Subagent driven development

Subagent-Driven Development

<!-- Adapted from obra/superpowers — Tessera graph discipline injected into subagent prompts -->

Execute plan by dispatching fresh subagent per task, with two-stage review after each: spec compliance review first, then code quality review.

Core principle: Fresh subagent per task + two-stage review = high quality, fast iteration.

Continuous execution: Do not pause between tasks unless BLOCKED, genuinely ambiguous, or all tasks complete.

When to Use

Use when you have a written implementation plan with independent tasks and want to stay in the current session. For parallel-session execution, use executing-plans instead.

Tessera Coordinator Discipline

At the start of each dispatch round (before dispatching any subagent) — if tessera MCP is configured:

1. graph_continue (mandatory first call for the coordinator too)
2. graph_retrieve with the current task's key terms

This keeps the coordinator's routing accurate as files change during execution.

Tessera Subagent Discipline

Every implementer subagent prompt you dispatch MUST include these instructions:

Tessera graph policy (if tessera MCP is configured in this session):
1. Call graph_continue as your FIRST tool call
2. Call graph_retrieve with this task's key terms
3. Read all recommended_files via graph_read before exploring
4. After each file edit, call graph_register_edit with file::symbol notation
   - If graph_continue returned an active_checklist, include the checklist_item_id
     for the task you just completed
5. Call graph_lock_decision when you make an architectural choice

This ensures Tessera's compliance tracker stays updated as each subagent works.

The Process

Per Task Loop

1. Dispatch implementer subagent

   • Craft a focused, self-contained prompt with exactly the context needed
   • Include the Tessera graph policy block above
   • Include the specific task steps, file paths, and expected test outcomes
   • Include: "Do NOT inherit session history. Work only from this prompt."

2. Implementer questions?

   • Answer questions, provide context, re-dispatch

3. Implementer implements, tests, commits, self-reviews

4. Dispatch spec reviewer subagent

   • Provide: the plan's task spec + the actual diff (git diff HEAD~1)
   • Ask: Does code match spec? List any gaps.

5. Spec gaps?

   • Dispatch implementer to fix gaps

6. Dispatch code quality reviewer subagent

   • Provide: the diff + project coding standards from CLAUDE.md
   • Ask: Code quality, security, type safety, test adequacy. Severity: CRITICAL/IMPORTANT/MINOR.

7. Quality issues?

   • Fix CRITICAL and IMPORTANT before proceeding

8. Mark task complete — if tessera MCP is configured, graph_register_edit marks the checklist item done

9. Next task

After All Tasks Complete

If tessera MCP is configured: run tessera-verify to confirm plan compliance before finishing.

All checklist items done? → tessera-verify passes → proceed
Missing items? → complete them first

Then invoke finishing-a-development-branch skill.

Subagent Prompt Template

## Task: [Task name from plan]

**Goal:** [One sentence]

**Files to modify:**
- Create: `path/to/file`
- Modify: `path/to/existing`

**Steps:**
[Paste exact steps from plan]

**Verification:**
[Exact test commands and expected output]

**Tessera graph policy (if tessera MCP is configured):**
- Call graph_continue as your FIRST tool call
- Read recommended_files via graph_read
- After each edit: graph_register_edit(files=["file::symbol"], summary="...")
- Lock architectural decisions: graph_lock_decision(...)

**Do NOT inherit session history. Work only from this prompt.**
**Return: Summary of what you implemented and what tests pass.**

Red Flags

Never:

• Skip the two-stage review (spec then quality)
• Proceed with unfixed CRITICAL issues
• Give subagents your session's full history
• Pause to ask "should I continue?" between tasks unless genuinely blocked