Skip to content

Repository files navigation

AIAgentMinder

License: MIT Version

AIAgentMinder (AAM) is an opinionated governance layer for Claude Code — a minder for your AI coding agent. Built for the solo developer, it gives Claude Code the engineering discipline a team would normally supply, so AI-assisted work stays tested, reviewed, and in scope. It's a set of skills, subagents, and hooks that let you enforce your standards without repeating your instructions, stop a session before context degradation sets in, and run autonomous sprints that span hours and multiple sessions without losing the thread.

Basically, I wanted Claude Code to have the skills I use daily as a developer — and figured a few product-manager tasks belonged in there too. This plugin lets you effectively manage AI development of large, complex projects that can't be finished in one to five sessions. When a project needs 100+ PRs over the span of months, you need consistency from your tooling, self-updating memory, the ability to change direction quickly, fast quashing of the defects that inevitably arise, and deterministic enforcement of quality.

Installs easily into existing or new repos. Works from a back-of-the-napkin sketch of an idea, or a fully spec'd plan.


The 30-second case

  • Run a sprint, walk away. Start a sprint for Phase 2. AAM plans, takes your approval at the spec gate, then executes item-by-item in isolated git worktrees until the PRs are ready for review or even merged.
  • A different model reviews your code before you do. /aiagentminder:self-review runs the change through five specialist lenses (security, performance, API design, cost, UX friction) plus a cross-model judge pass.
  • Clean cross-session resume. /aiagentminder:handoff writes the resume state into Claude Code's native Auto Memory; in the next session, "resume work" picks up exactly where you stopped.
  • Plan and decision discipline by default. scope-check keeps work aligned with the roadmap, DECISIONS.md logging keeps the record honest, BACKLOG.md captures everything you can't do right now.

Quick Start

Step 1 - Install In Claude Code:

/plugin marketplace add lwalden/AIAgentMinder
/plugin install aiagentminder@lwalden-aiagentminder

Step 2 - Configure In your target project:

/aiagentminder:setup

/aiagentminder:setup runs interactively — it fingerprints your codebase, asks for project identity, writes a starter CLAUDE.md, DECISIONS.md, BACKLOG.md, and docs/strategy-roadmap.md, and seeds .claude/rules/ and .gitignore.


Suggested Next Steps

  1. /aiagentminder:brief to draft a product brief and strategy roadmap interactively.
  2. Start building. Start a new session with 'claude --agent sprint-master' then prompt 'start sprint 1'

Reference

Skills (slash commands)
Command Purpose
/aiagentminder:setup Bootstrap a project (one-shot install / re-sync)
/aiagentminder:brief Interview-driven product brief and strategy roadmap
/aiagentminder:revise Mid-stream roadmap revision with DECISIONS.md logging
/aiagentminder:scope-check Compare proposed work against the roadmap; promote to backlog if out of scope
/aiagentminder:milestone Phase health assessment — progress, timeline, scope drift
/aiagentminder:tdd Guided TDD — plan, tracer bullet, RED-GREEN loop, refactor
/aiagentminder:grill Plan interrogation — walk every decision branch before implementing
/aiagentminder:triage Structured bug triage — reproduce, diagnose, fix plan, GitHub issue
/aiagentminder:quality-gate Pre-PR quality checklist (build, tests, coverage, lint, security)
/aiagentminder:self-review Pre-PR review via 5 specialist sub-agents + cross-model judge pass
/aiagentminder:pr-pipeline Autonomous PR review-fix-test-merge pipeline
/aiagentminder:handoff Session checkpoint — writes "Next Session" into Auto Memory
/aiagentminder:retrospective Sprint metrics and adaptive sizing guidance
/aiagentminder:backlog Capture, list, promote BACKLOG.md items
/aiagentminder:sync-issues Optional GitHub Issues bridge
Sub-agents (Task tool dispatch)
Agent Phase / role
sprint-master Orchestrator — drives the sprint state machine
sprint-planner PLAN phase — decomposes roadmap items into a sprint
sprint-speccer SPEC phase — per-item implementation spec
item-executor EXECUTE phase — TDD-driven implementation in a worktree
quality-reviewer REVIEW phase — runs the quality gate
pr-pipeliner Drives the autonomous PR pipeline post-create
sprint-retro Sprint retrospective and adaptive sizing
api-reviewer, cost-reviewer, performance-reviewer, security-reviewer, ux-reviewer Five specialist review lenses
dev, qa, debug, hotfix Optional session profile agents (different permissions / risk tolerances)
Hooks and scripts

All scripts ship in bin/ and are on PATH while the plugin is enabled. Hooks register via hooks/hooks.json.

