ace-review 0.49.0

data/CHANGELOG.md

@@ -0,0 +1,1780 @@
+# Changelog
+
+All notable changes to ace-review will be documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [Unreleased]
+
+## [0.49.0] - 2026-03-24
+
+### Fixed
+- Added missing `ace-review-feedback` executable to gemspec — gem consumers now get both CLIs installed.
+
+### Changed
+- Removed stale task-symlink documentation from feedback workflow guide (task-linking was removed in v0.46.0).
+- Updated stale code comment in review manager referencing removed symlink architecture.
+- Re-recorded getting-started demo with live review execution.
+
+## [0.48.1] - 2026-03-23
+
+### Changed
+- Refreshed the package README overview, section flow, and quick links to align with the current ACE package layout pattern.
+
+## [0.48.0] - 2026-03-18
+
+### Changed
+- Refined `TS-REVIEW-001` E2E verifier goal structure with explicit impact-first validation and check sections across all review workflow goals.
+
+## [0.47.1] - 2026-03-18
+
+### Changed
+- Migrated CLI namespace from `Ace::Core::CLI::*` to `Ace::Support::Cli::*` (ace-support-cli is now the canonical home for CLI infrastructure).
+
+
+## [0.47.0] - 2026-03-18
+
+### Changed
+- Removed legacy backward-compatibility behavior as part of the 0.10 cleanup release.
+
+
+## [0.46.1] - 2026-03-18
+
+### Fixed
+- Hardened feedback synthesis to preserve raw output, clean common malformed JSON, and run a bounded JSON-repair pass before failing extraction.
+- Switched the default feedback synthesis model to `gemini:flash-latest@ro` so synthesis uses the Gemini CLI read-only provider path.
+
+### Technical
+- Isolated review-driven synthesis artifacts under a dedicated `feedback-synthesis` session subdirectory to avoid prompt-file collisions with the parent review session.
+- Added regression coverage for malformed JSON cleanup/repair and review-manager session propagation during feedback extraction.
+
+## [0.46.0] - 2026-03-18
+
+### Changed
+- Removed `--task` and `--no-auto-save` CLI options and all review-to-task linking integration (symlink creation, branch-based task detection, `auto_save` config).
+- Removed `task_reference` state, `link_session_to_task`, `link_session_to_task_if_requested`, `auto_link_session_if_enabled`, and dead-code task feedback methods from `ReviewManager`.
+- Simplified `PrTaskSpecResolver.extract_from_branch` to use inline branch-prefix extraction instead of external `TaskPatternExtractor`.
+
+### Technical
+- Removed 15 task-linking tests and cleaned up stale `@task_reference` references in remaining tests.
+
+## [0.45.4] - 2026-03-18
+
+### Fixed
+- Fixed review CLI option precedence so `auto_execute` now uses config default when the flag is omitted and `--dry-run` remains authoritative.
+
+## [0.45.3] - 2026-03-17
+
+### Fixed
+- Fixed local PR diff fallback to fetch and compare against the PR head ref instead of caller `HEAD`.
+- Unified review model timeout handling and raised the default to 900 seconds.
+
+### Technical
+- Exposed structured Claude CLI response metadata when successful subprocess output is empty.
+- Added regression coverage for timeout forwarding and PR diff fallback behavior.
+
+## [0.45.2] - 2026-03-17
+
+### Fixed
+- Updated Claude context limits from 200K to 1M tokens in hardcoded fallback table.
+- Added GPT-5.x (1.05M) and o4 (1.05M) model patterns to context limit resolver.
+- Raised prompt size warning threshold from 160K to 800K tokens to eliminate spurious warnings with modern LLMs.
+- Updated default context limit for unknown models from 128K to 200K tokens.
+
+## [0.45.1] - 2026-03-17
+
+### Changed
+- Removed hardcoded `sandbox` injection from review CLI execution so review model presets control CLI flags consistently.
+- Standardized `code-valid`, `code-fit`, and `code-shine` presets to use `codex:gpt@ro` and moved thinking/behavior configuration into preset configuration.
+
+## [0.45.0] - 2026-03-15
+
+### Added
+- `DiffTooLargeError` for detecting GitHub API 406 responses on large PR diffs
+- Local git diff fallback in `GhPrFetcher` when GitHub API rejects diffs exceeding 300-file limit
+- Connection error retry guidance in PR review workflow (wait 30s, retry up to 2x)
+
+### Technical
+- New `GhPrFetcherTest` covering 406 detection, fallback success, and fallback error paths
+
+## [0.44.5] - 2026-03-15
+
+### Changed
+- Migrated CLI framework from dry-cli to ace-support-cli
+
+## [0.44.4] - 2026-03-13
+
+### Technical
+- Updated canonical review skills to rely on workspace workflow execution flow.
