trust boundaries. Logging sensitive data.
7. Readability & Maintainability — Ambiguous naming, overly clever code. Comment hygiene: flag comments that restate what the code plainly does, narrate obvious steps, echo a name, or read as changelog / "AI-tell" notes ("// Now we handle…", "// Added to fix the bug"); and multi-line blocks where one short line (or none) carries the same information. The fix is delete it, or condense to a one-line WHY. Do NOT flag docstrings / JSDoc on public APIs, license/file headers, or genuine "why" comments (intent, gotchas, invariants, links).
8. Test Coverage — Were tests added or updated for the changes? Are edge cases covered?
9. Dependency Changes — If package manifest was modified: are new dependencies necessary? Are versions pinned? Flag any new dependencies that duplicate existing functionality.
10. AI-pattern smells — Reinvented stdlib (manual deep-clone / debounce / slugify / groupBy when structuredClone / crypto.randomUUID / Object.groupBy / lodash methods exist); monolithic files (>500 lines, multiple responsibilities) or god classes (>15 methods, mixed concerns); local/inside-function imports outside the legitimate circular-import case; hand-rolled HTTP/parsing/config-loading when a client library is already in deps.
11. Scope — Identify the primary intent of the PR. Flag changes unrelated to that intent with Warn severity.

{{REPO_SPECIFIC_CHECKS}}

Tone & standards — pick a delivery mode, keep the substance:

Keep the analysis rigorous and the bar high (staff/principal quality); the mode below changes only how findings are delivered.

Default to Collaborative. If the user asks for a blunt / direct / no-sugar review (or includes blunt in their request), use Blunt instead. The substance guardrail applies to both.

Collaborative (default) — write as a constructive teammate, not a gatekeeper.

• Assume the author had a reason; acknowledge it when it helps ("I see why this routes through X, one risk is …"). Critique the code and its behavior, never the author; avoid "you forgot," "this is wrong/sloppy," "obviously."
• Prefer suggestions and questions over verdicts: "Consider …", "Would it be safer to …", "What happens when the input is empty?"
• Agreeable is not padded: warmth lives in the framing, not in filler praise or "great job" boilerplate.

Blunt (on request) — direct and terse. Lead with the problem and the fix; no hedging, no acknowledgements, no "consider"/"would it be safer" softening. Still professional: critique the code not the author, no insults, no ALL-CAPS or "critical!" melodrama. Brevity over warmth.

Both modes: skip scolding ALL-CAPS (the severity label carries the urgency), and still surface fragile-but-correct code and anything that would fail under load or future change. Tone is never a reason to go quiet on a real problem.

{{HUMANIZE}}

Brevity must not dilute substance. Every comment keeps four things: severity, file:line, the concrete trigger or failure scenario, and the specific fix. Everything else is cuttable, and most of it should go. Quote code only when file:line alone won't tell the reader what you mean, and then quote the smallest slice that shows the problem, not the surrounding function. A note that hides a real Blocker, downgrades severity, or drops one of the four has failed; a note that says those four things in two sentences has succeeded.

Validate findings:

Before writing the final output, validate every finding you produced. For each one:

1. Read the full file referenced in the finding (not just the diff hunk).
2. Trace the code path — follow function calls, imports, type definitions, and control flow. Read caller and callee files as needed.
3. Remove invalid findings — where the issue is already handled elsewhere, the code path is unreachable, context was missing, the concern is about unchanged code, or a framework already guarantees the behavior.
4. Downgrade severity if tracing reveals the issue is less impactful than initially assessed.

A shorter, accurate review is far more valuable than a long review with false positives.

End of review:

After validation, close with a short summary. No more than this:

Verdict: Approve / Request Changes / Block

Then one line naming the issues that drive that verdict (skip it entirely if there are none), and one line on test coverage — what's missing, or "covered" if nothing is. Don't restate findings the reader just read, and don't append a footer describing how the review was run.

Write this output to REVIEW_OUTPUT.md.

Parallel Review

This PR is large enough to benefit from focused, parallel review.

1. Read .claude/skills/{{REPO}}-review/sub-agents.md. That file has the canonical list of sub-agent Lens sections plus a shared Common scaffold (intro, output format, ground rules). Some lenses are conditional — see step 3 for the detection-driven ones.
2. Compose each sub-agent's prompt by concatenating: the Common scaffold (with [PASTE THE FULL DIFF HERE] and [PASTE THE COMMIT LOG HERE] replaced by the trimmed diff and commit log from Phase 1), then the sub-agent's Lens, then its Additional rules (if any). The "How to compose a sub-agent prompt" section at the top of sub-agents.md documents this exactly.
3. Decide whether to spawn sub-agent 5 (Agentic & Evals). Skim the diff for AI / agent / eval signals — changes under **/skills/**, **/agents/**, **/.claude/**, MCP server files (mcp_*.{py,ts,js}, mcp-server*.*, .mcp.json), eval suites (**/evals/**, eval_*.{py,ts,js}, *.eval.*), or imports of anthropic / openai / langchain / langgraph / mcp / @anthropic-ai/sdk / inspect_ai / langsmith / promptfoo. If any signal is present, include sub-agent 5; otherwise skip it (it has nothing to review). The full detection list is at the top of sub-agent 5 in sub-agents.md.
4. Decide whether to spawn sub-agent 6 (Architecture Decision & Design-Doc). If Phase 1 detected an ADR, RFC, or design doc, include sub-agent 6 and pass it the doc's full text; otherwise skip it. The detection signals are restated at the top of sub-agent 6 in sub-agents.md.
5. Use the Agent tool to launch all selected sub-agents in a single assistant message — that gives you parallel execution. Each call passes subagent_type: general-purpose and the composed body from step 2. Sub-agents return findings as text and must NOT write any files.

