plugin
Workflow infrastructure for [Claude Code][claude-code] agents. Without
structure, agent sessions drift: incomplete implementations, skipped reviews,
scope creep, lost context between sessions. plugin imposes a pipeline —
design, planning, TDD, orchestration, review, handoff — and a memory system
that persists decisions across sessions.
18 skills, 14 specialized sub-agents, 23 instruction fragments, 4 hooks, and
scripts for runbook preparation, session recovery, and batch operations.
Currently installed as a git submodule; converting to a [Claude Code
plugin][plugin-docs] (edify-plugin).
Requires [just] for task running. Scripts in bin/ use Python 3.
Installation
git submodule add <repo-url> plugin
cd plugin && just sync-to-parent
just sync-to-parent creates symlinks in the parent project's .claude/
directory so Claude Code discovers skills, agents, and hooks. It also
configures hooks in .claude/settings.json. Changes to plugin are
reflected immediately without re-syncing.
Set up your CLAUDE.md to reference fragments — use
templates/CLAUDE.template.md as a starting point, or add @ references to
an existing file:
@plugin/fragments/communication.md
@plugin/fragments/execution-routing.md
These are loaded into every conversation as ambient context.
Structured Workflow
The core pipeline:
/design → /runbook → [runbook-corrector] → /orchestrate → [corrector] → /handoff
/design triages complexity. Simple tasks execute directly. Moderate tasks skip
to planning. Complex tasks get Opus architectural design with outline iteration
and design review.
/runbook produces step-by-step execution plans with per-phase typing — TDD
cycles or general steps, mixed within a single runbook. runbook-corrector checks
TDD discipline, step quality, and LLM failure modes before execution.
/orchestrate dispatches each step to a sub-agent in isolated context.
Executing agents receive only their step file plus the design — no access to
other steps, so scope is enforced structurally rather than by prose
instructions. Review follows every production artifact.
/handoff captures completed work, pending tasks, and blockers for the next
session. /commit handles structured commit messages with gitmoji.
The pipeline is not mandatory. Any skill works independently. But the full
sequence prevents the failure mode where a task succeeds in one session and
leaves wreckage for the next.
Session Modes
Agents start in one of five modes, triggered by shortcuts:
| Shortcut | Mode | Behavior |
|---|
s | Status | List pending tasks with metadata, wait |
x | Execute | Resume in-progress or start first pending |
xc | Execute+Commit | Execute, then handoff and commit |
r | Resume | Strict: resume only, error if nothing in progress |
wt | Worktree | Set up parallel execution via git worktrees |
Memory Management
Claude Code sessions are stateless. The memory system persists project
knowledge in an agents/ directory in the consuming project:
- session.md — pending tasks, in-progress work, blockers. The handoff
document. Agents read it at startup, update it at
/handoff
- learnings.md — institutional knowledge from mistakes and discoveries.
Append-only, consolidated via
/codify when approaching the 80-line limit
- memory-index.md — keyword catalog pointing to detailed documentation.
Agents scan the index to decide what to load on demand
- decisions/ — permanent architectural and implementation decisions, one
heading per topic. Where learnings graduate to
- jobs.md — plan lifecycle tracking (
requirements → designed → planned
→ complete)
Validators (bin/validate-*.py) enforce cross-reference integrity, format
conventions, and key uniqueness across these files.
Skills
Skills are slash-command procedures that inject instructions into the current
conversation when invoked. Agents (next section) are separate processes with
isolated context, spawned via the Task tool. Each skill lives in
skills/<name>/SKILL.md.
Workflow pipeline:
| Skill | Purpose |
|---|
/design | Entry point — triages complexity, routes to appropriate workflow |
/runbook | Creates execution plans with per-phase TDD or general typing |
/review-plan | Reviews runbook quality, TDD discipline, LLM failure modes |
/orchestrate | Dispatches runbook steps to sub-agents |
/review | Reviews production artifacts for quality |
Session management:
| Skill | Purpose |
|---|
/handoff | Updates session.md with completed work and pending tasks |
/commit | Structured commit messages with gitmoji |
/next | Shows pending work when no context is loaded |
/shelve | Archives session context to task list, resets for new work |
/codify | Consolidates learnings into permanent documentation |
Specialized:
