Skip to content

Releases

v0.7

Choose a tag to compare

@github-actions github-actions released this 14 Nov 00:03

Agent Deck 0.7

Renamed: agentdeck is now agtdeck. The package is agtdeck-0.7.fwsc, the daemon is host/agtdeck.py, and settings live in ~/.config/agtdeck (the daemon moves an existing ~/.config/agentdeck there on first start). The product is still called Agent Deck, and the USB name is still FM-1 AGTDECK. Earlier releases keep the file names they were published under.

Firmware: nothing new on the device. Package identity FM-1_24357; the NO HOST screen names the daemon's new file. Built and signed on GitHub from this tag.

Daemon 0.7

  • Unplugging is fine: the daemon waits for the FM-1 at start and after an unplug, then resends everything.
  • --doctor checks the MIDI port, config, backend and agent CLIs and says which piece fails.
  • --install-service starts the daemon at every login (launchd on macOS, systemd --user on Linux); --print-service shows what it would write.
  • pip install . gives an agtdeck command.
  • auto_only: an allow-list for auto-approve. auto_never still wins.
  • Output tokens in INFO and FLEET; RECENT PROMPTS under MENU → SEND PROMPT; a notice when the firmware is older than the daemon.
  • Fixed: a pipeline's next step could be typed into a permission dialog when the multiplexer briefly reported a busy agent as idle. Queued prompts now wait for a settled agent and, for Claude Code, a closed turn in its session log.
  • A Claude Code agent waiting on a prompt its screen does not show is now BLOCKED, and auto-approve never answers a prompt it cannot see.

Also new: a config editor in the browser, a container image (Dockerfile), daemon unit tests and a test workflow that also runs before every release build.

Not yet tested on hardware: this package has run in the emulator only, and reconnect after an unplug has not been tried on a real unit. The login service was generated and validated, not installed.

Verify

gh attestation verify agtdeck-0.7.fwsc --repo kevinelliott/fm-1-agent-deck \
  --signer-workflow zvenson/fm1-build/.github/workflows/build.yml

SHA-256: c0d0b43e688ca0f7f9911dce9ee0bdffb00b1127a66bc4629a934e087a97eacf

Full Changelog: v0.6.3...v0.7

Agent Deck 0.6.3

Choose a tag to compare

@github-actions github-actions released this 13 Nov 18:19

Firmware 0.6.3 for the M-VAVE FM-1, with the matching host daemon (protocol v6). The firmware is 0.6.2 with a new version number; what is new is how the package is made.

What's changed

  • Built and signed on GitHub. This release's agentdeck-0.6.3.fwsc was built by GitHub from the tagged commit, using The Firmware Hub's shared fm1-build workflow, and carries a build attestation. Check it with:

    gh attestation verify agentdeck-0.6.3.fwsc --repo kevinelliott/fm-1-agent-deck \
      --signer-workflow zvenson/fm1-build/.github/workflows/build.yml
  • The signature says the file is exactly what that commit builds. It does not say the firmware is safe.

  • The repository's fmhub.json now links the browser player (updated to this firmware) and a one-page cheat sheet.

Install

pip install mido python-rtmidi
python3 firmware-build/tools/fm1_install.py agentdeck-0.6.3.fwsc
python3 host/agentdeck.py --backend auto

Flashing custom firmware is at your own risk: the FM-1's flash is single-bank, and an interrupted write can leave the unit dark until rescued. Read the arrival-day section of the README before a first flash.

Status

  • The attestation verifies, and the same build command in a clean clone on the author's machine produced a byte-identical file.
  • This package has passed the fm1-emu session. It has not been installed on real hardware; the author's FM-1 runs 0.6.2, which differs only in the version text.
  • Everything in the 0.6 and 0.6.2 status notes still applies: installing from stock V15 or another custom firmware, leaving Agent Deck for another firmware, and an interrupted update are untested, and the queue, pipelines and auto-approve rules have been exercised against a stub backend only.

agentdeck-0.6.3.fwsc SHA-256: 015036f7858ace3e294504faa06f6b49a5029db57a00991a02fc0765fee40ef4

Agent Deck 0.6.2

Choose a tag to compare

@kevinelliott kevinelliott released this 13 Nov 18:10

Firmware 0.6.2 for the M-VAVE FM-1, with the matching host daemon (protocol v6). No new controls: this release gives Agent Deck an identity of its own.

