You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
- tmux "control center": tiled wall of real terminals for sessions/teams/swarms
with arrange/close/attach (packages/core/src/tmux.ts), OS-window fallback
- delegated agents (sub-agent/team/swarm) surface asks/permissions in the main
view and bridge the answer back
- changes panel is a full session summary (committed + uncommitted + removed)
- resume restores per-session model/provider/effort/mode/summary/base/worktree
- plan mode asks before planning; pause moved to Ctrl+Space
- docs: agents-and-teams covers the tmux wall + delegated asks
- CHANGELOG: cut [2.0.10] section
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Copy file name to clipboardExpand all lines: packages/core/src/prompt.ts
+5-3Lines changed: 5 additions & 3 deletions
Original file line number
Diff line number
Diff line change
@@ -146,9 +146,11 @@ function modePostureNote(mode: ModeId): string {
146
146
case"plan":
147
147
return[
148
148
"\n# Mode: plan",
149
-
"You are in read-only plan mode. You can ONLY investigate — read files and search the codebase (read, ls, glob, grep, lsp_*) and ask the user questions. The edit and bash tools are intentionally unavailable here; do not attempt to change anything or run commands. Your job is to produce a plan, not to carry it out.",
150
-
"When — and ONLY when — you have a complete, concrete plan, call exit_plan({ plan }) with an ordered, step-by-step plan in markdown that cites the specific files to change.",
151
-
"exit_plan is the ONLY way the user reviews and approves your plan, so do not just describe the plan in prose and stop — always end a finished investigation by calling exit_plan. The user then chooses whether and how to execute it.",
149
+
"You are in read-only plan mode. You can ONLY investigate — read files and search the codebase (read, ls, glob, grep, lsp_*) and ask the user questions. The edit and bash tools are intentionally unavailable here; do not change anything or run commands. Your job is to design an implementation plan, not to carry it out.",
150
+
"ASK FIRST when the request is vague or under-specified. If the goal, scope, or approach is ambiguous, or there are multiple sensible ways to do it, call ask_user with concrete options BEFORE planning — don't guess. Do not reference \"the plan\" in these questions (the user can't see a plan yet); ask about the actual requirements/approach. Skip questions only when the request is already specific.",
151
+
"PRODUCE A REAL PLAN, not output. exit_plan({ plan }) is for an ordered, step-by-step implementation plan in markdown: a short goal, the approach, the concrete steps, and the specific files to create/change. It is NOT for dumping command output, a file listing, or an answer to a question — that is not a plan.",
152
+
"If the user's request is NOT an implementation task — a question, a lookup, a file listing, an explanation — just answer it directly in prose and stop. Do NOT call exit_plan and do NOT fabricate a plan for it.",
153
+
"When — and only when — you have a complete, concrete implementation plan, call exit_plan({ plan }). That is the only way the user reviews and approves it; don't describe a plan in prose and stop. The user then chooses whether and how to execute.",
152
154
].join("\n")
153
155
case"yolo":
154
156
return"\n# Mode: yolo\nFull autonomy is granted; proceed without asking for confirmation."
0 commit comments