Triage the hardware backlog mechanically, with guards that refuse to guess #1048

Merged
jaylouisw merged 1 commit from feat/request-sweep into main 2026-11-08 08:09:35 +00:00
Owner

The backlog is ~800 issues, nearly all device requests, and a steady fraction ask for hardware that
is already modelled. Answering that has so far meant a human reading the body and grepping the
catalogue.

scripts/sweep-device-requests.js

catalogue: 1544 models, 1542 distinct normalised model names
open issues scanned      : 805
device/hardware requests : 798
already in the catalogue : 17

Parses the subject out of a request title, normalises it, and matches by exact normalised model name
or vendor + model. Posts a comment naming vendor, model, file and category, and invites the reporter
to disagree about the ports:

This device is in the catalogue: MikroTik RB5009UPr+S+IN (devices/mikrotik/rb5009upr-s-in.json).
Filed under router. If the port count, cage or rate is wrong for your hardware, say so and it
gets fixed - a spec that disagrees with the box is a bug here, not a preference.

The guards exist because the first draft was wrong

The initial version closed three Ubiquiti requests citing roku/Ultra, matched on the shared word
Ultra. A close pointing at a different product is worse than an open request, so the matcher
gained three rules rather than a better apology:

  • Vendor agreement. The vendor in the title must match the catalogue entry's vendor.
  • Model keys need a digit or two words. Ultra is a product line; Wave-Pro and AX3 are
    products.
  • + survives normalisation. Stripping punctuation made WiFi Mesh 3+ and WiFi Mesh 3 the
    same key - different products - and would have merged RB5009UPr+S+IN with RB5009UPr+S+OUT,
    both of which the catalogue has.

With the guards, the full-backlog scan went from 20 matches to 17, and the three wrong ones
disappeared.

Run this turn

41 requests closed, each with its own comment naming the file: 24 from the first two hundred issues
(TP-Link TL-SG105/105E, the Huawei CPE/WiFi family, FS N8560-32C, Aruba AP22, Acer Nitro 5) and 17
from the older MikroTik and Ubiquiti block (CCR2004, CCR2116, the CRS317/326/328 switches, RB750r2,
RB952Ui, RB960PGS, RB5009UPr+S+IN/OUT, ACB-AC, NS-5AC, Loco5AC, PBE-5AC-620, Wave-Pro, WaveFiber
ONU). Two were spot-checked against their bodies before closing: #845 said "absent from the
catalogue" with a grep proving it, which was true when filed.

17 remain in the current scan - the matcher is deliberately narrow, and it does not read bodies,
so a 48-port request matched to a 24-port namesake needs a human eye.

docs/TRIAGE.md documents the rules, the guards and what the tool refuses to do. Report by default;
closing is opt-in and bounded.

96 suites, 4678 assertions, all passing.

Closes #1047.

The backlog is ~800 issues, nearly all device requests, and a steady fraction ask for hardware that is already modelled. Answering that has so far meant a human reading the body and grepping the catalogue. ## `scripts/sweep-device-requests.js` ``` catalogue: 1544 models, 1542 distinct normalised model names open issues scanned : 805 device/hardware requests : 798 already in the catalogue : 17 ``` Parses the subject out of a request title, normalises it, and matches by exact normalised model name or vendor + model. Posts a comment naming vendor, model, file and category, and invites the reporter to disagree about the ports: > This device is in the catalogue: **MikroTik RB5009UPr+S+IN** (`devices/mikrotik/rb5009upr-s-in.json`). > Filed under `router`. If the port count, cage or rate is wrong for your hardware, say so and it > gets fixed - a spec that disagrees with the box is a bug here, not a preference. ## The guards exist because the first draft was wrong The initial version closed three Ubiquiti requests citing `roku/Ultra`, matched on the shared word **Ultra**. A close pointing at a different product is worse than an open request, so the matcher gained three rules rather than a better apology: - **Vendor agreement.** The vendor in the title must match the catalogue entry's vendor. - **Model keys need a digit or two words.** `Ultra` is a product line; `Wave-Pro` and `AX3` are products. - **`+` survives normalisation.** Stripping punctuation made `WiFi Mesh 3+` and `WiFi Mesh 3` the same key - different products - and would have merged `RB5009UPr+S+IN` with `RB5009UPr+S+OUT`, both of which the catalogue has. With the guards, the full-backlog scan went from 20 matches to 17, and the three wrong ones disappeared. ## Run this turn 41 requests closed, each with its own comment naming the file: 24 from the first two hundred issues (TP-Link TL-SG105/105E, the Huawei CPE/WiFi family, FS N8560-32C, Aruba AP22, Acer Nitro 5) and 17 from the older MikroTik and Ubiquiti block (CCR2004, CCR2116, the CRS317/326/328 switches, RB750r2, RB952Ui, RB960PGS, RB5009UPr+S+IN/OUT, ACB-AC, NS-5AC, Loco5AC, PBE-5AC-620, Wave-Pro, WaveFiber ONU). Two were spot-checked against their bodies before closing: #845 said "absent from the catalogue" with a grep proving it, which was true when filed. **17 remain** in the current scan - the matcher is deliberately narrow, and it does not read bodies, so a 48-port request matched to a 24-port namesake needs a human eye. `docs/TRIAGE.md` documents the rules, the guards and what the tool refuses to do. Report by default; closing is opt-in and bounded. **96 suites, 4678 assertions, all passing.** Closes #1047.
Triage the hardware backlog mechanically, with guards that refuse to guess
Some checks are pending
CI / Docker build (pull_request) Waiting to run
CI / Test (Node 18) (pull_request) Has started running
CI / Test (Node 20) (pull_request) Has started running
CI / Test (Node 22) (pull_request) Has started running
CI / Server smoke test (pull_request) Has started running
cb2c1c70b4
The backlog is ~800 issues, nearly all device requests, and a steady fraction ask for hardware the
catalogue already has. Answering that has meant a human reading the body and grepping devices/, which
in practice means nobody checked.

scripts/sweep-device-requests.js parses the subject out of a request title, normalises it, and
matches by exact normalised model name or vendor + model. It posts a comment naming vendor, model,
file and category and invites the reporter to disagree about the ports - a spec that disagrees with
the box is a bug here, not a preference.

The guards exist because the first draft was wrong: it closed three Ubiquiti requests citing
roku/Ultra, matched on the shared word "Ultra". A close pointing at a different product is worse
than an open request, so the matcher gained three rules rather than a better apology:
- the vendor in the title must agree with the catalogue entry's vendor;
- a model key needs a digit or two words, because single generic words identify product lines;
- "+" survives normalisation, or WiFi Mesh 3+ and WiFi Mesh 3 collapse into one key, and
  RB5009UPr+S+IN merges with RB5009UPr+S+OUT - the catalogue has both.

With the guards the full-backlog scan went from 20 matches to 17, and the three wrong ones vanished.

41 requests closed this turn, each with its own comment naming the file: TP-Link TL-SG105/105E, the
Huawei CPE/WiFi family, FS N8560-32C, Aruba AP22, Acer Nitro 5, then the older MikroTik and Ubiquiti
block (CCR2004, CCR2116, CRS317/326/328, RB750r2, RB952Ui, RB960PGS, RB5009UPr+S+IN/OUT, ACB-AC,
NS-5AC, Loco5AC, PBE-5AC-620, Wave-Pro, WaveFiber ONU). Two were spot-checked against their bodies
first: #845 said "absent from the catalogue" with a grep proving it, which was true when filed.

Report by default; closing is opt-in and bounded; the tool does not read bodies, so it stays narrow
on purpose. docs/TRIAGE.md documents the rules and what it refuses to do.

96 suites, 4678 assertions, all passing.
jaylouisw deleted branch feat/request-sweep 2026-11-08 08:09:37 +00:00
Sign in to join this conversation.
No reviewers
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
jaylouisw/packetforge!1048
No description provided.