yukta
CLAUDE.md1 starsChanged 20 days agoApache-2.0View on GitHub
This is the global repository guide for Yukta. Yukta is a reactive, DAG-based workflow orchestration server: define workflows once (JSON nodes/edges) and run them via REST, a web UI, or an MCP server for AI agents. It ships with built-in Trigger/Processor/Terminal plugins (CI/CD, data pipelines, quality gates, etc.) and is extensible via plugin-api. Code-quality-gate enforcement for AI-driven development is one use case, not the whole product. Java 25 (toolchain) · Spring Boot (WebFlux; version in gradle/libs.versions.toml) · Gradle 9.8 (wrapper)…
In this repository21 other files it gives its agents.
- architecture/CLAUDE.mdCLAUDE.md
- boot/CLAUDE.mdCLAUDE.md
- build-logic/CLAUDE.mdCLAUDE.md
- .claude/CLAUDE.mdCLAUDE.md
- cli/CLAUDE.mdCLAUDE.md
- core/CLAUDE.mdCLAUDE.md
- mcp/CLAUDE.mdCLAUDE.md
- messaging/CLAUDE.mdCLAUDE.md
- plugin-api/CLAUDE.mdCLAUDE.md
- plugins/CLAUDE.mdCLAUDE.md
- plus/CLAUDE.mdCLAUDE.md
- ui/CLAUDE.mdCLAUDE.md
- web/CLAUDE.mdCLAUDE.md
- ide-index-mcpSkill
- pickupSkill
- playwrightSkill
- testsSkill
- ticketSkill
- yuktaSkill
- yukta-supervisorSkill
- yukta-workflowsSkill
The file
We split this CLAUDE.md into the rules it gives an agent
11 rules. About 0 are common, 2 are distinctive, and 9 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
9
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 17
architecture: test-only ArchUnit rules (module layering, plugin independence, coding conventions). - Very specific: line 50
Every run ends with one RESULT status=… failed_node=… line.
- Distinctive: line 51
Run the smallest scope that covers your change (a test class, then a package, then a module's tasks).
- Very specific: line 61
Check gh api rate_limit --jq .resources.graphql before a batch of project-board calls.
- Very specific: line 63
Prefer REST (gh issue list, gh issue view, gh pr list) when it answers the question.
- Very specific: line 85
All commits follow Conventional Commits format.
- Very specific: line 102
Use StepVerifier for testing reactive streams.
- Very specific: line 110
Default: yk gradle :<mod>:test --tests '<FQCN>' (repeat --tests per class.
- Very specific: line 112
Run spotlessApply once at the end, not after every edit.
- Very specific: line 113
Run long builds with run_in_background.
- Distinctive: line 119
Description: Use for generating unit tests (Java, JUnit 5, Mockito, AssertJ).
≈ “Common” is an estimate: at least 1 in 5 of the 27,641 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Yukta Project CLAUDE.md
This is the global repository guide for Yukta.
Project Overview
Yukta is a reactive, DAG-based workflow orchestration server: define workflows once (JSON nodes/edges) and run them via REST, a web UI, or an MCP server for AI agents. It ships with built-in Trigger/Processor/Terminal plugins (CI/CD, data pipelines, quality gates, etc.) and is extensible via plugin-api. Code-quality-gate enforcement for AI-driven development is one use case, not the whole product.
Tech Stack
Java 25 (toolchain) · Spring Boot (WebFlux; version in gradle/libs.versions.toml) · Gradle 9.8 (wrapper) · Project Reactor · GraalVM native image · Spring AI MCP Server.
Repository Roadmap & Layout
This is a multi-module Gradle monorepo. Launch Claude from the specific package folder for efficiency — each module below has its own CLAUDE.md with module-specific build commands and architecture notes; this file covers only what's true repo-wide.
architecture: test-only ArchUnit rules (module layering, plugin independence, coding conventions);plus/plus-architecturedoes the same for Plus. See.claude/rules/architecture.md.boot: Spring Boot application entry point and GraalVM native config.build-logic: Gradle build conventions (Java, Quality, JaCoCo).cli: Lightweight Go-based CLI for remote server interaction.core: Core services, DAG orchestration engine, and MCP tools.mcp: MCP server implementation.messaging: Shared messaging abstractions.plugin-api: Abstraction interfaces for Yukta plugins.plugins: Implementations of Processors, Triggers, and Terminals.ui: React SPA (Vite, Tailwind CSS) built into a static JAR forboot.web: REST controllers and web-layer mapping.
Global Commands
Builds, tests and lint (Gradle, Go, pnpm/npx) run through Yukta, never ./gradlew directly (the yukta-dogfood mod blocks it and names the command to
use). yk is the one of the sibling yukta-dogfood checkout: run scripts/dogfood-setup.sh once (it clones the ref pinned in .dogfood-ref; on PATH in Claude Code sessions after that); yk gradle and yk tool run every call as
the tool Yukta workflow on the dev server: built from origin/main, on http://127.0.0.1:9858, started
on demand. See the yukta skill.
yk gradle :core:compileJava :core:compileTestJava
yk gradle :core:test --tests com.infenia.yukta.service.orchestrator.WorkflowOrchestratorTest
yk gradle spotlessApply # format (before commit)
yk gradle :architecture:test :plus:plus-architecture:test # the ArchUnit architecture rules
yk check # the scoped, self-healing gate for your changes (yk check --full before a PR); yk plan shows what it would run
yk tool -C cli go test ./... # Go (cli/)
yk tool -C ui npx --no-install prettier --check . # UI (ui/): pnpm/npx/node_modules/.bin commands
# yk, heal, plan, hooks (TypeScript on Bun) are tested in the yukta-dogfood repository, not here
# Run the application you are developing (not a build: stays direct; Swagger UI at http://localhost:9858/swagger-ui.html)
./gradlew bootRunEvery run ends with one RESULT status=… failed_node=… line; read that and the ERROR lines before it, not logs.
