Skip to content

Commit 508930e

Browse files
committed
docs(planning): add pool-invalidation manual test playbook and Aug 20 results
Signed-off-by: crimsonsunset <jsangio1@gmail.com>
1 parent dcc2977 commit 508930e

1 file changed

Lines changed: 201 additions & 0 deletions

File tree

Lines changed: 201 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,201 @@
1+
# Pool Invalidation + Session Survival — Manual Test Playbook
2+
3+
**For:** verifying Phases 1–4 on `root-resolution`
4+
**Last Updated:** Aug 20, 2026
5+
**Implements:** [`pool-invalidation-and-session-survival.md`](./pool-invalidation-and-session-survival.md)
6+
**Shipped:** `dcc2977` — config-save `reconnect_fresh`, stdio OAuth refuse, hold `X-Mcpmux-Workspace` until session id
7+
**Parent playbook:** [`backend-connection-resilience-test.md`](./backend-connection-resilience-test.md) (Cases A/B already PASS)
8+
9+
---
10+
11+
## Results (Aug 20, 2026)
12+
13+
Ran live against `127.0.0.1:45818` (`{"status":"ok","version":"0.5.0"}`). Binary `target/debug/mcpmux` contains the new strings (`held until mcp-session-id`, `stdio cannot reconnect via OAuth`, `reconnect_fresh after config update`). Process up since 11:36:44 local / recycle log `17:37:21Z`. Mux only via `user-mcpmux`. `set_workspace_root` not used.
14+
15+
| Case | Result |
16+
| ---- | ------ |
17+
| E header pin | **PASS** — this repo `ready` (wakatime, HA, …). Session `60e5261f…` header `~/Desktop/Repos/Personal/mcp-mux` → pin `/Users/joe/Desktop/Repos/Personal/mcp-mux`. Log: `held until mcp-session-id exists` then `pinned explicit workspace root`. Not 6-way `bindable`. |
18+
| F stdio refuse / Case B regression | **PASS**`wakatime` / `wakatime_wakatime_summaries`. Killed child pids 74227/74257 at 17:39:38Z. Second invoke succeeded. `trigger=transport_closed`, `reconnect_fresh completed ok=true` (1544ms). No `Unexpected STDIO`, no OAuth reconnect. Source grep of that string: 0. |
19+
| G config-save reconnect | **BLOCKED** — admin `:45819/api/v1/health` is 401, no CF probe headers in this shell. Did not sqlite-edit (no event). Did not mutate `wakatime` env. Needs Configure UI or a working admin token. |
20+
| H rebuild 404 | **PASS** (same-day recycle, no second rebuild) — new handler start at `17:37:21Z` (line 42 = this binary). `/health` 200 after. New session ids + Case E pin within 3s. 404s are expected-noise (not warned), so no 404 warn lines. Did not `pkill` / rebuild again (would 404 this chat). |
21+
22+
---
23+
24+
## Answer first
25+
26+
Same rules as the parent playbook. **Do not** call `mcpmux_set_workspace_root` to "fix" a reconnect. **Do not** `pkill -f mcpmux`. **Do not** rebuild mid-case unless you are *in* Case H, and Case H goes last because it 404s this chat.
27+
28+
| Action | Use for this test? |
29+
| --- | --- |
30+
| Same Cursor chat, same workspace (`/Users/joe/Desktop/Repos/Personal/mcp-mux`) | **Yes — required** |
31+
| Reload MCP before starting (only if `/mcp` is already 404) | Yes, once, then stop |
32+
| `set_workspace_root` as a reconnect | **No** |
33+
| `set_workspace_root` only if Case E shows 6-way `bindable` *and* the workspace header is empty | Last resort, note it as a confounder |
34+
| `pnpm dev:stop` / rebuild | **Case H only**, last |
35+
| `pkill -f mcpmux` | **No** |
36+
37+
---
38+
39+
## Snapshot
40+
41+
```
42+
DB=~/Library/Application\ Support/com.mcpmux.desktop/mcpmux.db
43+
LOG=~/Library/Application\ Support/com.mcpmux.desktop/logs/mcpmux.$(date +%Y-%m-%d).log
44+
```
45+
46+
| Fact | Value |
47+
| ---- | ----- |
48+
