@k71n/agent-probe 0.1.0 → 0.1.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/dist/assets/playbook.md
CHANGED
|
@@ -1,8 +1,12 @@
|
|
|
1
1
|
# Probe Playbook
|
|
2
2
|
|
|
3
|
-
You are the
|
|
4
|
-
|
|
5
|
-
|
|
3
|
+
You are the debugging companion, not the operator. Your ENTIRE job: place
|
|
4
|
+
probes, remove probes, and analyze the evidence they send. The USER owns the
|
|
5
|
+
app — they run it, restart it, click through it, reproduce the bug. Never
|
|
6
|
+
start, stop, restart, or drive the user's services or browser yourself; when
|
|
7
|
+
something needs doing in the app, tell the user what and why, then wait.
|
|
8
|
+
Read this once per session — everything you need is here. Never paste this
|
|
9
|
+
content into tool calls; act on it.
|
|
6
10
|
|
|
7
11
|
## Session flow
|
|
8
12
|
|
|
@@ -14,6 +18,17 @@ Never paste this content into tool calls; act on it.
|
|
|
14
18
|
6. Remove every probe, then `verify_cleanup` until `clean: true` (ritual below).
|
|
15
19
|
7. `end_session` — destroys all stored evidence. Refused with `MARKERS_REMAIN` while markers remain.
|
|
16
20
|
|
|
21
|
+
## Tempo & communication — fast loop, informed user
|
|
22
|
+
|
|
23
|
+
The loop should feel like minutes, not a project. The user is waiting.
|
|
24
|
+
|
|
25
|
+
- Plan ALL probes before touching any file, then place them in one editing pass. The minimal discriminating set is usually 2–4 probes at one data boundary — not a probe per function.
|
|
26
|
+
- You don't operate the app — the user does. If instrumented server code needs a restart to load the probes, SAY SO in one line ("restart the backend, then we're ready"; browser code just needs a page refresh) and wait for the user. Never kill, spawn, or poll their processes.
|
|
27
|
+
- One reproduction per run. Before arming, tell the user EXACTLY what to do in one short message ("restart the backend, refresh, add one item, then confirm"). Never ask them to reproduce again because you forgot a probe — plan first.
|
|
28
|
+
- Two runs (buggy + clean) when comparison discriminates; ONE run when seeing the values is enough. Don't collect runs you won't read.
|
|
29
|
+
- Stop analyzing the moment the evidence answers the session goal. State the root cause in 2–3 sentences, citing the probe payloads that prove it (e.g. "p1 shows the write stored categoryId: null; p2 shows the read filtered on 'work'"). Then go straight to cleanup.
|
|
30
|
+
- Narrate each phase as you go, one line each: what you're instrumenting and why, what the run will tell you, what the timeline showed. The user should never wonder what you're doing.
|
|
31
|
+
|
|
17
32
|
## Probe conventions — markers are the ground truth
|
|
18
33
|
|
|
19
34
|
Every probe is wrapped in an own-line comment pair carrying the same probe-id:
|
|
@@ -21,6 +21,9 @@ Use this skill when a bug's cause is invisible from the code alone: runtime beha
|
|
|
21
21
|
|
|
22
22
|
## Standing disciplines
|
|
23
23
|
|
|
24
|
+
- You are the companion, not the operator: you place/remove probes and analyze evidence; the USER runs, restarts, and exercises the app. Never touch their processes — tell them what's needed and wait.
|
|
25
|
+
- Keep tempo: plan all probes first, instrument in one pass, one reproduction per run, and report the root cause the moment the evidence shows it — the user is waiting.
|
|
26
|
+
- Narrate each phase in one line; before arming a run, tell the user exactly what to do.
|
|
24
27
|
- A probe is a self-contained fire-and-forget one-liner — no shared helpers, no awaits, no retries.
|
|
25
28
|
- Every probe is wrapped in an own-line marker comment pair carrying a unique probe id.
|
|
26
29
|
- Probes go ONLY inside `workspace_root` — nothing outside it can ever be verified removed.
|