Agent detection for Atomic Agent (tested manifest and captured screens included)

[sosidudku1]

Hi! I am building Atomic Agent, an open source local-first coding agent with an Ink TUI: https://github.com/AtomicBot-ai/atomic-agent

It already runs inside a herdr pane today. The terminal model is the same as Claude Code: alt screen, composer at the bottom, hotkey strip pinned to the last row. So I did the homework and prepared everything for first-class detection, hoping to live in the agents table next to Claude Code, Codex and Hermes.

What I bring

1. A ready screen manifest in your engine format (included below). Engine v2 features only, 13 rules covering idle, working (footer, steer/queue meta row, stop button, braille spinner), blocked (approval strip and approval banner, plan handoff), the first-run wizard, self-update, and a terminal too small guard with skip_state_update.
2. Real detection snapshots captured from a herdr pane with herdr pane read --source detection (below: idle, working, blocked; home paths anonymized).
3. Verification with your own engine: I ran herdr agent explain --file <screen> --json over 16 captured screens covering a wide terminal, a 55-column terminal (footer chips shed and the last row clips), the first-run wizard, the plan handoff offer and the queue-mode working variant. 16/16 classify correctly; working screens match four independent rules at once and the approval strip keeps both of its un-sheddable anchors at 55 columns.

Verification summary

screens	terminal	expected	result	rules matched
idle, post-turn idle	~120x40	idle	idle	idle_footer
working x3	~120x40	working	working	running_footer, running_meta_row, stop_button, spinner_activity
approval x3	~120x40	blocked	blocked	approval_strip, approval_banner
narrow idle	55x20	idle	idle	idle_footer
narrow working x3	55x20	working	working	all four working rules
narrow approval	55x20	blocked	blocked	approval_strip
first-run wizard intro	100x30	idle	idle	wizard_idle
plan handoff (run/dismiss offer)	120x40	blocked	blocked	plan_handoff
working, queue mode (ctrl+t)	120x40	working	working	all four working rules, queue variant

Facts for process lookup

• Executables: atomic-agent (SEA binary or node bin shim), plus the short alias atag. npm package atomic-agent. Via node the cmdline contains dist/cli/index.js.
• Typical launch: atomic-agent tui --cwd <project>.
• Suggested id: atomic, aliases atomic-agent, atag. No OSC title is emitted, so screen rules only.

Maintenance promise

We will keep the manifest in sync with TUI changes and can re-capture fixtures whenever you need them.

I also tried the custom lifecycle route (pane report-agent --source custom:atomic) against a live pane and it works nicely; native self-report from our runtime is something we want to ship as well. If you would rather see that first, tell me what the bar looks like.

Happy to adjust anything about the manifest to fit your conventions. Thanks for building herdr, the "leave them running" model is exactly how we think agents should be operated.

atomic.toml (the manifest)

id = "atomic"
version = "2026.09.15.1"
min_engine_version = 2
updated_at = "2026-09-15T00:00:00Z"
aliases = ["atomic-agent", "atag"]

# Atomic Agent (https://github.com/AtomicBot-ai/atomic-agent): local-first coding agent, Ink TUI.
# Executables: atomic-agent (SEA binary or node bin shim), atag (alias), dist/cli/index.js.
# The TUI runs on the alt screen and pins a hotkey strip to the last terminal row.
# Chips render as "[key] label" joined by "  ·  " (U+00B7). No OSC title is emitted.

# --- overlays that must not flip state ---

[[rules]]
id = "terminal_too_small"
state = "unknown"
priority = 1000
region = "bottom_non_empty_lines(5)"
skip_state_update = true
contains = ["terminal too small"]

# --- blocked ---

# Tool approval strip: "[ctrl+y] approve  ·  [ctrl+d] deny  ·  [esc] abort run".
[[rules]]
id = "approval_strip"
state = "blocked"
priority = 980
region = "bottom_non_empty_lines(2)"
visible_blocker = true
contains = ["[ctrl+y] approve", "[ctrl+d] deny"]

