ana-and-new-team 0.2.0

Slash-command-driven MCP server with Ana and team (Cham, Ark, Fey, Bud, Tess) for Jakarta EE (koseiin-*) delivery.

ana-and-new-team

Slash-command-driven MCP server that turns Excel design specs into a coordinated development pipeline for the koseiin-* Jakarta EE / WildFly stack.

Six file-bus agents — Ana (Analyzer), Cham (QA), Ark (Architect), Fey (Frontend), Bud (Backend), Tess (Tester) — collaborate via artifacts under .ant-context/. The server itself is host-agnostic; it speaks MCP stdio and works with Claude Code, opencode, and Cline.

Windows-first. The deploy gate calls domain.bat. A cross-OS code path is retained for development on macOS / Linux but production usage is Windows.

• Source: https://gitlab.enlinka.co/work-project/ana-and-new-team
• Issues: https://gitlab.enlinka.co/work-project/ana-and-new-team/-/issues
• PyPI: https://pypi.org/project/ana-and-new-team/

---

What it does

Three slash commands, all delivered as MCP prompts:

Slash command	Purpose
/ant-new-screen	New screen from design Excels. Halts if プログラム設計書(簡易) is not attached.
/ant-new-ita	ITA (issue-tracker) fix. Each Ana finding = code-issue / root-cause / solution / why-resolves.
/ant-change-request	CR analysis + impact + delivery plan.

Every pipeline:

1. Reads every sheet of every attached .xlsx — never samples.
2. Drops strikethrough text from analysis (cell and run level), but logs it.
3. Resolves dependency jars through pom.xml → ~/.m2/repository/....
4. Halts at the deploy gate if WildFly is unreachable or target/*.war is missing, surfacing the exact mvn / domain.bat commands the user must approve.
5. Routes work to six personas via a DispatchPlan keyed by file paths under .ant-context/.
6. Versions every revision with an -r{N} suffix so re-runs never clobber prior work.

---

Quick install

Once ana-and-new-team is on PyPI, the fastest path is:

uvx ana-and-new-team install --all

That single command:

1. Detects which AI hosts (Claude Code, opencode, Cline) are installed on this machine.
2. Patches their config files in place (atomic write, .bak left next to the original) — never touching other servers' blocks. The edit is surgical: your existing keys, // comments, and formatting are preserved byte-for-byte, and only the ana-and-new-team block is added or removed. For opencode it updates your existing config — opencode.jsonc, opencode.json, config.jsonc, or config.json — rather than creating a duplicate file.
3. Registers the server's launch command as uvx ana-and-new-team, so each host always pulls the latest published wheel.

To inspect what would change without touching disk:

uvx ana-and-new-team install --all --dry-run

To overwrite an existing — but different — block:

uvx ana-and-new-team install --all --force

To uninstall just our block from every host:

uvx ana-and-new-team uninstall --all

You can scope to a single host with --host claude-code, --host opencode, or --host cline (repeatable; mutually exclusive with --all).

Windows install

ana-and-new-team is the same Python package on every OS, but the bootstrap on Windows is different from macOS/Linux because Windows ships neither Python nor uv by default. Pick one of the three paths below.

Path A — uv + uvx (recommended)

This mirrors the macOS quick-install flow exactly. uvx fetches the latest published wheel into a temporary, isolated environment each run, so you never have to manage a venv yourself.

1. Open PowerShell (regular, not elevated) and install uv:

   powershell -ExecutionPolicy Bypass -c "irm https://astral.sh/uv/install.ps1 | iex"
   

   Close and reopen PowerShell so the new PATH entry is picked up. Verify:

   uv --version
   uvx --version
   

2. Install / register:

   uvx ana-and-new-team install --all
   

3. Sanity-check:

   uvx ana-and-new-team --version
   uvx ana-and-new-team doctor
   

The install subcommand will detect whichever of Claude Code, opencode, Cline are present and patch each one's config file. The launch command written into those configs will be uvx ana-and-new-team, so every restart of your AI host pulls the latest published wheel.

Path B — pipx (persistent, no Python knowledge required)

Use this if you'd rather have one stable copy of the binary on disk instead of uvx resolving on every launch.

1. Install Python 3.10 or newer from https://www.python.org/downloads/windows/. On the installer, tick Add python.exe to PATH.

2. Install pipx and put it on PATH:

   py -m pip install --user pipx
   py -m pipx ensurepath
   

   Close and reopen PowerShell.

3. Install the package:

   pipx install ana-and-new-team
   

4. Register with your AI hosts:

   ana-and-new-team install --all
   

   With this path the command written into host configs is the absolute path of ana-and-new-team.exe under %USERPROFILE%\.local\bin\ (or wherever pipx put it). To stay on the latest version run pipx upgrade ana-and-new-team.

