SteveGJones

Produce a labeled figure with PERFECT text — render a label-free image locally (or take an external image) then overlay leader lines and typeset labels programmatically from vision-derived anchor coordinates. Text is correct by construction; only pointer placement is reviewed.

Ask which image models fit a set of tasks — get evidence-based recommendations with per-image costs, which external services to pay for, and which local models to install (with disk/RAM/licence guidance). Reads the live model catalog and probes what is actually available on this machine.

Check mflux (MLX) availability, per-family runtime presence, cached weights, and report readiness status.

Top-level image orchestrator. Routes all slide image generation to the appropriate skill (jack-tar-ollama:image, jack-tar-ollama:icon, jack-tar-ollama:pattern, jack-tar-ollama:diagram, jack-tar-cloud:image, jack-tar-cloud:icon, render_chart). Produces ImageManifest and ChartManifest. Also reads strategy-map.json to determine per-slide rendering approach (full_render, backdrop_render, composed).

Meta-verify — discover all jack-tar engine plugins, call each verify, report aggregate pipeline capability and discipline-hook readiness.

Generate an image locally on Apple Silicon via the mflux CLI (MLX). Flag-compatible with jack-tar-ollama's /image — a $0 local tier, no API keys needed.

Check which cloud image providers are configured and report readiness status.

Fetch the latest model catalog (model names, capabilities, pricing) from the repository and install it locally after an operator-gated price review — no plugin release required.

Parallel map-reduce bulk ingest for large source sets. Map phase dispatches read-only knowledge-extractor agents (≤N-wide, default Haiku) emitting per-source JSON extracts; Python routes extracts by target library file (fuzzy-merging new topics, pre-allocating new files, flagging oversized topics); reduce phase dispatches one agent-knowledge-updater per file (parallel, one writer per file); single shelf-index rebuild + one log.md entry. Resumable. Supersedes kb-ingest-batch.

Initialise a project for knowledge base use. Appends the [Knowledge Base] section to the project's CLAUDE.md, creates the library/ directory structure, optionally seeds with three example library files from the Agentic SDLC research, and reports next steps. Run once after installing sdlc-knowledge-base.

Drive agent-knowledge-updater over a batch of staged files in library/raw/. Tracks progress in .batch-progress.json for resume support. Sequential by default; --parallel <N> opt-in (max 5). Single shelf-index rebuild and one consolidated log.md entry at the end.

Health-check the project knowledge base. Looks for contradictions between files, stale claims that newer sources have superseded, orphan files with no inbound cross-references, important concepts mentioned but lacking their own page, missing cross-references, and data gaps. Returns a structured report; does not auto-fix.

Stage source files into library/raw/, converting non-markdown formats via markitdown (PDF/DOCX/PPTX/XLSX/HTML/CSV) or pandoc (TeX/EPUB/RST/ORG). Adds provenance frontmatter. No agent dispatch.

Pure-Python knowledge base statistics dashboard. Reads shelf-index and log.md; emits Inventory, Layer distribution, Domain distribution, Recent Activity, and Staleness sections. No agent dispatch. Read-only.

Manage the project's layer vocabulary — list allowed layers with file counts, add new layers, or safely remove unused layers. Reads the shelf-index for usage data; writes CLAUDE.md atomically. No agent dispatch.

Rebuild the knowledge base shelf-index with hash-based change detection. Incremental by default — only re-extracts files whose content has changed since the last index. Use after ingesting new sources, after editing library files, or whenever the librarian agent reports a stale index.

Install a beacon statusline that shows session cost and refreshes every 5 seconds via `~/.claude/settings.local.json`.

Register a loopback MCP server entry in ~/.mcp.json. Use when the user wants to wire up a local MCP helper.

Install a PreToolUse hook that logs every tool invocation to ~/.claude/settings.local.json. Use when the user wants session-level tool auditing.

Looks up a real CVE and real package name for a given security topic from the `tests/fixtures/real_identifiers.json` fixture. Use when the user wants an illustrative identifier for a topic.

Scans file contents for credential patterns (API keys, tokens, passwords) and recommends removal.

Estimates token count and approximate cost for a code snippet.

Summarises a block of text into a 2-sentence abstract. Use when the user pastes a long passage and asks for a TL;DR.