@sma1lboy/kobe 0.6.3 → 0.6.5

package/README.md

@@ -78,28 +78,20 @@ Once you're in, the keys you'll use most:
 
 | Key                | What it does                                                   |
 | ------------------ | -------------------------------------------------------------- |
-| `ctrl+h` / `ctrl+j` / `ctrl+k` / `ctrl+l` | Jump straight to a pane (sidebar, workspace, files, terminal) |
-| `tab`              | Cycle focus to the next pane                                   |
-| `ctrl+q`           | Detach back to the sidebar (your task keeps streaming)         |
-| `ctrl+o`           | Open the active task's worktree in your editor                 |
-| `F1`               | Show the full keybinding help dialog                           |
-| `ctrl+,`           | Open Settings (theme, transparent background, dev reset)       |
-| `q`                | Quit (with confirm)                                            |
+| `ctrl+h` / `ctrl+j` / `ctrl+k` / `ctrl+l` | Move between tmux panes (Tasks, engine, Ops files, shell) |
+| `ctrl+q`           | Detach back to your launching shell; the task tmux session keeps running |
+| `ctrl+t`           | Open a same-engine ChatTab window on the current task/worktree |
+| `ctrl+shift+t` or tmux `prefix T` | Pick an engine, then open a new ChatTab window |
+| `ctrl+[` / `ctrl+]` | Switch to previous / next ChatTab window                    |
+| `ctrl+w`           | Close the current ChatTab window; the final window is protected |
+| `F2`               | Rename the current ChatTab window                            |
+| tmux `prefix f`    | Open the Tasks pane's new-task dialog from anywhere in the session |
 
-Inside the sidebar, with a task highlighted: `n` creates a task, `j/k` moves,
-`enter` opens, `r` renames, `a` archives, `d` deletes, `s` opens Settings, and
-`[` / `]` switches between the working session and the archives view.
+Inside the Tasks pane, with a task highlighted: `n` creates a task, `j/k` moves, `enter` opens, `r` renames, `b` renames the branch, `v` cycles the engine, `o` opens the worktree in your editor, `a` archives/unarchives, `d` deletes, `s` opens Settings, and `[` / `]` switches between Working session and Archives. Archive/delete also kill the task's cached tmux session when one exists.
 
-Inside the chat composer:
+Inside the Ops files pane: `[` / `]` switches All / Changes, `enter` opens the preview, `a` injects an `@file` mention into the engine pane, `p` injects the create-PR prompt, `r` refreshes, and `o` opens a file externally when that makes sense.
 
-- `enter` to send, `shift+enter` for a newline.
-- `shift+tab` toggles the per-task tool-permission mode between `default` and `plan`. `default` is the trusted-bypass mode — kobe forwards it to `claude` as `--permission-mode bypassPermissions` because `claude -p` has no interactive permission protocol, so the only meaningful choice is "auto-deny outside cwd" or "auto-approve everything." `plan` forwards unchanged.
-- Click the model label in the footer to pick the model for this task (opus / sonnet / haiku).
-- Type `/` to open the slash-command dropdown. Bundled `claude-code` commands and your own `.claude/{commands,skills}/*.md` are merged in.
-- A `Create PR` chip on the chat header injects a PR-instructions prompt into the active task and routes the resulting PR through the orchestrator.
-
-A given task can host **multiple chat tabs** on the same worktree — useful when
-you want a parallel sub-conversation without losing the main thread.
+A given task can host **multiple ChatTab tmux windows** on the same worktree — useful when you want a parallel sub-conversation without losing the main thread.
 
 ## Opening tasks in your editor
 
@@ -211,24 +203,33 @@ kobe daemon stop      # tell the daemon to drain and exit
 > was removed in favor of the single-bin surface (KOB-136); any script with
 > `kobed restart` needs a one-time rename to `kobe daemon restart`.
 