+
+## [0.44.3] - 2026-03-13
+
+### Changed
+- Updated canonical review skills to explicitly run bundled workflows in the current project and execute them end-to-end.
+
+## [0.44.2] - 2026-03-13
+
+### Changed
+- Updated the review E2E error-handling scenario to distinguish real dry-run validation failures from invalid-model dry-run preparation behavior.
+
+## [0.44.1] - 2026-03-12
+
+### Changed
+- Updated review workflow instructions to reference bundle-first review flows instead of slash-command examples.
+
+## [0.44.0] - 2026-03-10
+
+### Added
+- Added canonical handbook-owned review skills and the new `wfi://review/package` workflow for package-level review.
+
+
+## [0.43.8] - 2026-03-08
+
+### Fixed
+- Resolve review preset/config discovery from the caller's explicit `project_root` and gem-default preset files, preventing temp-dir context extraction from crashing after project-root hardening in `ace-support-fs`.
+
+### Technical
+- Add regression coverage for gem-default preset discovery and temp-dir file-path context extraction.
+
+## [0.43.7] - 2026-03-06
+
+### Fixed
+- Make `--dry-run` disable automatic model execution even when project defaults enable `auto_execute`.
+- Allow `@` in review model identifiers so preset models can use `@ro`, `@rw`, and `@yolo` suffixes.
+
+### Changed
+- Reduce the shipped review preset surface to five explicit single-file presets: `code-valid`, `code-fit`, `code-shine`, `docs`, and `spec`.
+- Align both package defaults and repo-level `.ace/review` overrides on the explicit five-preset review model.
+
+## [0.43.6] - 2026-03-05
+
+### Changed
+- Renamed config key `auto_save_branch_patterns` → `task_branch_patterns` to reflect broader usage (task spec resolution, not just auto-save)
+
+## [0.43.5] - 2026-03-04
+
+### Fixed
+- Feedback session discovery now prefers `.ace-local/review/sessions` with legacy `.cache` fallback.
+
+
+## [0.43.4] - 2026-03-04
+
+### Fixed
+- Review workflow instructions (apply-feedback, pr, run) corrected to use `.ace-local/review/` (not `.ace-local/ace-review/`)
+
+## [0.43.3] - 2026-03-04
+
+### Fixed
+- `feedback-workflow.md` session path examples corrected to short-name convention (`.ace-local/review/` not `.ace-local/ace-review/`)
+
+## [0.43.2] - 2026-03-04
+
+### Fixed
+- Reverted feedback synthesis workspace handling to standard `Dir.mktmpdir` behavior instead of project-local temporary workspaces.
+
+## [0.43.1] - 2026-03-04
+
+### Fixed
+- README session storage path examples corrected to short-name convention (`.ace-local/review/` not `.ace-local/ace-review/`)
+
+## [0.43.0] - 2026-03-04
+
+### Changed
+- Default session directory migrated from `.cache/ace-review/sessions` to `.ace-local/review/sessions`
+- Feedback synthesis workspace now uses `Ace::Support::Items::Atoms::TmpWorkspace` for deterministic `.ace-local/tmp` paths
+
+## [0.42.7] - 2026-03-02
+
+### Changed
+- Replace `ace-taskflow` dependency with `ace-task` — migrate `TaskResolver`, `SubjectExtractor`, and `PresetManager` to ace-task APIs
+- Remove dead `save_to_release` method from `TaskReportSaver` (used `ReleaseManager` from ace-taskflow, never called in production)
+- Update CLI subprocess call from `ace-taskflow task REF --path` to `ace-task show REF --path`
+
+## [0.42.6] - 2026-02-25
+
+### Technical
+- Bump runtime dependency constraint from `ace-git ~> 0.10` to `ace-git ~> 0.11`.
+
+## [0.42.5] - 2026-02-25
+
+### Technical
+- Add test case for `nil` return from `extract_task_reference` when PR metadata contains no task reference
+
+## [0.42.4] - 2026-02-25
+
+### Fixed
+- Safely extract task reference from PR text using safe navigation to avoid NoMethodError when regex match returns nil
+
+## [0.42.3] - 2026-02-25
+
+### Added
+- Automatically include the task behavioral spec (`.s.md`) in PR review context when a task can be discovered from PR branch metadata or PR text references.
+- Add `PrTaskSpecResolver` to map PR metadata to a single primary task spec file with graceful fallback.
