feat: ship the Hermes Agent mark, register WorkBuddy's generic one

[yujiezhang-ops]

Addresses the actionable part of #133. One of the four got a real mark; three did not, and I did not invent one for them.

Hermes Agent — shipped

lobe-icons publishes a Hermes Agent mark under the same MIT licence the four existing vendor marks come from. I verified the LICENSE file hashes identical to the copy already in this repository, so it is the same audited basis rather than a new one to evaluate.

It is byte-identical to lobe-icons' nousresearch.svg but for the <title>. That is correct, not a mix-up: Hermes Agent is Nous Research's product, so the company mark identifies it. Recorded in the manifest entry so nobody later reads the reuse as carelessness.

Verified in the running app: the row now renders data-mark-kind="asset" and inherits rgb(58,58,60) from its neighbours, so currentColor resolves against the page rather than painting black.

Aider, WorkBuddy — deliberately nothing

Neither has a mark in lobe-icons. Aider has no published brand kit, an open logo/branding issue from 2024, and an Apache-2.0 code licence that grants no trademark rights. WorkBuddy publishes no vector at all.

Copying either favicon would ship vendor artwork with no redistribution basis, which is exactly what asset-rights.json exists to prevent. So they keep generic symbols — and WorkBuddy is now a registered generic rather than reaching AgentIcon's unknown-Agent fallback. Both render the same glyph, so the two states were indistinguishable; "evaluated, nothing redistributable exists" now reads differently from "never looked at".

ChatGPT Desktop already had a mark in the app — that gap is site-only, tracked in #133.

Two things I nearly got wrong

The 19.7KB asset. Every existing mark is ≤1.7KB, and 3132 of Hermes' 4096 coordinates carry three decimals — 0.00075px at 18px. I was going to reduce the precision. That would have been a modification under MIT requiring a declared change, and it would have silently invalidated the recorded hash. Shipped byte-identical instead, so modified stays absent.

The measurement that almost misled me. I rasterised the mark and found 30% of pixels at 18px were mid-grey, and nearly concluded the asset was too detailed for the UI. Then I measured the existing codex mark as a control: 47% grey at the same size. Antialiasing at 18px is universal, not a defect in this asset. Without the control I'd have rejected a perfectly good mark.

New tests

ships each asset byte-identical to its recorded hash — nothing verified that claim before. I confirmed it catches the precision reduction I'd contemplated. It iterates the manifest rather than Agent ids, because chatgpt-desktop reuses codex's rights object and has no file of its own.

registers a mark for every Agent and desktop Agent — so a newly added Agent cannot quietly land on the fallback.

Note the hash test uses path.join, not new URL(..., import.meta.url): Vite statically rewrites that form into an asset reference, which replaced the interpolated filename with undefined at transform time. That cost me a few iterations.

Verification

330 frontend tests (43 files), go test ./..., 6 e2e, check-docs.py, and the release-compliance gate — which caught the new asset and required regenerating third_party/; --check now passes. NOTICE updated: six assets, five from lobe-icons unmodified, and the Aider/WorkBuddy reasoning stated.

Sources: lobe-icons · Aider · Aider logo issue

🤖 Generated with Claude Code