Developing Flink Agents Applications

Core Model

Model every application as four connected parts:

1. Resources: named models, prompts, tools, skills, vector stores, MCP servers, and connections.
2. Actions: event handlers that use Resources and emit events.
3. Orchestration: trigger conditions and emitted-event graph.
4. Implementation: Python/Java functions, types, and runner.

Skills, vector stores, MCP, RAG, and memory are combinations of these parts.

The Agent API language and custom Action/Tool language do not constrain every Resource implementation. Chat-model connections/setups, embedding-model connections/setups, and vector stores support Python/Java bridging in both directions in the bundled snapshot. For those Resource types, offer every target-version implementation supported natively or through the bridge, regardless of whether the application uses YAML, direct Python, or direct Java. Treat the selected implementation's language as part of that Resource choice; do not add a separate application-wide cross-language confirmation. Do not generalize this rule to Resource types that lack a verified bridge.

Source Authority

This skill must work from its installed directory; never assume the user cloned the Flink Agents repository. Resolve bundled paths relative to this SKILL.md.

Inspect the target version and conventions. Use sources in this order:

1. Target application code, dependency metadata, installed package/JAR APIs, and tests.
2. Version-matching Flink Agents schema, docs, examples, or source when available.
3. This skill's YAML contract manifest, matching bundled schema, and references as the offline baseline.

Select contracts by the target Flink Agents version, not by the version that published this skill. The manifest maps released versions to complete schemas and marks versions without YAML support. The unversioned main schema describes only the repository revision that published this skill. Never validate a released application against that schema unless the manifest maps the selected version to it.

Do not invent APIs, versions, or commands from memory.

Before presenting code as runnable, resolve every nontrivial import, constructor, method, descriptor argument, and dependency coordinate against those sources. If a contract cannot be resolved, either use a documented alternative or label the fragment as pseudocode and state exactly what remains unresolved. Apply Provider Contract Triage below to external descriptor-backed Providers before loading bundled references. This blocked Provider path is different from an unresolved business contract, which still receives a user-fillable skeleton.

Preflight

Before loading a platform adapter, asking a decision gate, or generating files:

1. Confirm that the target is a Flink Agents application or is being converted into one. Leave ordinary Flink jobs outside this workflow.
2. Inspect the complete user request, existing project, dependency metadata, source, configuration, and tests. Resolve every framework decision already answered by that evidence. The latest explicit user choice takes precedence; treat a conflict with existing project metadata as an intentional requested migration and report it rather than silently restoring the old value.
3. Record each required framework decision as confirmed or unresolved: Flink Agents version, Flink version, API, YAML application language when applicable, each descriptor-backed Resource implementation, and the Python environment when Python is required.
4. Run Provider Contract Triage for every concrete external descriptor-backed Provider already named by the request or project. When the implementation is an unresolved gate, run the triage immediately after the user selects it.
5. Build the ordered gate list from only the unresolved decisions. If the list is empty, skip host detection and interaction adapters and proceed directly to contract resolution and implementation.
6. For YAML, select the exact target-version schema through yaml-contracts.yaml. Filter the API choices when the manifest says that the selected version has no YAML API. If a requested YAML version has no matching schema, stop YAML generation and request the exact target-version schema or source; never substitute the main schema.

Keep this decision record current after every answer. It prevents repeated questions and ensures that later references operate only on confirmed versions and capabilities.

Provider Contract Triage

This is an early-exit pass, not a later Build Workflow step. Run it as soon as a concrete external descriptor-backed Provider class or alias is confirmed and before reading any file under references/.

Inspect only evidence that can establish that Provider's contract:

1. Explicit user-supplied JARs, packages, source, tests, or documentation.
2. The target project's dependency metadata and the exact installed artifact named by that metadata.
3. Version-matching authoritative Provider source or documentation already available at a known location.

