Pitboss

Pitboss

**Operating pitboss from an AI agent?** See [`AGENTS.md`](AGENTS.md) — schema reference, decision tree, canonical examples, and the rules for translating natural-language requests into valid manifests.

Documentation

Full operator guide, MCP tool reference, Security section (threat model, defense-in-depth, Rule of Two), approval policy reference, and a cookbook of working scenarios: **https://sds-mode.github.io/pitboss/**

To browse offline:

cd book/
cargo install mdbook  # one time
mdbook serve --open

**v0.17.0** is a **stabilization release** focused on budget enforcement correctness and parser coverage for Claude CLI 2.x. The headline fix (#602) closes a double-count in `compute_total_spend` for terminated sub-leads: pre-fix, a hierarchical run with reconciled sub-trees saw the live budget watcher trip at ~2× the actual run spend (a $1.65 run tripped a $4.00 cap at $4.71 in validation). Post-fix, the live watcher's view matches `summary.json::spend_breakdown.total_usd` byte-for-byte. **Parser support for extended-thinking content blocks** (#601) — real Claude CLI emits `{"type":"thinking", ...}` content blocks inside assistant messages under haiku-4-5 / sonnet-4-5 / opus-4-x; pre-fix these were silently dropped, and on thinking-only turns the line was discarded entirely (losing `Event::AssistantUsage` so mid-run budget enforcement saw stale spend). The fix surfaces `Event::AssistantThinking`, reorders `parse_assistant` so `AssistantUsage` always lands, and renders thinking content with a `~` prefix in `pitboss attach` / TUI. **Parser drift detector** (#599) — the parser now emits `tracing::warn!` on unknown stream-json `type` fields and unknown assistant content blocks, so future Claude CLI wire-format changes surface in production tracing the first time they fire (this very detector caught the thinking-block issue on the first live run after merge). **Concurrency hardening** — six medium-severity audit findings closed: nested `