context-check

Context Check

Analyze the recent conversation to detect facts worth persisting.

Usage

/context check

What to extract

Scan the last 10 human/assistant messages for concrete facts in these categories:

insight — completed work:

• Code changes with context (what was done + why)
• Patterns applied or techniques used
• Gotchas, pitfalls, or surprising behaviors discovered
• Decisions made with reasoning

agent_edit — AI behavior changes:

• Agent guards, skill descriptions, hook logic edits
• Prompt template changes
• CLAUDE.md or plugin config updates
• Directives that control how the assistant works

task — ONLY unstarted intentions:

• "I need to refactor X", "TODO: add Y"
• Work the user plans but has NOT begun

Fact quality

Good facts (concise, high-level, with WHY):

• "ContextDomain is a new field for matching block outputs by domain context"
• "guard findSessionJsonl against undefined sessionId — was matching real files"
• "skills should describe pure workflow steps — moved delegation details to agent frontmatter"

Bad facts (skip these):

• Too granular: "added field to BObjectSpec" (what is it FOR?)
• Too vague: "fixed authentication issues"
• Routine: "ran tests", "committed code", "read a file"
• Reasoning: "this is because...", "the issue stems from..."

Process

1. Read the current session transcript
2. Extract facts grouped by topic and classification
3. Save immediately using the context-save skill — do NOT ask for confirmation

Configuration

Read insights_root from ~/.claude/memory-keeper.local.md YAML frontmatter.

Output

If no facts found, return "nothing worth saving". If facts found, save them immediately and report what was saved (topic + classification + file).