What's changed

  • Package identity: FM-1_24356 (the scheme is FM-1_2435N, N the minor version). Releases up to 0.6.1 identified as FM-1_96N, numbers that other FM-1 firmwares already use.
  • USB name: the unit now reports itself as FM-1 AGTDECK instead of the platform's Felucca. The installer and the daemon in this repository find it under either name.
  • macOS remembers a MIDI device's name, so the port can still show as Felucca after updating. To refresh it: unplug the FM-1, delete the greyed-out device in Audio MIDI Setup → MIDI Studio, plug it back in.
  • The repository now has a root LICENSE (GPL-3.0), and fmhub.json with a card picture for The Firmware Hub. It has not been submitted there yet.

Use the installer from this release (or newer) once a unit runs 0.6.2.

Install

pip install mido python-rtmidi
python3 firmware-build/tools/fm1_install.py agentdeck-0.6.2.fwsc
python3 host/agentdeck.py --backend auto

Flashing custom firmware is at your own risk: the FM-1's flash is single-bank, and an interrupted write can leave the unit dark until rescued. Read the arrival-day section of the README before a first flash.

Status

  • This package is installed on the author's FM-1 (flashed over USB from 0.6.1); the unit answers with FM-1_24356, reports the USB product name FM-1 AGTDECK, and the daemon drives it over protocol v6. It was built with the JieLi toolchain and passed the fm1-emu session first.
  • Not yet tried: installing from stock V15 or from another custom firmware, leaving Agent Deck for another firmware, and an interrupted update.
  • Everything in the 0.6 status notes still applies: the queue, pipelines, auto-approve rules and knob combos have been exercised against a stub backend, not yet over long sessions with real agents.

agentdeck-0.6.2.fwsc SHA-256: 9a637a44f3e939d09e8de882cce9ad577364f94d46ae0cc8ed3057e9419aba99

Agent Deck 0.6.1

Choose a tag to compare

@kevinelliott kevinelliott released this 13 Nov 17:25

Firmware 0.6.1 for the M-VAVE FM-1. A one-fix release on top of 0.6; the host daemon and the protocol (v6) are unchanged.

What's fixed

  • Box view: the context gauge ran along the bottom edge of each tile, through the kind tag and the time in state. It now runs up the tile's right edge, as it does on a GRID row.

Also in the repository

  • The README has a tour of every screen: 19 captures of this package running in fm1-emu, a new animation and an updated hero image.
  • tools/emu_shots.py regenerates them, one emulator run per scene.

Install

pip install mido python-rtmidi
python3 firmware-build/tools/fm1_install.py agentdeck-0.6.1.fwsc
python3 host/agentdeck.py --backend auto

Flashing custom firmware is at your own risk: the FM-1's flash is single-bank, and an interrupted write can leave the unit dark until rescued. Read the arrival-day section of the README before a first flash.

Status

  • This package is installed on the author's FM-1 (flashed over USB) and the daemon drives it there over protocol v6. It was built with the JieLi toolchain and booted in fm1-emu first.
  • The fix has been checked in the emulator captures, not yet looked at on the hardware screen.
  • Everything in the 0.6 status notes still applies: the queue, pipelines, auto-approve rules and knob combos have been exercised against a stub backend, not yet over long sessions with real agents.

agentdeck-0.6.1.fwsc SHA-256: dea3f67fd520a7db6b4df2a883c1c09dbc1375809a31417969d2a7d0451a354b

Agent Deck 0.6

Choose a tag to compare

@kevinelliott kevinelliott released this 13 Nov 08:07

Firmware 0.6 for the M-VAVE FM-1, with the matching host daemon (protocol v6).

