dsh-domain-watch
harness shell 中的状态徽标，以及四个代理工具：域名是否确实已在 REGISTRY 注册、记录中的注册商是谁、何时到期，以及那个告诉你一切正常的仪表板是否在说真话。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:samorachumbi/dsh-domain-watch#ac30b85683307ded325deec682e042c651ce8d44Domain Watch 以 dsh-domain-watch 发布，当前版本为 0.2.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Ask the registry, not the vendor's dashboard. A registrar's control panel is a cache of the registry, and this plugin answers the only question that matters when money has changed hands: is the domain actually mine?
curl 'http://127.0.0.1:3080/domain-watch/check?domain=example.co.ke&claimed=paid&json=1'
On 1 October 2026 a hosting invoice was marked PAID. The order showed active in the vendor's dashboard. There was a receipt, an invoice number, and a customer who believed they owned a domain.
KENIC's registry had no record of it at all.
The panel was not lying, exactly. It was a cache, and the cache had not been written to the registry. Days went by waiting on a queue that was never going to complete, an escalation clock ran, and a refund had to be requested — for a domain that could have been registered correctly in the first ten minutes.
Three days later the same thing happened in the other direction: the panel said Pending for a
domain the registry had already made active, and it took a registry query to know that nothing was
wrong.
A cache is wrong in both directions. The fix is not a better dashboard. The fix is asking the authority, and refusing to average the two answers into a comfortable middle.
The domain board, dark theme The same board, light theme
Four agent tools and one host-rendered page.
| Tool | Answers |
|---|---|
domain_check | Is it registered? Registrar of record, created, expiry, days remaining, transfer lock — and, if you pass claimed, whether the panel and the registry disagree |
domain_watch_add | Start watching: one registry query, then remember the answer |
domain_watch_remove | Stop watching. Removes the record only; never touches the domain |
domain_watch_list | Every watched domain, soonest to expire first, countdown recomputed live |
The board is at /domain-watch/ui — a server-rendered page in its own tab, no client bundle, no
JavaScript required. Countdowns are computed per request from the remembered registry answers, so the
numbers move even when nobody clicks anything. That is the point: the failure being guarded against
is time passing.
The pill in the shell: bottom-left beside the apps pill, with the cost bubble at bottom-right
A draggable bubble in the harness shell, beside the apps pill. It exists because the board answers what is happening while the pill answers is it on, and is it working — a different question, asked far more often, by someone who is not going to open a tab to ask it.
Always visible, on purpose: a thing that is always present makes its own absence meaningful, and an indicator that comes and goes can never tell you it has broken.
| State | Reads | What a non-developer does with it |
|---|---|---|
| calm | domains fine ✓ | nothing |
| invoked | domains checking itikia.co.ke… | sees that it heard, before the answer arrives |
| just finished | domains itikia.co.ke ✓ | nothing |
| expiring | domains 2 expiring | look, soon |
| expired | domains needs renewing | renew today |
| not registered | domains not registered | act now — you do not own this |
| unreachable | domains ? | cannot check right now |
The chip just after a successful check: the domain, then a tick
Three rules make it trustworthy, and all three are about the person reading it rather than the code:
It can say "I don't know." A failed poll shows ?, never a tick. A status light that lies would be
this plugin's own thesis broken by the thing built to enforce it.
Colour changes on invocation, not on completion. The worst thing an assistant can do is go quiet — so the acknowledgement cannot wait for the result. This one is enforced with a floor rather than hoped for; see below for why that distinction cost us a day.
Every alarm carries a verb. A developer sees red and checks the logs. Everyone else sees red and needs to be told what to do, or the light is just anxiety.
The level, the words and that sentence are all decided host-side by summarize(), and the board
header paints the same object — so the bubble and the page can never disagree.
Clicking the pill re-checks every watched domain and acknowledges the click at once. It does not
open the board: a status light's click should do the cheap, frequent thing and answer in the shell.
The board is one click away in the apps pill, or at /domain-watch/ui.
The acknowledgement is held for a minimum of 1.2 s (ACK_MS), and that floor is the whole point.
A .ke whois answers in 24–74 ms while the pill polls once a second, so the first version
rendered its blue checking… in roughly one invocation in thirteen — and, because a click opened
a tab rather than checking anything, in practice it never rendered at all. Rule 1 above was a claim
the code did not keep.
The floor is a bound, not a delay: when it elapses the truth takes over whatever it is, so a slow or failing check can never hide behind an acknowledgement. In a plugin whose entire thesis is that a status light must not lie, that bound is the only reason the floor is allowed to exist.