# Approval banner above the composer: approve/deny buttons + abort hint near its bottom edge.
[[rules]]
id = "approval_banner"
state = "blocked"
priority = 975
region = "bottom_non_empty_lines(20)"
visible_blocker = true
contains = ["esc abort run · ctrl+c stop everything"]
any = [
  { contains = ["✓ approve · ctrl+y"] },
  { contains = ["✗ deny · ctrl+d"] },
]

# Plan handoff offer: agent finished planning and waits for a run/dismiss decision.
[[rules]]
id = "plan_handoff"
state = "blocked"
priority = 970
region = "bottom_non_empty_lines(2)"
visible_blocker = true
contains = ["[ctrl+y] run it · auto"]

# --- working ---

# Running footer keeps both abort chips: "[esc] abort … [ctrl+c] abort".
# The approval strip has "[esc] abort run" but never a "[ctrl+c] abort" chip.
[[rules]]
id = "running_footer"
state = "working"
priority = 950
region = "bottom_non_empty_lines(2)"
visible_working = true
contains = ["[esc] abort", "[ctrl+c] abort"]

# Composer meta row shows a steer/queue marker only while a turn is in flight.
[[rules]]
id = "running_meta_row"
state = "working"
priority = 940
region = "bottom_non_empty_lines(4)"
visible_working = true
any = [
  { contains = ["⏎ steer (ctrl+t)"] },
  { contains = ["⏎ queue (ctrl+t)"] },
]

# Composer stop button, only visible while a turn is in flight: "■ stop   send →".
[[rules]]
id = "stop_button"
state = "working"
priority = 935
region = "bottom_non_empty_lines(8)"
visible_working = true
contains = ["■ stop"]

# Thinking indicator right above the composer: braille spinner + activity + elapsed.
# e.g. "⠋ thinking · 12s", "⠹ writing reply · 1m05s", "⠴ os.fs.read · src/x.ts · 14s".
[[rules]]
id = "spinner_activity"
state = "working"
priority = 930
region = "bottom_non_empty_lines(12)"
visible_working = true
line_regex = ['^\s*[\x{280B}\x{2819}\x{2839}\x{2838}\x{283C}\x{2834}\x{2826}\x{2827}\x{2807}\x{280F}]\s+\S.*·\s*\d+(?:s|m\d{2}s)\s*$']

# Self-update runs to completion and restarts; treat as busy.
[[rules]]
id = "self_update"
state = "working"
priority = 920
region = "bottom_non_empty_lines(12)"
visible_working = true
contains = ["updating atomic-agent"]

# First-run wizard while a local model download is in progress.
[[rules]]
id = "wizard_download"
state = "working"
priority = 910
region = "bottom_non_empty_lines(3)"
visible_working = true
contains = ["c set up cloud meanwhile", "s skip to the agent"]

# --- idle ---

# The "[ctrl+c] quit" chip only exists when no turn is running and nothing is pending
# (while running the chip reads "[ctrl+c] abort", during approvals the strip has no ctrl+c chip).
[[rules]]
id = "idle_footer"
state = "idle"
priority = 250
region = "bottom_non_empty_lines(2)"
visible_idle = true
contains = ["[ctrl+c] quit"]
not = [
  { contains = ["[esc] abort"] },
  { contains = ["[ctrl+y] approve"] },
]

# First-run wizard footers (plain chips, no brackets): "… esc back   ctrl+c quit".
[[rules]]
id = "wizard_idle"
state = "idle"
priority = 240
region = "bottom_non_empty_lines(2)"
visible_idle = true
any = [
  { contains = ["ctrl+c quit"] },
  { contains = ["[ press any key to continue ]"] },
]
not = [
  { contains = ["c set up cloud meanwhile"] },
]

Screen: idle (bottom of the capture)

         add a health c…    ↑                   #########+++
         why is the bui…    ↑                   ########+++
         review my chan…    ↑                   #######+++
         fix the failin…    ↑                     +++++++
   ↓ 1 more
                                        • /help                   list all slash commands
   TASKS   0 running     + new          • /sessions               switch to a previous thread
   (no active tasks)                    • /new                    start a fresh session
                                        • /model                  change the chat model
                                        • /tasks                  jump to the Tasks tab (cron + ingress UI)
                                        • /import                 bring sessions and skills over from another agent




