Skip to content

Release v2.1.0

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 20 May 04:24
· 3 commits to main since this release
v2.1.0
4833c7d

Release Notes v2.1.0

Released: 2026-05-20

This release introduces a critic verification pass that validates generated artifacts against repository evidence, structured review findings with GitHub inline publishing, and a suite of new analysis tools. The focus is on accuracy, cost efficiency through Anthropic prompt caching, and deeper codebase exploration.

✨ Highlights

Critic Verification Pass

New verify capability reviews generated artifacts (commits, PRs, changelogs, reviews) for unsupported claims, incorrect citations, and scope overstatement. When issues are found, Iris regenerates with guidance. Material errors like critical severity issues trigger automatic revision; style or wording differences are ignored.

Structured Review Findings with GitHub Inline Comments

Reviews now return typed Finding objects with severity, confidence, category, file location, and evidence references. Findings gate at 70% confidence (configurable via DEFAULT_MIN_FINDING_CONFIDENCE). Publishing to GitHub places inline comments directly on PR diff lines with permalinks in the review body.

Repository Map for Codebase Orientation

New repo_map tool builds a compact, ranked view of source files with definitions and imports. Scoring prioritizes mentioned files (+1000), recently changed files (+250), and shallow paths. Token budget clamps to 50-8000 tokens; max files caps at 200.

Anthropic Prompt Caching

Automatic caching now applies to both main agent and subagents. Cached input tokens bill at ~90% discount, cutting costs for multi-turn tool loops where the same context/system prompt is re-sent.

Configurable Subagent Turn Budget

Parallel subagents now accept a max_turns argument (1-100) to cap tool loops. Precedence: per-invocation argument, then tool instance config, then default of 20. Failures now return descriptive error messages instead of empty strings.

🔧 New Tools

  • repo_map: Ranked codebase overview with definitions/imports extracted from 14 file types. Respects .gitignore, excludes files >400KB. Parameters: token_budget (default 2000), max_files (default 60), mentioned_files for explicit prioritization.
  • static_analysis: Runs installed linters (clippy, ruff, biome/oxlint, golangci-lint/go vet) without package installation. Auto-detects project type from config files. Parameters: analyzer, timeout_secs (1-600), max_output_chars (512-40000).
  • git_show: Inspects historical commit metadata, stats, and patch. Parameters: commit (required), files (optional filter), max_output_chars (1000-50000).
  • git_blame: Line-level history with author, date, and summary. Also fetches recent commits touching the file. Parameters: file, start_line, end_line, recent_commits (1-10).

🤖 Agent Framework

  • Added verify.toml capability with severity levels: critical, high, medium, low
  • Critique struct returns requires_revision, issues[], revision_prompt, and confidence
  • Critic executes only for supported output types: GeneratedMessage, Review, MarkdownPullRequest, MarkdownChangelog, MarkdownReleaseNotes
  • Graceful degradation: critic failures skip verification and return original artifact

🔎 Review Improvements

  • Review struct now includes metadata (risk level, strategy, specialist passes), findings[], and stats
  • Finding fields: id, severity, confidence, file, start_line, end_line, category, title, body, suggested_fix, evidence[]
  • Flexible confidence parsing handles integer, float, and percentage string formats
  • ReviewMetadata surfaces the agentic review strategy for transparency
  • visible_findings() filters to 70%+ confidence by default

🐙 GitHub Integration

  • publish_structured_review() posts inline comments on PR diff lines
  • Validates comment candidates against reviewable lines before submission
  • Review body includes GitHub permalinks section linking to each finding location
  • Multi-line findings use start_line/end_line ranges in GitHub API

⚡ Performance & Cost

  • Anthropic with_automatic_caching() applied via shared anthropic_agent_builder()
  • CompletionProfile differentiates OpenAI reasoning effort: MainAgent (medium), Subagent (low), StatusMessage (none). This applies only to OpenAI models with reasoning effort support.
  • Parallel subagent errors now populate SubagentResult.error with details instead of silent failures
  • ParallelAnalyzeResult includes successful and failed counts for visibility

⬆️ Dependencies

  • Bumped rig-core to 0.37 (crate renamed to rig in imports)
  • Edition updated to Rust 2024

📝 Documentation

  • Refreshed docs covering GitHub publishing, --amend flag, XDG config paths, Homebrew installation
  • Added ROADMAP.md replacing the companion design document
  • Updated SKILL.md with branch usage patterns
  • Expanded architecture docs for new tools, critic pass, and studio refactor

Upgrade Notes

  • The rig-core crate is now imported as rig. Update any direct Rig dependencies if extending Git-Iris.
  • Critic verification is enabled by default. Disable with critic_enabled: false in config if you prefer no revision passes.
  • Review findings below 70% confidence are hidden by default. Adjust DEFAULT_MIN_FINDING_CONFIDENCE if needed.
  • Subagent turn budget defaults to 20. For broad searches, increase via max_turns argument; for cost control, decrease it.
Assets 9
Loading