Scope every inspection command to the target application's dependency files, configuration, source roots, and explicitly named artifacts. Exclude the installed Skill directory containing this SKILL.md, including project-local .agents/, and exclude host metadata such as .codex/, .claude/, tool caches, and unrelated home or package caches. Do not run an unscoped recursive search over . when it would traverse those paths.

Do not load general application, YAML, language, local-development, or verification references to compensate for a missing Provider artifact. A general YAML schema can validate descriptor shape but cannot establish a Provider's dependency coordinate, constructor, or forwarded mandatory arguments. When the user explicitly states that no Provider artifact or contract source is available, treat that as evidence; check the target project metadata for an existing dependency, but do not search unrelated caches or examples.

The triage must establish the implementation class or alias, dependency coordinate or package, constructor/factory contract, and every mandatory descriptor argument. If any element remains unverifiable:

• preserve all working dependency, source, and configuration files;
• stop the dependent integration immediately, before the normal Build Workflow;
• do not generate a descriptor, adapter, wrapper, placeholder dependency, or tests;
• report the exact sources checked, each missing contract element, and the minimum JAR, package, source, test, or documentation needed to continue.

List a source under checked only when its content was actually inspected during the triage. Do not claim that a bundled schema, reference, cache, or documentation was checked merely because SKILL.md names it.

Return that blocked report for a Provider-only request. Continue other work only when the user explicitly requested an independent change that does not depend on the blocked Provider. If the complete Provider contract is verified, proceed with the normal workflow using that evidence.

Definition Strategy

Situation	Recommendation
New Workflow Agent or rewired event graph	Offer YAML plus Python/Java implementation files
Existing YAML application	Preserve and extend YAML
Existing programmatic Agent	Preserve its current API unless conversion is requested
ReActAgent, unsupported YAML surface, or explicit code-only request	Direct Python or Java API

For a new application, derive the available API choices from the confirmed target version. Present YAML, direct Python API, and direct Java API when all three are supported; otherwise omit unsupported surfaces. A recommendation explains the tradeoff; it is not permission to select the API for the user.

YAML does not declare an agent type. Do not add type: workflow or type: react; type selects an implementation language where the schema allows it.

YAML does not choose the business implementation language either. For a new YAML application, ask the user to choose Python or Java before generating files. For an existing application, detect and preserve the language from build metadata, source files, function references, and explicit YAML type fields. Do not default to Python merely because omitted YAML type fields currently resolve to Python.

Workflow Agents already include built-in chat, tool-call, and context-retrieval Actions. A model reasoning/tool loop can therefore remain YAML-defined. Choose ReActAgent only when the user explicitly wants that programmatic abstraction, the existing application already uses it, or a required ReAct-specific surface is not available through YAML.

Scaffolding Boundary

Complete framework-owned wiring, but do not invent business behavior. Before implementing each custom Action, Tool, domain client, data transformation, Prompt, or runtime Skill, classify these contract elements from only explicit user requirements, existing code, and existing tests:

Contract element	supplied means
Input	Required fields and their meaning are defined
Output	Result fields or emitted Events are defined
Transformation	Input-to-output or message construction is defined
Side effects	External calls, state changes, and authentication boundary are defined
Errors	Required failure and fallback behavior is defined

Anything not established by those sources is unresolved; a high-level capability name is not a supplied contract. When behavior depends on an unresolved element, generate only an importable or compilable signature skeleton with focused TODOs and an explicit NotImplementedError or UnsupportedOperationException. Preserve user-provided names, parameters, types, and descriptions. When those are absent, derive a stable capability name and use the narrowest framework-compatible signature; for a function Tool or domain client with no typed contract, use one opaque string request and string result.

An unresolved skeleton must not emit business Events, transform payloads, compose Prompt or chat messages, call a backend, return business data, or provide fallback behavior. Tests may check import/compilation, signature and YAML reference resolution, and explicit failure, but must not assert business behavior that was not supplied.