+
+### Changed
+- Extend task resolution metadata with `spec_path` for direct behavioral spec selection.
+- Extend PR metadata fetch fields to include PR `body` for task reference fallback detection.
+
+## [0.42.2] - 2026-02-24
+
+### Technical
+- Strengthen TS-REVIEW-001 preset-composition E2E runner/verifier instructions by requiring explicit dry-run subject input and resilient artifact-based verification criteria.
+
+## [0.42.1] - 2026-02-23
+
+### Changed
+- Centralized ContextComposerError and UnknownStrategyError into Ace::Review::Errors module
+- Narrowed exception handling in feedback_file_reader, feedback_file_writer, and task_report_saver
+
+### Technical
+- Updated internal dependency version constraints to current releases
+
+## [0.42.0] - 2026-02-22
+
+### Changed
+- **Breaking:** Migrated from multi-command Registry to single-command pattern (task 278)
+  - Removed `review` subcommand: `ace-review review --preset pr` → `ace-review --preset pr`
+  - Removed `list-presets` subcommand: `ace-review list-presets` → `ace-review --list-presets`
+  - Removed `list-prompts` subcommand: `ace-review list-prompts` → `ace-review --list-prompts`
+  - Removed `version`/`help` subcommands: use `--version`/`--help` flags only
+  - Added `--version`, `--list-presets`, `--list-prompts` flags to main command
+  - Simplified `preprocess_array_options` (no command name detection needed)
+  - No backward compatibility (per ADR-024)
+
+## [0.41.2] - 2026-02-22
+
+### Fixed
+- Add flag variants (`--help`, `-h`, `--version`) to KNOWN_COMMAND_NAMES for preprocessing safety
+
+## [0.41.1] - 2026-02-22
+
+### Fixed
+- Include built-in help/version commands in KNOWN_COMMAND_NAMES for correct array option preprocessing
+
+## [0.41.0] - 2026-02-22
+
+### Changed
+- Migrate ace-review-feedback CLI to standard help pattern (task 278.13)
+- Remove DWIM default routing — no args now shows help instead of running list command
+
+## [0.40.5] - 2026-02-22
+
+### Fixed
+- Migrate to standard help pattern with explicit `review` subcommand
+- Remove DWIM default routing — no args now shows help instead of running review
+
+## [0.40.3] - 2026-02-22
+
+### Changed
+- Migrate skill naming and invocation references to hyphenated `ace-*` format (no underscores).
+
+## [0.40.2] - 2026-02-19
+
+### Technical
+- Namespace workflow instructions into review/ subdirectory with updated wfi:// URIs
+
+## [0.40.1] - 2026-02-19
+
+### Fixed
+- Architecture-reflection preset diff range changed from `HEAD~1..HEAD` to `origin/main..HEAD` to capture all implementation commits
+
+## [0.40.0] - 2026-02-19
+
+### Added
+- New `architecture-reflection` review preset for pre-PR self-assessment — combines ATOM compliance, over-engineering detection, and missing abstraction analysis
+- New `reflection` focus prompt (`focus/architecture/reflection.md`) with refactor/accept/skip categorization output format
+
+### Fixed
+- **Feedback extraction failures no longer silently swallowed**: When synthesis LLM call fails, error is now propagated to CLI output and response hash instead of being silently dropped
+- **CLI reports feedback status**: Multi-model review output now shows feedback item count on success or error message on failure
+
+### Added
+- **Fallback model support for feedback synthesis**: Configure `feedback.fallback_models` in config.yml to try alternative models when primary synthesis model fails
+- Default fallback model `claude:glm` added to config
+
+## [0.39.3] - 2026-02-17
+
+### Changed
+- Clarify review workflow execution guards to prevent premature follow-up actions:
+  - `review-pr.wf.md`: explicitly require process exit before feedback commands
+  - `review.wf.md`: explicitly require process exit before feedback commands
+- Standardize timeout handling guidance for review workflows (10-minute timeout with stop-on-timeout behavior)
+- Add explicit precondition + session-targeted feedback examples to reduce session ambiguity in feedback listing
+
+## [0.39.2] - 2026-02-16
+
+### Added
+- **Phased review presets**: `code-valid` (correctness), `code-fit` (quality), `code-shine` (polish) for laser-focused review cycles
+- **Backward-compatible** `code-deep` preset as composition of `code` with detailed format
+- **Focus prompts**: `phase/correctness.md`, `phase/quality.md`, `phase/polish.md` with explicit scope and DO NOT Review sections