| Space | `00000000-0000-0000-0000-000000000001` |
49+
| This root | `/Users/joe/Desktop/Repos/Personal/mcp-mux` |
50+
| Stdio for F | `wakatime` / `wakatime_wakatime_summaries` |
51+
| Config-save target for G | enabled stdio or HTTP that is `ready` — prefer a harmless `env_overrides` / `extra_headers` bump, revert after |
52+
| Pool stats | in-memory only |
53+
54+
Health: `curl -sf http://127.0.0.1:45818/health``{"status":"ok",…}`.
55+
56+
---
57+
58+
## Preconditions
59+
60+
1. Debug gateway on `:45818`. This binary must include `dcc2977` (or later on `root-resolution`). If `tauri dev` has not rebuilt since that commit, Cases G/E will exercise the *old* evict-only / pin-skip code — mark INCONCLUSIVE and rebuild *before* starting, not mid-case.
61+
2. Mux only via `user-mcpmux`.
62+
3. `mcpmux_list_servers` is not all-`bindable`. If it is, Case E failed (or header is empty). Do not paper over it with `set_workspace_root` until E is scored.
63+
4. Grep guard for the deleted landmine (any time):
64+
65+
```bash
66+
rg "Unexpected STDIO transport for OAuth reconnection" \
67+
crates/mcpmux-gateway/src "$LOG"
68+
```
69+
70+
Must be **zero** hits in source. Historical log lines from before `dcc2977` do not fail this run.
71+
72+
---
73+
74+
## Case E — header pin, no `set_workspace_root` (~2 min)
75+
76+
**Goal:** a non-empty `X-Mcpmux-Workspace` makes this repo bound. A 6-folder `roots/list` is not a binding.
77+
78+
1. `mcpmux_list_servers`.
79+
2. Grep the last minute of `$LOG`:
80+
81+
```bash
82+
rg "SessionRoots|X-Mcpmux-Workspace|pinned explicit workspace|held until mcp-session-id|PendingRoots|multiple roots reported" "$LOG" | tail -n 80
83+
```
84+
85+
**Pass:** at least one server is `ready` (or `bound` with a real block, not "unpinned roots"). Log shows a pin (`pinned explicit workspace root` or `held until mcp-session-id` then pin), not `multiple roots reported, no pinned header — PendingRoots`.
86+
**Fail:** every server `bindable` *and* the log shows a non-empty workspace header that was skipped (`pin skipped` / `present without mcp-session-id` without a later pin).
87+
**Inconclusive:** header is empty (`present but empty`) — Agents-window hole, out of scope. Note it.
88+
89+
Do **not** call `set_workspace_root` before scoring.
90+
91+
---
92+
93+
## Case F — stdio OAuth refuse + transport-closed still reconnects (~5 min)
94+
95+
**Goal:** killing a stdio child still uses `reconnect_fresh`. It must not go through `reconnect_after_oauth` / HTTP.
96+
97+
1. `mcpmux_search_tools({ "server_id": "wakatime", "mode": "browse", "limit": 5 })`. Need `server_readiness: ready`.
98+
2. Invoke `wakatime` / `wakatime_wakatime_summaries` with explicit `start`/`end` for today. Warm the instance.
99+
3. Note `date -u +%H:%M:%S`.
100+
4. Kill **only the child**:
101+
102+
```bash
103+
pgrep -lf wakatime
104+
# kill <child-pid>
105+
# Do not kill target/debug/mcpmux.
106+
```
107+
108+
5. Same chat, same invoke. No `set_workspace_root`.
109+
6. Grep from step 3:
110+
111+
```bash
112+
rg "backend call_tool failed|reconnect attempted after call_tool failure|reconnect_fresh completed|reconnect_after_oauth|stdio cannot reconnect via OAuth|Unexpected STDIO" "$LOG"
113+
```
114+
115+
**Pass:** invoke succeeds (or a normal tool error). Logs: `trigger=transport_closed`, `reconnect_fresh completed ok=true`. No `Unexpected STDIO`, no OAuth reconnect for this kill.
116+
**Fail:** raw `Transport closed` / `-32000` to the agent, or HTTP OAuth reconnect for stdio.
117+
118+
---
119+
120+
## Case G — config save reconnects (~5 min)
121+
122+
