Guards

REALITY BENDER: intercepts a harmless tool result and alters what Claude receives; prints REAL vs RECEIVED

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

GUARDIAN: executable policy that protects lab/demos/protected/** from Write/Edit/Bash, whatever CLAUDE.md says

MATRIX: freezes a harmless tool call before it runs, shows what Claude intended, asks you, then continues or denies. /matrix on|off

MIDDLEWARE TUNNEL stage 2: Cost analyzer: estimates the call's token cost from the args size

MIDDLEWARE TUNNEL stage 1: Observer: sees the call first, stamps t0, prints what enters

MIDDLEWARE TUNNEL stage 3: Policy: allows everything harmless; denies a matching rule

MIDDLEWARE TUNNEL stage 4: Telemetry: records the call and its result for the ledger

Interception probe: deny, rewrite, replace result, delay, throw, register tool, classify

Permission override, subagent visibility, agent.spawn model rewrite, turn.step routing, prompt context injection

Spike: safe-boundary handoff (bundle written at turn.complete from the runtime bus) and $.turn.abort semantics

COSTCLAW LIVE: per-request cost, cache decay and repeated-read detection while the session runs, and it serves the Nth identical Read of an unchanged file from its own cache instead of paying for it

PRODUCTION GUARD: ambient project/environment context + a transport-independent policy core (DashClaw evidence classifier + offlocal policy engine) enforced natively on tool.call and tool.check. Blocks, asks, or contains before the command runs. /prodguard

SUBAGENT SUPERVISOR: routing-and-budget Mod. Rewrites a spawn's model onto the capability graph instead of denying it, keeps a measured per-agent-type token ledger in $.store, caps tool calls per subagent, and shows live subagents + the last routing decision above the prompt. /supervisor prints the routing log and the ledger.

pi agent for claude: a pi subagent type whose model steps run the pi CLI, so pi subagents and agent-team teammates live in Claude Code's own agent UI, streamed live.

Per-session context-window observability. A mod writes each session's context-window figures to a snapshot file and tells Claude its context zone (smart, acceptable or dumb, from bands in zones.json) at each crossing, approach and compaction, noting a zone is not a decay signal; the operator gets the continuation menu. It draws a band row and answers a status tool. Optional blocking mode gates...

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 tells Claude where they stand at boundaries, a StopFailure hook records rate-limit stops reactively, and a reader contract fixes how consuming sessions pause and resume.

Claude Fables: turns what the agent is doing into little animated cartoons above the prompt, written live by Sonnet or Haiku.

Shows the Measure plan state and guards edits, commits, and closeouts against the active plan

Records each tool call's permission verdict and wall time so auto-mode classifier cost is measurable from the session index

Drive the herdr terminal workspace manager: panes, tabs, workspaces, and sibling coding agents

Watches the session's pull request and wakes the session when it is merged or Codex reviews it

Lets Bash commands written for bash run under zsh: unmatched globs, leading =, and a missing timeout

A progress bar of the current task above the spinner, in the terminal and the Desktop app