Mesh-join prompt: three copies on the laptop, a decision does not dismiss the others, the second click is refused as 'no resolveNonce' (toast tag too long, decided prompt replayed)

by adom

Adom Bridge unleashes your AI onto your full PC: full power, full safety. The AI breaks out of the container it runs in and onto your real machine, managing and running your entire laptop as you: launch apps, move files, drive any window, control KiCad, Fusion 360 or your real signed-in browser. Works with any AI, cloud or local (Claude Desktop, Claude Code, Codex), no Hydrogen required. Bridge and its bridges are an operating system for AI, with managed Node and Python runtimes, and a human-onl

Overview Files Skills Discover Packages Releases Dependencies Issues20 Breadcrumbs Pull Requests1 Activity

Closed bug report

John Lauer · 1mo ago ·closed by John Lauer

Reported by John on AdomLapper (2.1.137, 2026-09-20 21:05Z): a mesh-join prompt ("Trust this machine", winvm rejoining with a new node identity, origin ConfRoomROG) showed up as several copies on the laptop; clicking Trust on one did not take the others down, and clicking the next one produced a Windows toast "Could not decide it on ConfRoomROG: Refused: this decision did not carry the one-shot resolveNonce the origin stamped into the mirrored prompt, so it cannot have come from a human reading the mirrored prompt".

From the two boxes' logs, three bugs:

Toast tag too long. The mirrored card's toast tag is remote:ConfRoomROG:mesh-join:<64-hex node id> (92 chars); Windows caps a toast tag at 64. The first toast spawn logged Exception setting "Tag": The size of the notification tag is too large but still SHOWED the toast (untagged, so no dismiss can ever find it), then ab's retry showed a second, tagged copy. Two toasts for one prompt; only one can be dismissed.
A decided prompt is replayed. One second after ConfRoomROG accepted the Trust, its replay_to_peer(AdomLapper) ("link formed after it was raised", #197) re-sent the same pending prompt to the laptop, so a fresh card appeared for a prompt that was already decided.
A stale duplicate click is reported as a refusal. The second click hit the consumed one-shot nonce (correct security-wise) but the human is told "Could not decide it" with the nonce prose. When the id already has a recorded decision, the right answer is "already decided: trusted on AdomLapper at 21:06:07" and the stale card/toast on the clicking box should simply go away.

Fix plan: (1) hash any toast tag over 64 chars deterministically (same mapping for show and dismiss) and make a Tag failure abort the show instead of showing untagged; (2) replay_to_peer skips prompts that already have a decision (and the decision path removes the prompt from the live registry before any fan-out or grant carry); (3) perm_mirror_resolve answers already_decided with the recorded decision/box/time (from permission_history, #191) for a consumed nonce, and the clicking box dismisses its copy instead of toasting a refusal. Also confirm a decision on ANY surface dismisses every copy of a mesh-join card on the same box (window card, toast, Hydrogen).

1 Reply

John Lauer · 1mo ago

Fixed in insiders 2.1.138 (all four boxes pick it up on their tick; arav-rog took it through the real staged-installer path in 15 s).

What was wrong, from the two logs

The mirrored card's toast tag remote:ConfRoomROG:mesh-join:<64-hex node id> is 92 chars; Windows caps a toast tag at 64. $toast.Tag threw, PowerShell's default is non-terminating, so Show still ran and the toast appeared untagged (no dismiss can ever find it); ab's retry then showed a second, also untagged, copy. Two toasts for one prompt.
One second after your Trust landed on ConfRoomROG, its replay_to_peer(AdomLapper) ("link formed after the prompt was raised") re-sent the same prompt from a snapshot taken before the decision, so a fresh card appeared for a prompt that was already decided.
Your second click hit the consumed one-shot nonce (correct security-wise), but the answer was the nonce lecture as a "Could not decide it" toast instead of "already decided".

What changed

toast_tag_fit: any tag over 64 chars keeps its readable prefix and gets a stable 16-hex hash tail; applied on show AND on dismiss so they agree. The toast script now starts with $ErrorActionPreference = 'Stop', so a bad Tag aborts before Show instead of showing an untagged toast.
replay_to_peer re-checks the live registry at send time; a prompt decided meanwhile is not replayed.
perm_mirror_resolve on a consumed nonce for a prompt that has a recorded decision (from permission_history, #191) answers errorCode: already_decided {decision, decidedOn, decidedAt} and re-fans the dismiss so the clicker's leftover copies drop; the clicking box shows no warning toast for that case. An unknown or still-live id is still refused as before.

Proof (arav-rog on 2.1.138)

(a) notify_user with a 93-char id -> ab log:
    Toast PowerShell process spawned (tag=remote:ConfRoomROG:mesh-join:ababababababababab~6d44abe108ab2162)
    no "Toast child FAILED", no "Toast RETRY" (the 21:05Z lines just above it in the same log show the old failure + retry)

(b) perm_mirror_resolve {id: <a decided prompt>, resolveNonce: "bogus-duplicate-click"}:
    {"errorCode":"already_decided","decision":"timeout","decidedOn":"arav-rog","decidedAt":"2026-09-19T19:35:38.590Z",
     "error":"Already decided: timeout on arav-rog at 2026-09-19T19:35:38.590Z"}
    perm_mirror_resolve {id: "never-existed-207", ...} -> still refused (no decision on record)

unit test: toast_tag_tests::long_tags_fit_and_stay_stable (<=64 chars, prefix kept, deterministic) passes

One cosmetic leftover I will fold into the next build: because that reply's error text contains the decision word ("timeout"), the generic status stamp labels the reply status: timeout with the "raise timeoutSeconds" hint, which is wrong for already_decided; the reply will carry its own status so the stamp leaves it alone.

Your three clicks were the untagged toast, its untagged retry and the replayed card; on 2.1.138 there is one tagged toast, it goes away on the first decision anywhere, and nothing is replayed after a decision.

John Lauer closed this · 1mo ago