v0.12.0

• Grok status line, trim and guard. limitcast install wraps [ui.status_line] in ~/.grok/config.toml. --guard writes ~/.grok/hooks/limitcast.json for spawn_subagent. limitcast trim --grok measures the system prompt and unused MCP servers (--apply / --undo like Codex). Grok stores spend, not a % used: the status line shows grok 5h $X · 7d $Y from usage.json; limitcast budget grok <usd> (Pro) turns that into a forecast the guard can block on.

• limitcast report --why now splits subagent tokens by type and model and shows all new tokens by model. A subagent's type comes from the .meta.json file next to its transcript. Models may count differently against your limit, so these are token counts, not window shares.

• limitcast report --jobs adds a cold-starts column per job and a "Cold starts" section: how many runs started without the prompt cache, how much more that cost than the same job's warm runs, and whether starting a job's runs closer together would help.

• New (Pro): limitcast report --jobs lists the scheduled (headless) jobs whose runs spend the most new tokens just starting. Runs of one job are grouped by the first 45 characters of their prompt, with runs, average new tokens at start, average starting prompt, share of all new tokens and calls per run, plus how to launch a job leaner. report --why points to it when there are headless runs.

• New (Pro, install --alerts): a PostToolUse hook tells you, once per tool per session, when a single tool result is very large (about 20k tokens or more, from its size; LIMITCAST_TOOL_TOKENS changes the threshold), with what to ask for next time. It runs on Bash, Read, Grep, WebFetch and MCP tools, shows the message to you only, and adds about 30 ms per call. Re-run limitcast install --alerts (or doctor --fix) to add it to an existing install.

• limitcast report --why ends with "What to try": how many single steps added over 20k new tokens and what share of tool-result tokens they were, and advice for the sources whose results average 2k+ new tokens.

• New (Pro, with the guard): before a subagent starts you see what that kind has cost you, e.g. "Explore subagents have cost you a median ≈1.2% (max ≈3.6%) of a 5-hour window over 7 runs; this window is 38% used". It appears once there are 3 priced runs, uses all your subagents when that kind has fewer, and stays quiet when a typical run is under 1% of a window and the window is not NEAR. It never blocks.

• New: limitcast report --why shows where your new tokens went (subagents, session starts, cache rebuilds after idle, tool results by tool, model output) and the largest single steps. It reads your Claude Code transcripts once and caches them, so only the first run over a few GB is slow (3 to 40 seconds, depending on whether the disk cache is warm) and later runs only read what changed.

• limitcast report now ranks projects, models and sessions by new tokens (input, cache writes, output) instead of total tokens, so cached reads no longer decide the order.