Run the smallest scope that covers your change (a test class, then a package, then a module's tasks).
Rebuild in progress (see the yukta-dogfood README): only
yk gradle,yk tool,yk planandyk check(with healing) are exposed so far. The previous guide, with issue pickup, steering and the heal reports, is in git history (examples/11-dogfood/backup/CLAUDE.md).
GitHub API budget
The GraphQL quota is shared per user by every Claude session and tool on the machine. gh project item-list pages the whole board and is the
heaviest call; the pickup tick runs it on every round.
- Check
gh api rate_limit --jq .resources.graphqlbefore a batch of project-board calls. Never callgh project item-listin a loop or only to verify one item. - On "API rate limit exceeded": do not retry. Report the
resettime (date -d @<reset>) and wait. - Prefer REST (
gh issue list,gh issue view,gh pr list) when it answers the question.
Files Claude cannot Read
Checked-in generated or vendored files that would fill the context are denied to Claude's Read (and to cat/grep/head on them):
| Pattern | What | Use instead |
**/pnpm-lock.yaml | UI lockfile (210 KB) | pnpm why, pnpm list via yk tool -C ui |
**/openapi.json | API snapshots ui/openapi.json, docs/api/openapi.json (200 KB each) | the controllers; docs/api/*.mdx |
**/*.generated.* | ui/src/types/api.generated.ts (180 KB) | the hand-written types in ui/src/types/ |
**/META-INF/native-image/**/* | boot reflection config (115 KB, regenerate with the agent, see boot/CLAUDE.md) | @RegisterReflectionForBinding |
**/dogfood/session.json | generated by scripts/gen-*.py | the generators and heal.json |
build/, dist/ and node_modules are not denied: they are gitignored (search skips them) and build/test-results is read on purpose.
The rules are in .claude/settings.json; Claude Code reads the shared settings of the directory it starts in, not of the repository
root, so ui/, boot/ and docs/ carry their own .claude/settings.json with the rules for their files. Add a
rule to the root file and to the module that holds the path. Worktree sessions get them with the checkout. To work on a denied file,
remove its rule in your own checkout for the session (a deny cannot be overridden by allow or settings.local.json).
Commit Message Conventions
All commits follow Conventional Commits format. See .gitmessage template (shown when committing), .git-commit-guide.md, and .claude/rules/git-workflow.md for full details. The .git/hooks/commit-msg hook validates format automatically.
Worktrees
claude --worktree and subagent worktree isolation create worktrees under .claude/worktrees/. .claude/settings.json sets
worktree.symlinkDirectories to ui/node_modules, so each one links to the main checkout's copy (about 440 MB) instead of installing its
own. yk tool -C ui (tsc, vitest) works through the link. Caveats: a pnpm install (Gradle's ui tasks run one) in such a worktree
writes into the shared copy, so run it deliberately when a branch changes ui/package.json or ui/pnpm-lock.yaml; worktrees made with
git worktree add (the pickup skill, yk runs) do not get the link.
worktree.sparsePaths is not set (measured 2026-10-31): the tracked tree is 16 MB and a full worktree is created in 0.2 s (0.1 s sparse),
and a sparse tree breaks the build: Gradle fails with "Configuring project ':plugins:processors:internal' without an existing directory is not
allowed" because settings.gradle.kts includes every module, and yk itself lives in the sibling yukta-dogfood checkout. Keeping all modules and examples
leaves only docs, design and .github out, a few MB.
Testing Strategy
- Unit tests use JUnit 5 + Mockito (for reactive:
reactor.test). - Use
StepVerifierfor testing reactive streams. - Quality gates (Checkstyle, PMD, SpotBugs, OpenGrep) are enforced via
build-logic— see.claude/rules/coding-standards.mdfor details. - Minimum coverage thresholds are module-specific.
Test Execution Policy (speed)
Waiting on tests is the main development bottleneck. Run the smallest scope that covers your change.
- Quick compile feedback:
yk gradle :<mod>:compileJava :<mod>:compileTestJava. - Default:
yk gradle :<mod>:test --tests '<FQCN>'(repeat--testsper class;Class.methodwhile iterating). - A public contract or many classes in a package:
--tests 'com.infenia.yukta.<package>.*'. - Run
spotlessApplyonce at the end, not after every edit. Don't usecleanor--rerun-tasks. - Run long builds with
run_in_background; on failure read only the failing test's report inbuild/test-results. - Locate tests via
<Class>Testnaming or codebase-memory, not repo-wide grep.
Test Skill
Load the below skill for test generation and test improvement.
- Name: tests
- Description: Use for generating unit tests (Java, JUnit 5, Mockito, AssertJ).
Imports
The yukta skill is imported so it is always loaded (also after /compact and in subagents), not only when its description matches.
@.claude/skills/yukta/SKILL.md @.claude/rules/coding-standards.md @.claude/rules/git-workflow.md
IMPORTANT: Navigate code with the codebase-memory MCP tools first (see .claude/CLAUDE.md); when the IDE is open, the intellij-index MCP tools are fine for refactoring.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

