structured-return

name	structured-return
description	Choose compact, machine-readable commands across common test, lint, build, and log workflows. Use when raw terminal output would be noisy and the agent should prefer json, junit/xml, porcelain, quiet, no-pretty, or narrow-scope command forms.

structured-return

Prefer better output at the source.

What structured_return does

structured_return is a tool — use it instead of bash for lint, test, and build commands. It runs the command, stores the full log, and returns a compact parsed result to the model. When a parser matches, only the signal (failures, errors, summary) comes back into context. When no parser matches, it falls back to the last 200 lines of output plus a log path — still better than dumping everything into context via bash.

Rules

1. Before running any lint, test, or build command, check for project-defined scripts (package.json, Makefile, pyproject.toml, etc.). If a script exists for the task, inspect it to identify the underlying tool and the scope/paths it uses. Then invoke that tool directly with machine-readable flags and the same scope — rather than running through the script wrapper.
2. Use structured_return instead of bash for lint, test, and build commands. For tools with a built-in parser, use the command form shown in the examples below. For tools without one, use known machine-readable flags (--json, --format json, --message-format=json, --console=plain, --porcelain, etc.) and let the tail fallback handle the result.
3. Otherwise prefer quiet, plain, no-color, and narrow-scope modes.
4. Prefer file/test-specific runs before full-suite runs.
5. Prefer bounded log reads over full log dumps.
6. Do not invent flags. Use known patterns from examples below or your own knowledge of the tool.
7. When using structured_return, pass explicit parser hints when known.
8. If a repo defines project-local parser mappings in .pi/structured-return.json, prefer those mappings.

Examples

pytest

• structured_return({ command: "pytest [any pytest args] --junitxml=.tmp/report.xml", parseAs: "junit-xml", artifactPaths: [".tmp/report.xml"] }) — --junitxml is built into pytest, no extra dependencies needed; scope, filters, markers, etc. go in [any pytest args]

ruff check

• structured_return({ command: "ruff check [any ruff args] --output-format=json", parseAs: "ruff-json" }) — scope, selects, ignores, etc. go in [any ruff args]

mypy

• structured_return({ command: "mypy [any mypy args] --output json", parseAs: "mypy-json" }) — --output json is built into mypy (1.0+); outputs NDJSON to stderr with file, line, column, message, error code, and severity; notes are folded into their parent error's message

htmlhint

• structured_return({ command: "npx htmlhint --format json [any htmlhint args]", parseAs: "htmlhint-json" }) — --format json outputs structured array; strips ANSI codes, source evidence, rule descriptions, and URLs

isort

• structured_return({ command: "isort --check --diff [any isort args]", parseAs: "isort-text" }) — parses --check output; strips diff hunks, absolute paths, timestamps; lists files with unsorted imports

npm audit

• structured_return({ command: "npm audit --json", parseAs: "npm-audit-json" }) — --json is built into npm; strips advisory URLs, CVSS vectors, fix instructions; advisory titles joined per package; severity breakdown in summary

jsonlint

• structured_return({ command: "npx jsonlint [file]", parseAs: "jsonlint-text" }) — strips stack trace and source pointer; extracts line number and "Expecting X, got Y" message

tidy (HTML Tidy)

• structured_return({ command: "tidy -errors [any tidy args]", parseAs: "tidy-text" }) — parses line N column N - Warning/Error: message from stderr; strips remediation advice, accessibility tips, reformatted HTML output, and Info lines

vale

• structured_return({ command: "vale --output JSON [any vale args]", parseAs: "vale-json" }) — --output JSON is built into vale; strips ANSI codes, Action/Span metadata, column-aligned formatting; severity breakdown in summary

prettier

• structured_return({ command: "prettier --check [any prettier args]", parseAs: "prettier-text" }) — parses --check output; strips "Checking formatting..." preamble, [warn] prefixes, and "Run Prettier with --write to fix" footer

markdownlint

• structured_return({ command: "markdownlint --json [any markdownlint args]", parseAs: "markdownlint-json" }) — --json outputs structured array; strips context quotes, rule info URLs, fix info, and error ranges; error details folded into description

black

• structured_return({ command: "black --check [any black args]", parseAs: "black-text" }) — parses --check output; strips diff hunks, emoji banners, timestamps; lists files needing reformatting; also handles --check --diff mode

