Add jcode to agent detection

[gloiiire]

jcode (github.com/1jehuang/jcode) runs as a plain jcode process, so Herdr treats it as an unknown terminal program today — no state, no rollup, no waits.

I built the detection side and validated it against live panes: https://github.com/gloiiire/herdr/tree/feat/detect-jcode

• adds Agent::Jcode plus a screen manifest, no integration/hooks
• jcode pins one composer line to the bottom: <n> plus a marker glyph — > ready, … turn in flight, » skill armed, $ shell mode. Only … means busy, and the marker is authored in one place (input_prompt() in jcode's TUI), so it should hold across jcode releases.
• second working signal from the status line above the composer (braille spinner or the animated tool bar)
• no blocked rule on purpose: jcode has no tool-approval prompt, so Herdr's idle fallback is the correct state for every interaction it shows

Worth flagging: jcode emits no OSC 9;4 progress, and its OSC title is only a session emoji plus the session title, so neither carries state. A previously closed PR (#1822) added OSC 9 rules matching payloads jcode does not emit.

11 files, ~90 lines. cargo fmt --check, clippy and cargo nextest run (2953 tests) green. Verified on jcode 0.61.1 in a debug Herdr server: idle -> working -> done.

Related: #1848 covers the integration/hook side instead of detection.

Happy to adjust or drop this if it is not a direction you want.

[gloiiire]

Follow-up: validated against live sessions, not fixtures.

Ran the branch on macOS with three unrelated jcode panes alongside two claude panes on one server:

w1:p7    jcode    idle
wE:pC    jcode    working
wD:pE    jcode    idle

The discrimination comes straight off the composer line — bottom row of the detection buffer:

idle      2>
working  22…

Second working signal one row above, the animated tool bar: ●·· bg ··● · Attendre les merges · 6m 13s.

No per-session tuning was needed — the same rule held across three jcode sessions at different message counts and in different repos. jcode 0.61.1, Herdr 0.7.5.

[ppronobis]

I re-checked your branch description against jcode 0.65.0 on Linux, since you verified it on 0.61.1 on macOS. The composer marker holds across both versions and platforms.

All four states from a freshly started pane, last line of herdr pane read <pane> --source detection --format text:

1>     fresh start, empty
2…     mid-answer
6$     after !, status line above reads "shell mode · Enter runs on server"
6»     after activating a skill via /herdr

The leading number is the turn counter, not part of the marker. I watched a full idle -> working -> idle cycle in that pane, and … only ever showed up while a turn was in flight.

Two things to add, one of which corrects myself:

No OSC 9, confirmed. The 0.65 binary contains no ]9;4 and no other OSC 9 payload, only OSC 0 for the title. My title reads 🌐 jcode/pagoda Duck, so emoji plus model plus session name, no state in there. The approach from #1822 does not hold on current jcode either.

On the missing blocked rule. I set out to contradict you, because the binary does contain Approve, Deny and "Agent is waiting for this decision". Following that up: those strings belong to jcode permissions, a separate CLI subcommand for ambient requests, not to the TUI of a running session. Inside a session I could not provoke a prompt with rm -rf or a write to /etc/hosts; jcode either refuses or just runs the command. So your reasoning is right for the TUI and my objection was wrong. Worth recording in case someone else trips over the same strings later.

One edge case that might save you some work: shell mode ($) is not working, but it is not ordinary idle either. The pane waits for input, but the next Enter goes to the shell rather than the model. Idle is probably still the right answer, it should just be a deliberate call rather than a fallback.

Tested on jcode 0.65.0 (d847d62e4) with herdr 0.7.5, Linux, session over SSH.