v0.4.0

Highlights

Opt-in reviewer debug output capture and streaming (#15)

• New diffwarden review --debug-reviewer-output: CLI-transport reviewer artifacts gain a bounded debug_output transcript (separate stdout/stderr, 256 KiB per stream per reviewer, UTF-8-safe truncation; retried runs share the budget so failing first attempts survive). With --ndjson, bounded reviewer_debug_output events stream while reviewers run. Defaults are byte-identical without the flag; debug output is non-authoritative and never affects results, validation, gating, or exit codes.

Cross-engine structured-output repair pipeline (#14)

• Schema-parse failures now trigger a short same-engine repair request before falling back to one labeled full re-run; engine-reported spend is summed across invocations.

Interactive setup and reviewer management (#6, #7, #13)

• Host-aware reviewers discover plus interactive-by-default init and reviewers add/edit/remove/set flows (@clack/prompts) with atomic writes and concurrency guards.
• Interactive model catalog picker in setup, backed by live engine listModels.

Claude reviewer improvements (#8–#12)

• xhigh effort passthrough with catalog-aware top-tier clamp; diffwarden-default effort high on claude reviewers.
• Diffwarden review contract delivered as a system prompt on claude transports.
• First-class CLAUDE_CODE_OAUTH_TOKEN support.
• fallbackModel, maxTurns, and maxBudgetUsd reviewer settings.

Review UX

• Diff-backed focus review lanes (--focus), reshaped output modes with a human review display, and batch report aggregation fixes.

Full changelog: v0.3.2...v0.4.0