knowledge

Review captured learnings and memory inbox items, propose durable create/update diffs, and apply approved changes. User-run only; nothing is written without explicit approval.

Diff a saved session context snapshot against its archived history versions for the current Codex project.

Generate a concise context snapshot or structured handoff for the current Codex session or project so another session can continue the work.

List available session context snapshots for the current Codex project.

Load a previously saved session context snapshot into the current Codex session.

Remove a saved session context snapshot for the current Codex project.

Search the contents of session context snapshots across local projects. Use when the user wants to find which snapshot or project mentions a topic, keyword, or decision.

Share a saved session context snapshot with another named tmux pane.

Verify a structured handoff's local file and commit evidence against an explicitly bound Git repository, without changing stores or executing recorded commands.

This skill should be used when creating new documentation files or updating existing ones, especially when documenting system architecture, features, integrations, workflows, or technical decisions. Also use when asked to "document", "write docs for", or "create a doc about" any system component. Trigger this skill whenever the user wants to capture knowledge about how something works, even if they don't explicitly say "documentation" — phrases like "explain how X works and save it", "write up the auth flow", or "I need a reference for the API" all qualify.

This skill should be used to independently verify the accuracy of documentation after it has been written or updated — it checks that every referenced file path, function, table, endpoint, and cross-link actually exists in the codebase, and runs the plugin's validation scripts. Trigger it after creating or editing docs (e.g. "review the docs I just wrote", "verify docs/ARCHITECTURE.md is accurate", "check the documentation for stale references") or as the verification step of the docs-create workflow. It reads and reports only; it does not edit files.

Diagnose knowledge-store health across docs, memory, context, AGENTS.md, and provider capability (strictly read-only, cross-store).

Search local docs, memory, and context together, with separate source, authority, and lifetime labels (read-only).

Explicit-link knowledge graph over memory backlinks: neighbors, reverse links, orphans, components, and whole-graph renders.

Bootstrap a new .agents/memory/ store for this repository through a reviewable two-call protocol.

Choose the right knowledge command for project docs, durable memory, or session context. Read before using a knowledge command for store boundaries and role rules.

Lint the memory store's frontmatter, schema, and index for defects (read-only).

Promote stable context or memory into a durable destination, with separate confirmation before deleting the source. User-run only; nothing is written without explicit approval.

Agent-facing recall: return slug citations and bounded snippets as untrusted fallible context.

Capture a low-friction memory candidate into the inbox for later consolidation, list pending candidates, or purge explicitly selected candidates.

Deterministic lexical ranked search over the memory store (read-only).