v0.11.0

Features

• Preflight now checks your toolchain, not just your CLI provider. styre run fails fast, before any spend, if a command your repo's build/test/check steps need isn't installed, and the design phase now runs provision first so environment problems surface before design work begins.
• Missing or outdated claude/codex CLIs are now caught immediately. If your configured agent CLI is missing or below the version styre requires, styre run, styre setup, and --resume all fail fast with a clear message telling you to install or upgrade it, instead of retrying three times and failing with a cryptic error.
• Run outcomes are easier to read. styre run now leads with a plain-language outcome sentence, always prints the PR URL when there is one, names the pending signal when a ticket is waiting on you, and shows loopback history as a readable timeline instead of raw internal terms. Errors render as a single clear message with a suggested fix instead of a stack trace.
• Escalations to a human are now reported distinctly from dead ends. A run that hands off to a person now shows the outcome escalated (exit code 75) everywhere — the terminal summary, Slack, and analytics — instead of being lumped in with blocked.
• Telemetry now includes run identity and honest cost totals. Every exported event carries a stable run_id so you can tell two runs of the same ticket apart (and --resume keeps the same id), summary cost/token fields are null when nothing reported them (instead of a misleading 0), and a new usage_coverage field tells you how complete a summary total is. This is a breaking change to the NDJSON export schema (v1 → v2).
• styre notify and legibility docs. Documentation now fully covers the checks/arbiter subsystem, all config keys, exit codes, environment variables, and prompt templates, so the reference material actually matches shipping behavior.

Bug Fixes

• A missing or broken test launcher no longer counts as a passed check. If the test framework can't even start (exit code 126/127, e.g. the launcher isn't installed), styre no longer marks the acceptance criterion as covered — it's retried loudly instead.
• A check that errors with no output no longer counts as covered. Previously a check that couldn't run at all (no test framework detected, no interpreter, or a timeout) and produced no output was silently marked as covering its acceptance criterion; it's now routed to a loud retry with a reason.
• Discarded-file detection now works across more languages. The guard that catches a check silently broken by an undeclared, discarded file now understands Go, Rust, JVM, Ruby, and PHP output (previously Python/Node only), and Rust/Ruby/PHP symbol matching is now anchored so a test's own error text can't be mistaken for a real compiler error.
• Discarded Python support files now produce a clear message. If a discarded __init__.py or conftest.py breaks test collection, styre now names the real cause instead of silently installing a broken check.
• Telemetry events can now be joined back to the dispatch that caused them. loopback and escalated events now carry the originating dispatch_id, which was previously always null.
• Early failures (bad config, wrong adapter, running outside a git repo) are now counted in analytics. Previously only failures after config was fully resolved were recorded, so early exits were invisible to usage tracking.