+- Project-level preset configuration files for monorepo discovery
+
+### Changed
+- **Multi-model reviewers**: All 3 phased presets use `claude:opus`, `codex:max`, `gemini:pro-latest`
+- **Focus distribution**: All 6 focus modules from `code` preset distributed across the 3 phased presets (security→code-fit, tests+ruby→code-fit, docs→code-shine)
+
+## [0.38.1] - 2026-02-15
+
+### Changed
+- **LlmExecutor**: Pass `sandbox: "read-only"` to `QueryInterface.query()` to enforce non-agentic mode for CLI providers during reviews
+
+## [0.38.0] - 2026-02-08
+
+### Added
+- Unified feedback verification workflow with single `verify` command supporting three modes
+- `--skip` option to `feedback verify` for marking items as skipped (correct but not fixing)
+- Support for skipping both draft and pending status items (previously skip only worked on draft)
+- Enhanced documentation with clear examples for when to use --valid, --invalid, or --skip
+- Comprehensive test coverage for --skip mode and backward compatibility
+
+### Changed
+- `FeedbackManager.verify()` now supports `skip:` mode in addition to `valid:` parameter
+- `FeedbackManager.skip()` now delegates to `verify()` with `skip: true` (backward compatible)
+- Enhanced verification workflow documentation with clear decision criteria
+- Improved CLI help text with examples for each verification mode
+
+### Technical
+- Added tests for skip mode transition from both draft and pending states
+- Added tests for mutually exclusive validation in verify command
+- Ensured backward compatibility for existing skip command usage
+
+### Changed
+- Embed CLI help reference in apply-feedback and verify-feedback workflows via bundle frontmatter
+- Agents now see exact flag syntax per subcommand (verify/resolve/skip) when workflows load
+
+## [0.37.3] - 2026-02-04
+
+### Added
+- Priority range filtering with `+` suffix for `feedback list` command
+- `--priority medium+` filters for medium, high, and critical items
+- `--priority high+` filters for high and critical items
+- New `PriorityFilter` atom for priority filtering logic
+
+## [0.37.2] - 2026-02-04
+
+### Added
+- `SubjectFilter` molecule extracting file filtering logic from Reviewer model
+- README documentation for Subject Strategy Configuration (strategy types, adaptive mode)
+- README documentation for Reviewers Format (attributes, file patterns, migration)
+- `ContextLimitResolver` integration with ace-llm provider config for context limits
+
+### Changed
+- `Reviewer` model now delegates filtering to `SubjectFilter` molecule (ATOM architecture)
+
+### Technical
+- Comprehensive tests for `SubjectFilter` molecule
+
+## [0.37.1] - 2026-02-04
+
+### Fixed
+- Graceful error handling for preset composition failures (circular deps, missing refs)
+- E2E test MT-REVIEW-001 preset data now includes required `instructions:` key
+- E2E test MT-REVIEW-001 uses valid model provider format
+
+## [0.37.0] - 2026-02-04
+
+### Added
+- Token estimation foundation with `TokenEstimator` (chars/4 heuristic) and `ContextLimitResolver`
+- `DiffBoundaryFinder` atom for parsing unified diffs into file blocks
+- Subject strategy system with `SubjectStrategy` factory and registry
+- `FullStrategy` for complete subject pass-through when content fits context
+- `ChunkedStrategy` for file-boundary diff splitting of large subjects
+- `AdaptiveStrategy` for auto-selecting between Full/Chunked based on token estimate
+- `Reviewer` entity model with configurable focus, file patterns, and prompt additions
+
+### Changed
+- `FullStrategy` accepts `headroom` config option to customize context margin (aligns with `AdaptiveStrategy`)
+
+### Fixed
+- GPT-4 variant handling for 32k and preview models (gpt-4-32k, gpt-4-NNNN)
+- Extended diff header preservation during truncation (rename operations)
+- Chunk edge cases with non-positive available tokens
+
+## [0.36.17] - 2026-02-04
+
+### Added
+- `--session all` flag for `feedback list` to aggregate feedback from all sessions
+- `find_all_sessions` helper in `SessionDiscovery` module
+- SESSION column in table output when viewing all sessions
+- `session` field in JSON output when using `--session all`
+
+## [0.36.16] - 2026-02-03
+
+### Added
+- `SessionDiscovery` shared module for DRY session resolution across feedback CLI commands
