Status line

/ci-watch : polls gh pr checks every minute in the status line, no tokens.

Fleet P1/P2 counts from ~/.continuity/FLEET-STATE.md in the status line; /fleet and /ports commands.

Failing tests spawn a pixel boss; every run that fixes tests lands a hit, zero failures is a KO with loot

KOZMOS Marquee: a status-line ticker. Model and effort, context fill, 5-hour and weekly limits with the reset countdown, session cost, git branch with ahead and dirty counts, and the tool running now, refreshed every second; it scrolls like a marquee when longer than its width. /marquee toggles it.

TypeSafe's Jev decides how Claude Code works on each prompt: the main conversation's reasoning effort (low to xhigh, raised to max mid-turn when tool calls keep failing), each subagent's model, whether to split the work across subagents, and the one skill the prompt needs. Brings in custom OpenRouter models (a junior coder reviewed by Opus, or cheap bulk work) and Codex or OpenCode as...

Shrinks each Bash result before the model reads it: per-command filters for git, test runners, linters, package managers, containers and file tools, the full output kept in a file.

Shows the subscription usage limits under the prompt or in the shared sidebar with a reset countdown and a projection, opens a /limit-watch pane, logs once when a limit passes 80% and 95%, and stops sampling entirely with /limit-watch off.

Loads MEMORY.md into the session and saves project learnings to it after every turn through a tool-less fork in the background, without blocking the stop.

Shows each subagent of the session with its turns, time, model and tokens, the costliest first, yellow while it runs and green once it answered, and marks one that passed the token limit.

Shows how long the prompt cache stays warm, and when it expires, exactly how many tokens your next message will re-send uncached. No tokens spent.

Asks before Claude edits a file another chat on this machine changed in the last 30 minutes. Each chat keeps its own small ledger; no model calls, no network.

Lists skills you have not used lately in this project by name only, so their descriptions stop riding along on every request; the Skill tool and /name still load any of them. No model calls.

Moves tools you have not used lately behind ToolSearch, so their descriptions stop riding along on every request; core tools and anything you used in your last five sessions stay loaded.

Shows you, above the prompt, the critical decisions the agent made without being asked: red to rethink, yellow to know, each with a rethink button. Works out of the box for any kind of project; /critical-areas auto lets the agent learn what is critical in yours.

Experimental kgai add-on for Claude Code with function hooks: checks each code-editing turn for a structural decision on the side, records it without a line in the transcript, and draws kgai's own tool calls as one line each. Requires the kgai plugin.

Opinionated pipeline from a feature, bug, or refactor description to a verified, publish-ready implementation: scoped interview, wave-based plan, parallelized subagent execution, cross-model branch review, on-device UI verification via claude-in-chrome, and GitHub reconcile — layered on the superpowers skill set, with a doctor/learn loop that benchmarks and improves the pipeline from session...

Test-first workflow for coding agents, without the spec ceremony. A PreToolUse hook keeps source edits behind a mirror test and protects existing assertions; the flow drives one vertical slice at a time, red to green, through the project's own gate. Ships its own tdd and grill skills; includes the Jev layer (model and effort routing, compaction, review) and the interface (usage row, gate...

See the decisions Claude made in each turn, why it made them, and steer future turns with keep and avoid rules.

Keeps an idle session's prompt cache warm with a short prompt 2 min before it expires; capped per model, the last one a handoff

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

Swear at Claude Code all you like: every prompt is cleaned before Claude reads it. Typos, translit, the wrong keyboard layout, swearing and CAPS go; the meaning, paths and numbers stay; your original rides along so nothing is lost.

Hides secrets in tool results before Claude reads them: the secret values of every .env and .env.* from your session's folder up to the drive root, and anything that looks like a secret (KEY=VALUE under a secret name, passwords in URLs, private keys, JWTs, AWS keys, Kubernetes Secrets); tells Claude how to use them without printing them

Status line score 0-100 for how good a moment it is to /compact now, with reasons; offers a ready /compact that keeps goal, decisions, leftovers and paths, and adds that template to every compaction. Never compacts by itself.

Keeps long sessions going through API errors and rate limits: resubmits after a backoff when idle, waits out rate limits with a countdown, never loops, and stops the moment you type.