ki-housekeeping-claude
Standard: the hygiene of accumulated Claude state across the surfaces where it collects — the areas and the skill↔server pairing in the Claude-state standard; the memory area's file format in the auto-memory standard; line-by-line criteria in the rubric; tracked provenance in sources. The standards embed the two exact canonical shapes (directory/index and frontmatter), so this skill intentionally needs no separate exemplars file.
What this skill owns
The standard and judgment over the state Claude accumulates on a machine, across three surfaces — Claude Desktop / Cowork sessions, Claude Code (~/.claude/), and VSCode chat sessions — spanning the areas: sessions, artifacts / outputs, backups, plugins, project cache, and auto-memory. Full model in the Claude-state standard.
The mechanical arm is split by area:
- Native memory — governed locally when a readable native settings record establishes a selected contained directory. An absent or malformed settings record, disabled or unsupported override, or out-of-bounds override is reported unavailable; the rubric never silently falls back to its default path. Once selection is established, the index/file contract (every
memory/*.md listed in MEMORY.md, every entry resolving to a file), frontmatter schema, four-type doctrine, and promote-then-delete reconciliation are checked by ki repo audit --skill ki-housekeeping-claude. It never enumerates, reports, or writes another repository's memory. Detail in the auto-memory standard.
- Headroom output and every other area — a
headroom:learn block is rendered-file evidence only; it does not prove a Headroom database, version, installation, or executed learn action. The paired mcp-housekeeping-claude server is a separate tool source: a source checkout or inventory declaration does not prove registration, access exposure, or an executed audit. The skill applies judgment only to independently obtained server audit evidence; it never re-implements those tools.
AI session acquisition
Use the provider-neutral lifecycle: acquire → stage → harvest → durable knowledge → archive/delete source. For one exact physical repository, mcp-housekeeping-claude exposes read-only claude_code_sessions_discover, claude_code_sessions_list, claude_code_session_read, and claude_code_sessions_checkpoint operations. list and checkpoint are content-minimised provenance; the repository/session form of session_read is the faithful source payload. Its legacy project/session form remains a bounded preview.
The MCP does not write Knowledge Islands state. ki space acquire claude import will own inbound staging and incremental checkpoint persistence. Do not archive or delete a Claude session until acquisition, review, and harvesting have passed their later safety checkpoint.
Operating modes
Carries the universal AUDIT · CONFORM · EDUCATE · REFRESH. Invoked as help / -h / ?, it explains itself and stops — the generated HELP block (name, purpose, invocation, modes, off-ramps), taking no action. With no mode it does the same, then, in an interactive session only, offers the mode choice via AskUserQuestion, prompting for any argument-hint target the chosen mode shows.
| Mode |
What it does |
| AUDIT |
Run ki repo audit --skill ki-housekeeping-claude for selected native-memory evidence and its bounded store. Report non-memory server state unavailable unless registration, access exposure, and an executed server audit are separately evidenced; then apply the judgment criteria in rubric.md. Procedure in mode-audit.md. |
| CONFORM |
Run ki repo conform --skill ki-housekeeping-claude; it proposes only safe transactional repairs in an already selected bounded memory directory. Any non-memory cleanup requires independently established server access and authority. Re-AUDIT until the local evidence is clean. Procedure in mode-conform.md. |
| EDUCATE |
Run ki repo educate --skill ki-housekeeping-claude to declare the repository's bounded user-home evidence; add it with ki skill add ki-housekeeping-claude when it should also be installed for every supported Claude Code agent. User activation installs a skill only — it does not audit or conform it. |
| REFRESH |
Re-check the standard against its sources: Headroom's memory-feature behavior for the memory format, and the mcp-housekeeping-claude server's tool surface for the other areas, per mode-refresh.md. |
Notes
- The state this skill governs lives outside the repo tree — under
~/.claude/, ~/Library/Application Support/Claude/, and VSCode's workspaceStorage/. Its structured session is restricted to the selected repository's Claude project memory; the session / artifact / storage areas are machine-level and audited directly through the server, not by repository audit.
- A selected native directory with no
memory/ directory yet is an NA, not a FAIL. Uncertain selection is a FAIL, not an NA.
- Route Markdown formatting to the separately applicable
ki-authoring standard. The local ki-skills:rubric shared module is compile-time packaging, not a governance dependency.
- Does not assume any particular personal
~/.claude/CLAUDE.md content — the doctrines checked are the systems' own universal instructions, not one user's private elaboration of them.
