Code Forge — Port

Port a project to a new target language by batch-generating implementation plans from shared feature specs.

When to Use

• Have a documentation project with feature specs (docs/features/*.md) and want to implement in a new language
• Have an existing implementation in one language and want to create another language version
• Need to batch-plan multiple features for a new language SDK

Command Format

/code-forge:port @<docs-project> --ref <reference-impl> --lang <target-language>

Example:

/code-forge:port @../apcore --ref apcore-python --lang java

Parameters

Missing required parameters → use AskUserQuestion to collect interactively.

Workflow

Step 0 → 1 → 2 → 3 → 4 → 5 → 6 (sub-agent loop) → 7 → 8

Context Management

Step 2 and Step 6 are offloaded to sub-agents. Step 2 uses a single sub-agent to analyze the reference implementation. Step 6 dispatches one sub-agent per feature (serial) to generate plans. The main context retains only concise summaries.

Detailed Steps

Step 0: Configuration Detection and Loading

@../shared/configuration.md

Port-specific additions to Step 0:

• 0.2 additional defaults: port.source_docs = "", port.reference_impl = "", port.target_lang = ""
• 0.4 note: During Steps 1-4 (before target project exists), config loading runs against the current working directory. After Step 5 creates the target project, subsequent steps use the target project’s config.

Step 1: Parse Arguments and Discover Feature Specs

1.1 Resolve Docs Project Path

Parse the @<path> argument:

1. Resolve to absolute path
2. Validate directory exists
3. Look for feature specs at <docs-project>/docs/features/*.md
4. If none found, try <docs-project>/features/*.md, then <docs-project>/*.md
5. If still none: display error with directory contents and stop

Store docs_project_path and feature_specs[] (list of absolute file paths).

1.2 Resolve Reference Implementation

If --ref <name> is provided:

1. Try sibling directory: <docs-project>/../<name>/
2. If not found, try as absolute path
3. Validate: directory exists and contains planning/ with at least one */state.json
4. If invalid: warn "Reference '{name}' not found or has no planning data. Continuing without reference." and set ref_project_path = null

If --ref not provided: set ref_project_path = null.

1.3 Resolve Target Language

Validate --lang value. Recognized identifiers:

• java, typescript (alias: ts), go (alias: golang), rust, python, csharp (alias: cs), kotlin, swift
• Unrecognized value: warn and continue (do not reject — new languages are valid)

1.4 Derive Target Project Path

• Default: <docs-project>/../<docs-name>-<lang>/ (e.g., ../apcore-java/)
• Display derived path and use AskUserQuestion:
  • “Use {derived-path} (Recommended)” — proceed
  • “Custom path” — user provides path

1.5 Display Discovery Summary

Docs project:      ../apcore (7 feature specs)
Reference impl:    ../apcore-python (planning/ found)
Target language:   java
Target project:    ../apcore-java

Proceed directly — no confirmation needed.

Step 2: Analyze Reference Implementation (Sub-agent)

Skip if ref_project_path is null.

Spawn a Task tool call with:

• subagent_type: "general-purpose"
• description: "Analyze reference implementation"

Sub-agent prompt:

Analyze the reference implementation at {ref_project_path} for a cross-language porting operation.

Read the following files:
1. {ref_project_path}/planning/overview.md — project-level overview with dependency graph
2. All {ref_project_path}/planning/*/plan.md — implementation plans per feature
3. All {ref_project_path}/planning/*/state.json — completion status per feature

Return ONLY a structured summary in this exact format:

REFERENCE_IMPL: {project-name}
REFERENCE_LANG: {detected language from build files or plan content}
OVERALL_STATUS: {completed-count}/{total-count} features complete

DEPENDENCY_ORDER:
- {feature-1} (no deps)
- {feature-2} (depends on: feature-1)
...

PER_FEATURE:
- {feature-name}:
  STATUS: completed | in_progress | pending
  TASK_COUNT: {N}
  KEY_DECISIONS: {2-3 bullet points of important architecture decisions from plan.md}
  LESSONS: {any notable patterns, gotchas, or non-obvious choices}

Keep the summary concise — target ~2-3KB total.

