AI Agent Workflow Collection

Use when the user is about to perform a risky, large, or hard-to-reverse change so the pre-change context is recoverable. Invokes mcpclxclxcheckpoint to snapshot the current session. Examples: "checkpoint before refactor", "snapshot the current state before migration", "save context before squash". Do NOT use on every assistant turn (that is auto-summarize, off by default), or for single small facts (use clx-remember instead).

Use when CLX recall returns no results, looks empty, or behaves unexpectedly; diagnoses provider, embedding, DB, and config issues via shell commands (clx doctor, clx embeddings status, clx providers ping) and log inspection. Examples: "why isn't recall working", "clx returns empty", "diagnose embedding model mismatch", "check provider fallback status". Do NOT invoke for normal "no relevant context" outcomes; only for outright failures or recall results that look structurally wrong (e.g., wrong model id, missing snapshots).

Use when the user references earlier sessions, prior decisions, or asks "what did I do" / "did we already" about past work in this repo. Invokes mcpclxclxrecall to search the local CLX snapshot store. Examples: "look up our Azure backend discussion", "find checkpoints from yesterday", "what did we decide about retries". Do NOT use for current-session context (already in the transcript), real-time data, or web search; do not use for secrets retrieval.

Use when the user states a durable preference, decision, or fact worth carrying across sessions on this machine. Invokes mcpclxclxremember to persist a single short fact. Examples: "remember that we use Conventional Commits", "save this decision about Azure tenant routing", "note that prod uses M-series only". Do NOT use for transient state, draft code, speculation, secrets, PII, or large in-flight context (use clx-checkpoint for snapshots instead).

Use when the user wants to resume earlier work on a specific topic, ticket, or file area from a prior session. Orchestrates mcpclxclxrecall plus snapshot reads, then confirms before acting. Examples: "resume the recall accuracy work", "pick up where we left off on the Azure fallback", "continue the dashboard refactor". Do NOT use to start brand-new work, or to auto-replay prior steps without user confirmation; recalled context may be stale.

Use when the user asks "what rules apply here", after a long session where CLAUDE.md instructions may have decayed from context, or when rule violations seem likely. Invokes mcpclxclxrules to refresh project, user, and org rules in precedence order. Examples: "what coding rules apply", "refresh project rules", "remind me of the commit convention". Do NOT invoke at the start of every turn; rules are already injected via CLAUDE.md on session start.

Use when working in a Claude Code session with CLX installed and persistent memory is relevant — specifically when the user references prior work ("earlier", "we discussed", "before", "last time"), when a decision or preference surfaces that should survive across sessions, before a risky or hard-to-reverse change, or when deciding whether to call clxrecall, clxremember, clxcheckpoint, or clxrules. Covers CLX's memory model, MCP tool triggers, query craft, result interpretation, and how CLX memory relates to Claude's native auto-memory and context compression.