bandit

• structured_return({ command: "bandit -f json [any bandit args]", parseAs: "bandit-json" }) — -f json is built into bandit; strips source snippets, CWE URLs, run metrics, and confidence labels; severity breakdown in summary

gcc / clang

• structured_return({ command: "gcc -c [any gcc args]", parseAs: "clang-text" }) — works with gcc, g++, clang, clang++; parses file:line:col: error: message from stderr; source snippets and caret indicators stripped; warning flags preserved as rule

dotnet build

• structured_return({ command: "dotnet build [any dotnet args]", parseAs: "dotnet-build-text" }) — parses MSBuild file(line,col): error CODE: message format; strips restore/timing noise; deduplicates the inline+summary error repetition; absolute paths relativized

javac

• structured_return({ command: "javac [any javac args]", parseAs: "javac-text" }) — parses file:line: error: message from stderr; source snippets and caret indicators stripped; cannot find symbol errors fold the symbol continuation line into the message

pyright

• structured_return({ command: "pyright --outputjson [any pyright args]", parseAs: "pyright-json" }) — --outputjson is built into pyright; JSON output includes file, line, message, and rule code; multi-line detail messages collapsed; warnings excluded from failures

tsc

• structured_return({ command: "tsc --noEmit --pretty false [any tsc args]", parseAs: "tsc-text" }) — --pretty false suppresses source snippets and ANSI codes; parser extracts file, line, TS error code, and message from the compact file(line,col): error TSXXXX: message format

pylint

• structured_return({ command: "pylint [any pylint args] --output-format=json", parseAs: "pylint-json" }) — --output-format=json is built into pylint; rule includes both message-id and symbol name (e.g. W0612(unused-variable))

shellcheck

• structured_return({ command: "shellcheck [any shellcheck args] --format=json", parseAs: "shellcheck-json" }) — --format=json is built into shellcheck; strips source snippets, caret indicators, "Did you mean" suggestions, and wiki URLs

rubocop

• structured_return({ command: "rubocop [any rubocop args] --format json", parseAs: "rubocop-json" }) — --format json is built into rubocop; cop name prefix stripped from messages; source snippets and caret indicators removed

swiftc

• structured_return({ command: "swiftc -typecheck [any swiftc args]", parseAs: "swiftc-text" }) — parses file:line:col: error: message from stderr; source annotations and duplicate error lines deduplicated; warnings filtered out

hadolint

• structured_return({ command: "hadolint [any hadolint args] --format json", parseAs: "hadolint-json" }) — --format json is built into hadolint; strips ANSI color codes and level labels from text output

stylelint

• structured_return({ command: "stylelint [any stylelint args] --formatter json", parseAs: "stylelint-json" }) — --formatter json is built into stylelint; rule name suffix stripped from message text; requires a .stylelintrc config

node --test

• structured_return({ command: "node --test [any node test args]", parseAs: "node-test-text" }) — parses the default spec reporter output; strips full stack traces, assertion error internals, timing data; preserves expected/actual values from assertion failures

ava

• structured_return({ command: "npx ava [any ava args] --no-color", parseAs: "ava-text" }) — parses default text output from stderr; assertion failures extract expected/actual from diff lines and file:line; runtime errors extract message and file:line from stack trace

mocha

• structured_return({ command: "mocha [any mocha args] --reporter json", parseAs: "mocha-json" }) — --reporter json is built into mocha; assertion failures surface explicit expected/actual values; runtime errors surface message and file:line from stack trace

Python unittest

• structured_return({ command: "python3 -m unittest [any unittest args]", parseAs: "unittest-text" }) — no special flags needed; parses the default verbose traceback output from stderr; assertion messages extracted from AssertionError lines; file:line from traceback frames

go test

• structured_return({ command: "go test -json [any go test args]", parseAs: "go-test-json" }) — -json is built into go test; NDJSON output; assertion failures extract file:line and message from t.Error/t.Errorf output; panics extract message and user-code file:line from stack trace; 97% reduction vs raw NDJSON

vitest

• structured_return({ command: "vitest run [any vitest args] --reporter=json", parseAs: "vitest-json" }) — file paths, filters, etc. go in [any vitest args]

eslint

• structured_return({ command: "eslint [any eslint args] -f json", parseAs: "eslint-json" }) — paths, configs, rules, etc. go in [any eslint args]

rspec

• structured_return({ command: "bundle exec rspec [any rspec args] --format json", parseAs: "rspec-json" }) — --format json is required; without it output is not parseable

minitest

• structured_return({ command: "ruby [any minitest args]", parseAs: "minitest-text" }) — no format flags needed; works with plain ruby invocation

dbt (run, test, compile)

