| name | zhihu-pp-ai-slop-cleaner |
| description | Use for Zhihu++ maintenance work that scans Kotlin main sources for low-call functions, structurally similar function bodies, dead code, pure forwarding wrappers, pointless abstractions, repeated helpers, and cross-platform duplicate glue before refactoring or PR cleanup. Trigger when the user asks to clean AI slop, remove low-call wrappers, find similar or duplicated code, review duplicated helper functions, or audit functions with call count at most 2 in /Users/zhaoliyan/IdeaProjects/Zhihu. |
Zhihu++ AI Slop Cleaner
强制验证门禁
- 单生产调用点的过滤、映射或测试专用 helper 优先内联，不得为注入时间或制造可测性扩大生产 API。
- 回归测试必须先在基线得到红测，再在修复后用同一生产路径得到绿测；测试新 helper 不算回归证据。
- seed、默认初始化和本地快照不能冒充真实线上数据库写入，必须使用真实管理接口或数据库读回。
Core Rule
Treat low call count as a queue for review, not proof of deletion.
Use non-test production line count as a hard refactor check, measured per independent change cluster rather than
only across the whole commit. Deleting one large client must not hide line growth caused by new owners, adapters,
constructor parameters, or initialization glue elsewhere. A cleanup that preserves behavior but materially grows
the production implementation is presumed wrong unless the added lines encode a demonstrated new contract.
Do not treat unit-test reachability as a production contract. Before merging or preserving tests for a pure
function, inspect whether the test caused production code to expose an internal helper, accept extra parameters,
or retain test-only branches and fallback values. When the behavior is simple and already exercised through its
real caller, delete the direct helper test, inline or make the helper private, and remove parameters or special
cases that no production path needs. Optimize the production API first, then keep only tests for observable
contracts and regression-prone boundaries; reducing test-method count while leaving test-shaped production code is
not a successful cleanup.
Do not create a State object merely because several local values belong to the same screen, feature, or naming family, or because their count crosses a mechanical threshold. A state object is justified only when its members participate in the same invariant, transition, or lifecycle; independent dialog flags, display modes, loading flags, and navigation flags must remain independent even when there are many of them. Count-based review rules identify places to inspect, not fields to bundle. Example: a reading mode flag and a next-item loading flag may both appear on a content page, but changing one does not constrain the other, so grouping them only hides unrelated state behind a vague container.
Delete or inline a function only after inspecting its declaration, every real call site, and the surrounding contract. Keep small functions that are framework entry points, stable UI/test selectors, platform contracts, interface defaults, Room/serialization hooks, navigation hooks, or meaningful domain boundaries.
Runtime / UI State dependency bundles are migration debt, not contracts. Search for *Runtime, *State, remember*Runtime, and similar dependency-bundle names as a first-class cleanup target instead of waiting for low-call evidence. When a screen or shared UI exposes one of these objects and it aggregates ViewModels, settings snapshots, environment methods, callbacks, network/database access, or platform helpers, dismantle it unconditionally. Move each responsibility to the real owner: common UI selects common ViewModels directly, common code calls cross-platform network/database helpers directly, and platform-only effects use narrow expect/actual functions or existing composition locals. Do not preserve a runtime field just because it currently carries login, update, install metadata, debug flags, or another “platform service”; split that service into the smallest real platform primitive or inline it at the caller.
When a runtime/capability object or composition local already owns a cross-platform helper, do not also thread that same capability through sibling parameters. Treat runtime + duplicate capability parameter or composition local + forwarded parameter as a half-cleaned dependency bundle: remove the duplicate parameter and read the capability at the lowest composable/helper that actually uses it, or from the runtime if the runtime remains temporarily. Example: an item renderer should not receive image save/share functions merely to pass them into an image menu when those functions are already cross-platform composition helpers available to that menu.
When merging duplicated platform implementations, do not stop at moving the duplicated body into an environment/interface default if there is still only one real semantic caller. If a ViewModel is the only place that owns the workflow, put the request logic in that ViewModel and let the environment expose only lower-level capabilities such as authenticated cookies and signed requests. Example: a question follow action should live in the question feed ViewModel that catches its errors, not as a one-call environment.follow...() wrapper that only chooses POST or DELETE.
