I am experimenting with separating writing code from authorizing code. The coding agent can push a branch and open a pull request. It cannot produce the cryptographic evidence required to merge that pull request, and it does not hold the merge credential.
This is the smallest useful version of the design behind Herdacat. The included demo is deliberately local and forge-independent so you can run it without creating bots, apps, or cloud infrastructure.
This is a synthetic-data template of the personal finance system I run with Claude. All merchants, names, and amounts here are fake; the structure is the real thing.
taxonomy.yaml # scopes + categories (the vocabulary)
| # Spawn codex's app-server on a free port and attach a TUI to it via --remote. | |
| # By default, codex's TUI bundles its own app-server. Splitting them lets other | |
| # TUIs or tools plug into the same running session. | |
| # | |
| # Usage: | |
| # cr # random free port, server dies with TUI | |
| # cr --keep # leave the server running after TUI exits | |
| # cr 8765 # use port 8765 (start server if not already up) | |
| # cr ws://host:8765 # attach TUI to an already-running app-server | |
| codex-remote() { |
Source: Codex VSCode session (~/.codex/sessions/2026/04/25/rollout-...jsonl).
Date: 2026-04-25 (single session, ~7h elapsed wall clock — ~2h to first working build on phone). Turns: 18 user prompts.
Filtered: user prompt + final assistant response per turn (running status commentary stripped).
A two-line statusline for Claude Code showing:
Line 1 — working dir, git branch, context window bar with breakdown (system / messages / free / autocompact buffer), per-call token buckets (new / cache write / cache read / output), session totals, lines changed.
Line 2 — 5-hour rate-limit %, 7-day rate-limit %, session $ cost with per-bucket breakdown, 5h-window $ total, 7-day $ total, wall-clock time.
| #!/bin/bash | |
| input=$(cat) | |
| # One-shot debug dump: write the latest received JSON so we can verify what | |
| # Claude Code actually sends. Safe to leave on; it's a single file overwrite. | |
| echo "$input" > /tmp/claude-statusline-last.json | |
| # ── Colors ──────────────────────────────────────────────────────── | |
| RESET="\033[0m" | |
| GREEN="\033[38;2;74;222;128m" | |
| YELLOW="\033[33m" |
| #!/usr/bin/env python3 | |
| """ | |
| cc_cost — quick token / $-cost queries against ~/.claude/projects. | |
| For LIVE current-session info, prefer the authoritative numbers from | |
| Claude Code's statusline JSON (`cost.total_cost_usd`, `rate_limits.*`). | |
| This script is for HISTORICAL queries (per-day, per-window, since-X) that | |
| Claude Code doesn't surface directly. | |
| Usage: |
| #!/usr/bin/env python3 | |
| """ | |
| cc_cost — quick token / $-cost queries against ~/.claude/projects. | |
| For LIVE current-session info, prefer the authoritative numbers from | |
| Claude Code's statusline JSON (`cost.total_cost_usd`, `rate_limits.*`). | |
| This script is for HISTORICAL queries (per-day, per-window, since-X) that | |
| Claude Code doesn't surface directly. | |
| Usage: |
User: @dui_toledo (Max 20x / $200 plan)
Models in dataset: claude-opus-4-5 (Jan 20 – Feb 5) → claude-opus-4-6 (Feb 6 → present), reasoning effort high throughout. Both share the same price tier ($5 / $6.25 / $0.50 / $25 per MTok).
Timezone: UTC-3 (BRT, São Paulo)
Original report: 2026-04-06 (since corrected — see below)
Latest revision: 2026-04-07 evening (added $-cost view, then found a second over-count bug — duplicate streaming chunks + cross-file replication. Numbers are now ~14x lower than the prior revision.)