None of this is covered by "it looked fine" — the decision was hoisted out of the React function
body into a pure pillState() precisely so it could be tested, and reintroducing each defect fails
the suite.
Motion is a slow breath (~2.5s), never a blink: a permanently blinking element becomes wallpaper
within a week and then hides the morning it actually mattered. It stops entirely under
prefers-reduced-motion, and the breath lives in the chip's tint rather than the pill's overall
opacity — fading the whole pill faded its words to 3.09:1, under the 4.5:1 floor. Fading the tint is
a pulse; fading the text is an unreadable status light. Measured after the change: 5.21:1.
domain_check does not just look things up. Given what a panel told you, it names the disagreement:
| Verdict | Meaning |
|---|---|
confirmed | Both agree. |
registry-ahead | The registry already has it; the panel is a lagging cache. Do not re-order. |
registry-silent | The panel says it is yours and the registry has no record. You do not own this name yet. |
in-progress | Neither shows it registered. The order has not landed. |
registry-only | No claim was given, so this is the registry answer alone. |
A failed query is never reported as "not registered". A network error and a missing domain are indistinguishable to a caller that only sees an empty string, and conflating them is how you pay twice for a domain you already own. Errors are reported as errors.
An unknown TLD is an error, not a guess. Verisign serves .com and .net; it does not serve
.org, .io or .dev. Ask it about one of those and it answers "No match" — which reads as
available and is really you asked the wrong server. A registry that is not in the map returns an
explicit error, because a wrong answer is worse than no answer: the wrong answer is the one you act
on. Add TLDs from the IANA root database; never infer one from a similar TLD.
A put merges. Re-checking a domain must not erase your note, and adding a note must not erase
the registry answer. Every write reads-then-merges through one helper, so undefined means leave
it and never clear it.
DEST="$DSH_HOME/profiles/web/node_modules/dsh-domain-watch"
mkdir -p "$DEST"
cp -r lib package.json cordis.patch.yml "$DEST"/
# $DSH_HOME/profiles/web/cordis.patch.yml (append)
- insert:
- id: domain-watch
name: 'dsh-domain-watch'
config:
domains:
- example.co.ke
# Registry policy, not universal law. These defaults are KENIC's.
suspendAfterDays: 14
deleteAfterDays: 90
warnWithinDays: 60
Then restart the host so the new entry is composed. @deepseek-ai/* imports resolve through the
shared profile tree, so nothing needs installing.
Do not use
dsh plugin addfor this. It forwards to a package manager inside the profile directory, which materialises a second physical copy of@deepseek-ai/dsh-tools— and two copies mean two module-levelSymbols, so every tool call in the session dies withCannot read properties of undefined (reading 'prepare'). Copying real files avoids it entirely.
| Key | Default | Meaning |
|---|---|---|
domains | [] | Domains to seed into the watch list. Seeded without a network query, so a slow registry can never hold up a boot |
suspendAfterDays | 14 | Days after expiry before the name stops resolving |
deleteAfterDays | 90 | Days after expiry before anyone else may register it |
warnWithinDays | 60 | When a domain starts showing as expiring |
timeoutMs | 15000 | Per-query whois timeout |
The grace windows are configurable because they are registry policy, not universal law. The
defaults are KENIC's, from the .ke Third Level Policy: suspended 14 days after expiry, deleted 90
days after. A .com behaves differently, and a plugin that hard-coded one registry's rules while
claiming to answer for any TLD would be lying quietly.
npm install
npm run typecheck # tsc, strict
npm test # 38 tests, no network — real registry replies as fixtures
npm run build
node scripts/probe.mjs example.co.ke # LIVE: a real registry round trip
The test suite runs against verbatim registry replies captured on 2026-10-01 — one for a name
that exists, one for a name that does not, one carrying a transfer lock. With no network in the
suite, those fixtures are the registry, which makes them the most important file in tests/.
The probe is the other half: recorded replies prove the parsing, and only a live query proves the transport, the registry map, and the round trip.
Built by Samora Chumbi — @samorachumbi — for a harness that had to stop believing vendor dashboards. I build DSH plugins; this is the fourth.
To reach me — about a bug, a registry this map gets wrong, or a plugin you wish existed — open an issue: https://github.com/samorachumbi/dsh-domain-watch/issues. If the map is wrong for a TLD you actually use, that is the most useful bug you can file: adding a TLD from the IANA root database is a small change, and it makes the plugin correct for the next person too.
Other plugins from the same workshop:
omdsh-dev/dsh-daily-progress).MIT.