+
+### Changed
+- Refactored 6 feedback CLI commands to use shared `SessionDiscovery` module
+- Updated CLI descriptions and error messages to remove obsolete `--task` references
+- Added warn statement for lock file cleanup failures in `FeedbackFileWriter`
+
+### Removed
+- 5 deprecated auto-save methods from `ReviewManager` (~150 lines of dead code):
+  - `save_multi_model_to_task`
+  - `auto_save_review_if_enabled`
+  - `auto_save_multi_model_reports`
+  - `auto_save_to_release`
+  - `save_to_task_if_requested`
+- 11 deprecated auto-save tests (~270 lines)
+
+### Fixed
+- Documentation referencing outdated `--no-synthesize` flag in README.md
+- Documentation referencing removed `--task` flag in feedback-workflow.md
+
+## [0.36.15] - 2026-02-03
+
+### Added
+- Feedback list shows archived count when no active items exist with helpful hint to use `--archived`
+- Status-based sorting for feedback list: draft → pending → done → skip → invalid (then by ID)
+
+### Changed
+- Feedback list summary now shows archived count hint when viewing active items only
+
+## [0.36.14] - 2026-02-03
+
+### Fixed
+- Feedback CLI commands failing from subdirectories - now uses `ProjectRootFinder` for session discovery
+
+## [0.36.13] - 2026-02-03
+
+### Fixed
+- Workflow documentation referencing removed `--task` CLI flag in session discovery
+
+### Changed
+- Improved feedback categorization guide in review workflows with clearer skip/defer semantics
+
+## [0.36.12] - 2026-02-03
+
+### Fixed
+- Non-unique feedback IDs when multiple items created in rapid succession within same millisecond
+
+### Added
+- `FeedbackIdGenerator.generate_sequence(count)` for batch unique ID generation using `encode_sequence`
+
+### Changed
+- `FeedbackSynthesizer.parse_synthesis_response` pre-generates unique sequential IDs for all findings
+- `FeedbackSynthesizer.create_feedback_item` accepts optional `id:` keyword argument
+
+## [0.36.11] - 2026-02-03
+
+### Fixed
+- Missing `feedback.synthesis_model` configuration in default config files
+
+## [0.36.10] - 2026-02-03
+
+### Fixed
+- JSON parsing for Claude Opus responses that include text before JSON code fence
+- `FeedbackSynthesizer.parse_synthesis_response` now handles "Based on my analysis..." preamble
+
+### Technical
+- Extracted `extract_json_from_response` helper for robust JSON extraction from LLM responses
+- Added test case for text-before-JSON-fence pattern
+
+## [0.36.9] - 2026-02-03
+
+### Removed
+- `extraction_model` config option (legacy alias) - use `synthesis_model` only
+- `extract-feedback.system.md` prompt - unified into `synthesize-feedback.system.md`
+
+### Changed
+- Feedback model configuration simplified to single `synthesis_model` option
+- `FeedbackSynthesizer.default_synthesis_model` cascade simplified to remove `extraction_model` fallback
+- `ReviewManager.extract_feedback` cascade simplified to use `synthesis_model` only
+
+### Technical
+- Removed ~140 lines of legacy prompt code
+- Simplified configuration cascade - reduces confusion from duplicate config options
+
+## [0.36.8] - 2026-02-03
+
+**BREAKING CHANGES**: Feedback CLI now session-scoped only (--task option removed)
+
+### Removed
+- `FeedbackContextResolver` molecule - task-based path resolution no longer needed
+- `--task` option from all feedback commands (create, list, show, verify, skip, resolve)
+
+### Changed
+- **Feedback is now session-scoped only** - all commands use `--session` flag with latest session default
+- Simplified path resolution across all feedback commands to use consistent session-based pattern
+- Updated test helper to remove `with_context_resolver_stub` helper
+- All feedback command tests updated to use `session:` option directly
+
+### Technical
+- Removed ~546 lines of overengineered task-based resolution code
+- Simplified feedback path resolution to: explicit session → latest session cache
+
+### Fixed
+- Documentation drift: ID format updated from "10-char" to "8-char" in feedback-workflow.md
+- Test base class violations: FeedbackFileWriter, FeedbackFileReader, FeedbackDirectoryManager tests now inherit from AceReviewTest
+- Missing trailing newline in .ace-defaults/review/config.yml
+
+## [0.36.6] - 2026-02-03
+
+### Removed
+- `FeedbackDeduplicator` atom - superseded by LLM-based deduplication in FeedbackSynthesizer