Hook Event Purpose
context-monitor.sh statusLine Writes .context-usage with token thresholds
context-warning-hook.sh Stop Advisory warning when over threshold (v5.1+)
sprint-phase-guard.sh PreToolUse (matcher: "Agent") Blocks sub-agent dispatches that don't match SPRINT.md phase
pre-pr-gate-hook.sh PreToolUse (Bash + mcp__github__create_pull_request) Blocks PR creation (gh pr create or the GitHub MCP tool) until the quality gate passed, the review judge didn't block, and the diff adds no high-confidence secret; fail-open, AAM_PR_GATE_BYPASS=1 to skip
sprint-phase-reminder.sh Stop One-line per-turn phase reminder during an active sprint
sprint-stop-guard.sh Stop Blocks premature turn endings during sprint execution
session-start-cycle-reset.sh SessionStart Wipes stale .context-usage
session-start-hook.sh SessionStart Detects active sprints; surfaces a one-line reminder
stop-failure-hook.sh StopFailure Logs API errors
hlpm-ping.sh SessionStart/End Optional HLPM executive layer integration
Bash script Purpose
sprint-update.sh Zero-token SPRINT.md status/phase updates
decisions-log.sh Zero-token DECISIONS.md append
backlog-capture.sh Zero-token BACKLOG.md add/list/promote/detail/count
sprint-metrics.sh Sprint metrics collection
version-bump.sh Multi-file version bump
sprint-runner.sh / .ps1 Optional wrapper that auto-restarts Claude on /exit (unattended runs)
aam-bootstrap.sh Internal — /aiagentminder:setup helper

Zero-token-cost scripts replace LLM file I/O with deterministic bash. Significant token savings during long sprints.

Files installed in your project
File Auto-loaded? Purpose
CLAUDE.md Yes Project identity and behavioral rules
.claude/rules/git-workflow.md Yes Branch naming, commit discipline, PR-only workflow
.claude/rules/tool-first.md Yes Use tools, don't ask the user
.claude/rules/shell-and-files.md Yes Write/Edit for files (not shell); platform-native shell — avoids Windows Git Bash file-write hangs
.claude/rules/context-warnings.md Yes How to respond to the context-warning hook
DECISIONS.md On-demand (@DECISIONS.md) Architectural decision log
docs/strategy-roadmap.md On-demand Product brief and phase plan
SPRINT.md On-demand Active sprint header (when in a sprint)
BACKLOG.md On-demand Unscheduled work inbox
.pr-pipeline.json n/a (config) PR pipeline config

Total always-loaded baseline: ~1,000 tokens per session.


Minimum Claude Code version

v5.0.x+ needs Claude Code 2.1.139+ for worktree-isolated agent execution (isolation: "worktree"), the /goal command, and ${CLAUDE_PLUGIN_ROOT} hook variable substitution.


Where this works

Environment Plugin works? Notes
Claude Code CLI (macOS / Linux / Windows) ✅ Full Primary target.
VS Code with the Claude Code extension ✅ Full Extension hosts a local Claude Code; same surface as CLI.
JetBrains IDE extension ✅ Full Same as VS Code.
claude.ai/code (browser web sessions) ❌ No Web sessions don't load plugins. No /aiagentminder:* commands, no sub-agents, no hooks, no bin/ scripts on PATH. Only the artifacts that /aiagentminder:setup previously copied into the repo (CLAUDE.md, .claude/rules/*.md, SPRINT.md, BACKLOG.md, DECISIONS.md, docs/strategy-roadmap.md) remain available, since they live in the repo itself.

A project bootstrapped from a local session is still readable from web — Claude will load CLAUDE.md and rules natively — but plugin-only workflows are CLI / IDE only. Mobile and Claude desktop app parity is not a goal.


What you actually use

1. Autonomous sprint execution

Trigger: Run clause with the sprint-master agent: claude --agent sprint-master Start a sprint for Phase 2. What you get: sprint-master orchestrates a state machine — PLAN → SPEC → APPROVE → [per item: EXECUTE → TEST → REVIEW → MERGE → VALIDATE] → COMPLETE. Each item runs in its own git worktree, isolated from other in-flight work. TDD is mandatory. Quality gate runs before every PR. You approve once at the spec gate; the rest is autonomous.