Main context retains: Only the returned summary. Store as reference_summary.

Step 3: Display Feature List and User Selection

Build a feature display table combining discovered specs with reference data (if available).

With reference:

Feature specs discovered: 7 (from ../apcore/docs/features/)
Reference: apcore-python (7/7 complete)

  #  Feature              Ref Status    Ref Tasks
  1  acl-system           ✅ complete   5 tasks
  2  core-executor        ✅ complete   5 tasks
  3  decorator-bindings   ✅ complete   6 tasks
  4  middleware-system     ✅ complete   4 tasks
  5  observability        ✅ complete   7 tasks
  6  registry-system      ✅ complete   8 tasks
  7  schema-system        ✅ complete   7 tasks

Without reference:

Feature specs discovered: 7 (from ../apcore/docs/features/)

  #  Feature
  1  acl-system
  2  core-executor
  ...

Use AskUserQuestion with multiSelect: true:

• Question: “Which features do you want to port to {lang}?”
• First option: “All features (Recommended)”
• Remaining options: one per feature

Store selected_features[].

Step 4: Confirm Target Tech Stack

Use a single AskUserQuestion with up to 3 questions based on target language. Skip questions that have obvious single answers.

For Java:

• Question 1 — Build tool: “Maven (Recommended)” / “Gradle”
• Question 2 — Java version: “Java 17+ (Recommended)” / “Java 21+” / “Java 11+”
• Question 3 — Test framework: “JUnit 5 (Recommended)” / “TestNG”

For TypeScript:

• Question 1 — Runtime: “Node.js (Recommended)” / “Deno” / “Bun”
• Question 2 — Package manager: “pnpm (Recommended)” / “npm” / “yarn”
• Question 3 — Test framework: “Vitest (Recommended)” / “Jest”

For Go:

• Question 1 — Go version: “1.21+ (Recommended)” / “1.22+”
• Question 2 — Test extras: “Standard testing (Recommended)” / “testify”

For Rust:

• Question 1 — Async runtime: “tokio (Recommended)” / “async-std” / “None (sync only)”
• Question 2 — Serialization: “serde (Recommended)” / “Other”

For other languages: Ask a single open-ended question: “What tech stack and key libraries should be used for {lang}?”

Store tech_stack decisions.

Step 5: Initialize Target Project Skeleton

5.1 Create Target Directory

1. If target directory already exists:
   • Use AskUserQuestion: “Target directory {path} already exists.”
     • “Update config only (Recommended)” — overwrite .code-forge.json, keep everything else
     • “Use as-is” — skip all initialization, jump to Step 6
     • “Cancel” — stop
2. If not exists: mkdir -p <target-path>

Do NOT copy feature specs: Never copy docs/features/ or any feature spec files from the docs project into the target project. Feature specs are accessed from the source docs project via relative paths configured in directories.input. The target project must NOT contain its own copy of feature specs.

If the user explicitly requests local copies of feature specs, use AskUserQuestion to confirm before copying:

• “Copy feature specs locally (creates docs/features/ in target)” — copy and update directories.input to "docs/features/"
• “Keep remote references (Recommended)” — do not copy, keep relative path in config

5.2 Generate .code-forge.json

Write to <target-path>/.code-forge.json:

{
  "_tool": {
    "name": "code-forge",
    "description": "Transform documentation into actionable development plans with task breakdown and status tracking",
    "url": "https://github.com/tercel/code-forge",
    "skills_collection": "https://github.com/tercel/claude-code-skills"
  },
  "directories": {
    "base": "./",
    "input": "<relative-path-to-docs>/docs/features",
    "output": "planning/"
  },
  "reference_docs": {
    "sources": ["<relative-path-to-ref>/planning/*/plan.md"]
  },
  "port": {
    "source_docs": "<relative-path-to-docs>",
    "reference_impl": "<relative-path-to-ref>",
    "target_lang": "<lang>"
  },
  "execution": {
    "default_mode": "ask",
    "auto_tdd": true,
    "task_granularity": "medium"
  }
}

Use relative paths from the target project to the docs and reference projects.

If no reference: omit reference_docs.sources and port.reference_impl.

5.3 Generate Project Skeleton (Sub-agent)

Spawn a Task sub-agent (subagent_type: "general-purpose"):

Sub-agent prompt:

Initialize a {lang} project skeleton at {target_path}.

Project name: {project-name}
Language: {lang}
Tech stack: {tech_stack decisions from Step 4}

Create the following files:
1. Build file — {pom.xml | package.json | go.mod | Cargo.toml | etc.} with project metadata and minimal dependencies
2. .gitignore — language-appropriate patterns
3. README.md — minimal: project name, one-line description ("apcore SDK for {lang}"), link to docs project

Do NOT create src/ or test/ directories — those are created by feature tasks during implementation.
Do NOT copy docs/features/ or any feature spec files into this project.

You MUST create all three files listed above (build file, .gitignore, README.md). Return the list of files created.

Verify skeleton files: After sub-agent completes, check that the following files exist in {target_path}:

• Build file (pom.xml / package.json / go.mod / Cargo.toml / etc.)
• .gitignore
• README.md

If any are missing, create them directly in the main context. Do NOT skip this verification.

Also verify that docs/features/ does NOT exist in the target project. If it was created, delete it immediately.

5.4 Initialize Git Repository

cd <target-path> && git init && git add . && git commit -m "chore: initialize {project-name} project skeleton"

Only if not already a git repo.

Step 6: Batch Generate Plans (Sub-agent per Feature, Serial)

Process each selected feature in order. Use dependency order from Step 2 reference summary if available; otherwise alphabetical.

6.0 Display Batch Header

Generating plans for {count} features...

6.1 Per-Feature Loop

For each feature in selected_features[]:

6.1.1 Check Existing Plan

If <target>/planning/<feature>/state.json exists:

• Display: [{i}/{total}] {feature} — skipped (plan already exists)
• Continue to next feature

6.1.2 Create Feature Directory

mkdir -p <target>/planning/<feature>/tasks/

6.1.3 Dispatch Plan Sub-agent

Spawn a Task sub-agent (subagent_type: "general-purpose"):

Sub-agent prompt:

Generate an implementation plan for porting the "{feature}" feature to {lang}.

## Input
- Feature spec: {docs_project_path}/docs/features/{feature}.md (read this file)
- Target project: {target_path}
- Target language: {lang}
- Tech stack: {tech_stack}
{if type_mapping_exists:}
- Type mapping reference: {docs_project_path}/docs/spec/type-mapping.md (read this file for cross-language type translations)
{end if}

## Reference Context (from existing {ref_lang} implementation)
{reference_summary — the per-feature section for this feature from Step 2}

## Output Files
Write ALL of the following files:

### 1. {target}/planning/{feature}/plan.md
Required sections:
- **Goal** — one sentence
- **Architecture Design** — component structure, data flow, technology choices with rationale
- **Task Breakdown** — mermaid dependency graph + task list with estimated time and dependencies
- **Risks and Considerations** — technical challenges
- **Acceptance Criteria** — checklist
- **References** — related docs

Ensure the plan uses {lang}-idiomatic patterns (not a line-by-line translation of the reference).

**Task ID naming:** Task IDs must be descriptive names **without numeric prefixes**. Use `setup`, `models`, `api` — NOT `01-setup`, `02-models`. Execution order is defined in overview.md, not by filename ordering.

### 2. {target}/planning/{feature}/tasks/{name}.md (one per task)
Each task file must include:
- **Goal** — what this task accomplishes
- **Files Involved** — files to create/modify
- **Steps** — numbered, TDD-first (write tests → run → implement → verify), with {lang}-specific commands and code examples
- **Acceptance Criteria** — checklist
- **Dependencies** — depends on / required by
- **Estimated Time**

**Naming (critical):** Use descriptive filenames — `setup.md`, `models.md`, `api.md`. **NO numeric prefixes** (`01-setup.md`, `02-models.md` are WRONG). Execution order is controlled by `overview.md` Task Execution Order table and `state.json`, never by filename ordering.

### 3. {target}/planning/{feature}/overview.md
Sections:
- **Overview** — feature summary
- **Scope** — included/excluded
- **Technology Stack** — language, framework, key deps, test tools
- **Task Execution Order** — table: #, Task File, Description, Status
- **Progress** — counts
- **Reference Documents** — link to source spec

## Return Format
Return ONLY a concise summary:
FEATURE: {feature}
STATUS: planned
TASK_COUNT: <N>
TASKS:
- <id>: <title> (~<estimate>)
EXECUTION_ORDER: <id1>, <id2>, ...

6.1.3.1 Verify Plan Output

After sub-agent completes, verify the following files were created:

• {target}/planning/{feature}/plan.md — must exist and be non-empty
• {target}/planning/{feature}/tasks/ — must contain at least one .md file
• {target}/planning/{feature}/overview.md — must exist and be non-empty

If any file is missing, this is a sub-agent failure. Follow the error handling in 6.2 (display failure, ask user to skip/retry/stop). Do NOT proceed to state.json creation with missing plan files.

6.1.4 Initialize state.json

After sub-agent completes, parse its summary and create <target>/planning/<feature>/state.json:

{
  "feature": "{feature}",
  "created": "{ISO timestamp}",
  "updated": "{ISO timestamp}",
  "status": "pending",
  "execution_order": ["{id1}", "{id2}", "..."],
  "progress": {
    "total_tasks": {N},
    "completed": 0,
    "in_progress": 0,
    "pending": {N}
  },
  "tasks": [
    {
      "id": "{task-id}",
      "file": "tasks/{task-id}.md",
      "title": "{task-title}",
      "status": "pending",
      "started_at": null,
      "completed_at": null,
      "assignee": null,
      "commits": []
    }
  ],
  "metadata": {
    "source_doc": "{relative-path-to-feature-spec}",
    "created_by": "code-forge:port",
    "version": "1.0",
    "ported_from": {
      "reference_impl": "{ref-project-name or null}",
      "reference_lang": "{ref-lang or null}",
      "target_lang": "{lang}"
    }
  }
}

6.1.5 Display Progress

[{i}/{total}] {feature} planned ({task_count} tasks, ~{estimate})

6.2 Error Handling

If a sub-agent fails for a feature:

• Display: [{i}/{total}] {feature} — FAILED: {error summary}
• Use AskUserQuestion: “Feature planning failed.”
  • “Skip and continue” — mark as skipped, continue to next feature
  • “Retry” — re-dispatch sub-agent
  • “Stop” — exit the batch loop

Step 7: Generate Project Overview

@../shared/overview-generation.md

Scan <target>/planning/*/state.json and generate <target>/planning/overview.md.

Display: Project overview generated: planning/overview.md

Step 8: Display Results and Next Steps

Port completed!

Target project: {target_path}
Features planned: {planned}/{selected} ({skipped} skipped)
Total tasks: {total_task_count}

Feature Summary:
  #  Feature              Tasks  Estimated
  1  schema-system        7      ~14h
  2  core-executor        5      ~10h
  ...

Next steps:
  cd {target_path}
  /code-forge:status                         View project dashboard
  /code-forge:impl {first-feature}           Start implementing first feature

Coordination with Other Skills

• After port: cd to target project, use /code-forge:impl {feature} to execute tasks
• With /code-forge:status: View overall progress across all ported features
• With /code-forge:review: Review completed features
• With /code-forge:fixbug: Debug issues in ported code
• Port does NOT invoke impl — the user controls when to start implementation

Notes

1. Serial execution: Features are planned one at a time for cross-feature consistency and user control
2. Reference is optional: Port works without --ref, just without architecture reference context
3. Idiomatic porting: Sub-agents are instructed to use target-language idioms, not translate line-by-line
4. Type mapping: If docs/spec/type-mapping.md exists in the docs project, it’s provided to sub-agents for cross-language type translation guidance
5. Resumable: Re-running port skips features that already have state.json — safe to resume after interruption
6. Standard output: Generated plans are indistinguishable from regular /code-forge:plan output — all downstream skills work unchanged