measuring-what-a-kiro-agent-loads-into-contextlisted
Install: claude install-skill sblattj/agent-chandlery
## More triggers
kiro-context-probe (old name)
# measuring-what-a-kiro-agent-loads-into-context
Answers "which instructions does this Kiro agent start with?" from the session log, not from the docs (they disagreed once). Examples use `lean`, a trimmed custom agent; substitute your own agent names.
## Do this
```bash
P=${SKILL_DIR:-~/.claude/skills/measuring-what-a-kiro-agent-loads-into-context}/probe.sh
$P --agent lean --mark "<a heading from your rules>" CANARY_X # V3: lists steering docs + chars, counts each mark
$P --mark "<a heading from your rules>" # the default agent (chat.defaultAgent)
$P --engine v2 --recall # V2 logs no steering: ask the model what it was told
```
To test a mode, drop a canary into `~/.kiro/steering/` (`---\ninclusion: manual\n---\nCANARY_MANUAL_1\n`), probe, then **delete it right away**: it is live global steering for every Kiro session on the machine.
## Why it's built this way (friction mined)
- `kiro-cli chat --no-interactive` runs the **V2** engine unless you pass `--agent-engine v3`. V2 writes `~/.kiro/sessions/cli/*.jsonl` with no steering content, so a grep there finds nothing and proves nothing.
- V3 sessions live in `~/.kiro/sessions/<hash>/sess_*/messages.jsonl`. Find yours by `workspacePaths` in `session.json` (the script runs in a fresh `/tmp` dir and matches on its realpath, which on macOS is under `/private`). `ls -t` picked another session's live 700 KB log.
- V3 behavior, m