+- `FeedbackExtractor` molecule - replaced by FeedbackSynthesizer which handles multi-report synthesis
+
+### Technical
+- Removed ~600 lines of dead code and tests that were not used in the current feedback architecture
+
+## [0.36.5] - 2026-02-03
+
+### Added
+- Prompt size warning before LLM execution - warns when prompt exceeds 80% of typical context window (~160K tokens)
+- `--session` flag for `feedback list` command - allows explicit session directory path
+- Documentation for session-scoped feedback context in workflow instructions
+
+### Fixed
+- Lock file (.feedback.lock) now cleaned up after writes - prevents clutter in feedback directories
+
+### Technical
+- Added LlmExecutor and MultiModelExecutor prompt size warning with formatted token count
+- Updated FeedbackFileWriter with ensure block for lock file cleanup
+- Added tests for prompt size warning, lock file cleanup, and --session flag
+
+## [0.36.4] - 2026-02-03
+
+**BREAKING CHANGES**: Feedback-first architecture replaces synthesis-report.md output
+
+### Changed
+- **Feedback is now the primary output** - synthesis-report.md has been removed entirely
+- FeedbackIdGenerator now uses 8-char millisecond timestamp format (was 10-char with random suffix)
+- `should_extract_feedback?` always returns true if results exist (no config toggle)
+- Workflow instructions updated to use feedback commands instead of synthesis
+
+### Removed
+- **ReportSynthesizer molecule** - no longer needed with feedback-first architecture
+- `ace-review synthesize` CLI command removed
+- `--synthesize`, `--no-synthesize`, `--synthesis-model` CLI flags removed
+- `synthesis:` config section removed from config files
+- `feedback.enabled` config removed (feedback always runs)
+- `synthesize`, `no_synthesize`, `synthesis_model` options from ReviewOptions
+
+### Technical
+- Updated review.wf.md and review-pr.wf.md with feedback verification workflow
+- Simplified ReviewManager by removing synthesis-related methods
+- Tests updated for new 8-char feedback IDs and removed synthesis tests
+
+## [0.36.3] - 2026-02-03
+
+### Fixed
+- Synthesis now disabled by default (feedback files are primary output per task 227 spec)
+- Added `--synthesize` flag to opt-in to synthesis-report.md generation
+- Added explicit `feedback.enabled: true` default to config
+
+### Changed
+- `should_synthesize?` now defaults to false, checks for --synthesize CLI flag
+- ReviewOptions now includes :synthesize attribute
+
+### Technical
+- Updated default config to clarify synthesis is opt-in
+
+## [0.36.2] - 2026-02-03
+
+### Added
+- FeedbackSynthesizer molecule for multi-reviewer consensus analysis
+- Support for multiple reviewers per FeedbackItem (reviewers array format)
+- Consensus detection: marks items agreed upon by 3+ models
+- Synthesize feedback prompt template at handbook/prompts/synthesize-feedback.system.md
+
+### Changed
+- FeedbackItem now stores multiple reviewers with consensus flag
+- FeedbackFileWriter handles new multi-reviewer format
+- FeedbackManager integrates synthesis workflow
+- ReviewManager adds synthesis step for multi-model reviews
+
+### Technical
+- Added FeedbackSynthesizer with LLM-based consensus detection
+- Updated tests for multi-reviewer workflow
+
+## [0.36.1] - 2026-02-03
+
+### Changed
+- **Session Symlink Architecture**: Task reviews now symlink to session directories instead of copying files
+  - Multiple review sessions can be linked to the same task
+  - All session artifacts (prompts, metadata, feedback) accessible via symlink
+  - Feedback stays in session directory (gitignored via .cache)
+  - New methods: `link_session_to_task`, `link_session_to_task_if_requested`, `auto_link_session_if_enabled`
+  - Deprecated: `save_to_task_if_requested`, `save_multi_model_to_task`, `save_synthesis_to_task`
+
+### Technical
+- Simplified `determine_feedback_path` to always return session directory
+- Updated feedback-workflow.md documentation for symlink architecture
+
+## [0.36.0] - 2026-02-03
+
+### Added
+- **Feedback System**: Replace monolithic synthesis reports with tracked feedback items
+  - `FeedbackItem` model with full lifecycle support (task 227.01)
+  - File I/O with atomic writes and locking (task 227.02)
+  - LLM-based feedback extraction with deduplication (task 227.03)
