popoffvg/context-check

Context Check

Analyze the current conversation for facts worth persisting. Save immediately without asking.

Usage

/context check

What to extract

Scan the conversation for concrete, reusable knowledge:

insight — completed work worth remembering:

• How a system works (architecture, data flow, API behavior)
• Patterns applied and why they work
• Gotchas, pitfalls, surprising behaviors
• Decisions made with reasoning

agent_edit — AI behavior changes:

• Hook logic, prompt changes, skill updates
• CLAUDE.md or plugin config changes
• Directives that change how the assistant behaves

task — ONLY unstarted intentions:

• "I need to refactor X", "TODO: add Y"
• Work planned but not started

none — skip silently:

• Routine: "ran tests", "committed code", "read a file"
• Too vague: "fixed authentication issues"
• Too granular: "added field X to struct Y" (what is it FOR?)

Entry Format

## <Keyword-rich title, 3-7 words> — YYYY-MM-DD HH:MM

<Lead: direct statement of what this is and why it matters. 1-3 sentences.
Include concrete identifiers: file paths, function names, config keys.
No preamble. Write for someone with zero memory of this session.>

### Subsection (only when topic has multiple distinct aspects)
<Details>

Heading = the search key. Must be findable by someone who forgot the topic:

• Good: Go context: cancel propagates through goroutine tree
• Good: better-sqlite3 native module: breaks on Node version upgrade
• Bad: Context stuff, Fixed bug, Important pattern

Lead sentence = complete picture. Self-contained, states facts directly:

• Good: Pi's shutdown() awaits all session_shutdown handlers before process.exit(0), giving handlers unlimited time for cleanup or LLM calls.
• Bad: In this session we discovered that the shutdown handler was being called...

Process

1. Scan the conversation — identify distinct topics with reusable value
2. For each topic: classify, write entry in wiki format
3. Check active task in <insights_root>/_tasks/pending.md — route insight entries to task notes if active task exists
4. Deduplicate against target file (exact heading, semantic overlap, superset/subset)
5. Save immediately — do NOT ask for confirmation
6. Report: topic + classification + file saved to

If nothing worth saving: return "nothing worth saving".

Autoresearch rules

Eval checklist:

1. Did the scan extract at least one concrete, reusable fact (not a vague observation)?
2. Was every extracted insight saved immediately without asking the user?
3. Were zero duplicate insights saved (already exists in memory)?
4. Did the extraction focus on debugging discoveries, architectural decisions, or corrections — not routine actions?

Test inputs:

• "Session where user debugged a tricky Kubernetes pod failure"
• "Session with only routine file edits, nothing novel"
• "Session where an architectural decision was made and rationale explained"

Can change: extraction criteria, insight categories, save format, scanning depth Cannot change: save-without-asking behavior, QMD tool usage for persistence Min sessions before eval: 5 Runs per experiment: 3