Do not author domain rules, REST endpoints, diagnostic procedures, prompts, data models, Tool results, runtime Skill instructions, or test doubles from a high-level application idea. Do not propose a standard REST API, MCP server, or mock backend as the implementation of a custom Tool. Do not ask the user to lock a business input identity schema, deployment platform, service API, authentication design, or log/ metric backend merely to scaffold the application. Generate the neutral skeleton and leave those choices as TODOs for the user. Built-in Flink Agents Actions are framework behavior and need no generated implementation.

Interaction Discipline

This section is the sole authority for interaction capability detection, fallback selection, retry behavior, and host-mode handling. Platform adapters only describe how to encode a question for a tool that this section selected; other references must point here instead of restating this policy.

Process the unresolved decisions from Preflight as sequential gates. Ask only the current gate, wait for the answer, update the decision record, and continue. Never send one proposed baseline that bundles versions, API, implementation language, runtime version, Resource implementations, business backends, and mock behavior.

Only when at least one unresolved closed gate remains, identify the current host from explicit system and tool context. Use this table to select the sole candidate tool and adapter for the gate:

Host	Candidate structured tool	Adapter
Codex	request_user_input	codex.md
Claude Code	AskUserQuestion	claude-code.md
Gemini CLI	ask_user communication tool	gemini-cli.md
Qoder	Explicitly exposed structured single-select tool, if any	qoder.md
Unknown or unsupported	Explicitly exposed structured single-select tool, if any	generic.md

Treat the current tool contract as the only capability signal. When the table's candidate is exposed and callable, read its adapter to encode one question. An adapter must not perform capability discovery or select a fallback. If the candidate is absent, returns an availability error, or cannot represent the complete valid option set without altering or obscuring it, immediately read generic.md and use its numbered fallback. Do not retry the tool, ask the user to change modes, require a slash command or keyboard shortcut, perform a post-interview handoff, add host metadata, or install host configuration in the generated application.

Never use an open-ended text question when the valid options are already known. When a gate has a recommendation, place it first and label it (Recommended), but do not preselect or continue without the user's answer. The YAML implementation-language gate intentionally has no recommendation: present Python and Java as equal peer options with parallel descriptions and no (Recommended) label.

Use this order:

1. If unresolved, ask for the Flink Agents version and wait. Then offer only compatible Flink versions and wait for the Flink choice. Do not mention Python/JDK versions, an Agent API, a model provider, or business architecture in these version questions.
2. If unresolved, ask the user to choose among the APIs supported by the confirmed Flink Agents version. Offer YAML only when its exact contract is available. Wait.
3. Only when YAML is selected, ask whether custom Actions, function Tools, and the Flink entry point use Python or Java. Resource implementations are selected independently at their own gates. Direct Python or Java API already determines the application-code choice. Do not recommend, preselect, or imply a preference for either YAML implementation language. Only now resolve a compatible Python or JDK version.
4. Inventory the Resources required by the user's stated design. Resolve real descriptor-backed framework integrations one at a time and wait before moving to the next integration. Assign a deterministic internal name, ask for its implementation class or documented alias, and stop the interview for that Resource. Immediately run Provider Contract Triage after the implementation choice; it verifies the target-version constructor and descriptor. If it passes, generate every mandatory configuration key as TODO_REQUIRED_<FIELD> for the user to fill. Do not ask about another Resource or load general references before the triage result. For chat-model connections/setups, embedding-model connections/setups, and vector stores, do not filter candidates by the Agent API or application-code language. Present all verified Python and Java implementations, label each candidate's implementation language when needed for disambiguation, and generate the matching bridge declaration and runtime dependencies after selection. Do not ask for model identifiers, endpoints, credential values/mechanisms, provider options, Skill source paths/URLs/packages, or other Resource arguments. Custom Actions, function Tools, domain clients, Prompt content, runtime Skill instructions, business input schemas, and backend platforms are not integration gates: scaffold them without a business interview. Do not ask the user to name a single Resource or repeat a generated reference. Ask about naming only when multiple Resources need semantic disambiguation, an existing external reference constrains the name, or the user requested a naming convention.
5. As soon as the confirmed design first requires a Python runtime, pause before environment creation or dependency installation. Unless the existing project already declares its environment unambiguously, inspect compatible local Python executables/environments and ask whether to reuse one of them or create a project-local .venv. Wait for the choice and use the selected interpreter for every install, import, test, and local run. This conditional gate fires immediately after the choice that introduces Python; do not delay it until the end of the Resource interview.
6. After every framework decision is confirmed, design the Action graph and generate the project. Do not add a host-specific handoff.