1---2name: ki-housekeeping-claude3description: Governs accumulated Claude state from Desktop, Cowork, Claude Code (`~/.claude/`), and VSCode chat: sessions, artifacts, backups, plugins, project cache, and selected native auto-memory. It guides repository-scoped Claude Code session acquisition and later housekeeping: discover, list, faithful read, checkpoint, staging, harvest, and safe cleanup. Triggers: "acquire Claude Code sessions", "import Claude sessions", "audit Claude memory", "Claude memory hygiene", "clean up Claude storage", "obsolete Cowork sessions", "Claude housekeeping audit", "check ~/.claude". It is not a Knowledge Islands base memory cascade (`ki-repo-kb`) or context cost (`ki-tokenomics`).4---56# ki-housekeeping-claude78**Standard:** the hygiene of accumulated Claude state across the surfaces where it collects — the areas and the skill↔server pairing in [the Claude-state standard](references/standards-claude-state.md); the memory area's file format in [the auto-memory standard](references/standards-auto-memory.md); line-by-line criteria in [the rubric](references/rubric.md); tracked provenance in [sources](references/sources.md). The standards embed the two exact canonical shapes (directory/index and frontmatter), so this skill intentionally needs no separate exemplars file.910## What this skill owns1112The **standard and judgment** over the state Claude accumulates on a machine, across three surfaces — Claude Desktop / Cowork sessions, Claude Code (`~/.claude/`), and VSCode chat sessions — spanning the areas: **sessions**, **artifacts / outputs**, **backups**, **plugins**, **project cache**, and **auto-memory**. Full model in [the Claude-state standard](references/standards-claude-state.md).1314The **mechanical arm** is split by area:15161. **Native memory** — governed locally when a readable native settings record establishes a selected contained directory. An absent or malformed settings record, disabled or unsupported override, or out-of-bounds override is reported unavailable; the rubric never silently falls back to its default path. Once selection is established, the index/file contract (every `memory/*.md` listed in `MEMORY.md`, every entry resolving to a file), frontmatter schema, four-type doctrine, and promote-then-delete reconciliation are checked by `ki repo audit --skill ki-housekeeping-claude`. It never enumerates, reports, or writes another repository's memory. Detail in [the auto-memory standard](references/standards-auto-memory.md).172. **Headroom output and every other area** — a `headroom:learn` block is rendered-file evidence only; it does not prove a Headroom database, version, installation, or executed learn action. The paired **`mcp-housekeeping-claude`** server is a separate tool source: a source checkout or inventory declaration does not prove registration, access exposure, or an executed audit. The skill applies judgment only to independently obtained server audit evidence; it never re-implements those tools.1819## AI session acquisition2021Use the provider-neutral lifecycle: **acquire → stage → harvest → durable knowledge → archive/delete source**. For one exact physical repository, `mcp-housekeeping-claude` exposes read-only `claude_code_sessions_discover`, `claude_code_sessions_list`, `claude_code_session_read`, and `claude_code_sessions_checkpoint` operations. `list` and `checkpoint` are content-minimised provenance; the repository/session form of `session_read` is the faithful source payload. Its legacy project/session form remains a bounded preview.2223The MCP does not write Knowledge Islands state. `ki space acquire claude import` will own inbound staging and incremental checkpoint persistence. Do not archive or delete a Claude session until acquisition, review, and harvesting have passed their later safety checkpoint.2425## Operating modes2627Carries the universal **AUDIT · CONFORM · EDUCATE · REFRESH**. Invoked as `help` / `-h` / `?`, it explains itself and stops — the generated HELP block (name, purpose, invocation, modes, off-ramps), taking no action. With no mode it does the same, then, in an interactive session only, offers the mode choice via `AskUserQuestion`, prompting for any `argument-hint` target the chosen mode shows.2829| Mode | What it does |30| --- | --- |31| AUDIT | Run `ki repo audit --skill ki-housekeeping-claude` for selected native-memory evidence and its bounded store. Report non-memory server state unavailable unless registration, access exposure, and an executed server audit are separately evidenced; then apply the judgment criteria in [rubric.md](references/rubric.md). Procedure in [mode-audit.md](references/mode-audit.md). |32| CONFORM | Run `ki repo conform --skill ki-housekeeping-claude`; it proposes only safe transactional repairs in an already selected bounded memory directory. Any non-memory cleanup requires independently established server access and authority. Re-AUDIT until the local evidence is clean. Procedure in [mode-conform.md](references/mode-conform.md). |33| EDUCATE | Run `ki repo educate --skill ki-housekeeping-claude` to declare the repository's bounded user-home evidence; add it with `ki skill add ki-housekeeping-claude` when it should also be installed for every supported Claude Code agent. User activation installs a skill only — it does not audit or conform it. |34| REFRESH | Re-check the standard against its sources: Headroom's memory-feature behavior for the memory format, and the `mcp-housekeeping-claude` server's tool surface for the other areas, per [mode-refresh.md](references/mode-refresh.md). |3536## Notes3738- The state this skill governs lives **outside the repo tree** — under `~/.claude/`, `~/Library/Application Support/Claude/`, and VSCode's `workspaceStorage/`. Its structured session is restricted to the selected repository's Claude project memory; the session / artifact / storage areas are machine-level and audited directly through the server, not by repository audit.39- A selected native directory with no `memory/` directory yet is an **NA**, not a FAIL. Uncertain selection is a **FAIL**, not an NA.40- Route Markdown formatting to the separately applicable `ki-authoring` standard. The local `ki-skills:rubric` shared module is compile-time packaging, not a governance dependency.41- Does not assume any particular personal `~/.claude/CLAUDE.md` content — the doctrines checked are the systems' own universal instructions, not one user's private elaboration of them.
knowledgeislands