**Goal:** `ServerConfigUpdated` runs `reconnect_fresh` for an enabled server. Next invoke does not say "bound but not connected." Do **not** click Retry Connection.
123+
124+
Harmless write (revert after):
125+
126+
1. Snapshot the row (`extra_headers` / `env_overrides`) for the target server.
127+
2. `PUT` admin `save_server_inputs` (or Configure save in the UI) adding one dummy key, e.g. `MCPMUX_TEST_PIN=1` in `env_overrides`. Do not change command/url.
128+
3. Do **not** call retry/reconnect from the UI.
129+
4. Invoke a cheap tool on that server.
130+
5. Grep:
131+
132+
```bash
133+
rg "ServerConfigHandler|reconnect_fresh after config update|re-resolve failed, evicting only" "$LOG"
134+
```
135+
136+
**Pass:** `[ServerConfigHandler] reconnect_fresh after config update` with `ok=true` (or `ok=false` if the server cannot spawn, but the stale instance is gone and the error is a connect failure, not "bound but not connected"). Invoke works if the server is healthy.
137+
**Fail:** invoke returns "bound but not connected" / `diagnose_server`, or the handler only logs evict with no `reconnect_fresh`.
138+
6. Revert the dummy key the same way.
139+
140+
If admin write is unavailable (no token / desktop-only), use the Configure UI. If neither is possible, **BLOCKED** — do not sqlite-edit the row (no event).
141+
142+
---
143+
144+
## Case H — rebuild 404, last (~5 min + MCP reload)
145+
146+
**Goal:** process death → inbound 404 is expected. `/health` is the liveness check. Recovery is Reload MCP. Do not persist sessions.
147+
148+
This **will** 404 the current Cursor MCP session. Run it last.
149+
150+
1. `curl -sf http://127.0.0.1:45818/health` → ok.
151+
2. Trigger a Rust rebuild (`touch` a gateway `.rs` or wait for `tauri dev` to recycle). Do **not** `pkill -f mcpmux`.
152+
3. While the new process is up: `/health` is 200. A `POST /mcp` with the old `Mcp-Session-Id` is 404. Logs do not `warn` the 404 (`expected session noise`).
153+
4. Reload MCP **once**.
154+
5. Re-run Case E (list_servers). Expect bound, not 6-way `bindable`, without `set_workspace_root` if the header is a real path.
155+
156+
**Pass:** health 200 throughout the new process; chats 404 until Reload MCP; after reload, Case E still passes.
157+
**Fail:** `/health` down and we treat 404 as "gateway dead"; or after reload + real header, everything is `bindable` (Case E fail).
158+
**Skip if:** you cannot afford to 404 this chat. Score from an earlier same-day rebuild in `$LOG` if the lines are unambiguous; otherwise SKIPPED.
159+
160+
---
161+
162+
## Do not
163+
164+
- `set_workspace_root` to recover a closed backend.
165+
- `pkill -f mcpmux`.
166+
- Rebuild except Case H.
167+
- Leave the Case G dummy env/header in place.
168+
- Widen the transport-closed matcher if F fails with a new string — log `trigger=` and stop.
169+
170+
---
171+
172+
## Report back
173+
174+
```
175+
Case E header pin: PASS | FAIL | INCONCLUSIVE
176+
list_servers: ready/bound/all-bindable
177+
pin log:
178+
set_workspace_root used: no
179+
180+
Case F stdio refuse: PASS | FAIL | INCONCLUSIVE | SKIPPED
181+
server_id / tool:
182+
first invoke:
183+
child pid:
184+
second invoke:
185+
trigger=:
186+
reconnect_fresh ok=:
187+
Unexpected STDIO / OAuth: yes/no
188+
189+
Case G config save: PASS | FAIL | BLOCKED
190+
server_id:
191+
write path (admin/UI):
192+
handler log:
193+
invoke after save:
194+
reverted: yes/no
195+
196+
Case H rebuild 404: PASS | FAIL | SKIPPED
197+
health during/after:
198+
404 observed:
199+
reload MCP:
200+
Case E after reload:
201+
```

0 commit comments

Comments
 (0)