Home

Inspired by the Noisy Cricket — agent primitives that punch far above their weight.

_{Works with Claude Code + Antigravity — see compatibility}

Crickets is the noisy cricket — small, focused agent customizations that punch above their weight. Skills, hooks, sub-agents, bundles, MCP servers, slash commands, status lines, output styles, workflows, rules, snippets, settings-fragments. The primitives you carry into any project to make Agent M (agentic-harness) effective.

This wiki is the contributor-facing documentation for Crickets itself. Every page is written for a single reader intent (learn / do / look up / understand) per the Diátaxis convention.

Note

Repo: github.com/alexherrero/agent-toolkit. Public, MIT-licensed. PII guardrails (script + skill + pre-push hook) are foundational — see Purpose and scope for why.

📚 New here? Learn by doing.

• Tutorial 1 — Your first customization — add a hello-world skill, install it into a scratch project, see it land at three host paths. ~10 minutes.

🔧 Trying to do something specific?

• Install agent-toolkit into a project — drop the shipped customizations into a target project.
• Add a skill — add a new standalone skill.
• Add a bundle — package multiple primitives that ship together.
• Use the evaluator — dispatch the evaluator sub-agent for PASS / NEEDS_WORK grading against a precise rubric.
• Use the base hooks — kill-switch, steer, commit-on-stop. Operator-precision control for long-running Claude Code sessions.
• Use the evidence-tracker hook — default-FAIL evidence enforcement on /work task closeouts (4th base hook; agent must Read evidence files before flipping PLAN.md [x]).
• Use the design skill — /design author walks the 10-section template, /design translate splits into parts, /design sequence generates a PLAN.md per part. Hand-off to harness /work + /review for execution.
• Use the memory skill — /memory save captures durable preferences / workflows / fixes; /memory evolve supersedes existing entries (preserves audit trail). Recall (auto-injection at SessionStart + UserPromptSubmit), reflection sidecar, idea ledger, discovery come in subsequent parts of the MemoryVault design.
• Use the diataxis-author skill — /diataxis author / check / repair / migrate / classify covering the full Diátaxis-wiki lifecycle. See the diataxis-author design.
• Use the quality-gates bundle — one-command install of the 4 base primitives (evaluator + kill-switch + steer + commit-on-stop + evidence-tracker) most harness /work sessions want.

📖 Looking up a detail?

• Customization Types — what each of the 11 kinds means and where to put them.
• Manifest Schema — YAML frontmatter contract.
• Per-Host Paths — destination paths per kind per host.
• Installer CLI — flags, prereqs, exit codes.
• Compatibility — supported hosts (Claude Code, Antigravity) + OS matrix + per-manifest supported_hosts contract.

💡 Want to know why?

• Purpose and scope — what this repo is for, what it's not.
• Cross-Repo Memory Protocol — how agentic-harness reads from + writes to the toolkit-side /memory skill at phase boundaries.
• V3 Retrospective — what shipped, what we learned, what's next (closes the V3 arc; ships with harness v3.0.0 + toolkit v1.0.0).
• Agent Memory Evolution: From ContextVault to V4 — V1→V4 HLD of the AgentMemory architecture (V3 ships with this release; V4 is roadmap-deferred).

Architecture decisions

• ADR 0001 — agent-toolkit purpose, scope, public-with-PII-guardrails
• ADR 0002 — evaluator sub-agent design
• ADR 0003 — base operator-control hooks
• ADR 0004 — design skill: human-facing design pipeline → agent execution handoff
• ADR 0006 — Gemini CLI host removal
• ADR 0007 — MemoryVault Discovery + Mining
• ADR 0008 — diataxis-author skill
• ADR 0009 — evidence-tracker hook
• ADR 0010 — quality-gates bundle + sibling-reference dispatch

Designs

The canonical "Why we built X" entry points (published via the /design skill, surfaced here when the parent design's last queued part hits Status: done and harness /release transitions the design to launched).

• MemoryVault — permanent agent memory via Obsidian-vault-folder + reflection sidecar — single memory skill in agent-toolkit with sub-commands save / evolve / reflect / search; SessionStart + UserPromptSubmit recall hooks; reflection sidecar with tri-modal confidence routing; two-tier idea capture (Ideas.md surface + _idea-incubator/ deep research); planned in 6 parts across plans #7a (5 parts) + #7b (1 part). See parts: write-primitives, recall-loop, reflection-and-recovery, idea-ledger, seed-pass, discovery-mining.
• diataxis-author — Diátaxis wiki authoring + maintenance for any repo — second major skill after memory; subsumes the harness's migrate-to-diataxis predecessor; AgentMemory-integrated conventions; planned in 5 parts. See parts: skill-scaffold, author-classify, check-repair, migrate-subsume, agentmemory-docs-release.

Conventions

Page templates, filename rules, and the Diátaxis four-mode split mirror the sibling repo's documentation convention — see agentic-harness/harness/documentation.md. The mode-purity lint (scripts/check-wiki.py --strict) is the same.