Before adding a cross-platform file-storage interface, inspect and use the Kotlin multiplatform filesystem API already available to the project. Platform code may supply the application-private root path when the operating system determines it, but it must not wrap ordinary read, write, exists, or delete operations behind environment methods that merely forward to the common file object. Example: a feature cache should receive or derive a common Path under the app-private directory and perform JSON file I/O in common code, instead of adding four readPrivateTextFile and writePrivateTextFile methods to an environment and duplicating those operations in every actual implementation.
Do not push platform or storage dependencies upward just because the current accessor is only convenient from UI code. If a lower-level navigation or filtering component owns the query, make the cross-platform dependency available at that level instead of threading a database or platform handle through screen and ViewModel calls. Example: answer switching should ask its own support layer for already-opened content, not force the article loading call to accept a database parameter that exists only to be forwarded.
Do not delete runtime guards just to make tests pass or to simplify state. If a throttle, retry limiter, debounce, permission gate, or cache invalidation guard protects production behavior, keep it and make tests reset or inject the relevant state explicitly. Example: an authenticated request refresh throttle should be reset in tests; removing the throttle changes runtime semantics.
Default to whole-file or helper-family cleanup batches. Do not wait for the user to ask for a “second stage” before leaving tiny one-helper edits behind. If the current checkout is not already a dedicated cleanup branch or worktree, put broad cleanup in its own worktree so the main checkout remains available, then process whole-file helper clusters at once. A valid broad batch should remove or inline at least one complete file's helper cluster and cover at least 30 helper functions unless the user explicitly narrows the scope. If one suspicious helper is found, inspect its surrounding file-level cluster and remove adjacent useless wrappers together instead of deleting one isolated function per commit.
Workflow
- Record the current time and inspect
git status --short --branch.
- Run the low-call scan:
python3 .agents/skills/zhihu-pp-ai-slop-cleaner/scripts/scan_low_call_functions.py \
--root . \
--max-calls 2 \
--output /tmp/zhihu_low_functions.tsv
- Run the similar-function scan in every cleanup pass; duplicated wrappers often hide outside low-call results:
python3 .agents/skills/zhihu-pp-ai-slop-cleaner/scripts/find_similar_kotlin_functions.py \
--root . \
--threshold 0.82 \
--output /tmp/zhihu_similar_functions.tsv
- Review the TSVs from highest-confidence candidates first.
- For each candidate, verify with
rg before editing:
rg -n "\bFUNCTION_NAME\b" app shared desktopApp -g '*.kt'
-
Classify each candidate:
delete: no production caller and no framework/test entry value.inline: function body is a pure forwarding call, local one-use wrapper, or renaming shell.merge: repeated helpers perform the same operation; replace with one shared helper or add a parameter.keep: function is a real contract, domain boundary, parser step, platform actual, override, DAO method, serializer, stable test tag, or improves readability of a complex expression.- Runtime/state glue is not classified as
keep; delete or inline the whole bundle and all of its actual/expect shells.
-
For every candidate, switch to a file-level pass before editing. Read the nearby functions and classify the whole helper cluster together instead of deleting one rg hit at a time. Example: if one signed request helper in an environment file looks unnecessary, inspect adjacent signed helpers in the same file; keep multi-call primitives, but inline a single-call wrapper that only forwards to the lower-level client and adds no contract.
-
Edit only after classification. Prefer the nearest existing API over creating a new helper.
-
Re-run the relevant scan after each batch and check that deleted or merged names disappeared.
-
Run project verification in the required order before committing:
./gradlew assembleLiteDebug
./gradlew ktlintFormat
- Run a final review pass:
git diff --check
git diff --stat
rg -n "TODO|unused|deprecated wrapper|pure forwarding" app shared desktopApp -g '*.kt'
What To Remove Aggressively
- Private or local functions with one call and a body under 10 lines that simply call another function.
- Public/internal functions used only by tests when production should instead test the lower-level encoder, decoder, URL builder, or data type directly.