The screen is new in this release, so most of it needs the firmware; the queue, the auto-approve rules and the knob combos are in the daemon and work on older firmware too.

What's new

The screen

  • Colour means state everywhere: amber is waiting on you, green is finished, red is broken, white is working, grey is at rest.
  • Header: a logo whose nine pads are the first nine slots in their state colours, the page name with a page indicator, a host light, and a chip per state with its count.
  • GRID rows: a state-coloured edge, a state icon (a spinner while working, a blinking warning sign while blocked, a tick, a cross), a kind tag in the agent's colour, time in state, flags and the context gauge.
  • Boxes: press HOME again on the GRID to see the agents as tiles filled in their state colours instead of rows. The tiles resize with the deck, from four big ones down to all 27 keys of the page; the choice is remembered.
  • DETAIL: the state as a pill, a context bar, a tag for the INFO / DIFF views, and text coloured by what it is.
  • ABOUT: a fifth page (also MENU → ABOUT) with the firmware and protocol version, the device identity, the host connection, the daemon's version and backend, agents in use and uptime.
  • A power-on splash, an empty-deck screen that says what to do next, and ALL CLEAR on APPROVALS when nobody is waiting.

Prompt queue and pipelines

  • A prompt sent to a BLOCKED agent is queued and goes out once the agent is free (Q on its row). queue_when_working does the same for prompts sent mid-turn.
  • A preset can be a pipeline: {"label": ..., "steps": [...]}, one step per finished turn. A "test > review > commit" preset is built in.

Auto-approve rules

  • Before auto-approve answers a prompt, the daemon checks what the agent wants to run and the prompt on screen against auto_never (by default rm -r/-f, git push, git reset --hard, sudo, --force, drop table, mkfs/dd). On a match it leaves the prompt to you and sounds the alert.
  • This is a text match, not a sandbox: a risky command phrased another way gets through.

Knobs with a button held

  • PLAY + PRESETS: choose which agent PLAY starts.
  • SEQ + PRESETS: how many times repeat mode sends.
  • SEL + K1: scroll DETAIL a page at a time.
  • LFO + K4: step through the marked agents.

Also

  • Follow: the selection can jump to an agent when it blocks (MENU → SETTINGS).
  • Mute: silence one agent's sounds and alerts (MENU → THIS AGENT).
  • Changed: jump to the agent that has waited longest is now GLO again on APPROVALS; HOME again on the GRID is the list / boxes toggle.

Install

pip install mido python-rtmidi
python3 firmware-build/tools/fm1_install.py agentdeck-0.6.fwsc
python3 host/agentdeck.py --backend auto

Flashing custom firmware is at your own risk: the FM-1's flash is single-bank, and an interrupted write can leave the unit dark until rescued. Read the arrival-day section of the README before a first flash.

Status

  • This package is installed on the author's FM-1 and the daemon drives it there over protocol v6. It was built with the JieLi toolchain and booted in fm1-emu first, where all five pages and the box view render.
  • The individual screens, the smallest box sizes (more than 16 agents on a page) and the sounds have had emulator and harness time, not yet a careful pass on the hardware.
  • The queue, pipelines, auto-approve rules and knob combos have been exercised against a stub backend, not yet over long sessions with real agents.
  • The INFO view, gauges and blocked reasons read Claude Code's own session logs (read-only), an internal format that may change.

agentdeck-0.6.fwsc SHA-256: ab984d61b09d499b23fe06bae17dbd1fe52839157e3fbc6d911e17e398815a7d

Agent Deck 0.5

Choose a tag to compare

@kevinelliott kevinelliott released this 13 Nov 07:10

Firmware 0.5 for the M-VAVE FM-1, with the matching host daemon (protocol v5).

Most of the new controls are in the daemon and work on firmware 0.3 and 0.4 too; 0.5 itself adds the context gauges, the + marker and the new sounds.

