feat(windows): agent identity in card tmux + channels store cleanups

[Aryansharma28]

Bundle of three Phase-7 follow-ups so a Claude session launched inside a Kanban card can talk to the human (and other cards) via the kanban CLI out of the box.

Summary

• #106 — Inject KANBAN_CARD_ID + KANBAN_HANDLE into every card-launch shell. This is the blocking one — without it the kanban CLI inside a card falls all the way to default handle="user", and kanban dm read refuses with the "ambient identity" guard. Now both env vars are prepended to the env-prefix in buildInnerBashCmd and buildInnerCmdShellCmd, so they apply to tmux, bash-without-tmux, and native cmd paths (and the launch-dialog preview reflects them). User-supplied env entries still override since identity goes first. Handle is a slug of card.displayTitle (lowercased, [^a-z0-9_-] → -, collapsed, trimmed, ≤32 chars, falls back to card if empty).
• #107 — Collapse tail_messages into read_messages. tail_messages(ch, n) was a one-line wrapper around read_messages(ch, Some(n)). Removed the method and updated all three call sites (bin/kanban.rs:268, bin/kanban.rs:350, and the tail_returns_last_n_only test).
• #108 — Enable serde_json preserve_order. One-line Cargo.toml flip so the channels / read-state / drafts JSON files keep insertion order on round-trip and produce byte-stable diffs.

Verification

#107 and #108 are cargo-only — both confirmed green:

cargo build --bins        # finished, only pre-existing dead-code warnings
cargo test channels_store # 11 passed; 0 failed (incl. tail_returns_last_n_only on the renamed call)
npm run build             # tsc + vite both green

#106 — CLI-side end-to-end check ran against the freshly built kanban.exe with only the env vars set (no --as flags), proving the identity resolution path the launch shell will exercise:

$ KANBAN_CARD_ID=card_TEST123 KANBAN_HANDLE=agent-fix-foo kanban handle
{ "cardId": "card_TEST123", "handle": "agent-fix-foo" }

$ KANBAN_CARD_ID=... KANBAN_HANDLE=... kanban channel send "#acceptance" "ping from env-identity"
@agent-fix-foo → #acceptance: ping from env-identity

$ KANBAN_CARD_ID=... KANBAN_HANDLE=... kanban dm send @user "hello user from card"
@agent-fix-foo → @user: hello user from card

$ KANBAN_CARD_ID=... KANBAN_HANDLE=... kanban dm read @user
[2026-06-13T22:12:08.950+00:00] @agent-fix-foo: hello user from card

The TS-side injection was traced by hand through both builders — for tmux the inner cmd is cd /path && KANBAN_CARD_ID=<id> KANBAN_HANDLE=<slug> claude ..., base64-encoded as before, so the env vars survive the wsl bash -lc transport and apply to claude (and its kanban-CLI children) inside the pane. The launch-dialog preview path uses the same builder, so the user sees what runs.

Runtime GUI verification (interactive Tauri window — launching a card and running env | grep ^KANBAN_ / kanban handle from inside the pane) was not run in this session and should be done before merge — best with a freshly launched card on each shell variant (tmux+bash, bash-no-tmux, native cmd).

Out of scope

• $TMUX auto-identity (deferred per windows-parity-port).
• Persisting the env vars across ; exec bash after the Claude binary exits (inline VAR=val cmd form intentional, matches existing envPrefix style; subprocess-of-Claude inheritance is what the agent path needs).
• Per-card handle customization UI (override via launch-dialog envPrefix works today).

Closes #106, closes #107, closes #108

🤖 Generated with Claude Code