v0.2.0

Added

• tripline init installs the hook for you. It covers Claude Code (~/.claude/settings.json, or --project) and Codex CLI (--agent codex, ~/.codex/hooks.json). The rest of the file is preserved, the previous version is kept as .bak, and the write is atomic. Running it again updates the entry. --remove uninstalls.
• Codex CLI support. tripline hook --agent codex guards Codex 0.124+ through its PreToolUse hook. scan and blame read Codex sessions, including apply_patch edits.
• Cursor support in scan and blame, read from ~/.cursor/projects/*/agent-transcripts. Cursor records no timestamps or results, so times come from the file.
• Per-rule actions. --ask RULES lets you decide through the agent's permission prompt. --warn RULES lets the call run and tells the agent why it looks wrong.
• Shell-edit detection. Commands that write files (sed -i, > redirects, tee, mv, git checkout, formatters, package managers, PowerShell Set-Content, …) now show up in blame as (inferred), and reset the loop and reread rules.

Changed

• budget fires once at each multiple of the budget ($5, $10, …). Previously it blocked every call after the first time the budget was reached, so the session couldn't continue.
• scan shows cost as n/a for models without known prices (such as GPT models), instead of pricing them as Sonnet.
• scan without a file picks the most recent session across all supported agents.

Fixed

• loop no longer fires when the agent edits files through shell commands between test runs.
• The hook strips a UTF-8 byte order mark from its input. Some Windows hosts send one, and it made the hook fail open silently.
• The pending call's token usage is no longer dropped when estimating the budget.
• go install builds report their real version.

Not yet

• Cursor hook: Cursor's response semantics for allow and ask couldn't be verified. Guessing wrong would either auto-approve commands or block everything.
• opencode: it stores sessions in SQLite since v1.2, which needs a new dependency or its CLI.