Do not infer OpenAI, Ollama, any model name, an environment-variable credential, Skill distribution, MCP, a vector store, or a domain-service protocol. After a framework implementation is selected, do not run a second configuration interview. Generate a declaration or builder scaffold that names every verified mandatory argument and leaves its value explicit for the user. For YAML string fields, use a clear placeholder such as TODO_REQUIRED_API_KEY; for direct APIs, generate a compilable factory skeleton that lists the required arguments and fails explicitly until they are filled. If the user already supplied a value or explicitly requested plaintext in local YAML, use that instruction; keep supplied secrets out of tracked files and output. Never claim ${ENV_VAR} is interpolated unless the target loader or provider actually implements it.

The conditional Python-environment gate is dependency management, not Resource configuration. It still runs when a selected Resource introduces Python, but it asks only which concrete compatible Python environment to use, never provider values.

Build Workflow

1. Run Preflight, including Provider Contract Triage for every concrete external Provider already named by the request or project. If a Provider is blocked, issue its report and stop before reading any bundled reference.
2. Read application patterns, then ask only the ordered unresolved gates. Do not repeat decisions already confirmed by the user or project.
3. When versions are unresolved, complete the version gates in local development: Flink Agents first, then a compatible Flink version. Do not generate files or combine later decisions into these questions.
4. Complete the API gate with only APIs supported by the confirmed target version. For YAML, read YAML patterns, select the exact schema through yaml-contracts.yaml, and reject generation when no exact contract is available.
5. If YAML was selected, complete the implementation-language gate. Then read Python patterns or Java patterns for the application code. When a bridge-supported Resource uses the other language, also read that language's cross-language Resource section; selecting that Resource is already explicit confirmation.
6. Inventory each Resource and assign stable names automatically. Ask only which documented implementation alias/class to use for actual framework integrations. For bridge-supported types, build this choice from both Python and Java implementations instead of the application-code language alone. Inspect the selected implementation through Provider Contract Triage before loading more references or moving to another integration. When it passes, add its integration and bridge dependencies and scaffold all mandatory configuration keys without asking for their values. Classify every custom Tool/Action/Prompt/runtime-Skill business contract before generating its skeleton; do not ask the user to choose a domain platform or complete an unresolved input contract. Generate and record all cross-references without asking the user to repeat internal identifiers. Never select a provider, model, endpoint, credential source, business backend, or integration from the application domain alone.
7. When runtime Skills are requested, preserve an explicit source in an existing application. For a new application, generate a minimal runtime SKILL.md business scaffold plus a source-configuration TODO that lists the valid paths, urls, Python package, and Java classpath forms. Do not ask for loading paths or distribution, and never inspect, copy, or offer to reuse Skills installed in the coding-agent host, such as a local flink-diag; coding-agent Skills and Flink runtime Skills are different artifacts. Read YAML patterns.
8. Draw only framework Action graph edges established by built-in contracts or supplied business contracts. Where an emitted Event, payload transformation, branch, output, or error contract is unresolved, generate the Action signature and explicit failure but do not claim or implement that edge. Do not turn those TODOs into more gates.
9. Use only documented built-in Actions. Generate a resolvable, correctly typed signature skeleton for every other Action and reference it as <module-or-class>:<qualname>; leave its business body for the user unless they explicitly requested implementation.
10. Resolve concrete API calls and dependency coordinates from the target-version artifacts before generating the complete executable project. For Java, generate a Maven project with flink-agents-api, flink-agents-plan, flink-agents-runtime, and only the integrations actually used; declare all Flink Agents and Flink dependencies as provided. Whenever the design requires Python, generate source and dependency files, resolve the Python environment choice when not already declared, and install the resolved Flink Agents, PyFlink, and integration dependencies into the selected existing environment or project-local .venv. Read local development. Preserve versions already selected by the target project; never guess one.
11. Connect the Agent to a Flink DataStream or Table through the public factory backed by RemoteExecutionEnvironment. Local validation submits that same remote-style job to a MiniCluster. Never use a local Agents environment, a no-argument factory, from_list/to_list, or their Java equivalents.
12. Run the checks in verification before claiming the application is valid or runnable.