+  - `FeedbackManager` organism for lifecycle orchestration (task 227.04)
+  - Integration with review pipeline via `--no-feedback` flag (task 227.05)
+  - CLI commands: `feedback list`, `show`, `verify`, `skip`, `resolve` (task 227.06)
+  - Task integration with automatic directory creation (task 227.07)
+  - Single-model feedback extraction (previously only multi-model supported)
+  - Documentation at `docs/feedback-workflow.md` (task 227.08)
+
+### Fixed
+- Use blocking lock for reliable concurrent feedback file access
+- Expand branch pattern to match `task-` prefix for auto-save
+- Merge files on deduplication instead of discarding
+- Use `extraction_model` config key for consistency
+- Use dynamic distance threshold for deduplication
+- Use dedicated lock file for concurrent write coordination
+- Add random suffix to FeedbackIdGenerator for uniqueness
+
+### Technical
+- Update ID format documentation from 6-char to 10-char
+
+## [0.35.6] - 2026-02-02
+
+### Added
+
+- Batch review preset for consolidated task evaluation
+- Integration test for prompt path resolution
+
+### Changed
+
+- Review workflows now default to "medium and higher" priority threshold for coworker automation
+- Removed AskUserQuestion from review skill allowed-tools (no longer needed)
+- Migrated integration tests to E2E format with consolidated DeepMerger
+
+### Fixed
+
+- Guarantee 0 exit code for help requests
+
+## [0.35.5] - 2026-01-31
+
+### Performance
+
+- Optimized test suite from 14.68s to 5.52s (62% reduction)
+  - Added `SharedTempDir` module for opt-in per-class temp directory sharing
+  - Reduces setup/teardown overhead from ~10ms per test to near-zero
+  - Test classes opt-in with `def self.use_shared_temp_dir?; true; end`
+
+### Changed
+
+- Migrated 58 integration-style molecule tests to E2E format
+  - Created MT-REVIEW-004 (GitHub CLI integration) - 7 test cases
+  - Created MT-REVIEW-005 (multi-model executor) - 5 test cases
+  - Removed gh_cli_executor_test.rb (13 tests)
+  - Removed gh_pr_fetcher_test.rb (13 tests)
+  - Removed gh_pr_comment_fetcher_test.rb (19 tests)
+  - Removed multi_model_executor_test.rb (13 tests)
+
+## [0.35.4] - 2026-01-31
+
+### Refactored
+
+- Consolidated `deep_merge_hash` utility into centralized `Ace::Support::Config::Atoms::DeepMerger`
+  - Removed duplicate implementations from `PresetManager` and `ReviewManager`
+  - Both now use `DeepMerger.merge(base, overlay, array_strategy: :union)`
+  - Improves maintainability and aligns with ATOM architecture (ADR-011)
+
+### Performance
+
+- Moved 8 integration test files to E2E test suite
+  - Created MT-REVIEW-001 (preset composition), MT-REVIEW-002 (multi-subject), MT-REVIEW-003 (auto-save workflow)
+  - E2E tests run via `/ace:run-e2e-test ace-review MT-REVIEW-001`
+  - Test execution time reduced from 19.52s to ~13.5s (31% reduction)
+  - Remaining time due to legitimate timeout tests in molecules layer
+
+### Changed
+
+- Removed integration test files (replaced by E2E tests):
+  - preset_composition_integration_test.rb (12 tests)
+  - multi_subject_integration_test.rb (8 tests)
+  - auto_save_integration_test.rb (E2E flow moved, unit tests covered elsewhere)
+  - preset_diff_integration_test.rb
+  - multi_model_cli_test.rb
+  - full_prompt_generation_test.rb
+  - pr_diff_generation_test.rb
+  - synthesis_test.rb
+
+## [0.35.3] - 2026-01-29
+
+### Fixed
+- Multi-model executor hanging on slow CLI providers by adding timeout to Thread.join
+- Added deadline-based join to ensure total wait is bounded regardless of stuck threads
+- Suppressed IOError exceptions from killed threads for cleaner output
+- Added warning display for threads killed after timeout
+
+## [0.35.2] - 2026-01-29
+
+### Added
+- Enhanced review synthesis prompts with accuracy and conflict resolution guidelines
+- Added "Verification Required" section for unverifiable model claims
+- Added "Future Considerations" section to separate speculation from action items
+- Added severity classification and scope boundary guidance to base review prompt
+
+### Fixed
+- Exception-based CLI error handling for consistent exit codes
+
+## [0.35.1] - 2026-01-16
+
+### Changed
+- Rename context: to bundle: keys in configuration files
+
+## [0.35.0] - 2026-01-15
+
+### Changed
+- **Dependency Migration**: Replaced ace-context dependency with ace-bundle (~> 0.29)
