[KuSh]

The bug

detect_task picked the gradle task as "the last argument that does not start with -, ignoring clean". It has no notion of a flag consuming the next token, so a flag's value became the task:

rtk gradlew test --tests com.example.Foo   → task = "com.example.Foo" → GradlewTask::Other → unfiltered passthrough
rtk gradlew assembleDebug -p ../other      → task = "../other"        → GradlewTask::Other → unfiltered passthrough

Routing is substring-based (task.contains("test")), so some wrong values still land on the right branch by luck (--tests MyTest contains "test"); the reproducers above are where the luck runs out. Confirmed on the develop binary against a ./gradlew stub: both print the daemon banner and the PASSED line, i.e. no filter ran at all.

Flag facts, verified against real gradle

gradle --help on gradle:jdk21 (9.7.1) and gradle:8.14-jdk21, plus actual runs against a throwaway project, both under LC_ALL=C:

form	result
test --tests com.example.Foo	value consumed (no "Task not found"); without --tests, Selection failed: Task 'com.example.Foo' not found
-P myprop=v / -Pmyprop=v	both set the property
-D sysprop=v / -Dsysprop=v	both set the property
--console plain, -x compileJava, -M sha256, --update-locks org.foo:bar	value consumed
dependencies --configuration compileClasspath, help --task showProps	value consumed
tasks --group build, run --args x, dependencyInsight --dependency x, init --test-framework testng	value consumed (gradle help --task <name> lists them; init alone has eight)
-qm	clusters (quiet + dry-run)
-mx compileJava	rejected: "No argument was provided for command-line option '-x'"
-xcompileJava	accepted
-b/-c	present (deprecated) in 8.14, removed in 9.7.1 — kept in the table for 8.x wrappers

Dialect

Gradle's parser is org.gradle.cli — gradle:jdk21 ships gradle-cli-9.7.1.jar and no commons-cli — and no other tool uses it, so under #4008's preset naming rule it gets no preset. It is a one-axis derivation of Posix, composed as a const at its call site in gradlew_cmd.rs:

/// `org.gradle.cli` is gradle's own parser, shared with no other tool, so it gets no preset.
const GRADLE: Dialect = Dialect {
    dash_dash: DashDashRole::EndsGlobalOptions,
    ..Dialect::Posix
};

Every other axis is Posix, and each is measured: gradle -qi tasks works, so short flags cluster; only = attaches; every short is single-char; there is no /flag spelling. Shorts use ValueSpec::solo_only() — gradle -qp /w tasks is rejected while gradle -q -p /w and gradle -q -p=/w work, which is exactly what solo_only models, and the same shape git's -n already uses.

DashDashRole::EndsGlobalOptions is the axis value gradle needs. gradle --help calls -- "Signals the end of built-in options. Parses subsequent parameters as tasks or task options only", and that is measurably what it does — gradle -- test --tests com.example.Foo still consumes the filter value, while gradle -- showProps --offline fails. That is neither EndsOptions (git, cargo, rg) nor Forwards (dotnet).

So task detection is one pass under GRADLE. A repeated -- needs one extra rule: where a task is expected it is a task name (gradle -- -- help → Task '--' not found), but after a task it separates that task's options (gradle -- help -- → BUILD SUCCESSFUL), so a -- is dropped from the task names. wants_verbose_logging still slices with before_dashdash, because gradle rejects its own global options past --. #[allow(dead_code)] comes off DashDashRole::EndsGlobalOptions, which gradlew is the in-tree caller of.

Also migrated

The verbose-logging bypass was an arg == "--flag" scan. It now runs on tokens, which gains two things:

• gradle's short spellings -i/-d/-s/-S turn filtering off the same way --info etc. already did;
• --info sitting in a value position (--exclude-task --info) no longer trips it.

restore_double_dash is now called: clap strips the user's -- when it is the first token of the trailing region (measured: rtk gradlew -- test … reached the stub as test …, while rtk gradlew test -- … kept its --). Gradle's -- is meaningful, so dropping a token the user typed is a silent argv mutation.

gradlew is the only command routing through this module — gradle, ./gradlew and gradlew.bat all rewrite to rtk gradlew (src/discover/rules.rs), so nothing is left behind.

When no task resolves

Consuming a flag's value can leave no task name at all, and gradle then runs its default help task. Bare rtk gradlew and flags-only invocations (-p ../other, --console plain, -g <home>, -I <script>) therefore stay on passthrough instead of the build filter, which would strip that output.

The value-taking task options cover the built-in tasks rtk routes on rather than claiming to be exhaustive: plugins define task options freely.

Tests

tests/gradlew_arg_grammar_test.rs stubs ./gradlew with a script that records its argv and prints gradle-shaped output, then asserts on the resolved routing (did a filter run) and the child argv, not on the predicate, which reports green for this whole bug family. It spawns rtk through common::rtk_command(), so it never writes to the developer's real ~/.local/share/rtk/. Unit tests cover detect_task's resolved task, including the repeated -- and the no-task cases.

Each behavioural test fails against the previous implementation and passes after. test_detect_last_task_still_wins and test_detect_attached_project_prop_is_not_the_task pass both ways by design: they pin the existing "last task wins" intent and a case the old starts_with('-') scan already got right.

[rtk-wshm-sync-bot]

wshm · Automated triage by AI

📊 Automated PR Analysis

🐛 Type	bug-fix
🟡 Risk	medium

Summary

Fixes gradlew task detection to use a proper argument tokenizer instead of a naive 'last non-dash argument' heuristic, so flag values (e.g. --tests com.example.Foo, -p ../other) are no longer mistaken for the gradle task name and silently bypass output filtering. Also migrates verbose-flag detection to the same tokenizer and adds handling for gradle's -- semantics, which differ from both existing Dialect models.

Review Checklist

• Tests present
• Breaking change
• Docs updated

---

Analyzed automatically by wshm · This is an automated analysis, not a human review.