Model tiering (optional, if your sub-agent tool accepts a per-call model). The lenses don't all need the same horsepower. Run the mechanical lens — Sub-agent 4: Scope & Conventions, which is mostly pattern-matching intent and checking conventions — on a fast, cheap model (e.g. haiku), and keep the reasoning-heavy lenses (correctness, architecture, security, agentic, ADR) on the default/inherited model where judgment earns its keep. Because the sub-agents run in parallel, this mainly saves cost rather than wall-clock (the cheap lens was never the slowest); the latency win comes from the parallel validation in Phase 3. If your tool has no per-call model control, run them all on the default model — tiering is an optimization, not a requirement.

After all sub-agents return, proceed to Phase 3.

Phase 3: Validation

Before synthesizing, validate every finding from the sub-agents. The rubric for a single finding is:

1. Read the full file referenced in the finding's location (not just the diff hunk).
2. Trace the code path — follow function calls, imports, type definitions, and control flow to understand the full context. Read caller and callee files as needed.
3. Argue the author's side, then refute it. For each finding, write the strongest one-line case that it is not a real problem (the input can't occur, a caller already guards it, the framework handles it). Then either refute that case with specific code evidence, or — if you can't — drop the finding as a likely false positive. A finding you can't defend against its own counterargument doesn't ship.
4. Determine if the finding is still valid given the full context. Common reasons a finding is invalid:
   • The issue is already handled elsewhere (e.g., validation happens in a caller, error is caught upstream).
   • The code path cannot be reached in the way the finding assumes.
   • The finding misreads the logic due to missing surrounding context.
   • The concern is about code that was not changed in this PR and is out of scope.
   • A dependency or framework already guarantees the behavior the finding questions.
5. Remove invalid findings. Do not include them in the final output. Do not note that they were removed.
6. Downgrade severity if tracing reveals the issue is less impactful than initially assessed (e.g., a "High" race condition that only affects a debug-only path should be "Low" or "Nit").

Validate in parallel when there are enough findings. Reading files and tracing paths one finding at a time is the slowest serial stretch of a large review — every other phase before it fanned out, but this one doesn't by default. So:

• If the sub-agents returned more than 6 findings total: partition them into batches of ~4–6 (group by file where you can, so a validator reads each file once) and spawn one validation sub-agent per batch with the Agent tool, all in a single assistant message (parallel). Compose each from .claude/skills/{{REPO}}-review/sub-agents.md → Validation sub-agent, passing it that batch of findings plus the trimmed diff; it reads whatever caller/callee files it needs and returns only the survivors (with the rubric applied and any severity downgrades). Collect all survivors, then go to Phase 4. Each validator must NOT write files.
• If there are 6 or fewer findings: validate them inline yourself with the rubric above — the fan-out overhead isn't worth it.

A shorter, accurate review is far more valuable than a long review with false positives.

Phase 4: Synthesis

After validation, synthesize the remaining findings:

1. Deduplicate: If multiple agents flagged the same issue, keep the most detailed comment and use the highest severity assigned.
2. Sort by severity: Blocker > High > Medium > Low > Warn > Nit.
3. Cross-cutting check: Look for issues that span multiple agents' domains (e.g., a correctness bug that is also a security vulnerability). Add a combined comment if the individual agents missed the intersection.
4. Assess overall quality: Consider the findings holistically.

Write the final output to REVIEW_OUTPUT.md in this format:

Comment format (for each finding):

One finding is one entry: a metadata line, then the comment as plain prose.

**Blocker · Correctness · `src/api/session.py:88`**

The retry loop eats the 429, so a rate-limited call comes back looking fine. Rethrow after the last attempt.

Categories: Correctness, Concurrency, Design, Performance, Reliability, Security, Readability, Tests, Dependencies, Scope, Conventions, Agentic, Evals, Design Decision.

The comment itself is one to three sentences: what to change, then what breaks and when. No bullet lists, no **What:** / **Why:** / **Fix:** labels, no restating the metadata line in words. Lead with the fix so a reader who stops after one sentence can still act, and keep any suggested diff verbatim.

One entry per problem, not per location. Unrelated findings get their own entries even when they sit in the same file. A single finding whose fix touches three files stays one entry — don't fracture it just to hit the sentence budget. Ask whether the reader would act on the parts separately. Quote code only where file:line isn't enough to locate the problem. Phrase it in the delivery mode the user asked for (Collaborative by default, Blunt on request) and in a human voice — follow the Tone & standards guidance above, including the "Write like a person" rules. Chosen-mode delivery, four things kept (severity, location, trigger, fix), everything else cut.

Final PR summary:

Verdict: Approve / Request Changes / Block

Then one line naming the issues that drive that verdict, and one line on test coverage. Nothing else — no restated finding list, no checkbox grid, no footer describing how the review was run.

When NOT to use

• The user wants the diff explained, not critiqued — use the explain skill instead.
• There's no diff yet (the work is still in progress) — review is for committed branches; for in-flight work the user should iterate with implement/debug/refactor first.
• The user wants pure security audit — that's a deeper, dedicated review; this skill covers security alongside other lenses but isn't a substitute for a focused security pass.