WilcoRakhorst The five root causes of context loss
1. Re-reads
The same instruction in multiple files means the agent reloads it every time it's referenced. Cross-cutting content belongs exclusively in references/*.md. Skills link to them — they never copy them.
User tip: Use the edit button on a mistaken prompt instead of sending a correction as a new message. This keeps the error out of context entirely.
2. Reading entire files instead of snippets
- Always use
startLine+endLinewithread_file - Use
grep_searchto locate the right lines before callingread_file - Never use
list_dirrecursively on the workspace root — deep structures (e.g.node_modules,.git) index thousands of paths and fill context fast - Use
semantic_searchsparingly — broad queries load a lot of irrelevant content; usegrep_searchfor exact terms - Never write back an entire file after a change — use only the changed section or a diff format
3. Too many tools active
- Load deferred tools only via
tool_searchwhen actually needed - Use the
Exploresubagent for read-only research — keeps the main context clean - Keep
copilot-instructions.mdshort; defer detail toreferences/
4. Working too long in one session
- Start a new chat after 15–20 messages or after completing a distinct task
- At the end of a complex task, ask: "Summarize our progress as a short status update." Paste that as the first message in the new chat
- Save intermediate results to session memory (
/memories/session/) or to files - Use
manage_todo_listto track progress — so you never need to scroll back
5. Trial-and-error tool use
Require the agent to write a Plan of Action before any tools are invoked on complex tasks — which steps, which tools, in which order. Use manage_todo_list when the plan has multiple steps.
Pre-task checklist
| Check | Action |
|---|---|
| Is this info already in context? | Don't reload it (→ cause 1) |
| Which skills do I actually need? | Load only those — not preemptively (→ cause 3) |
| Am I reading a large file? | Use startLine/endLine (→ cause 2) |
| Searching with a broad term? | Use grep_search instead of semantic_search (→ cause 2) |
| Complex task with multiple tools? | Write a Plan of Action first (→ cause 5) |
| Been in this session a long time? | Consider new chat + session memory (→ cause 4) |
| Agent repeating a question? | Session is full — save state and restart (→ cause 4) |
| Output getting too long? | Ask for "diff format" or only the changed section (→ cause 2) |
| Search results slow or irrelevant? | Check files.exclude and search.exclude |
VS Code: prevent silent re-indexing
Add to .vscode/settings.json:
"files.exclude": { "out/": true, "Output/": true },
"search.exclude": { "out/": true, "Output/": true }
VS Code indexes your entire workspace — including generated output folders with HTML reports, screenshots, and logs. These end up as irrelevant context when the agent runs a search. .gitignore does not prevent this. Only files.exclude and search.exclude do.
References (loaded on demand)
- terminal-protocol.md — required transparency rules before running terminal commands
- terminal-risk-classification.md — 🟢/🟡/🔴/⛔ risk matrix per command type