Review: $ARGUMENTS
The arguments are a ticket number and a base ref; a word such as spec or standards after them changes nothing, every run checks both. When one is missing, return the expected form and stop. You did not write this code and do not know why it looks the way it does; judge it only against its sources, and quote them. Change nothing: no edits, commits or comments, and run no tests or checks: the implementer ran ci.command before handing over, and the gate runs it again on the pull request.
Budget
Every turn costs the whole context again, so read aimed, never exploratory. At most 20 tool calls; chain read-only commands in one call. Open only files in the diff, the ticket's sources and the hits of a duplication search. When the budget runs out, return what is checked and name each changed file not yet checked as not checked: <path>.
Done when the review returned within its budget, or named what it left unchecked.
1. Pin the change
- In one call:
git rev-parse <base>, git log <base>..HEAD --oneline and git diff <base>...HEAD. A bad ref or an empty diff ends the review with that as its only result.
- Read the tracker from
.claude/64x-lunicorn.yml. With issues.tracker: forge, read issues with gh issue view <n>. With local, issue <n> is the file in issues.path whose name starts with the number, zero-padded to four digits.
- Read the ticket. Its reference line starts with
> Part of Spec #<spec>. Architecture: #<architecture>. and may carry more parts such as Order:; numbers may be zero-padded. Read the Spec and the architecture issue it names.
Done when the diff, the commits and all three issues are known.
2. Check against the sources
This step owns scenarios, scope, the architecture issue's decisions and implementation notes, the Spec's domain rules and terms.
For every scenario of the ticket:
- A test named like the scenario runs with no pending marker (skip, todo or tag) left on it, at a seam from the implementation notes, and asserts the scenario's
Then.
- The code path produces the behaviour. Read the code; a test name proves nothing.
- A scenario without a test is a finding: quote the scenario and give the ticket as its location.
Across the whole diff:
- Behaviour a caller can observe that no scenario asks for: options, flags, endpoints, fields, fallbacks. Structure without observable behaviour, such as an unused parameter, is a smell for step 3.
- Departures from the Decisions or the ticket's implementation notes, and from the Spec's domain rules. Code that copies what a decision says to reuse is reported here as a departure from that decision.
- Terms that differ from the Spec, and from
CONTEXT.md when the project has one.
Done when every scenario and every changed file is checked.
3. Check against the standards
This step owns documented standards, duplication, smells and test quality; architecture decisions stay in step 2.
- Documented standards:
CONTRIBUTING.md, CLAUDE.md or AGENTS.md, ADRs and coding standard documents. Skip rules the project's linter or formatter configuration already enforces.
- Duplication: for every new function, type, constant, component and test helper, search the whole codebase once with
grep for its name and its job's key terms, and open only the hits. Repeated logic inside the diff counts as well; repeated literal test data does not. Code that copies what an architecture decision names for reuse is left to step 2, so it is reported once.
- Smells: check the diff against the smell baseline. A documented standard that endorses a pattern overrides the baseline.
- Tests: coupled to internals, mocking the project's own modules, expected values recomputed the way the code computes them, verification through a side channel instead of the interface.
- HTTP API: when the diff adds, changes or removes an HTTP endpoint, check it against the checklist of
design-http-api. An endpoint change without the matching change to the OpenAPI document and the Bruno collection, or a violated MUST rule without a recorded reason, is hard; a violated SHOULD rule is judgement.
Done when every changed file is checked against all four, and every changed HTTP endpoint against the API checklist.
4. Classify
- hard: a scenario missing, failing or wrong; scope creep; a departure from an architecture decision or domain rule; a violation of a documented standard; duplication of existing code; a test that cannot fail.
- judgement: smells and naming, unless a documented standard makes them hard.
- conflict: the sources contradict each other, so no change satisfies all of them, such as a decision that makes a scenario unreachable. Quote every source involved.
A failing test caused by a conflict belongs to that conflict, not to a separate hard finding. A hard finding whose fix cannot succeed until a conflict is decided says blocked by conflict <n>.
Hard findings are fixed without asking, judgement findings go to Daniel, and conflicts are never fixed by the implementer, because only Daniel can decide which source gives way. Related smells on the same code form one finding. Leave out praise, what tooling enforces, and anything you cannot quote.
Done when every finding has exactly one class.
5. Return
One entry per finding:
- class and
path:line, a line range where the finding spans several lines,
- the quoted code, plus the quoted existing code or source text it is compared with,
- the source: scenario, decision, domain rule, standard or smell name,
- a proposed fix, or for a conflict the options Daniel can choose between.
End with the number of hard, judgement and conflict findings and the files not checked, or none, and confirm that git status shows no change. Without findings, return No findings.
Done when the result is returned and nothing in the repository or tracker changed.