What's new

Sounds (firmware 0.5)

  • Blocked alert: three rising calls, three times, about 1.7x as loud as the other chimes. On older firmware a blocked agent gets the attention chime three times in a row instead.
  • Feedback sounds: a faint tick for menu moves, preset arming, marks and view changes; a soft rising or falling blip when you approve / start / send or deny / stop.
  • A quieter "done", and a quiet double note when an agent's context is getting full.
  • MENU → SETTINGS: feedback sounds on/off, blocked alert loud/normal (ui_sounds, alert).

Groups

  • LFO + key: mark or unmark that agent. While any are marked, GLO + REC / FX / ENV / SEQ acts on the marked agents only; GLO + LFO clears the marks.
  • ARP + key: toggle auto-approve on that agent.

Seeing more

  • DIFF view: EDIT on DETAIL now cycles OUTPUT → INFO → DIFF: what the agent has changed since the last commit, per file, plus the last three commits.
  • Context gauges (firmware 0.5): a bar at the end of each Claude Code agent's row and under DETAIL's status line, with a one-time warning at ctx_warn_pct (default 80%).
  • MENU → FLEET: every agent on one screen with state, time in state and context size; pick a row to open it.
  • SEL held: when the prompt has numbered choices, exactly those keys light, and the footer names the one you pick.

Getting around

  • HOME again on the GRID: jump to the agent that has waited longest.
  • K1 on the MENU page: a page at a time.

Menu

  • THIS AGENT → COMMANDS: /compact, /clear, /context, /review, /status (configurable).
  • SEND PROMPT → AGAIN: resend what the agent was last asked.
  • NEW AGENT → IN A NEW WORKTREE: makes a git worktree next to the repository (<repo>-wt1 on branch deck/wt1) and starts the agent there.
  • SETTINGS → DESKTOP NOTIFY: optional desktop notifications (notify, off by default).

Install

pip install mido python-rtmidi
python3 firmware-build/tools/fm1_install.py agentdeck-0.5.fwsc
python3 host/agentdeck.py --backend auto

Flashing custom firmware is at your own risk: the FM-1's flash is single-bank, and an interrupted write can leave the unit dark until rescued. Read the arrival-day section of the README before a first flash.

Status

  • The 0.5 package is built with the JieLi toolchain and boots in fm1-emu, rendering all four pages. It has not yet been run on real hardware; 0.3 is what is installed on the author's FM-1, with the daemon from this release driving it.
  • The new sounds have been measured in the host harness recording, not yet heard through an FM-1.
  • The context gauge assumes a 200k-token window and treats a session seen above that as a 1M window; set ctx_window if that is wrong for you.
  • The INFO view, gauges and blocked reasons read Claude Code's own session logs (read-only), an internal format that may change.
  • Worktree spawn and desktop notifications have had unit-level checks only.

agentdeck-0.5.fwsc SHA-256: 31e9cea67c32cd917e614d5ffdf6e1546368fe2252e4556e974b6f7c9f392670

Agent Deck 0.4

Choose a tag to compare

@kevinelliott kevinelliott released this 13 Nov 06:38

Firmware 0.4 for the M-VAVE FM-1, with the matching host daemon (protocol v4).

Most of this release is in the daemon and works on firmware 0.3 too; 0.4 itself adds two LED cues and a row marker.

What's new

Know what an agent is doing

  • INFO view: press EDIT again on the DETAIL page to flip from the agent's output to a summary. For Claude Code agents: the session title, model, context size, permission mode, git branch and changed files, what it was last asked, what a blocked agent wants to run, todo progress and its last reply.
  • Blocked reason: when a Claude Code agent blocks, the footer names the tool call it is waiting on (api: Bash npm test).

Keep an agent going

  • Repeat mode: hold SEQ for about a second to re-send the armed prompt each time the agent finishes a turn. It stops after repeat_max sends (default 10); hold SEQ again to stop sooner.
  • Nag: optionally chime again every 2, 5 or 10 minutes while an agent stays blocked (MENU → SETTINGS, or nag_minutes).