Required Output

• Complete framework files or edits, including every custom function signature and a remote-style Flink job entry point.
• A runnable Maven project for Java and, whenever Python is required, pinned Python dependency input plus the user-selected existing environment or populated project-local .venv.
• User-confirmed Flink Agents and Flink versions. Never silently choose the bundled recommended versions for a new project.
• A user-confirmed API choice from the surfaces supported by the selected Flink Agents version.
• A user-confirmed Python or Java application-code language for every new YAML application, reflected consistently in project layout, custom Action/Tool function references, and the Flink entry point. Do not use it to filter bridge-supported Resource implementations.
• For applications using runtime Skills, a user-fillable Skill business scaffold and source-configuration TODO. Preserve existing source configuration, but do not ask a new-project user to choose distribution or reuse coding-agent host Skills.
• Resource declarations built from user-selected implementation aliases/classes, including independent Resource implementation language and bridge wiring where supported, with every verified mandatory provider key present and marked for user input. Keep model, endpoint, credential, and optional provider values unresolved instead of interviewing for them or choosing defaults.
• Deterministic Resource names and references generated by the coding agent. Do not require the user to name ordinary Connection, Setup, Prompt, Skill container, VectorStore, or MCP Resource identifiers when there is no ambiguity.
• When the user chooses plaintext credentials for local testing, a loader-compatible local YAML that is excluded from version control and actually used by the local run command. Do not replace this explicit choice with programmatic registration.
• Explicit user-fillable business skeletons for custom Actions, Tools, prompts, runtime Skills, domain clients, secrets, endpoints, and external data, backed by the supplied/unresolved contract classification. Keep framework wiring and signatures concrete instead of replacing them with fabricated business implementation.
• A final consolidated User must provide list for unresolved business input fields, platform clients, authentication, queries, response mapping, and domain behavior. These items must not block project scaffolding or become Agent workflow decisions unless the user explicitly asks for their implementation.
• Exact commands that match the target repository's build tooling; no guessed package, Flink, provider, model, or plugin versions.
• Framework snippets whose imports and API calls resolve in the target version. Business skeletons must import or compile, fail explicitly when invoked, and be labeled user-fillable rather than runnable behavior.
• For YAML, the selected contract key and exact schema path used for validation.
• Evidence separated into schema, load/compile, tests, and runtime.
• A clear statement for every check that was not run or requires an external service.

Quick Reference

Task	Read
Present closed choices in the current coding agent	Interaction Discipline, then its selected platform adapter
Select Agent/API shape; design Resources and event graph	application-patterns.md
Author or review YAML; resolve names and functions	yaml-patterns.md
Scaffold runtime Skill business and source TODOs	runtime Skills
Scaffold Python Actions, Tools, types, or runner	python-patterns.md
Scaffold Java Actions, Tools, resources, or runner	java-patterns.md
Select versions, generate dependencies, and submit to MiniCluster	local-development.md
Validate schema, references, imports, compilation, and execution claims	verification.md
Select and validate YAML without a source checkout	YAML contract manifest