code-execution

Code Execution

Overview

Execute coding subtasks with self-review and quality validation. Runs in isolated coder-agent context with pre-loaded standards.

Announce at start: "I'm using the code-execution skill to implement [subtask title]."

The Process

Step 1: Read Subtask JSON

Load the subtask file specified in $ARGUMENTS:

Read: .tmp/tasks/{feature}/subtask_{seq}.json

Extract:

• title — What to implement
• acceptance_criteria — What defines success
• deliverables — Files/endpoints to create
• context_files — Standards to apply
• reference_files — Existing code to study

Step 2: Load Context Files

Read each file in context_files:

Read: .opencode/context/core/standards/code-quality.md
Read: .opencode/context/core/standards/security-patterns.md

Understand:

• Project coding standards
• Naming conventions
• Security patterns
• Code quality requirements

Step 3: Load Reference Files

Read each file in reference_files:

Read: src/middleware/auth.middleware.ts

Study:

• Existing patterns
• Code structure
• Conventions in use

Step 4: Update Status to In Progress

Edit subtask JSON:

"status": "in_progress",
"agent_id": "coder-agent",
"started_at": "2026-02-16T00:00:00Z"

Step 5: Implement Deliverables

For each deliverable:

• Create or modify the specified file
• Follow acceptance criteria EXACTLY
• Apply standards from context_files
• Use patterns from reference_files
• Write clean, modular, functional code

Step 6: Run Self-Review (MANDATORY)

Type & Import Validation:

• ✅ Function signatures match usage
• ✅ All imports/exports exist
• ✅ No missing type annotations
• ✅ No circular dependencies

Anti-Pattern Scan:

grep "console.log" deliverables  # NO debug statements
grep "TODO\|FIXME" deliverables  # NO unfinished work
grep "api[_-]key\|secret" deliverables  # NO hardcoded secrets

Acceptance Criteria Check:

• Re-read acceptance_criteria array
• Confirm EACH criterion met
• Fix unmet criteria before proceeding

Step 7: Mark Complete

Update subtask status:

bash .opencode/skills/task-management/router.sh complete {feature} {seq} "{summary}"

Verify:

bash .opencode/skills/task-management/router.sh status {feature}

Step 8: Return Report

✅ Subtask {feature}-{seq} COMPLETED

Self-Review: ✅ Types | ✅ Imports | ✅ No debug code | ✅ Criteria met

Deliverables:
- src/auth/jwt.service.ts
- src/auth/jwt.service.test.ts

Summary: JWT service with RS256 signing and 15min token expiry

Error Handling

Missing subtask JSON:

• Main agent must create subtask before invoking this skill

Context files not found:

• Main agent must run /context-discovery first

Acceptance criteria unmet:

• DO NOT mark complete—fix issues first

Red Flags

If you think any of these, STOP and re-read this skill:

• "I know the pattern, I don't need to read the context files"
• "I'll do the self-review quickly at the end"
• "The acceptance criteria are obvious, I don't need to check them"
• "I'll mark it done and fix the edge cases later"

Common Rationalizations

Excuse	Reality
"I've seen this pattern before, context files will just confirm what I know"	Projects diverge from common patterns. One wrong assumption = rework. Read the files.
"Self-review is just checking my own work"	Self-review catches type errors, missing imports, and debug code before the main agent sees it.
"The criteria are implied by the task title"	Implied criteria are unverifiable. If it's not written, it's not a gate.
"I'll handle edge cases in a follow-up"	Edge cases left for follow-up become bugs in production. Handle them now.

Remember

• Context FIRST, code SECOND—always read context_files before implementing
• One subtask at a time—complete fully before moving on
• Self-review is MANDATORY—quality gate before marking done
• Functional, declarative, modular—clean code patterns only
• Return results to main agent—report completion for orchestration

Related

• context-discovery
• task-breakdown
• test-generation
• code-review

---

Task: Execute coding subtask: $ARGUMENTS