Path C — Editable clone (contributors)

For local development against the source tree:

git clone https://gitlab.enlinka.co/work-project/ana-and-new-team.git
cd ana-and-new-team
py -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -e .[dev]
ana-and-new-team --version
pytest -q

To wire up an AI host against this checkout (no PyPI hop):

ana-and-new-team install --all --command "uvx --from C:\path\to\ana-and-new-team ana-and-new-team"

Where Windows host configs actually live

The install subcommand finds these automatically; the table is here so you can hand-edit or troubleshoot.

Host	Config path on Windows
Claude Code	%USERPROFILE%\.claude.json
opencode	%APPDATA%\opencode\opencode.json
Cline	%APPDATA%\Code\User\settings.json (VS Code's settings)

Windows-specific gotchas

• Backslashes in env vars must be escaped inside JSON. Use C:\\src\\koseiin-common (or forward slashes — Windows accepts both in pathlib).
• WildFly's domain.bat is hard-coded into the deploy gate for Windows. Set ANT_JBOSS_HOME to the install root (e.g. C:\wildfly-31.0.0.Final), not to the bin\ directory.
• PowerShell vs. cmd quoting differs. The examples above use PowerShell. In cmd.exe, --command arguments containing spaces should be wrapped in double quotes the same way.
• Long-path support. If your koseiin checkouts are deep, enable long paths in Windows (gpedit → Computer Configuration → Administrative Templates → System → Filesystem → Enable Win32 long paths). The Maven and jar-inspection code paths handle long paths correctly once the OS allows them.
• Antivirus / Defender on first run can briefly delay uvx's extraction step. The first uvx ana-and-new-team install --all may take 30+ seconds; subsequent runs are cached.

Verifying everything works

ana-and-new-team doctor
ana-and-new-team inspect

doctor prints environment + persona + path-tool + host-detection report. inspect connects to the server in-process and lists the 8 tools, 3 prompts, and 7 persona resources.

Install (manual)

Requires Python 3.10+.

With pipx

pipx install ana-and-new-team

With uv tool

uv tool install ana-and-new-team

From a clone (development)

git clone https://gitlab.enlinka.co/work-project/ana-and-new-team.git
cd ana-and-new-team
pip install -e .[dev]

After install, the ana-and-new-team console script is on PATH. Subcommands:

Command	Purpose
ana-and-new-team (no args)	Run the MCP server over stdio (default).
ana-and-new-team serve	Same as above, explicit.
ana-and-new-team install [--all]	Patch host configs. --dry-run, --force, --command "<cmd>".
ana-and-new-team uninstall [--all]	Remove our block from host configs. --dry-run.
ana-and-new-team doctor	Read-only diagnostic. --json for machine-readable output.
ana-and-new-team inspect	Enumerate exposed tools/prompts/resources. --json supported.

---

Register with your AI host (manual fallback)

If you can't (or don't want to) use install, register manually.

Claude Code

claude mcp add ana-and-new-team -- ana-and-new-team

Slash commands appear as /ant-new-screen, /ant-new-ita, /ant-change-request.

opencode

Edit %USERPROFILE%\.config\opencode\opencode.json:

{
  "mcp": {
    "servers": {
      "ana-and-new-team": {
        "command": "ana-and-new-team",
        "args": [],
        "env": {
          "ANT_JBOSS_HOME": "C:\\wildfly-31.0.0.Final",
          "ANT_KOSEIIN_REPOS": "koseiin-common=C:\\src\\koseiin-common,koseiin-dao=C:\\src\\koseiin-dao,koseiin-backend=C:\\src\\koseiin-backend,koseiin-frontend=C:\\src\\koseiin-frontend"
        }
      }
    }
  }
}

Cline (VS Code)

In VS Code settings (Cline > MCP Servers):

{
  "cline.mcpServers": {
    "ana-and-new-team": {
      "transport": "stdio",
      "command": "ana-and-new-team",
      "args": [],
      "env": {
        "ANT_JBOSS_HOME": "C:\\wildfly-31.0.0.Final",
        "ANT_KOSEIIN_REPOS": "koseiin-common=C:\\src\\koseiin-common,koseiin-dao=C:\\src\\koseiin-dao,koseiin-backend=C:\\src\\koseiin-backend,koseiin-frontend=C:\\src\\koseiin-frontend"
      }
    }
  }
}

