devconcept

DevConcept

DevConcept is a lightweight senior-engineering workflow for Claude Code and Codex. It helps an agent choose the minimum process needed to understand, implement, verify, and hand off code changes correctly.

Philosophy: not ceremonial by default, but rigorous when the risk justifies it. Lean Mode stays out of the way; Standard adds alignment and verification; Full adds plan review and dispatch — and only when the work warrants it.

The plugin install slug is devconcept. The user-facing workflow name is DevConcept, and the default operating skill is devconcept-core.

Install

/plugin install devconcept@built-by-harry

Modes

DevConcept routes each task internally as Lean, Standard, or Full.

Mode	Use for	Behavior
Lean	Trivial, exact, mechanical, or read-only work	Inspect if needed, act directly, no alignment block, no plan file, no agents, narrow verification.
Standard	Normal code changes where correctness matters	Gather minimal context, align requirements, write a concise mini-plan, use TDD when practical, implement, verify, summarize.
Full	Risky, ambiguous, cross-surface, long-running, or multi-agent work	Plan, review the plan, make an explicit dispatch decision, dispatch only when allowed and separable, run an agent or inline review pass for broad/high-risk work, verify integrated behavior, hand off with residual risk.

Debugging, design alternatives, TDD, dispatch, and subagent review are playbooks inside these three modes rather than separate top-level modes.

Skills

• devconcept-core - Default DevConcept operating skill: mode router, correctness policy, plan review, deviation protocol, verification, and handoff contracts.
• using-devconcept - First-turn bootstrap that invokes devconcept-core before repo inspection so its rules persist through the session.
• aligning-requirements - Standard/Full Mode alignment block before non-trivial edits.
• planning-ledger - Lightweight durable planning for long-running, multi-phase, research-heavy, or context-compaction-prone work.
• dispatching-agents - Runtime-aware dispatch rules for Claude Code and Codex.
• tdd - Observable red-green-refactor workflow when practical behavior tests exist.
• systematic-debugging - Root-cause debugging loop for intermittent, cross-module, unclear, or shared-cause failures.
• design-alternatives - Three-option design exploration for interfaces, modules, and refactors with real tradeoffs.
• subagent-review - Two-stage review of implementer subagent output: spec compliance, then code quality.
• finishing-work - Final completion gate with changed, verified, risk, skills/agents used, dispatch/review, and skipped-workflow disclosure.
• compound-engineering - Captures durable repo-wide learnings into project instruction files.

DevConcept Agents

Claude Code can use the native agents in agents/:

• devconcept-explorer - read-only codebase mapping
• devconcept-plan-reviewer - requirement or plan review before risky implementation
• devconcept-worker - bounded, disjoint implementation slices
• devconcept-code-reviewer - independent diff review
• devconcept-debugger - failure diagnosis before fixes

Codex users can sync matching templates from templates/codex-agents/ into ~/.codex/agents:

scripts/sync-codex-agents.sh

Dispatch Differences

Claude Code should prefer named DevConcept agents when dispatch triggers hit. If the agents are unavailable, use the closest built-in agent or inline the same contract.

Codex does not spawn subagents automatically. When dispatch is appropriate, ask the user to choose between spawning agents and continuing inline unless they already explicitly requested subagents or parallel agent work. If spawning is chosen, name the DevConcept agent rather than the generic word "subagent":

• Have devconcept-explorer map <area> read-only and return Findings, Relevant files, Risks / unknowns, Recommended next step.
• Have devconcept-plan-reviewer review the plan above against the Plan Review Checklist before implementation.
• Have devconcept-code-reviewer review the diff for correctness, scope drift, and verification gaps.
• Have devconcept-debugger investigate the failing test before any fix; reproduce first, then identify root cause.

For parallel read-only exploration, spawn a batch by name:

Spawn one `devconcept-explorer` subagent per area, wait for all of them, and return a consolidated summary. Use read-only exploration only.

Areas:
1. <area A>
2. <area B>

Each subagent must return: Findings, Relevant files, Risks / unknowns, Recommended next step.

If the named DevConcept agents are not synced into Codex, run scripts/sync-codex-agents.sh (writes to ~/.codex/agents) and try again, or fall back to inline read-only exploration with the same return contract.