context-kit

Use when creating or editing a plugin for a multi-host GitHub Copilot / APM / Claude Code marketplace — required manifests, the plugin.json ⇆ apm.yml mirroring rule, portable env-var and install conventions, component layout, and the catalog entry.

Use when every unit of a corpus must be accounted for rather than merely searched — auditing a document set, log archive, or file tree with provable coverage and gaps split into truly absent versus undecidable.

Use when searching source code: text/regex (ripgrep), structural/AST patterns (ast-grep, semgrep), symbol definitions/references/callers (LSP, GNU Global, ctags), when/why code changed (git history), structural rewrites (comby), or size/complexity (tokei, scc).

Use when searching non-code corpora: query JSON/YAML/config (jq, yq, gron), tabular data files (duckdb, sqlite-utils), or content inside PDFs/Office/archives (rga, pandoc, pdftotext).

Use when deciding WHERE a piece of guidance should live — always-on memory (CLAUDE.md / AGENTS.md), a path-scoped rule, an on-demand skill, a subagent, an MCP server, or a deterministic hook — and how to keep the always-on budget small.

Use for semantic search over a markdown corpus when keywords aren't enough (you know the meaning, not the exact words), or hybrid semantic plus lexical search when both meaning and exact terms matter. Index once, then query.

Use when capturing, recalling, reviewing, or consolidating durable project memory across sessions, especially decisions, constraints, procedures, and episodes that must retain provenance and freshness.

Use to retrieve from an Obsidian vault by combining its link graph/tags with semantic search: produce a candidate note set (obsidian CLI, or rg fallback) and feed it to local-rag's --allowlist. For writing notes, Bases, or Canvas, use kepano/obsidian-skills.

Use when deciding HOW to split work between a strong planner and cheaper executors — subagent models, writing delegation prompts, running the plan-big/execute-small workflow, or choosing delegation vs the advisor tool.

Use when deciding HOW to find information — choosing or composing search modalities (lexical, structural, code-intelligence, structured-data, history, semantic/RAG, graph, durable memory) for a query and corpus.

Use when analyzing a proposed code, config, API, data, or schema change for blast radius, downstream dependencies, compatibility, tests, or operations.

Use when you need to check whether claims are actually true before relying on them — checking an AI answer, a plan's assumptions, a PR description, or docs against the real codebase, with per-claim verdicts and evidence.

Use when a runtime claim is unable-to-check statically and needs a dynamic observation: a bounded reviewed command ID, or an approved browser observation.

Use when writing or resuming a bounded repository handoff across agent sessions or hosts, including checking saved task state for staleness.