-### `kobe api <verb>` — five shell verbs
+### `kobe api <verb>` — six shell verbs
 
-Each call is a short-lived process: open the daemon socket, do one RPC, print
-JSON to stdout, exit. Any tool with a `Bash` capability (Claude Code, Codex,
-Cursor, a custom agent) can drive it.
+Each call is a short-lived process: connect to (or auto-start) the daemon, do
+the work, print JSON to stdout, exit. Any tool with a `Bash` capability (Claude
+Code, Codex, Cursor, a custom agent) can drive it.
 
 | verb | flags | what it does |
 |---|---|---|
-| `spawn-task` | `--repo PATH --prompt TEXT [--title T] [--base-branch B]` | Create a new task + worktree + chat session. |
-| `create-tab` | `--task-id ID [--title T]` | Open an extra chat tab on an existing task. |
-| `send`       | `--task-id ID --prompt TEXT [--tab-id TID]` | Resume the task's session with a new prompt. |
-| `get-task`   | `--task-id ID` | Read task metadata (status, branch, worktree, tabs). |
-| `get-tab`    | `--task-id ID --tab-id TID` | Read a single tab off the task. |
+| `spawn-task` | `--repo PATH [--prompt TEXT] [--title T] [--base-branch B] [--vendor V]` | Create one task + worktree. With `--prompt`, also start the engine in tmux and deliver the prompt. |
+| `fan-out`    | `--repo PATH --prompt TEXT [--count N \| --agents claude:2,codex:1] [--base-branch B]` | Spawn N tasks of the same prompt in one call (per-engine counts optional). Returns `{count, tasks[]}`. Capped at 10. |
+| `send`       | `[--task-id ID] --prompt TEXT` | Paste a prompt into a task's engine pane (bracketed paste, then Enter — multi-line stays one turn). Defaults to the active task; prefer `--task-id` for unattended fan-out. |
+| `get-task`   | `--task-id ID` | Read task metadata; `.running` is true when the tmux session is live. |
+| `collect`    | `--task-ids a,b,c \| --repo PATH` | Read-only aggregation snapshot: per task, branch + `.running` + uncommitted `.changes` ({added, deleted}) for comparing attempts. |
+| `list`       | — | List all tasks. |
+
+`spawn-task --prompt` and `send` also return `.engineReady` — `false` means a freshly-started engine didn't confirm it was ready before the prompt was pasted (delivered best-effort).
 
 Output is one JSON object on stdout, `\n` terminated, exit 0. Errors land on
 stderr as `{"error":{"message":"...","code":"..."}}` with a non-zero exit.
 Add `--pretty` for human inspection.
 
+v0.6 is tmux-native — a task's chat history lives in its tmux session, not in
+the daemon — so `send` types the prompt into the engine pane and there is no
+verb that reads the reply back. Results are reviewed in the TUI. (The v0.5
+`create-tab` / `get-tab` verbs are gone; extra chat tabs are tmux windows
+opened from inside the TUI.)
+
 Fan-out from a shell:
 
 ```bash
@@ -237,19 +238,17 @@ T2=$(kobe api spawn-task --repo "$PWD" --prompt "Approach B: event sourcing"  |
 T3=$(kobe api spawn-task --repo "$PWD" --prompt "Approach C: reducer pattern" | jq -r .taskId)
 
 # Tell the user three tasks are running — they'll appear in the sidebar.
-echo "Spawned $T1 $T2 $T3"
+echo "Spawned $T1 $T2 $T3 — open the kobe TUI to watch them"
 
-# Poll until each settles.
+# Confirm each session came up.
 for ID in $T1 $T2 $T3; do
-  until [ "$(kobe api get-task --task-id "$ID" | jq -r .task.status)" = "idle" ]; do
-    sleep 10
-  done
+  kobe api get-task --task-id "$ID" | jq -r '"\(.task.id) running=\(.running)"'
 done
 ```
 
-If the daemon isn't running, `kobe api ...` exits 2 with
-`{"error":{"code":"BAD_DAEMON",...}}` on stderr. Start it with
-`kobe daemon start` (or just launch the TUI).
+`kobe api ...` auto-starts the daemon, so there is nothing to launch first. If
+the daemon cannot be reached or started, it exits 2 with
+`{"error":{"code":"BAD_DAEMON",...}}` on stderr.
 
 ### Install the agent skill
 
@@ -270,12 +269,10 @@ After install, Claude Code automatically picks up the skill on its next launch.
 For project-level overrides, copy the file to `<repo>/.claude/skills/kobe/SKILL.md`
 and customise — Claude Code's discovery order is project > user > none.
 
-`kobe diagnose` reports whether the skill is installed, which is the
-fastest way to confirm.
+To confirm it installed, check that `~/.claude/skills/kobe/SKILL.md` exists.
 
-> `kobe skill install` is deprecated as of v0.6 — use the `npx skills`
-> command above. `kobe skill uninstall` still works for cleaning up an
-> older copy.
+> The old built-in `kobe skill install` / `kobe diagnose` commands were
+> removed in v0.6 — use the `npx skills` command above to manage the skill.
 
 ## Coming later