AgentDeck npm v1.0.24

AgentDeck npm v1.0.24

1.0.23 shipped the flashers. A hardware run over the fleet then found two
defects in them, and both had the same shape: a write that fully succeeded
reported as a failure, or was refused before it started. Neither could be seen
from a green test suite, because both live in the wire between the host and a
board.

The board now reboots into the firmware you just wrote

1.0.23's own notes called a real per-board reset open work. This closes it.

esptool-js's non-USB-OTG hard reset is, in full, sleep(100); setRTS(false) — a
release with no assert. That works only if whatever ran before left EN
asserted: --before default-reset does, and a finished writeFlash does not. So
every flash left the chip parked in the flasher stub. The firmware never ran, the
post-write device_info read-back could not succeed however long it waited, and
the CLI told users to power-cycle a board that had been written perfectly.

Both flashers now drive the reset themselves, from one sequence in the board
SSOT rather than one per flasher. Measured 2026-08-23 across three boards
spanning both of esptool-js's adapter strategies and the no-stub path:

sequence	T-Display-S3-Pro (native USB)	86 Box (CH340)	TTGO T-Display (CH340)
after(hard_reset)	parked	parked	parked
R1 W250 R0	booted	parked	—
D0 R1 W100 R0	booted	booted	booted

The D0 is load-bearing and only the CH340 boards reveal it: pulsing EN without
first driving IO0 high leaves them in download mode, which from outside is
indistinguishable from a board that failed to boot.

TTGO T-Display can actually be flashed

1.0.23 offered it in the browser picker. It could not be written there, or
anywhere: it was pinned --no-stub, the ROM loader has no compressed flash
mode, and the write died at Failed to enter compressed flash mode — before a
single byte landed. Making compression conditional moved the failure exactly one
step, to Failed to enter Flash download mode.

The evidence recorded for that board only ever claimed that it connects.
Connecting is not writing, and a board's verification now has to name a
completed write. With the stub it writes and boots end to end — 2.6 MB in 138
seconds, MD5 verified against the chip, and the board came up reporting its own
version. It also fixes the size guard as a side effect: stubless, this board's
flash id reads 0xffffff and no size can be checked at all.

esp32/platformio.ini keeps its --no-stub, deliberately. That pin drives
esptool.py, a different tool, and changing it on evidence gathered from
esptool-js would be the guess this whole round exists to stop making.

Still open

Stubless writing itself. Every --no-stub board measured so far fails before
writing a byte, and that is its own investigation. No board offered in the
browser is stubless now, so this affects only the CLI-only boards.

The browser flasher at /flash/
deploys from master and already carries both fixes; this release is the
terminal half.

The CLI, daemon, and Claude Code / Codex / OpenCode hook installers.
Optional — the macOS App Store app runs standalone without any of this.

Install

npx @agentdeck/setup

Packages

Package	Purpose
@agentdeck/setup	One-command installer
@agentdeck/bridge	agentdeck CLI — daemon hub + session bridge
@agentdeck/hooks	Agent lifecycle hook installers
@agentdeck/shared	Shared protocol types and layout engines

Full changelog: https://github.com/puritysb/AgentDeck/blob/npm-v1.0.24/CHANGELOG.md