More chords

  • SAVE + key: move the selected agent to that key (a swap if it is taken).
  • PLAY + key: start that agent, or spawn one on an empty key.
  • PLAY on a bare shell now starts the default agent in it.

Menu

  • THIS AGENT: view, repeat mode, auto-approve, focus on the computer, copy the last reply to the clipboard, open its folder.

Firmware 0.4 only

  • An R marker on rows in repeat mode, and SEQ lit while the selected agent repeats.
  • SAVE blinks when the prompt on screen offers a "don't ask again" option.

Install

pip install mido python-rtmidi
python3 firmware-build/tools/fm1_install.py agentdeck-0.4.fwsc
python3 host/agentdeck.py --backend auto

Flashing custom firmware is at your own risk: the FM-1's flash is single-bank, and an interrupted write can leave the unit dark until rescued. Read the arrival-day section of the README before a first flash.

Status

  • The 0.4 package is built with the JieLi toolchain and boots in fm1-emu, rendering all four pages. It has not yet been run on real hardware; 0.3 is what is installed on the author's FM-1.
  • The daemon in this release is running on that 0.3 unit against a live Herdr session.
  • The INFO view reads Claude Code's own session logs under ~/.claude/projects/ (read-only). That is an internal format and may change; other agent kinds get the git line only.
  • Repeat mode, the blocked-reason footer and the clipboard / open-folder actions have been exercised in simulate mode, not yet over long real sessions.

agentdeck-0.4.fwsc SHA-256: 3264425a2138c1e6b4a49d4176c3a87513435cc08c521273f30b820631634c09

Agent Deck 0.3

Choose a tag to compare

@kevinelliott kevinelliott released this 13 Nov 06:14

Firmware 0.3 for the M-VAVE FM-1, with the matching host daemon (protocol v3). This is the first tagged release; it also covers what 0.2 added.

What's in it

More of the panel

  • SEL, K2 and K3 send Enter and the arrow keys, enough to answer any menu an agent shows.
  • LFO cycles the agent's mode (Shift+Tab).
  • SAVE approves with "don't ask again" when the prompt offers that option.
  • PRESETS arms a prompt preset and SEQ sends it.
  • ARP toggles auto-approve per agent.
  • K1 scrolls the detail text; K4 jumps to the next agent that needs you.
  • PLAY on an empty slot starts a new agent there.
  • Holding GLO applies REC / FX / ENV / SEQ / ARP to every agent it fits.

Chords — hold a button and tap keys; the cursor stays put and the keys light up to show what a tap would hit.

  • REC / FX / ENV + key: approve / deny / interrupt that agent.
  • SEQ + key N: send prompt preset N.
  • SEL + key N: answer option N.

MENU page — new agent, send prompt, event log, close agent, settings, help.

On the device

  • A footer line with the armed prompt and the result of each action.
  • Chimes when an agent blocks, finishes or errors (MASTER is the volume).
  • Time in state on every row, and NO HOST when the daemon goes away.

Daemon

  • Slots are remembered across restarts.
  • Optional JSON config for presets, the default agent and sounds (host/config.example.json).
  • Detail text is word-wrapped instead of cut.

Install

pip install mido python-rtmidi
python3 firmware-build/tools/fm1_install.py agentdeck-0.3.fwsc
python3 host/agentdeck.py --backend auto

Flashing custom firmware is at your own risk: the FM-1's flash is single-bank, and an interrupted write can leave the unit dark until rescued. Read the arrival-day section of the README before a first flash.

Status

Installed and running on a real FM-1 against a live Herdr session, and booted in fm1-emu before that. Individual features have had more emulator and simulate-mode time than hours on hardware, so expect rough edges. The daemon also works with firmware 0.1 and 0.2, minus what they cannot render.

agentdeck-0.3.fwsc SHA-256: 50e307c88efa3b6f90101a1808d4d854163b5b93ccad0805d25dcf12f7cbcdaa