claude

agent-loop

Goal-driven loop harness: /agent-loop init|new|run|approve|reject|status|doctor|watch|canonize. Builds the 7-file Claude Code harness in any repo, then runs goal-spec-driven plan→act→verify→optimize loops with human plan approval, stall detection and scheduler integration — each pass proposes spec/verify improvements for the next.

2 on GitHubby belchmanv0.4.1MITupdated 4 months ago
/

Plugin agent-loop from marketplace belchman-claude-skills in belchman/claude-skills. Commands read from the repo's manifest — 21 days ago.

This plugin runs code on your machine — its hooks fire automatically on events like tool use. The exact commands are listed below; read them before installing.

What it installs

Skills (1)

  • agent-loop

    Goal-driven loop harness. Subcommands — init (build the 7-file harness in this repo), new (interview → goal spec), run (plan→act→verify iterations; plans await human approval), approve/reject (act on a pending plan), status (loop health), doctor (harness health), watch (live dashboard), canonize (pr

Subagents (5)

  • loop-critic

    agent-loop PLAN pressure-test — adversarial cross-model review of a proposed plan before the human sees it. Read-only. Invoked by /agent-loop run with a model override from goal.json critic_model, not directly.

  • loop-optimizer

    agent-loop OPTIMIZE step — fresh-context reviewer of one recorded iteration's evidence; proposes spec decisions, verify commands, planner guidance, and canon. Read-only, report-only. Invoked by /agent-loop run, not directly.

  • loop-planner

    agent-loop PLAN step — turns the goal spec + loop state into a bounded JSON task list with ALL assumptions declared. Read-only; never writes code. Invoked by /agent-loop run, not directly.

  • loop-verifier

    agent-loop VERIFY step — fresh-context judge of one iteration's diff against the goal rubric. Read-only. Invoked by /agent-loop run, not directly.

  • loop-worker

    agent-loop ACT step — executes exactly one planned task with minimal context. Invoked by /agent-loop run, not directly.

Hooks (3)

  • PreToolUse

    "${CLAUDE_PLUGIN_ROOT}"/hooks/guard-destructive.sh

  • PostToolUse

    "${CLAUDE_PLUGIN_ROOT}"/hooks/format-on-write.sh

  • Stop

    "${CLAUDE_PLUGIN_ROOT}"/hooks/loop-checkpoint.sh

View source on GitHub

More from belchman/claude-skills

All 4 plugins →

Composable skills for AI-agentic engineering: PRD → issues → spec → rubric → autonomous work pipeline, orchestrated end-to-end by /feature (3 human checkpoints, resumable) or run piece-by-piece. Plus TDD, architectural deepening, bug diagnosis, prototyping, domain interrogation (CONTEXT.md/ADR-aware), navigation, evolutionary search, CRAP-score reduction, and ARCHITECTURE.md mapping.

17 skills

2

Audit a Claude Code setup (global ~/.claude, project .claude, plugin scope) for the 9 token-waste patterns that silently inflate cost and burn through usage limits. Slash-only.

1 skill

2

PreToolUse hook that lints CLAUDE.md on every Edit/Write and blocks writes that introduce structural errors. Example of a plugin-scoped hook.

hooks

2

Maintain agent-loop?

This page updates itself from your repo. Add the badge to your README, or get it Featured to pin it above the organic results with a followed link to your site.

claude & codex plugins4 plugins