1---2name: review-change3description: Reviews a branch against a base in one run, checking ticket scenarios, Spec, architecture decisions and scope creep as well as documented standards, code smells, duplication across the codebase and test quality, within a fixed tool-call budget, and returns classified findings without changing anything. Use when a ticket's change is reviewed before its pull request, or when a branch or pull request needs a spec or standards review.4---56Review: $ARGUMENTS78The arguments are a ticket number and a base ref; a word such as `spec` or `standards` after them changes nothing, every run checks both. When one is missing, return the expected form and stop. You did not write this code and do not know why it looks the way it does; judge it only against its sources, and quote them. Change nothing: no edits, commits or comments, and run no tests or checks: the implementer ran `ci.command` before handing over, and the gate runs it again on the pull request.910## Budget1112Every turn costs the whole context again, so read aimed, never exploratory. At most 20 tool calls; chain read-only commands in one call. Open only files in the diff, the ticket's sources and the hits of a duplication search. When the budget runs out, return what is checked and name each changed file not yet checked as `not checked: <path>`.1314**Done when** the review returned within its budget, or named what it left unchecked.1516## 1. Pin the change1718- In one call: `git rev-parse <base>`, `git log <base>..HEAD --oneline` and `git diff <base>...HEAD`. A bad ref or an empty diff ends the review with that as its only result.19- Read the tracker from `.claude/64x-lunicorn.yml`. With `issues.tracker: forge`, read issues with `gh issue view <n>`. With `local`, issue `<n>` is the file in `issues.path` whose name starts with the number, zero-padded to four digits.20- Read the ticket. Its reference line starts with `> Part of Spec #<spec>. Architecture: #<architecture>.` and may carry more parts such as `Order:`; numbers may be zero-padded. Read the Spec and the architecture issue it names.2122**Done when** the diff, the commits and all three issues are known.2324## 2. Check against the sources2526This step owns scenarios, scope, the architecture issue's decisions and implementation notes, the Spec's domain rules and terms.2728For every scenario of the ticket:2930- A test named like the scenario runs with no pending marker (skip, todo or tag) left on it, at a seam from the implementation notes, and asserts the scenario's `Then`.31- The code path produces the behaviour. Read the code; a test name proves nothing.32- A scenario without a test is a finding: quote the scenario and give the ticket as its location.3334Across the whole diff:3536- Behaviour a caller can observe that no scenario asks for: options, flags, endpoints, fields, fallbacks. Structure without observable behaviour, such as an unused parameter, is a smell for step 3.37- Departures from the Decisions or the ticket's implementation notes, and from the Spec's domain rules. Code that copies what a decision says to reuse is reported here as a departure from that decision.38- Terms that differ from the Spec, and from `CONTEXT.md` when the project has one.3940**Done when** every scenario and every changed file is checked.4142## 3. Check against the standards4344This step owns documented standards, duplication, smells and test quality; architecture decisions stay in step 2.4546- **Documented standards:** `CONTRIBUTING.md`, `CLAUDE.md` or `AGENTS.md`, ADRs and coding standard documents. Skip rules the project's linter or formatter configuration already enforces.47- **Duplication:** for every new function, type, constant, component and test helper, search the whole codebase once with `grep` for its name and its job's key terms, and open only the hits. Repeated logic inside the diff counts as well; repeated literal test data does not. Code that copies what an architecture decision names for reuse is left to step 2, so it is reported once.48- **Smells:** check the diff against [the smell baseline](references/smells.md). A documented standard that endorses a pattern overrides the baseline.49- **Tests:** coupled to internals, mocking the project's own modules, expected values recomputed the way the code computes them, verification through a side channel instead of the interface.50- **HTTP API:** when the diff adds, changes or removes an HTTP endpoint, check it against the checklist of `design-http-api`. An endpoint change without the matching change to the OpenAPI document and the Bruno collection, or a violated MUST rule without a recorded reason, is hard; a violated SHOULD rule is judgement.5152**Done when** every changed file is checked against all four, and every changed HTTP endpoint against the API checklist.5354## 4. Classify5556- **hard:** a scenario missing, failing or wrong; scope creep; a departure from an architecture decision or domain rule; a violation of a documented standard; duplication of existing code; a test that cannot fail.57- **judgement:** smells and naming, unless a documented standard makes them hard.58- **conflict:** the sources contradict each other, so no change satisfies all of them, such as a decision that makes a scenario unreachable. Quote every source involved.5960A failing test caused by a conflict belongs to that conflict, not to a separate hard finding. A hard finding whose fix cannot succeed until a conflict is decided says `blocked by conflict <n>`.6162Hard findings are fixed without asking, judgement findings go to Daniel, and conflicts are never fixed by the implementer, because only Daniel can decide which source gives way. Related smells on the same code form one finding. Leave out praise, what tooling enforces, and anything you cannot quote.6364**Done when** every finding has exactly one class.6566## 5. Return6768One entry per finding:6970- class and `path:line`, a line range where the finding spans several lines,71- the quoted code, plus the quoted existing code or source text it is compared with,72- the source: scenario, decision, domain rule, standard or smell name,73- a proposed fix, or for a conflict the options Daniel can choose between.7475End with the number of hard, judgement and conflict findings and the files `not checked`, or `none`, and confirm that `git status` shows no change. Without findings, return `No findings.`7677**Done when** the result is returned and nothing in the repository or tracker changed.
64x-lunicorn