code plugin

The code plugin is the hub of ClosedLoop.ai. It implements the full autonomous SDLC — PRD → plan → critics → implementation → tests → visual QA → validation — through an orchestrator prompt and a roughly 1,351-line shell loop runner.

Commands

Command	Purpose
`/code:code`	Begin a ClosedLoop coding session. Flags: `[working-directory]`, `--prompt <name>`, `--prd <file>`, `--plan <file>`.
`/code:amend-plan`	Interactively amend a running plan. Flags: `--workdir <path>`, `--message "<text>"`, `--state-file <path>`.
`/code:cancel-code`	Stop a running loop by removing the state file.
`/code:plan-with-codex`	Run a Claude-vs-Codex debate loop to refine a plan. Flags: `--max-rounds N` (default 15), `--plan-file PATH`, `--codex-model MODEL` (default `gpt-5.4`), then a `<prompt>`.

Agents (23)

Planning

plan-draft-writer (opus, loop max 10) — draft from PRD
plan-writer (sonnet, loop max 5) — merge, finalize, gap modes
plan-validator (sonnet, loop) — semantic validation
plan-evaluator (sonnet) — simple-mode threshold
plan-importer (sonnet, loop max 3) — import external plan markdown
answered-questions-subagent (haiku) — merge Q&A into tasks
plan-agent (opus) — used by the plan-with-codex debate loop

Implementation

pre-explorer (haiku)
verification-subagent (sonnet)
implementation-subagent (sonnet, loop max 4)
build-validator (haiku)
code-reviewer (sonnet, loop max 5)

Cross-repo

cross-repo-coordinator (haiku)
generic-discovery (haiku)
repo-coordinator (haiku)
cross-repo-prd-writer (sonnet)
api-spec-writer (sonnet)

Support

visual-qa-subagent (sonnet) — Playwright
dev-environment (haiku)
learning-capture (sonnet)
amend-extractor (sonnet)
code-review-worker, code-review-guidelines

Skills (12)

plan-validate, plan-structure, plan-editing-conventions, extract-plan-md
critic-cache, build-status-cache, cross-repo-cache
closedloop-env, find-plugin-file
iterative-retrieval
prd-creator
codex-review

The orchestrator prompt

plugins/code/prompts/prompt.md (~271 lines) drives the phased workflow. It is read-only and never touches project files. It uses Bash (for directory listing, echo, mkdir), Task (for subagents), TodoWrite, AskUserQuestion, and SendMessage.

The first action is always a TodoWrite with the 18-entry phase list.

The loop runner

plugins/code/scripts/run-loop.sh relaunches claude -p with a fresh 200K-token context each iteration until the orchestrator writes <promise>COMPLETE</promise> or CLOSEDLOOP_MAX_ITERATIONS (default 10) is hit.

After each iteration, run-loop.sh runs an 11-step post-iteration pipeline (changed-files, relevance, goal eval, citation verification, outcomes merge, build-result merge, pattern processing, success-rate computation, closedloop-learnings export). See Self-learning.

Environment variables

All prefixed CLOSEDLOOP_* unless noted. Written to {WORKDIR}/.closedloop-ai/config.env by setup-closedloop.sh and injected into agent contexts via subagent-start-hook.sh:

Variable	Purpose
`CLOSEDLOOP_WORKDIR`	Absolute path to the work directory
`CLAUDE_PLUGIN_ROOT`	Plugin install dir
`CLOSEDLOOP_PRD_FILE`	PRD file path (auto-detected)
`CLOSEDLOOP_PLAN_FILE`	Pre-supplied plan file path
`CLOSEDLOOP_MAX_ITERATIONS`	Loop iteration cap (default 10)
`CLOSEDLOOP_AGENT_ID`	Current agent's unique ID
`CLOSEDLOOP_PROMPT_FILE`	Orchestrator prompt file
`CLOSEDLOOP_RUN_ID`	`YYYYMMDD-HHMMSS-<hex>`
`CLOSEDLOOP_ITERATION`	Current iteration
`CLOSEDLOOP_ACTIVE_GOAL`	Goal (default `reduce-failures`)
`CLOSEDLOOP_REPO_MAP`	Multi-repo mapping
`CLOSEDLOOP_ADD_DIRS`	Additional directories
`CLOSEDLOOP_ADD_DIR_NAMES`	Names for add dirs

Plus:

CLAUDE_BIN — claude executable path (set by closedloop-electron when launching)
POST_LOOP_REVIEW_CYCLES — default 2 post-loop review cycles

Artifacts produced

See Artifacts for the full list. The most important session artifacts: plan.json, plan.md, reviews/*.review.json, state.json, log.md, requirements-extract.json, code-map.json, visual-requirements.md, and the full .learnings/ tree.