• structured_return({ command: "dbt run [any dbt args] --log-format json", parseAs: "dbt-json" }) — --log-format json required; model selectors, flags, etc. go in [any dbt args]
• structured_return({ command: "dbt test [any dbt args] --log-format json", parseAs: "dbt-json" }) — same parser handles run and test; unit test diffs preserved verbatim
• structured_return({ command: "dbt compile [any dbt args] --log-format json", parseAs: "dbt-json" }) — returns compiled SQL; preamble stripped

cargo build

• structured_return({ command: "cargo build --message-format=json", parseAs: "cargo-build" }) — --message-format=json is built into cargo; errors include file, line, error code (E0308 etc.), and primary span label (expected X, found Y); warnings are filtered out

cargo test

• structured_return({ command: "cargo test", parseAs: "cargo-test" }) — no extra flags needed; assertion failures surface left/right values and file:line; panics surface the message and file:line; if compilation fails, the summary says so and tells the model to run cargo build --message-format=json for structured errors
• structured_return({ command: "cargo test [filter]", parseAs: "cargo-test" }) — filter by test name substring, module path, or --test integration_test_name

junit-xml

JUnit XML is the de facto standard output format across the JVM ecosystem and many others — Maven, Gradle, pytest (--junitxml), Playwright (--reporter=junit), PHPUnit/Pest (--log-junit), Go (go-junit-report), .NET (--logger trx with conversion), Jest (jest-junit), and more. If a tool can emit JUnit XML, junit-xml covers it without a custom parser.

artifactPaths supports glob patterns — the parser aggregates results across all matching files. Use globs for tools that write one XML per test class (Gradle, Maven/Surefire). Use a single path for tools that write one consolidated file (pytest, Jest, etc.).

• structured_return({ command: "[tool] [args] --junitxml=.tmp/report.xml", parseAs: "junit-xml", artifactPaths: [".tmp/report.xml"] }) — pytest, nose2, and any other tool that writes to a file via --junitxml
• structured_return({ command: "go test [any go test args] 2>&1 | go-junit-report > .tmp/report.xml", parseAs: "junit-xml", artifactPaths: [".tmp/report.xml"] }) — Go requires go-junit-report (go install github.com/jstemmer/go-junit-report/v2@latest); pipe go test -v output through it

Gradle

Gradle writes one XML file per test class — there is no single-file option. Always use a glob pattern.

• structured_return({ command: "gradle test", parseAs: "junit-xml", artifactPaths: ["build/test-results/test/TEST-*.xml"] }) — full test suite; results aggregated across all matched files
• structured_return({ command: "gradle test --tests 'com.example.FooTest'", parseAs: "junit-xml", artifactPaths: ["build/test-results/test/TEST-*.xml"] }) — single class; glob still works (matches the one file)
• For multi-module projects, use * in directory positions to match across modules: structured_return({ command: "gradle test", parseAs: "junit-xml", artifactPaths: ["*/build/test-results/test/TEST-*.xml"] }) — or scope to one module: structured_return({ command: "gradle :module:test", parseAs: "junit-xml", artifactPaths: ["module/build/test-results/test/TEST-*.xml"] })

Maven

Maven Surefire writes one XML per test class — same as Gradle. Always use a glob pattern.

• structured_return({ command: "mvn test", parseAs: "junit-xml", artifactPaths: ["target/surefire-reports/TEST-*.xml"] }) — full test suite; any maven args (e.g. -pl module, -Dtest=ClassName) go before test
• For multi-module projects, use * in directory positions: structured_return({ command: "mvn test", parseAs: "junit-xml", artifactPaths: ["*/target/surefire-reports/TEST-*.xml"] }) — or scope to one module: structured_return({ command: "mvn test -pl module", parseAs: "junit-xml", artifactPaths: ["module/target/surefire-reports/TEST-*.xml"] })

.NET / xUnit

dotnet test requires the JunitXml.TestLogger package (dotnet add package JunitXml.TestLogger). Pass the logger inline — no config file changes needed. Add a namespace to your test class or the logger will fall back to UnknownNamespace.UnknownType for classnames.

• structured_return({ command: "dotnet test --logger \"junit;LogFilePath=.tmp/report.xml\"", parseAs: "junit-xml", artifactPaths: [".tmp/report.xml"] }) — any dotnet test args (project path, --filter, etc.) go before --logger

Jest

Jest requires the jest-junit reporter (npm install --save-dev jest-junit). Pass it via --reporters on the CLI — no permanent config change needed.

• structured_return({ command: "jest [any jest args] --reporters=jest-junit", parseAs: "junit-xml", artifactPaths: [".tmp/junit.xml"] }) — set JEST_JUNIT_OUTPUT_FILE=.tmp/junit.xml or configure outputFile in jest-junit config; file paths, --testPathPattern, etc. go in [any jest args]

Playwright

Playwright has a built-in JUnit reporter. Set the output file inline so junit-xml has a concrete artifact to parse.

• structured_return({ command: "PLAYWRIGHT_JUNIT_OUTPUT_FILE=.tmp/report.xml playwright test [any playwright args] --reporter=junit", parseAs: "junit-xml", artifactPaths: [".tmp/report.xml"] }) — built-in JUnit reporter; file paths, --grep, project names, and shard flags go in [any playwright args]

PHPUnit

PHPUnit writes JUnit XML directly via --log-junit.

• structured_return({ command: "vendor/bin/phpunit [any phpunit args] --log-junit .tmp/report.xml", parseAs: "junit-xml", artifactPaths: [".tmp/report.xml"] }) — built-in JUnit logging; test paths, --filter, and suite selection flags go in [any phpunit args]

Pest

Pest supports the same JUnit XML logging flag as PHPUnit.

• structured_return({ command: "vendor/bin/pest [any pest args] --log-junit .tmp/report.xml", parseAs: "junit-xml", artifactPaths: [".tmp/report.xml"] }) — built-in JUnit logging; file paths, --filter, and parallelization flags go in [any pest args]

Swift / XCTest (Swift Package Manager)

swift test has native JUnit XML output via --xunit-output (Swift 5.7+). Use this for all SPM projects — no third-party reporter needed.

• structured_return({ command: "swift test --xunit-output .tmp/report.xml", parseAs: "junit-xml", artifactPaths: [".tmp/report.xml"] }) — full test suite
• structured_return({ command: "swift test --filter MathTests --xunit-output .tmp/report.xml", parseAs: "junit-xml", artifactPaths: [".tmp/report.xml"] }) — single test target or test case; --filter accepts a target name, class name, or ClassName/testMethodName

Swift / XCTest (Xcode projects)

xcodebuild output is high-volume and not directly parseable. Pipe through xcbeautify (install with brew install xcbeautify) to get JUnit XML:

• structured_return({ command: "xcodebuild test -scheme [scheme] -destination '[destination]' 2>&1 | xcbeautify --report junit --output .tmp", parseAs: "junit-xml", artifactPaths: [".tmp/report.junit.xml"] }) — xcbeautify writes report.junit.xml into the --output directory; run xcodebuild -showdestinations -scheme [scheme] first to find the correct [destination] string for the project

Already compact — use bash directly

These tools were evaluated for structured parsing but their output is already compact enough that a parser adds no token reduction. Use bash instead of structured_return.

• flake8 [any flake8 args] — file:line:col: CODE message, one line per violation
• ruff format --check . — just lists files needing formatting
• go build [any go build args] — file:line:col: message, one line per error
• go vet [any go vet args] — same format as go build
• golangci-lint run [any golangci-lint args] — file:line:col: message (linter), already minimal
• yamllint [any yamllint args] — file:line:col level message (rule), one line per issue
• pydocstyle [any pydocstyle args] — file:line context + CODE: message, two lines per issue
• vulture [any vulture args] — file:line: message (confidence%), one line per finding