                                   ╭──────────────────────────────────────────────────────────────────────────────────╮
                                   │                                                                                  │
                                   │ ❯  What are you working on today?                                        send →  │
                                   │                                                                                  │
                                   │                                                                                  │
                                   │ ● cloud · openrouter · anthropic/claude-fable-5.1                       default  │
                                   │                                                                                  │
    ☰ Menu                 esc    ╰──────────────────────────────────────────────────────────────────────────────────╯
                                   [enter] send  ·  [shift+enter] newline  ·  [esc] menu  ·  [ctrl+c] quit

Screen: working (bottom of the capture)

         explain how se…    ↑
         add a health c…    ↑
         why is the bui…    ↑
         review my chan…    ↑
   ↓ 2 more

   TASKS   0 running     + new
   (no active tasks)
                                     YOU
                                   │  Read README.md in this folder and write a two-line haiku about it. Then stop.
                                   │
                                      [copy] [try again]

                                      ⠙ thinking · 4s

                                   ╭──────────────────────────────────────────────────────────────────────────────────╮
                                   │                                                                                  │
                                   │ ❯  What are you working on today?                               ■ stop   send →  │
                                   │                                                                                  │
                                   │                                                                                  │
                                   │ ● cl…⏎ steer (ctrl+t)   context [        ] 1/27 tasks ·       7.0k/1M   default  │
                                   │                                                                                  │
    ☰ Menu                 esc    ╰──────────────────────────────────────────────────────────────────────────────────╯
                                   [⏎] steer  ·  [ctrl+t] queue mode  ·  [esc] abort  ·  [ctrl+c] abort

Screen: blocked, tool approval (bottom of the capture)

         explain how se…    ↑      ║ affects: /Users/dev/bizdev-collabs/herdr-integration/testproj/notes.txt    ║
         add a health c…    ↑      ║                                                                                  ║
         why is the bui…    ↑      ║  ✓ approve · ctrl+y   ✗ deny · ctrl+d                                            ║
         review my chan…    ↑      ║                                                                                  ║
   ↓ 2 more                        ║  allow file write · workspace this session · ctrl+f  edit target path… · ctrl+b  ║
                                   ║                                                                                  ║
   TASKS   0 running     + new     ║                                                                                  ║
   (no active tasks)               ║ esc abort run · ctrl+c stop everything                                           ║
                                   ║ approve covers this call once; the session grants last until the app exits       ║
                                   ║ (never persisted); loosen the standing stance with the coding-mode control in    ║
                                   ║ the composer · the composer stays live — type to answer the agent instead (enter ║
                                   ║  cancels this call and sends it)                                                 ║
                                   ╚══════════════════════════════════════════════════════════════════════════════════╝


                                   ╭──────────────────────────────────────────────────────────────────────────────────╮
                                   │                                                                                  │
                                   │ ❯  Type a message or `/` for commands…                                   send →  │
                                   │                                                                                  │
                                   │                                                                                  │
                                   │ ● cloud · openr…  · an… context [        ] 2/27 tasks ·      23.8k/1M   default  │
                                   │                                                                                  │
    ☰ Menu                 esc    ╰──────────────────────────────────────────────────────────────────────────────────╯
                                   [ctrl+y] approve  ·  [ctrl+d] deny  ·  [esc] abort run

[noelthekk]

Idea LGTM. I use both Atomic Agent and Herdr, and this would be a welcome feature for me at least. Especially given its very good support for local hosting and tinkering.

[sosidudku1]

Update: Atomic Agent v0.6.2 is out and ships native self-report for herdr. Inside a pane it now claims its own label through the custom lifecycle route (pane report-agent, source custom:atomic): working while a turn runs, blocked with the tool name while an approval waits for the human, idle otherwise, and it releases the label on exit.

Two implementation notes that may be useful for other integrators: sequence numbers start from epoch seconds, so a restart in the same pane outranks the previous run's reports, and the release call carries the next sequence number, since a stale-sequenced release is silently dropped.

So current Atomic versions are first-class in herdr out of the box. The manifest above still covers screen-only detection for older versions and stays maintained.