Prompt

A calibrated context sieve for Claude Code. Every large tool result is judged by a System One model before it enters context; irrelevant blocks are replaced with a summary and a recall key.

Prints which instruction files the engine loaded (tier, path, @ imports, size) once per context load, and hints when a project CLAUDE.md keeps its AGENTS.md from being read.

Measures a Claude Code session's fixed startup context payload per item at a pinned binary version, including per-tool attribution of the built-in tool pools that /context reports only as totals, derived by A/B deny differencing. Falls back from an SDK meter to a headless /context parser, then to an error rather than a wrong number. Keeps a before/after ledger per lever. Shows a live context...

Claude Code operations: audit-skill-visibility (skills the listing budget hides), inventory (every invocable surface), audit-install-state, audit-performance (slowness), audit-mod-posture (mods and the guard), audit-native-overlap (skills duplicating built-ins), observability (telemetry), known-issues, changelog, behavior-probes, prerequisites, check, machine-profile, plugins (update the...

KOZMOS Storyboard: Claude narrates its own progress. A chapter tool the model calls at each phase change drives a band above the prompt: the chapter title, a five-stage rail (explore, plan, build, verify, ship) with the active stage glowing, step 3/5 and a note. /storyboard toggles the band, /storyboard log opens the chapter history.

KOZMOS Thrift: a frugal mode. /thrift on|off|auto; while on, every prompt carries a short note asking Claude to be economical (targeted reads, no broad searches or subagents unless needed, concise answers). Auto turns it on when the 5-hour or weekly limit passes autoAt and off after the window resets. A band with a spinning coin, and /thrift compares the average cost of thrift turns with...

KOZMOS Tidewater: the live diff of the current turn as tides above the prompt. Every file Claude edits gets its own pool of rolling waves, a green swell for added lines and a red ebb for removed ones, with file chips and totals like '+128 −42 · 6 files'. From the first edit of a turn until the next prompt. /tidewater 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...

Runs proof-driven bug hunt rounds with /bughunt: each round proves one bug with a failing command the mod runs itself, fixes it, and proves the fix, and the loop stops on a blocked or unverified round. /bughunt collab runs a read-only scanner, planner and critic in sequence.

Names the uncommitted files to you at the end of each turn, and tells the model at the next prompt to commit each finished piece as it lands.

Parses each JSON, YAML, TOML or .env file an Edit or Write touches, notes a parse error at once instead of at the next build, and says so again when a later edit fixes it.

Checks each package the model installs against its registry and OSV.dev, and stops a missing, brand-new, look-alike, outdated or vulnerable one, naming the latest version.

Has a small model rate how hard each prompt is and runs that turn and its subagents at the matching effort, within the levels you allow, on the models whose prompt cache survives an effort change.

After each commit the model makes, names the env variables its added lines read that .env.example lacks, with the file and line of each.

After an edit that calls translation keys, names the keys the JSON, Laravel PHP, YAML or gettext locale files lack, and the languages that lack them.

After each commit the model makes, names the manifests whose dependencies it changed without their lockfile (npm, Composer, Cargo, Go, Python, Bundler, Dart, Mix).

Tells you when an MCP server failed to connect or dropped, with the engine's reason and a reconnect button, and when it is back; claude.ai connectors are left out.

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.

Learns the short prompts you send often in this project and draws them, with the ones you pin by hand, above the prompt; with the prompt box empty, a digit key sends one at once.

Keeps project and global memories in SQLite through a shared local daemon, hands the model the entries that match its file tool calls, prompts and subagent tasks, and saves new ones with a haiku consolidator after each main-loop turn.

After an edit that builds SQL by joining or interpolating strings, names each line, so the model passes the values as query parameters.

After an edit that keeps browser data in localStorage or sessionStorage, names each line, so the model stores the data in a cookie instead.

Submits a continue prompt while the task list has unfinished tasks, at most 99 times in a row by default, and turns the task tools on for every model unless CLAUDE_CODE_ENABLE_TODO_TOOLS is set.

After a compaction, adds one message with exact facts from before it: your latest prompt in full, up to 4 earlier ones, the files edited, the todo list and the last failed shell command. At most 4,000 characters. No model calls.