---

Configuration

Variable	Required	Default	Purpose
ANT_CONTEXT_ROOT	no	./.ant-context	File-bus root. Personas, analysis, QA, architecture, cache live here.
ANT_JBOSS_HOME	yes for /ant-new-screen and /ant-change-request	—	Path to WildFly/JBoss. Deploy gate calls ${ANT_JBOSS_HOME}/bin/domain.bat.
ANT_JBOSS_URL	no	http://localhost:9990	Health probe. Gate is ready when this responds 2xx-4xx within 2 s.
ANT_KOSEIIN_REPOS	yes for /ant-new-screen and /ant-change-request	—	Deploy targets. See Repo map format below.
ANT_QA_MAX_ROUNDS	no	6	Hard cap on Ana ⇄ Cham debate rounds.
ANT_DECOMPILER	no	unset	cfr or procyon. If set, koseiin-* jar deep-scans run a decompile.
ANT_JAR_DEEPSCAN_ALL	no	0	1 to deep-scan every jar (default: only those whose group/artifactId starts with koseiin).
MAVEN_REPO_LOCAL	no	~/.m2/repository	Override local Maven repo. <localRepository> in ~/.m2/settings.xml is also honored automatically.

ANT_KOSEIIN_ROOTS is not used. Earlier drafts referenced it; it has been removed entirely. Source jars come from pom.xml → ~/.m2/. Build/deploy targets come from ANT_KOSEIIN_REPOS.

Repo map format (ANT_KOSEIIN_REPOS)

All four keys are required: koseiin-common, koseiin-dao, koseiin-backend, koseiin-frontend. Each value must be an existing directory containing a pom.xml.

Comma list (recommended for env vars on Windows):

koseiin-common=C:\src\koseiin-common,koseiin-dao=C:\src\koseiin-dao,koseiin-backend=C:\src\koseiin-backend,koseiin-frontend=C:\src\koseiin-frontend

JSON (auto-detected by leading {):

{"koseiin-common":"C:\\src\\koseiin-common","koseiin-dao":"C:\\src\\koseiin-dao","koseiin-backend":"C:\\src\\koseiin-backend","koseiin-frontend":"C:\\src\\koseiin-frontend"}

Validation runs at deploy-gate entry, not at server startup — so the server keeps running for read-only inspection commands even when the deploy prerequisites are missing.

---

Deploy gate

/ant-new-screen and /ant-change-request consult the gate before any agent writes architecture. The gate is evaluative: it returns a structured directive listing the exact commands the user must approve. It never auto-runs.

Order is fixed:

1. Start WildFly/JBoss in domain mode — ${ANT_JBOSS_HOME}\bin\domain.bat (skipped if ANT_JBOSS_URL already responds).
2. mvn clean install in koseiin-common.
3. mvn clean install in koseiin-dao.
4. mvn clean wildfly:deploy in koseiin-backend.
5. mvn clean wildfly:deploy in koseiin-frontend.
6. Optional, always offered: mvn dependency:sources -q in the active workspace — pulls koseiin-* source jars into ~/.m2 so Bud and Fey can inspect real method bodies, not just constant pools.

Steps 2–5 are emitted only when target/*.war is missing in koseiin-backend or koseiin-frontend.

The host model surfaces the directive, asks for confirmation, then calls ant.deploy.execute with the user-approved list. Failure stops the chain at the first non-zero return code.

---

File-bus layout

All artifacts live under ${ANT_CONTEXT_ROOT} (default ./.ant-context/):

.ant-context/
├── personas/                                 # editable system prompts (on-disk wins)
│   ├── ana.md
│   ├── cham.md
│   ├── ark.md
│   ├── fey.md
│   ├── bud.md
│   ├── tess.md
│   └── shared_conventions.md
├── analysis/
│   └── NSS001-2026-06-03-r1.md               # Ana — overwritable within (date, r{N})
├── QA/
│   └── NSS001-2026-06-03-r1/
│       ├── round-1.md                        # Cham — append-only across rounds
│       ├── round-2.md
│       └── ...
├── architecture/
│   └── NSS001-2026-06-03-r1/
│       ├── roadmap/
│       │   └── roadmap.md                    # Ark
│       ├── frontend/
│       │   ├── NSS001Xhtml.md                # Fey, one md per artifact
│       │   ├── NSS001BackingBean.md
│       │   └── ...
│       ├── backend/
│       │   ├── NSS001Bl.md                   # Bud, one md per artifact
│       │   ├── NSS001BlImpl.md
│       │   └── ...
│       └── qa/
│           └── test-report.md                # Tess
└── .cache/
    ├── project-map.json                       # rebuilt on git HEAD/branch/dirty/pom change
    ├── jar-paths.json                         # GAV → resolved jar
    └── jars/{sha}/                            # per-jar deep-scan + optional decompile

Revision rule (-r{N})

Every (screenId, date) gets a fresh -r{N} suffix. The pipeline scans analysis/ for the highest existing -r{N} for the same screen+date and allocates N+1. This propagates through all directories so a re-run never clobbers prior QA rounds, architecture, or test reports.