2. Cross-model self-review

Trigger: /aiagentminder:self-review (also runs automatically in the REVIEW phase of every sprint item) What you get: Five specialist sub-agents — security-reviewer, performance-reviewer, api-reviewer, cost-reviewer, ux-reviewer — each give a focused-lens read. A judge pass picks real findings from noise. Configurable to use a different model than the one that wrote the code, so the second opinion actually catches things the original author missed.

3. Autonomous PR pipeline

Trigger: /aiagentminder:pr-pipeline <PR#> (also invoked in-sprint after PR creation) What you get: pr-pipeliner reviews the PR with full repo context (not just the diff), applies the fixes itself, waits on external CI, and merges when green. Escalates with a reason on cycle limit, high-risk files, or unresolvable blockers. Configurable per repo via .pr-pipeline.json.

4. Clean session handoff

Trigger: context status hook warning + /aiagentminder:handoff What you get: Scripting checks the sessions context usage after each turn. When a configurable limit is hit Claude warns the user the wrap up the session with the handoff skill; which writes a "Next Session" block into Claude Code's native Auto Memory — decisions made this session, the next concrete step, any blockers. In the next session, "resume work" picks up exactly where you stopped.

5. Plan and decision discipline

Triggers: /aiagentminder:scope-check, /aiagentminder:revise, /aiagentminder:backlog What you get: Before AAM lets new work into a sprint, scope-check compares it against docs/strategy-roadmap.md. If out of scope, you either revise the roadmap (logged in DECISIONS.md) or capture to BACKLOG.md. The plan, the record, and the work stay in sync.

The full feature inventory (15 skills, 16 sub-agents, 10 hooks) is in the Reference section.


See it in action

examples/demo-transcript.md walks through three sessions on a sample REST API: planning with /aiagentminder:brief, handing off mid-build with /aiagentminder:handoff, and resuming. Sample state files (examples/CLAUDE.md, examples/DECISIONS.md, examples/strategy-roadmap.md) show what a mid-project repo looks like.


When NOT to use this

  • Single-session projects. If your work fits in one Claude session and you don't need sprint governance or decision logging, plain CLAUDE.md is enough.
  • Ad-hoc exploration. AAM expects a roadmap. If you're poking around to figure out what to build, finish that first, then /aiagentminder:setup.
  • Multi-agent concurrency on overlapping work. sprint-master coordinates items sequentially in isolated worktrees. It does not orchestrate concurrent agents working on the same files.
  • Browser-only workflow. See Where this works — plugin features aren't loaded in claude.ai/code web sessions.

AAM adds structure. Only use it if the structure pays for itself.


Requirements

  • Claude Code 2.1.139+ — VS Code / JetBrains extension or CLI.
  • Bash — Windows users need Git Bash or WSL; macOS/Linux built-in.
  • jq — for context monitoring (winget install jqlang.jq / brew install jq / apt install jq). Hooks degrade gracefully without it.
  • Git — required.
  • GitHub CLI (gh) — optional, for PR pipeline and issue sync.

Works on Windows, macOS, and Linux. Node.js is no longer required (the v4.x npm CLI was retired in v5.0).


Troubleshooting

Symptom Fix
Skills not showing (VS Code) /reload-plugins, then close and reopen the Claude Code panel
Hooks not firing /plugin → check Errors tab. Verify the plugin is enabled, not just installed
/aiagentminder:setup not found Plugin not installed or not enabled. Run /plugin install aiagentminder@lwalden-aiagentminder
Claude re-debates a past decision Add it to DECISIONS.md; add @DECISIONS.md to CLAUDE.md to auto-load
Claude builds something out of scope Run /aiagentminder:scope-check
Claude asks you to do things manually Verify .claude/rules/tool-first.md exists in your project
Quality degrades late in session Verify jq is installed; otherwise context monitoring falls back to heuristics
Upgrading /plugin marketplace update lwalden-aiagentminder/reload-plugins → optionally /aiagentminder:setup to refresh .claude/rules/

Documentation


Related

HLPM is the executive layer above AAM — a Claude Code plugin for managing a portfolio of repos from a single meta-tracker, with /dispatch to hand autonomous work to consumer repos, cross-repo drift detection, and a review cadence system. AAM's hlpm-ping.sh hook integrates with it out of the box.


License

MIT — see LICENSE.

Works with Claude Code. Independent open-source project, not affiliated with Anthropic.

About

Project governance for AI-assisted development. Structured planning, sprint workflows, decision tracking, and scope enforcement — built as plain markdown files and slash commands on top of Claude Code.

Topics

Resources

Contributing

Stars

5 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages