Measured on zion after v1.22.41 shipped. The code half of the tab-spam fix (#2778/#2779) is live; the guidance half is not, and cannot be — skill edits made in this repo have no delivery path to installed machines.
Evidence
The published package contains no skills at all:
$ ls $(npm root -g)/@phnx-labs/agents-cli
CHANGELOG.md LICENSE README.md dist node_modules package.json scripts
$ find $(npm root -g)/@phnx-labs/agents-cli -name SKILL.md -path '*browser*'
(nothing)
What agents actually read is a different repo:
$ git -C ~/.agents/.system remote get-url origin
git@github.com:phnx-labs/.agents-system.git
$ git -C ~/.agents/.system rev-list --count HEAD..origin/main
0 # fully up to date
$ git -C ~/.agents/.system show origin/main:skills/browser/SKILL.md | grep -c tab-spam
0 # the guidance is not there
$ stat -f %Sm ~/.agents/.system/skills/browser/SKILL.md
Jun 12 18:56:22 2026 # two months old; contains no "navigate"
Meanwhile the edit exists in this repo and shipped in the tag:
$ git grep -c tab-spam v1.22.41 -- skills/browser/SKILL.md
1
agents sync does not close the gap — it reported "8 commands, 3 skills" and the file's mtime and content were unchanged.
Why it matters
PR #2778's behavior change is split across two halves:
- code — Arc recognized,
createPageTarget refuses instead of crashing. Live: dist/bin/agents contains Arc.app (3x) and the Comet, Chrome error string.
- guidance — the browser skill teaching agents to show a doc with
agents browser navigate --url file://… (one reused tab) instead of a raw open (a new tab every call). Not delivered.
Half 2 is the half that actually stops the tab-spam, because it changes what agents do. Without it the user's browser keeps accumulating duplicates: measured on zion the same day, 58 Arc tabs with 16 agent-opened file:// docs, monetize-agents-cli.html open x3 and plan-deep-dive-agent.html x2.
So a fix can be merged, released, and installed fleet-wide, and still not change agent behavior at all — with nothing in the release output indicating the shortfall.
Suggested direction
Either mirror skills/ from this repo into .agents-system as part of the release (and fail the release when they diverge), or move the browser skill's canonical copy to whichever repo actually delivers it and stop accepting edits in the other. Any option is fine; silently maintaining two copies where only one is read is not.
Refs #2778, #2779, #2786.
Measured on zion after v1.22.41 shipped. The code half of the tab-spam fix (#2778/#2779) is live; the guidance half is not, and cannot be — skill edits made in this repo have no delivery path to installed machines.
Evidence
The published package contains no skills at all:
What agents actually read is a different repo:
Meanwhile the edit exists in this repo and shipped in the tag:
agents syncdoes not close the gap — it reported "8 commands, 3 skills" and the file's mtime and content were unchanged.Why it matters
PR #2778's behavior change is split across two halves:
createPageTargetrefuses instead of crashing. Live:dist/bin/agentscontainsArc.app(3x) and theComet, Chromeerror string.agents browser navigate --url file://…(one reused tab) instead of a rawopen(a new tab every call). Not delivered.Half 2 is the half that actually stops the tab-spam, because it changes what agents do. Without it the user's browser keeps accumulating duplicates: measured on zion the same day, 58 Arc tabs with 16 agent-opened
file://docs,monetize-agents-cli.htmlopen x3 andplan-deep-dive-agent.htmlx2.So a fix can be merged, released, and installed fleet-wide, and still not change agent behavior at all — with nothing in the release output indicating the shortfall.
Suggested direction
Either mirror
skills/from this repo into.agents-systemas part of the release (and fail the release when they diverge), or move the browser skill's canonical copy to whichever repo actually delivers it and stop accepting edits in the other. Any option is fine; silently maintaining two copies where only one is read is not.Refs #2778, #2779, #2786.