docs: add agent harness instructions

This commit is contained in:
2026-08-07 11:18:58 +09:00
parent 8b0bda7a3f
commit d35c79217b
5 changed files with 348 additions and 0 deletions

21
.agent/profile.yml Normal file
View File

@@ -0,0 +1,21 @@
schema_version: 1
project:
id: kri-contact-server
name: KriContact Server
type: internal-service
layer: backend
legacy: false
stack:
- java-springboot
shared_contracts:
consumes:
- kri-common
databases:
- id: contact
engine: mysql
role: primary-application-data
access: read-write

30
.agent/project.md Normal file
View File

@@ -0,0 +1,30 @@
## KriContact Server Repository Rules
### Product Purpose and Current Phase
- This is the launched Java 17 and Spring Boot backend for KRI contact operations. Preserve first-release API and data behavior while delivering approved follow-up features.
- The domain includes projects, contact lists and headers/items, uploads and downloads, assignments, call sessions, call/event/update histories, progress, scenarios and images, result codes, users, companies, roles, message templates, reservations, SMS, and email workflows.
- Prefer domain correctness, operational safety, and backward-compatible evolution over broad refactoring.
### Domain Invariants and API Compatibility
- Preserve company, project, role, and user authorization boundaries for every query and mutation.
- Treat list schemas, uploaded columns, item values, filtering, reset/update operations, assignments, and all history records as one coordinated consistency boundary.
- Make call start, close, and hang-up state transitions explicit and idempotent where retries or duplicate requests are possible.
- Preserve external-list endpoint request, response, authentication, error, and pagination behavior as a consumer contract.
- Validate file names, paths, sizes, content handling, storage lifecycle, and cleanup scheduling at boundaries. Do not expose server paths or retain files beyond the approved lifecycle.
- Do not send real SMS or email, initiate real calls, or invoke production business integrations during development or validation.
### Shared Library and Integrations
- Consume `com.kri.common:starter:beta.1.6.3`. Do not change this approved version or duplicate an available KriCommon feature without separate explicit approval.
- Import public KriCommon APIs from `com.kri.common.feature.*` and keep application code out of KriCommon runtime internals.
- Treat IntranetExtension, NotiHub, ShortUrl, CTI, file storage, and other external dependencies as explicit boundaries with validated inputs, timeouts, error mapping, retry/idempotency decisions, and safe rollback behavior.
- Never log API keys, credentials, full message bodies, telephone numbers in bulk, uploaded list contents, or unrestricted personal information.
### Persistence and Verification
- Preserve MySQL transaction boundaries, QueryDSL/JPA ownership, locking, ordering, null, date/time, and generated-key behavior.
- Do not edit generated QueryDSL sources directly. Change source entities and regenerate only through the repository workflow when required.
- `AGENTS.md` and `CLAUDE.md` are generated. Edit `.agent/profile.yml` or `.agent/project.md`, then recompose instead of editing generated files.
- Use the Gradle Wrapper and run the narrowest safe compile and focused tests. Do not start the service, connect to a real database, or call external integrations without separate approval.

5
.gitattributes vendored
View File

@@ -1,3 +1,8 @@
/.gitattributes text eol=lf
/gradlew text eol=lf /gradlew text eol=lf
*.bat text eol=crlf *.bat text eol=crlf
*.jar binary *.jar binary
/.agent/** text eol=lf
/AGENTS.md text eol=lf
/CLAUDE.md text eol=lf

291
AGENTS.md Normal file
View File

@@ -0,0 +1,291 @@
<!--
GENERATED BY KRI AGENT HARNESS. DO NOT EDIT DIRECTLY.
Project: kri-contact-server
Schema version: 1
Edit the source fragments or repository .agent files, then recompose.
-->
# KRI Development Base Instructions
This document defines the shared working rules for KRI development repositories.
Keep project-specific facts and commands in the project profile and project rules.
## Instruction Resolution
Apply repository instructions in this order:
1. Base rules
2. Project-type rules
3. Technology-stack rules
4. Shared consumer contracts
5. Project-specific rules
6. The explicitly approved task scope
Later and more specific instructions may override earlier general instructions only within their stated scope.
- Apply a shared consumer contract only when `.agent/profile.yml` declares it.
- Do not silently resolve a material instruction conflict. Report the conflicting rules and request a decision before changing files.
- Rules involving destructive actions, credentials, external systems, production, deployment, databases, or Git publication require explicit authorization even when a later rule is less restrictive.
## Communication
- Respond to the user in Korean unless the user requests another language.
- Keep code identifiers, commands, paths, protocol names, and external API terms in their original form.
- Follow the repository's existing language conventions for code comments and user-facing messages.
- Lead with the result or decision, then provide only the detail needed to verify it.
- Do not create or retain a full session transcript.
## Work Authorization
- The default workflow is discussion first and explicit execution approval second.
- During discussion, perform only read-only investigation needed to understand the task and propose a concrete plan, unless the user prohibits inspection.
- Read-only investigation may include file inspection, code search, encoding checks, `git status`, `git diff`, `git log`, and sanitized official-documentation lookup.
- Before execution approval, do not edit files, install dependencies, run builds or tests, start services, connect to databases, or change local or external state.
- Before requesting approval, describe the intended scope, affected areas, validation plan, and material risks.
- Treat approval as limited to the presented scope. Pause and discuss again if the work materially expands.
- If the user explicitly requests immediate execution without prior discussion, the discussion gate may be skipped for the stated scope.
- Fast-track authorization does not implicitly authorize protected operations such as destructive commands, production changes, deployments, database mutations, Git publication, or external messages.
## Local Development Environment
- Primary OS: Windows 11.
- Primary shell: PowerShell.
- Use Git Bash only when the repository already depends on Bash or POSIX shell behavior.
- Prefer repository-provided wrappers and scripts over globally installed commands.
- Do not assume that GNU utilities, WSL, Docker, or a Unix-compatible shell are available unless the project profile confirms them.
- Detect Java, Node.js, .NET, Python, package-manager, and database-tool versions from repository configuration.
- Do not install or modify global toolchains as part of a project task.
Preferred human-facing tools:
- Classic ASP: Visual Studio Code
- Java and Spring Boot: IntelliJ IDEA
- C# and .NET: JetBrains Rider
- JavaScript, TypeScript, and Next.js: JetBrains WebStorm
- Python: JetBrains PyCharm
- Microsoft SQL Server: SQL Server Management Studio
- Other databases: JetBrains DataGrip
- These preferences do not authorize launching or controlling a GUI application.
- Prefer IDE-independent commands that can also run in CI.
- Do not modify global IDE settings.
- Do not create or modify `.idea`, `.vscode`, workspace, or run-configuration files unless the approved task requires it.
- Preserve existing shared IDE configuration when it is already tracked.
## PowerShell and Path Safety
- Quote paths containing spaces, brackets, parentheses, Korean characters, or PowerShell metacharacters.
- Prefer `-LiteralPath` when operating on an exact filesystem path.
- Use project-relative paths in committed configuration and documentation. Do not commit machine-specific paths such as `D:\_____Dev\...`.
- Do not pass filesystem targets between PowerShell, `cmd.exe`, and Bash for destructive or recursive operations.
- Keep a filesystem operation within one shell from target resolution through execution.
- Before a recursive delete or move, resolve and verify every absolute target.
- Never use a drive root, workspace root, user profile, unresolved variable, wildcard, or broad glob as a recursive delete target.
- Prefer recoverable operations when practical.
## Change Discipline
- Preserve all existing user changes, including uncommitted and untracked work.
- Keep changes minimal but complete for the approved objective.
- Avoid unrelated cleanup, formatting, renaming, dependency upgrades, and architectural refactoring.
- Do not overwrite or revert unrelated changes.
- Stop and report when the requested change overlaps existing work in a way that cannot be handled safely.
- Prefer changing an authoritative source over editing generated output.
- Treat files marked as generated as read-only unless regeneration is part of the approved plan.
- Do not edit vendored dependencies, build outputs, caches, or generated clients unless explicitly required.
- Do not regenerate a broad set of files for a narrow change without first explaining the expected output scope.
## Text Encoding and Korean Content Safety
- Treat the existing file encoding, BOM state, and line-ending style as part of the file's contract.
- Preserve Korean comments, user-facing messages, error messages, identifiers, and documentation exactly unless the task explicitly requires changing them.
- Do not assume that garbled terminal output means the file itself is corrupted. Verify bytes and encoding before modifying suspicious text.
- Before editing an existing text file, determine its current encoding and BOM state when repository configuration does not already guarantee them.
- Do not convert encoding, add or remove a BOM, or normalize line endings as a side effect of an unrelated change.
- Avoid whole-file rewrites for small changes.
- Do not rewrite source files through `Get-Content | Set-Content`, `Out-File`, shell redirection, or another method whose encoding behavior is implicit.
- Do not use a patching or formatting tool on a non-UTF-8 file unless it is known to preserve that file's encoding.
- Treat a console display problem and an on-disk encoding problem as separate issues.
- After editing Korean text, inspect the diff for replacement characters (`<60>`), mojibake, unexpected escapes, BOM changes, and unrelated line-ending changes.
- Run `.agent/tools/check-encoding.ps1` when it is available and the change touches Korean text or a non-default encoding.
- If encoding cannot be identified confidently, stop before writing and report the affected file and evidence.
- For a new text file, follow repository configuration. If no rule exists, use UTF-8 without BOM.
## Git Safety
- Use `git status`, `git diff`, and `git log` for read-only investigation.
- Do not run `commit`, `push`, `pull`, `fetch`, `merge`, `rebase`, `reset`, `checkout`, `switch`, `stash`, `clean`, `tag`, or force operations unless that specific operation is explicitly approved.
- Source-edit approval does not include commit or push approval.
- Never amend an existing commit unless explicitly requested.
- Before an approved commit, show the intended files, validation results, and proposed commit message.
- Never push directly to a protected branch unless the user names the target branch and explicitly authorizes the push.
## Dependencies and External Actions
- Use the package manager, lockfile, and toolchain already selected by the repository.
- Do not install global packages.
- Do not add, remove, or upgrade a dependency unless it was included in the approved plan.
- Pause for approval if an unplanned dependency becomes necessary.
- Keep lockfiles synchronized with approved dependency changes.
- Do not deploy, publish packages, create releases, send messages or email, call business APIs, or modify external services unless explicitly requested.
- Never upload proprietary source code, internal documents, credentials, customer data, or database contents to an external service.
- Remove secrets and unnecessary internal identifiers from external search queries.
## Database Safety
- Treat writing or reviewing SQL as source code separately from executing SQL against a database.
- SQL files, migrations, schema proposals, and query drafts may be created within an approved source-change scope.
- Do not connect to a shared, staging, or production database without explicit authorization for that environment.
- Read-only queries against a non-local database also require authorization because results may contain sensitive data.
- Do not execute DDL, DML, stored procedures, migrations, administrative commands, or data-correction scripts against a non-disposable database without separate authorization for the exact target and operation.
- Source-change approval does not authorize migration execution.
- Use a disposable local or test database for approved validation only after verifying its identity and disposable nature.
- Never place credentials, connection strings, or sensitive query results in source files, logs, generated documents, or chat output.
## Secrets and Sensitive Data
- Never add credentials, tokens, passwords, private keys, connection strings, session identifiers, or production data to source control.
- Read sensitive files only when the approved task requires it and no safer source exists.
- Prefer inspecting configuration keys and schemas without displaying secret values.
- Never echo secrets into terminal output, chat responses, logs, diffs, generated documents, test fixtures, or screenshots.
- Redact sensitive values when reporting errors or command output.
- If a possible secret is discovered, do not repeat it. Report only its location, type, and recommended containment action.
- Do not rotate, revoke, copy, or otherwise modify credentials unless that exact action is explicitly authorized.
## Validation
- Discover supported build, test, lint, type-check, and formatting commands from repository files and project instructions. Do not invent commands.
- Include intended validation in the discussion-phase plan.
- After approved changes, run the narrowest relevant validation first, then broader checks when appropriate.
- Limit formatting to files within the approved scope.
- Do not run tests that require a shared database, production service, paid external API, or sensitive data unless separately authorized.
- Do not claim that a check passed unless it was executed successfully.
- If a check cannot be run, state the exact reason.
- Distinguish a current-change failure from a pre-existing failure or an undetermined failure.
- Do not fix unrelated failures without expanding the approved scope.
- Inspect the worktree after validation for unexpected generated files, formatting churn, encoding changes, or lockfile changes.
## Artifacts and Session Records
- Do not create full-session transcripts, automatic HTML copies, analysis reports, or extra documentation unless explicitly requested.
- Store durable information in the repository's existing documentation structure.
- Do not place generated reports or temporary analysis files in the repository root.
- Remove task-specific temporary files created by the current task after validation when removal is safe.
## Completion Report
When implementation is complete, report in Korean:
- what changed and why;
- the main affected files or components;
- validation commands actually executed and their results;
- checks not executed and the reason;
- remaining risks, assumptions, or follow-up work;
- database, deployment, dependency, encoding, and compatibility impact when applicable.
- Do not report completion while required validation is failing or known encoding corruption is unresolved.
- Keep the report concise and do not include a full session transcript.
## Internal Service Project Rules
- Internal-only usage does not reduce security, privacy, compatibility, or validation requirements.
- Preserve authentication, authorization, role boundaries, auditability, and personal-data protections.
- Identify affected user workflows before changing visible behavior.
- Treat Frontend and Backend repositories as separate deployable systems with an explicit API contract between them.
- Coordinate request, response, authentication, error, and versioning changes across both sides when the approved scope includes both repositories.
- Do not silently require a simultaneous Frontend and Backend deployment.
- Prefer backward-compatible transitions when deployments may occur independently.
- Include migration and rollback considerations for database, API, and configuration changes.
- Do not modernize legacy code as an unrelated side effect.
- Apply Classic ASP text-safety rules only when the project profile declares the `classic-asp` stack.
- Do not access or modify production data merely because the service is internal.
## Java and Spring Boot
- Treat the configured JDK, Spring Boot, Gradle, plugin, dependency, and package versions as authoritative.
- Use the repository's Gradle Wrapper. Do not substitute a machine-global Gradle installation.
- Preserve the established package layout and controller, application/service, domain, and persistence boundaries.
- Keep controllers thin, place business rules in the established service/domain layer, and prefer constructor injection.
- Apply request and configuration validation at boundaries and preserve the project's exception and response conventions.
- Make transaction boundaries explicit and narrow; do not add broad transactions to hide lifecycle or query problems.
- Preserve authentication, authorization, auditing, logging, and secret-management conventions.
- Do not edit generated QueryDSL or other generated sources. Change their source definitions and regenerate only when the repository workflow requires it.
- Use the approved KriCommon contract when the project profile includes it; do not duplicate shared backend behavior locally.
- Run the narrowest relevant Gradle compile, test, static-analysis, and build tasks.
## MySQL
- Use MySQL syntax and behavior only for data sources whose declared engine is `mysql`.
- Preserve the configured version, schema, character set, collation, time zone, naming, migration, and access conventions.
- Parameterize values. Do not build SQL by concatenating untrusted input.
- Consider indexes and execution plans for changed filters, joins, ordering, grouping, and pagination without guessing from small data sets.
- Preserve transaction, locking, null, date/time, decimal, and generated-key semantics.
- Do not introduce MSSQL or other dialect syntax into MySQL paths.
- A source-code change approval authorizes SQL analysis and migration drafting only. Execute DDL, DML, procedures, or migrations only with separate explicit approval and the declared data-source access level.
## KriCommon Consumer Contract
### Dependency and API Boundary
- Consume KriCommon as `com.kri.common:starter:<approved-version>` using the consumer repository's established Gradle dependency and version-management conventions.
- Import service-facing APIs from `com.kri.common.feature.*`. Treat `com.kri.common.runtime.*` as internal and do not bind application code to it unless no public API exists and the exception is explicitly documented.
- Include the matching Spring starter for any optional Web, Security, JPA, Redis, AOP, or Swagger integration the service enables; KriCommon intentionally declares many integrations as compile-only.
- Prefer the shared feature or an approved extension point over duplicating authentication, HTTP, error, enum, Excel, Redis, JSON, date/time, JPA, or utility behavior locally.
### Configuration and Overrides
- KriCommon configuration uses only `kri.*`. Established groups include `kri.auth.jwt`, `kri.auth.apikey`, `kri.security`, `kri.swagger`, `kri.web`, `kri.web.mdc`, `kri.http.bypass`, `kri.redis`, `kri.scheduling`, and `kri.error.expose`.
- Read the version-matched README and configuration-property classes before setting a property; do not infer a property name or default from an older service.
- Supply production JWT secrets, API keys, credentials, allowlists, and service URLs through the consumer's approved secret and configuration system. Never rely on library example or default secrets in production.
- Preserve consumer override points. KriCommon's default security chain is created only when the consumer does not provide one, and overridable beans should remain explicit.
- Adding a path to API-key allowed paths scopes where API-key authentication is attempted; it does not make the path public or grant authorization.
### Major Optional Capabilities
- JWT and API-key authentication integrate with Spring Security. API-key JPA preload requires Spring Data JPA; a consumer-provided claims provider can be used without JPA.
- Common Web may register error handling, tracing/MDC, request-user resolution, enum conversion, and CORS behavior. Check for overlap before defining consumer-local equivalents.
- Swagger support requires springdoc and defines the shared Bearer/API-key schemes and grouped paths.
- Redis and scheduling are opt-in through their `kri.*.enabled` settings and matching runtime dependencies.
- HTTP bypass targets are server-configured. Never accept an arbitrary full target URL from a frontend; retain allowed-path checks and do not forward hop-by-hop headers.
- The external API client returns response bodies as documented, while bypass responses preserve raw bytes and target `Content-Type`.
- `ExternalApiClient` multipart requests generate their own boundary and ignore a caller-supplied `Content-Type`. Multipart headers and `formField` values use UTF-8, while file bodies preserve the supplied bytes.
- Multipart filenames, including non-ASCII names, are sent unchanged in the quoted `filename` parameter as raw UTF-8. No ASCII fallback or `filename*` parameter is added; CR/LF rejection and quote/backslash escaping still apply.
- `JsonUtil` requires the consumer's explicit `ObjectMapper`; shared date/time helpers define KST as `Asia/Seoul` and also expose UTC semantics.
### Upgrade and Verification
- Review KriCommon release notes, README changes, configuration defaults, auto-configuration conditions, and affected public APIs before upgrading.
- Commit dependency-lock or resolved-version changes according to the consumer repository's policy.
- Run the consumer's focused compile and tests after an upgrade, including application-context startup tests when auto-configuration or security behavior is affected.
- Do not modify KriCommon or another consumer repository unless it is explicitly included in the approved task scope.
## KriContact Server Repository Rules
### Product Purpose and Current Phase
- This is the launched Java 17 and Spring Boot backend for KRI contact operations. Preserve first-release API and data behavior while delivering approved follow-up features.
- The domain includes projects, contact lists and headers/items, uploads and downloads, assignments, call sessions, call/event/update histories, progress, scenarios and images, result codes, users, companies, roles, message templates, reservations, SMS, and email workflows.
- Prefer domain correctness, operational safety, and backward-compatible evolution over broad refactoring.
### Domain Invariants and API Compatibility
- Preserve company, project, role, and user authorization boundaries for every query and mutation.
- Treat list schemas, uploaded columns, item values, filtering, reset/update operations, assignments, and all history records as one coordinated consistency boundary.
- Make call start, close, and hang-up state transitions explicit and idempotent where retries or duplicate requests are possible.
- Preserve external-list endpoint request, response, authentication, error, and pagination behavior as a consumer contract.
- Validate file names, paths, sizes, content handling, storage lifecycle, and cleanup scheduling at boundaries. Do not expose server paths or retain files beyond the approved lifecycle.
- Do not send real SMS or email, initiate real calls, or invoke production business integrations during development or validation.
### Shared Library and Integrations
- Consume `com.kri.common:starter:beta.1.6.3`. Do not change this approved version or duplicate an available KriCommon feature without separate explicit approval.
- Import public KriCommon APIs from `com.kri.common.feature.*` and keep application code out of KriCommon runtime internals.
- Treat IntranetExtension, NotiHub, ShortUrl, CTI, file storage, and other external dependencies as explicit boundaries with validated inputs, timeouts, error mapping, retry/idempotency decisions, and safe rollback behavior.
- Never log API keys, credentials, full message bodies, telephone numbers in bulk, uploaded list contents, or unrestricted personal information.
### Persistence and Verification
- Preserve MySQL transaction boundaries, QueryDSL/JPA ownership, locking, ordering, null, date/time, and generated-key behavior.
- Do not edit generated QueryDSL sources directly. Change source entities and regenerate only through the repository workflow when required.
- `AGENTS.md` and `CLAUDE.md` are generated. Edit `.agent/profile.yml` or `.agent/project.md`, then recompose instead of editing generated files.
- Use the Gradle Wrapper and run the narrowest safe compile and focused tests. Do not start the service, connect to a real database, or call external integrations without separate approval.

1
CLAUDE.md Normal file
View File

@@ -0,0 +1 @@
@AGENTS.md