v0.2.0

Upgrade note. aigrd can now fail where 0.1.0 passed. A misconfigured
judge exits 1 instead of warning. Fix the judge, or set
on_unavailable = "warn" under [judge] to keep the old behaviour.

Changed

• A judge that could not run no longer reports a pass. A claude that ran
  and returned no usable verdict is now judge.failed, an error. It exits
  1 and blocks the Claude Code turn. Version 0.1.0 warned and exited 0,
  so a broken judge passed every file it was meant to check.
• judge.failed covers a crash, a timeout, spending past
  max_budget_usd, an is_error result, and a verdict outside the
  schema. A file aigrd cannot read joins it.
• judge.unavailable now means one thing: claude is not on PATH. It
  stays a warning and never blocks.
• A judge.failed file takes an attempt, so [judge].max_attempts bounds
  it. Three stops name the cause, then judge.gave-up releases the turn.
  A broken judge cannot trap the agent.

Added

• [judge].on_unavailable, either "error" (the default) or "warn".
  "warn" restores the 0.1.0 soft skip. Any other value is cfg.invalid
  and exits 2.

Fixed

• aigrd init and aigrd docs gave no range for max_thinking_tokens.
  The claude command-line tool accepts every value and raises a cap below
  1024 to 1024. Extended thinking is billed as output and spends against
  max_budget_usd, so a higher cap can exhaust the budget and lose the
  verdict. Both documents now say so.

v0.1.0

Added

• aigrd judge: one pass or fail per rubric criterion, with a quoted span
  as evidence and a fix note, from a second model run through the claude
  command-line tool.
• --harness claude: the Claude Code Stop-hook decision, with bounded
  attempts per session and file and a visible give-up.
• aigrd init, rubric derive, rubric check, rules and docs.
• aigrd hooks install, the 40-aigrd pre-commit fragment, and
  hooks claude.
• The run log .aigrd/runs.jsonl and aigrd runs; aigrd label, which
  writes portable evaluation cases beside the rubric; aigrd agreement.
• A verdict cache keyed on the file content, the rubric bytes and the
  judge command.
• [judge].max_thinking_tokens, a cap on extended thinking in the judge.
• Three example projects: an article, a landing page and an investment
  thesis, each with a rubric and a passing and a failing sample.