Agentic engineering kit: seven skills (context engineering, harness engineering, skill development, agent development, evals, sandboxing, fewer faster tests), most named by a short block a hook prints at the moment it applies; an agentic-engineer agent that starts with two of them loaded; /new-plugin, which makes your own plugin in a folder you pick; and /new-agent and /new-skill, which write the file into that plugin or the project from the skill's template, or hand your words to Claude with the skill loaded (a Claude Mod: needs CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1).
7 skills · 1 agent · hooks
Which session did what — a structural index over Claude Code's own transcripts (~/.claude/projects/*.jsonl), verbatim, no LLM pass. Its own command, `tracer` (bin/, on the Bash PATH while the plugin is on), with two groups: `tracer sessions` (ls · live · show · replay · turns · search — the conversations) and `tracer trace` (every tool call, plus turn · blame · index — what was done in them); every older spelling still runs, unlisted. With the ak CLI installed the same groups are also `ak sessions` and `ak trace`. Also the /tracer:trace, :blame and :index skills, and a Stop-hook index writer (it also launches the tool-call indexer this plugin ships under traces/, the store behind `tracer trace`). Its index lives under ~/.claude/plugins/data/conversation-index; it needs no vault, no ak and no other plugin.
3 skills · 1 agent · hooks
How a session works with you and with its subagents. Two session modes, each [on|off|status], answered with no model call, Claude told once at each switch and reminded by one line on each prompt while on, a label in the prompt footer: /orchestrate hands parts to subagents on another model (haiku · sonnet · opus); /ask answers and changes nothing, refusing Edit, Write and NotebookEdit while on. /modes:interview asks you rounds of AskUserQuestion until the direction is clear. /modes:align stops work in progress: it plays back what was built, names where it drifted from what you asked, and asks before any more changes. A Claude Mod: needs CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1.
2 skills
What past sessions learned. An observer runs at SessionEnd (the plugin agent agents/observer.md, run headless) and writes observations as rows in the vault's store (~/.ak/db/brain.db); the next session opens on a 'since you were last here' digest; `observer search → timeline → get` recalls them in three layers and `patterns · fold · forget` curate them; a Read hook says when a file has prior observations. CLI: `observer status|mute|unmute|prompts|show|try|store|types|search|timeline|get|patterns|fold|forget|shadow`, also mounted as `observer …` when ak is installed. Shares only the vault's data paths with the other kit plugins.
1 agent · hooks
The ak command, and ak setup, the kit's installer: it checks your machine, installs the plugins you pick, and takes a run back with ak setup undo. Its own verbs: plugin ls, inspect and log (what your plugins put in front of your agent) and hooks (which hook ran, when, how long); with the gateway, gateway, xray (the exact prompt a claude launch would send, as a dry run), sysprompt and shim. Every other kit plugin adds its verbs to it: the tracer's are ak sessions and ak trace. A uv tool, kept up to date when a session starts.
hooks
WebFetchRaw, a tool beside WebFetch: it fetches a URL and saves the page's own content to a file, with no model in between, and returns one line naming the file. WebFetch returns a small model's answer about the page. An HTML page is extracted to markdown, text or cleaned HTML by trafilatura; anything else (JSON, text, a PDF) is saved byte for byte. A Claude Mod (needs CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1) that runs a Python script through uv once per call. Shares nothing with the vault and needs no kit.