QA round rule

QA/.../round-{N}.md is append-only. Writing an explicit round_n that already exists raises FileExistsError. The host loops Ana ⇄ Cham until either two consecutive rounds report empty unresolved sections (convergence) or ANT_QA_MAX_ROUNDS is hit (escalation).

---

Editing personas

The seven persona prompts ship with the package. On first run, they are materialized to ${ANT_CONTEXT_ROOT}/personas/. From then on, the on-disk copy wins — edit them per project to tighten conventions or add domain rules.

To re-sync from the package, delete (or rename) the on-disk files and rerun any slash command.

The MCP resource URIs ant://persona/<name> always reflect the on-disk copy when present.

---

Tools exposed

The server exposes these tools (callable directly by the host model in addition to the slash-command prompts):

Tool	Purpose
ant.workspace.preflight	Ingest Excels, allocate revision, run pipeline-specific gates.
ant.deploy.evaluate	Return the deploy directive (no side effects).
ant.deploy.execute	Run an already-approved directive. Stops on first failure.
ant.codebase.scan	Refresh and return the project map.
ant.dispatch.plan	Build the agent dispatch plan (groups, parallelism, paths).
ant.fs.write_artifact	Write an artifact under .ant-context/ (idempotent).
ant.fs.read_artifact	Read an artifact by path.
ant.persona.load	Return persona text (on-disk wins, package fallback).

---

Non-negotiables (encoded in shared_conventions.md)

These rules are part of every persona's binding preamble:

• Total ingestion. Read every sheet of every attached .xlsx. Never sample.
• Strikethrough = invisible. At cell and run level. Logged under "Excluded (struck-through)" but excluded from analysis text.
• Required-field flagging. Every required UI field is flagged in the analysis and roadmap.
• ASSUMPTION: tag. Any inferred koseiin-* API or undocumented behaviour must be explicitly tagged. No fabricated method signatures.
• Read koseiin-* via resolved jars/sources. Multi-repo means the dependency surface is reached through ~/.m2, not sibling source trees.
• Idempotent revisioned writes. Within (screen, date, r{N}): analysis / roadmap / per-artifact md may be rewritten. QA rounds are append-only.
• Mermaid for diagrams. No ASCII art.
• Never self-deploy. The server emits a directive; the user runs it.
• Output format. H1 of the form # {ScreenId} — {Doc} (r{N}) and source-cell citations like (画面項目設計書 B14).

---

Troubleshooting

Run the diagnostic:

ana-and-new-team doctor

It reports:

• package version, Python interpreter, platform,
• every ANT_* env var (set / unset / value),
• parsed Settings (or ConfigError reason if ANT_KOSEIIN_REPOS is malformed),
• persona inventory (bundled + materialized to .ant-context/personas/),
• PATH availability of mvn, java, git, uvx, claude, code,
• config-file paths for each supported AI host (and whether they exist).

For machine-readable output (e.g. to attach to a bug report):

ana-and-new-team doctor --json

To enumerate the live MCP surface (tools, prompts, resources) without attaching a host:

ana-and-new-team inspect --json

---

Development

git clone https://gitlab.enlinka.co/work-project/ana-and-new-team.git
cd ana-and-new-team
pip install -e .[dev]
pytest -q

The test suite includes:

• synthetic XLSX builders that exercise both run-level rich-text strikethrough and cell-level font strikethrough,
• fs_state revisioning + non-clobbering QA round semantics,
• /ant-new-screen HALT on missing プログラム設計書(簡易),
• deploy-gate directive shape and ANT_KOSEIIN_REPOS parsing (comma + JSON, malformed input),
• screenId extraction precedence (filename > workbook scan > none),
• project-map cache invalidation on git HEAD / branch / dirty / pom changes,
• DispatchPlan snapshot for /ant-new-screen,
• in-process MCP server smoke test (initialize → list tools / prompts / resources → call tool),
• CLI argparse routing + doctor / inspect / install subcommands,
• per-OS host config path resolution (claude-code, opencode, cline),
• idempotent JSON merge engine + --dry-run + --force semantics + .bak retention.

All fixtures are synthetic — no real customer documents are committed.

See CONTRIBUTING.md for the release flow and CHANGELOG.md for version history.

---

License

MIT — see LICENSE.