- Same-file helper names that only rename an environment/platform method with no branch, state, permission, or lifecycle boundary.
- Cross-platform duplicate helpers that do identical formatting, parsing, URL normalization, or ID construction.
- Dead debug helpers, stale reset hooks, stale bulk methods, and comment-referenced methods with no production caller.
*Runtime, *State, remember*Runtime, and UI dependency-bundle fields, even when they still have multiple call sites. Treat each field as a forwarding layer to eliminate, not as a surface to preserve.
Merge Conflict Boundary
During slop cleanup, do not reintroduce platform/environment state just because master has a newer conflict-side implementation. If the cleanup intentionally pushed a cache/throttle into a lower-level helper to keep API surface small, preserve that direction when resolving conflicts. Example: when a lower-level authenticated request helper owns a short refresh throttle, do not add per-platform last... getters/setters back onto the environment interface merely to satisfy one merge side; keep the throttle at the lower layer and adapt the conflicting code around it.
After a user points out one bad conflict direction, audit the whole merge intersection before continuing. Check every file changed on both sides, search for deleted helper families that may have been reintroduced under the same or similar names, and verify new master-side behavior is still present. Example: if a conflict around an environment interface accidentally brings back one state accessor, also inspect adjacent files for old repositories, URL helpers, wrapper methods, and newly added feature interfaces instead of fixing only the named accessor.
What To Keep
- Non-runtime
override, expect, actual, @Serializable serializer methods, NavType, Room DAO, Room converter, migration, lifecycle callback, and JavaScript bridge entry points unless source inspection proves they are unused and removable.
- Interface default implementations that are part of a capability contract.
- Stable UI test tags and semantic tag builders, even if call count is one.
- Parser helpers where the name explains a non-obvious grammar or Markdown/HTML rule.
- UI composables that isolate a meaningful visual unit rather than only forwarding parameters.
- Platform code that intentionally differs by Android/JVM/native behavior.
- These keep rules do not protect runtime/state dependency bundles or their expect/actual constructors.
Scan Script Notes
scripts/scan_low_call_functions.py is intentionally conservative:
- It scans main source roots only by default.
- It excludes
test, androidTest, build output, and worktrees unless options change that.
- It groups
expect/actual declarations by name and arity so the multiplatform declaration family is counted once.
- It masks comments and strings before counting references.
- It still cannot know Kotlin reflection, framework dispatch, Java interop, or generated-code calls. Always inspect candidates manually.
Useful options:
python3 .agents/skills/zhihu-pp-ai-slop-cleaner/scripts/scan_low_call_functions.py --help
python3 .agents/skills/zhihu-pp-ai-slop-cleaner/scripts/scan_low_call_functions.py --include-tests
python3 .agents/skills/zhihu-pp-ai-slop-cleaner/scripts/scan_low_call_functions.py --max-calls 0
scripts/find_similar_kotlin_functions.py finds similar bodies by normalized token shingles:
- It normalizes non-keyword identifiers to
ID, numbers to NUM, and string/char literals to STR/CHAR.
- It compares function bodies with shingled Jaccard similarity.
- It defaults to main source roots, excludes tests, ignores tiny bodies, and outputs only high-score pairs.
- It is useful for repeated formatters, URL builders, platform stubs, and copied UI glue.
- It can false-positive on framework callbacks and similar Compose layout skeletons; inspect before merging.
Useful options:
python3 .agents/skills/zhihu-pp-ai-slop-cleaner/scripts/find_similar_kotlin_functions.py --help
python3 .agents/skills/zhihu-pp-ai-slop-cleaner/scripts/find_similar_kotlin_functions.py --threshold 0.9
python3 .agents/skills/zhihu-pp-ai-slop-cleaner/scripts/find_similar_kotlin_functions.py --min-lines 8 --min-tokens 60
PR Hygiene
When the cleanup becomes a PR:
- Do not use broad
git add .; stage explicit paths.
- Do not commit unrelated user changes.
- Keep PR title/body in Chinese and use
refactor: unless the user asks otherwise.
- Include the scan command, validation commands, and any deliberately kept categories in the PR body.
