bws-mods

Agents tab: every subagent of this session as a live, coloured card in the dashboard (status, type, model, task, elapsed, tool count, last tools, tokens, final answer), running ones first, with a summary strip and a stall flag when an agent makes no tool call for a while. /subagents opens it.

Blackboard for a team of Claude Code sessions in one repository: each posts what it works on and its role, sees the others, claims the files it edits so no one clips them, and asks the others for help (ask, take, done) with an optional urgent wake; lets one session at a time run git add/commit/push in a working tree refuses a commit that would sweep up a peer's claimed files, and refuses git commands (stash, reset --hard, checkout ., restore, clean, rm) that would overwrite a peer's changed files; keeps a history of who held, asked and woke what, and /board digest turns it into a team timeline page.

Dashboard pane (Board, Loop and Mods tabs, the Mods tab a switchboard that turns each bws mod on or off and hides or shows its tab) with a colour action bar above the prompt (progress bars for the session's tasks, the run checklist and a live loop), Editor & Device Guard for Unreal/Unity/Blender/adb, Finish-Line Guard, Project Locator, /handoff with a context nudge, and a worker model pin for subagents. A cache clock on the bar counts down the prompt cache and writes a handoff before it runs cold.

Claude delegates coding work to OpenAI Codex on your ChatGPT plan, and you chat with Codex on the side. codex_delegate starts a Codex job in its own git worktree on branch codex/<id>, sandboxed (workspace-write, no internet by default) behind Claude Code's permission prompt; the mod commits each turn, Claude reviews and merges the branch, and an idle Claude is woken when a job ends. The Codex tab in the cockpit dashboard lists the jobs (diff, follow-up, cancel, merge hint) and holds a side chat on the OpenAI API (with your key) or on Codex CLI, with a git capsule each turn, Pull from Claude and Send back.

When Claude asks you to pick a design direction, each option is drawn as a small rendered mockup above the picker.

Diagram pane: after a big turn a small model decides whether it is worth a diagram; if so a fork draws what happened (steps, files, new components, decisions) in a side pane with a legend. Mermaid blocks Claude writes are pinned there too. /diagram draws the whole session.

Drift detection: after a turn that edits files, Sonnet checks the work against the session's opening request (reading the whole conversation), its run-folder PLAN.md and the project's docs/ corpus (specs, PRD, requirements, architecture or design docs first; a GDD for a game) and says aligned, drifting or off course, and whether the agent wandered or the scope was widened. Verdicts are kept per project, so the Direction tab of the dashboard shows drift within the session and across sessions; /direction overall judges recent commits as a whole.

/enhance <rough prompt>, or a chord on the draft in the box: rewrites a vague prompt into a specific, actionable one using this project's instructions, installed skills and commands, MCP servers, the recent conversation and Anthropic's prompting guidance for the session's model, puts it in the prompt box, and offers Send / Revert in a small pane. Optional prompt bouncer: holds a vague prompt before it is sent, with the reason and the project glossary.

Inventory tab: what this conversation has used, as coloured cards in the dashboard: its todos (TodoWrite, the Task tools, the Desktop task list) with progress, the connectors (MCP servers) it called with call and error counts, the files it read, edited or wrote, and the skills it ran. /inventory opens it.

Screen-share masking for recordings and demos: /mask draws emails, phone numbers, money amounts, keys and tokens, your account name and your own terms as same-length bars in the transcript. Only the screen changes; Claude still reads the real data. A Mask tab in the dashboard shows the mode, what is hidden and what is not covered.

Memory: after a turn a small model decides whether the user said something worth remembering; if so a fork writes memory files in the background (redacted, de-duplicated) and a pane lists what changed. /remember <text>, /memories, /memory-undo. DreamPath: /dream merges duplicates, drops stale entries and rewrites for clarity (preview first, backup, undo), every 20 turns and before a compaction.

Windows toasts that say which project and what happened: permission asks, finished turns, finished subagents and errors, with the reply's first line. Replaces the modal MessageBox hook.

A showpiece for what a mod can draw: /prism opens a pane of animated generative art, painted cell by cell with Raster in the terminal and as animated Svg in the Desktop app.

Secret box: Claude asks for a secret with request_secret, you type it into a masked pane, it is saved outside the project, and Bash commands use it as {{secret:NAME}}. The value is read by the shell at run time and scrubbed from results, so it never reaches the transcript. /secrets lists and removes names.

Refuses Bash commands that will fail on Windows Git Bash for quoting or length reasons (cut-off long commands, trailing-backslash paths, inline Python whose backslashes the shell halves), with the exact fix, so Claude retries correctly at once.

Side chat in the right-hand pane: /fork (or /branch) and /btw open a copy of this conversation beside it, answered by a forked headless session so the main conversation is never touched. /explain turns the last reply's options into a decision tree you explore in forks.

Skill gating: on a conversation's first prompt a small model picks the skills that fit it and the skill listing the model reads is cut to those (plus any /name the prompt names); later prompts only add skills, so the prompt cache holds. /skillgate [off|on|status|explain].

Rolling conversation summary: each time Claude stops, Haiku updates a short summary of the whole conversation (goal, decisions, done, open items). The Summary tab of the dashboard shows it, with earlier summaries kept so a rollback never loses one. Session-only; never enters Claude's context.

Scratch: A/B the design-picker generation prompt across models.

Throwaway: logs surfaces() in claude -p and whether a mod-spawned console child gets a visible window

Throwaway: logs which events carry cross-session delivery notices, and sends a probe message

Throwaway: one $.session.send to a peer, for the Repo Sessions live receipt