Guards

The meeting Transcripted is recording, inside Claude Code: Claude gets the call as context, a live dashboard (now, asked of you, action items), prompt suggestions for questions aimed at you, and an after-call wrap-up that gains speaker names once Transcripted saves the meeting. Needs the transcripted-live helper running.

Shared rate-limit guard for loop lanes and Claude: a hooks module writes the subscription rate-limit windows to a fixed machine-scope file and toasts window changes (telling Claude too only with model_notices on, off by default), a StopFailure hook records rate-limit stops reactively, and a reader contract fixes how consuming sessions pause and resume.

BLACK BOX RECORDER: structured per-session timeline of prompts, model steps, tool calls, subagents and denials; /blackbox prints it

See inside a turn: a live Gantt timeline of model requests, tool calls and subagents, with tokens and cost per lane.

KOZMOS Blackbox: a flight recorder above the prompt. A live Gantt of the current turn: the main loop and each subagent on its own lane, every tool call a bar in its family color, model requests as ticks, a moving now cursor on an auto-scaling time axis; the last turn stays dimmed until the next. /blackbox toggles it.

Quality-first routing: picks main-loop effort (medium to max) per request, assigns model and effort to subagents by kind and weight, checks Workflow agent() staffing, records judgments locally; the top-model guard is opt-in

Measures the prompt cache for a team, in Claude Code and, with its command-line companion, Codex CLI and Devin CLI: hit rate per session, subagent and model, a warning before a cold resume re-sends a whole context, the quota each session used, and an anonymized export to pool cost and quota data. Records token counts and timings only, never prompts, code or paths.

Sets the reasoning effort of every native subagent call from Jev's rating of its task, leaving the model as it was, and keeps local records to compare efforts.

A goal pool inside the project: record goals, plan them with Claude, split them into sub-goals or tasks and run them one at a time; a side panel shows progress live.

Babysits a long Claude Code session so you don't have to: the context fill is watched, the recorder drafts the checkpoint, compaction happens at the right point with the state banked and the work is resumed, the rules and the checkpoint ride inside the compacted conversation and every subagent's prompt, tool results are trimmed and secrets redacted at the door, project memory and rules are a...

Unified workbench: a status band for subagents and mmrun reviews, plus a seven-page pane (Overview / Agents / Reviews / GPU / Timeline / Usage / Progress; Reviews, GPU and Progress show once they have data); the Progress page lists the per-project progress board, recorded after each typed turn that changed files or made a commit

The record of orchestrated work: runs of units and tasks, the subagents spawned for them, each gate's verdict and the findings reviewers report, summed up in the status line while there is something to say. It can follow a superpowers plan document as its input. Run /ledger.

Records which model and effort each subagent or background session ran at and how it went, with /routing-review to tune routing rules

사이드 패널에 진행 중인 작업·계획·스텝·서브에이전트를 보여주고, 코드를 고친 뒤 검증 없이 끝나면 경고하며, vibe 상태를 보여주고 검사를 돌린다

Transparency for native subagents: audit records, git-based file attribution, scope checks, chat receipts and a /subagents pane.

See inside a turn: a live Gantt timeline of model requests, tool calls and subagents, with tokens and cost per lane.