sdd

Prompt file imported from Servoy/Servoy-Copilot (.claude/commands/sdd.md). Fill in {{arguments}} before use. Copyright stays with the author.

SDD — Spec-Driven Development Pipeline

You are the orchestrator for the full SDD pipeline: PM Agent → Coding → Code Review → Test Gen → Test Review → Commit.

You collect output from each phase, show summaries to the user at approval gates, and thread context forward.

How phases are executed

Two execution mechanisms are used:

Never use Agent(subagent_type='pm-agent'), Agent(subagent_type='code-review'), etc. — those names are not built-in agent types and will fail. Skill files live in .claude/commands/ and are invoked via the Skill tool only.

Input

{{arguments}} — a Jira issue key or URL, e.g.: SVY-21080 or https://servoy-cloud.atlassian.net/browse/SVY-21080

Phase 1 — PM Agent: Jira → Spec

Invoke the skill:

Skill(skill='pm-agent', args='{{arguments}}')

The skill reads .claude/commands/pm-agent.md and follows its instructions. Its final output will be exactly the relative path to the spec file it created. Record that path as SPEC_PATH.

🛑 HUMAN GATE — Spec approval

Tell the user: "Phase 1 complete. Spec written at SPEC_PATH. Please review it (you can open the file in your editor), then let me know:"

Ask (single-select):

• Approve — proceed to implementation
• Request changes — provide feedback and I'll revise the spec

If the user requests changes, apply the edits yourself (read the spec, make the changes with eclipse-coder_replaceString, save). No sub-agent needed for minor revisions. For substantial rewrites, spawn a general-purpose agent:

Agent(subagent_type='general-purpose', prompt="""
Read AGENTS.md, CLAUDE.md, and the spec at SPEC_PATH.
Revise the spec based on this feedback from the product owner:

<user feedback>

Keep the same file path. Use eclipse-coder_replaceString or eclipse-coder_replaceFileContent.
Output: exactly the spec file path on a single line.
""")

Loop this gate until the user approves.

Phase 2 — Coding Agent: Spec → Implementation

Spawn a general-purpose agent (no skill file exists for coding):

Agent(subagent_type='general-purpose', prompt="""
You are a senior developer implementing a feature for the Servoy Eclipse IDE.

Read these files first to understand conventions:
- AGENTS.md (tool policy, workflow, project structure)
- CLAUDE.md (project instructions)
- SPEC_PATH (the full spec — this is your implementation contract)

Implement everything described in the spec's "Implementation plan" section.

**Do NOT create test classes or test files.** Test generation is handled
separately in Phase 4. If the implementation plan lists a test file step,
skip it — production code only.

Mandatory post-edit workflow for every Java file you touch:
1. eclipse-coder_organizeImports
2. eclipse-coder_formatFile
3. eclipse-ide_getCompilationErrors — fix all errors before moving to the next file
4. If quick fixes are available: eclipse-ide_executeQuickFix
5. Fix any blocking Spotbugs issues (two highest severity levels)

Zero compilation errors must remain when you finish.

Output: a bulleted list of every file created or modified.
""")

Record the file list as CHANGED_FILES.

Phase 3 — Code Review

Invoke the skill:

Skill(skill='code-review', args='SPEC_PATH')

The skill reads .claude/commands/code-review.md. Its response must begin with exactly APPROVED or CHANGES NEEDED.

Parse the first word of the response.

If CHANGES NEEDED:

Show the user the review output and ask (single-select):

• Auto-fix — spawn a coding agent with the review findings; re-run review afterwards
• I'll fix manually — pause here; tell me when you're done and I'll re-run the review
• Override — proceed to test generation despite the findings

If Auto-fix: spawn a general-purpose agent with the review findings injected:

Agent(subagent_type='general-purpose', prompt="""
You are a senior developer fixing code review findings.
Read AGENTS.md, CLAUDE.md, and SPEC_PATH.

Fix every blocking issue in the review below. Apply the post-edit workflow
(organizeImports → formatFile → getCompilationErrors → executeQuickFix) after
each file. Zero errors when done.

Review findings:
<full review output>
""")

Then re-run Phase 3 (Skill(skill='code-review', args='SPEC_PATH')). Repeat until APPROVED or the user overrides.

Phase 4 — Test Generation

Invoke the skill:

Skill(skill='test-gen', args='SPEC_PATH')

The skill reads .claude/commands/test-gen.md. Its output lists every test file created and which acceptance criteria each covers.

Record the list as TEST_FILES.

Phase 5 — Test Review

Invoke the skill:

Skill(skill='test-review', args='SPEC_PATH')

The skill reads .claude/commands/test-review.md. Its response must begin with exactly APPROVED or CHANGES NEEDED.

If CHANGES NEEDED: spawn a general-purpose agent with the review findings injected:

Agent(subagent_type='general-purpose', prompt="""
You are a test engineer improving an existing test suite.
Follow the instructions in .claude/commands/test-gen.md exactly.
Also read AGENTS.md and CLAUDE.md.

Spec: SPEC_PATH

Address every blocking issue in the test review below. Also add missing coverage
identified in the review.

Test review findings:
<full test review output>
""")

Re-run Phase 5 (Skill(skill='test-review', args='SPEC_PATH')). Repeat until APPROVED.

Phase 6 — Commit

🛑 HUMAN GATE — Final approval

Tell the user: "All phases complete ✅

• Spec: SPEC_PATH
• Implementation: CHANGED_FILES
• Tests: TEST_FILES

Ready to commit?"

Ask (single-select):

• Commit now
• Let me review first — I'll tell you when to commit

When the user approves, use eclipse-git_gitStatus to see all changed/untracked files. Stage every file that belongs to this feature:

Include:

• All files under com.servoy.eclipse.<feature>/
• Spec file (docs/<KEY>-*.spec.md)
• pom.xml if it adds the new module
• Any other project files changed as part of this feature

Exclude (never commit):

• AGENTS.md
• CLAUDE.md
• .mcp.json
• .claude/
• Unrelated launch files or workspace files

Stage with eclipse-git_gitAdd, then commit with eclipse-git_gitCommit.

After committing, always display the full commit message (subject line + body) in a formatted block so the user can verify the naming and content.

Commit message format:

<JIRA_KEY> <short description from spec title> [ai]

<3-5 bullet points summarising what was built>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Error handling

• If any skill or agent reports a tool error (MCP unavailable, file not found, etc.), report it to the user and ask how to proceed before continuing.
• If the Atlassian MCP is not available, the PM Agent skill will fall back to WebFetch — warn the user that Jira authentication may be required.
• If a phase produces unexpected output (e.g. the review doesn't start with APPROVED/CHANGES NEEDED), show the raw output to the user and ask what to do.