+  - Updated gemspec dependency from `ace-context ~> 0.9` to `ace-bundle ~> 0.29`
+  - Updated all `require 'ace/context'` to `require 'ace/bundle'`
+  - Updated all `Ace::Context` references to `Ace::Bundle`
+  - Renamed `load_context_via_ace_context` to `load_context_via_ace_bundle`
+  - Renamed `ace_context_preset_exists?` to `ace_bundle_preset_exists?`
+- Updated error messages and comments to reference ace-bundle
+
+### Technical
+- Updated test helpers to stub Ace::Bundle instead of Ace::Context
+- Updated all test files with ace-bundle references
+
+## [0.34.0] - 2026-01-14
+
+### Added
+- Support for Gemini provider in all review presets
+- Google Gemini 2.5 Flash as default model for code reviews
+
+### Changed
+- Updated LLM executor to handle gemini provider configuration
+- All review presets now use gemini:gemini-2.5-flash as default model
+
+## [0.33.2] - 2026-01-13
+
+### Fixed
+- Fix multi-model option handling: store parsed models in `:models` key (array) instead of `:model` key
+  - Resolves issue where multi-model reviews failed due to incorrect option key usage
+
+### Changed
+- Update dependencies: ace-support-nav (renamed from ace-nav), ace-support-config (renamed from ace-config)
+
+## [0.33.1] - 2026-01-09
+
+### Changed
+- **BREAKING**: Eliminate wrapper pattern in dry-cli commands
+  - Merged business logic directly into `ListPresets`, `ListPrompts`, `Review`, and `Synthesize` dry-cli command classes
+  - Deleted `list_presets_command.rb`, `list_prompts_command.rb`, `review_command.rb`, and `synthesize_command.rb` wrapper files
+  - Simplified architecture by removing unnecessary delegation layer
+
+## [0.33.0] - 2026-01-07
+
+### Changed
+- **BREAKING**: Migrated CLI framework from Thor to dry-cli (task 179.14)
+  - Replaced `thor` dependency with `dry-cli ~> 1.0`
+  - Converted CLI to dry-cli Registry pattern
+  - Created dry-cli command classes (list_presets, list_prompts, review, synthesize)
+
+## [0.32.1] - 2026-01-07
+
+### Added
+- Support `commit:HASH` subject type for reviewing individual commits
+  - Accepts commit hashes with 6-40 hexadecimal characters
+  - Validates hash format before git operations
+  - Generates diff using `COMMIT~1..COMMIT` syntax to show commit's changes
+  - Provides clear error messages for invalid formats
+  - Updated documentation with usage examples
+
+## [0.32.0] - 2026-01-07
+
+### Changed
+- **BREAKING**: Review report filenames changed from 14-character timestamps to 6-character Base36 compact IDs
+  - Example: `20251129-143000-model-preset-review.md` → `i50jj3-model-preset-review.md`
+  - Synthesis reports: `20251129-143000-synthesis.md` → `i50jj3-synthesis.md`
+- Migrate to Base36 compact IDs for session directories (via ace-timestamp)
+
+### Added
+- Dependency on ace-timestamp for compact ID generation
+
+## [0.31.1] - 2026-01-05
+
+### Technical
+- Clarified subject type prefix requirement in workflow documentation with explicit warnings and common mistakes examples
+
+## [0.31.0] - 2026-01-05
+
+### Added
+- Thor CLI migration with ConfigSummary display
+
+### Changed
+- Adopted Ace::Core::CLI::Base for standardized options
+
+
+## [0.30.0] - 2026-01-03
+
+### Changed
+- **BREAKING**: Minimum Ruby version raised to 3.3.0 (was 3.1.0)
+- Standardized gemspec file patterns with deterministic Dir.glob
+- Added MIT LICENSE file
+
+## [0.29.4] - 2026-01-03
+
+### Changed
+
+- **Test performance optimization**: Reduced test suite execution time from 7.15s to 1.77s (75% reduction)
+  - Removed unnecessary git init from MultiModelCliTest (tests only parse CLI options)
+  - Added shared `stub_synthesizer_prompt_path` helper to avoid ace-nav subprocess calls (~150-400ms per call)
+  - Optimized `mock_llm_synthesis` to use block-based stubbing pattern for better isolation
+
+### Technical
+
+- Extract shared prompt path stubbing to `AceReviewTest#stub_synthesizer_prompt_path`
+- Include `Ace::TestSupport::Fixtures::PromptHelpers` mixin in test helper
+- Remove unused `require "tmpdir"` from multi_model_cli_test.rb
+
+## [0.29.3] - 2026-01-01
+
+### Fixed
+
+- **Configurable timeout for gh CLI operations**: `execute_simple` and `execute` methods now use configurable timeout from config instead of hardcoded values
+  - Added `gh_simple_timeout` config option (default: 10s) for simple commands like `--version`, `auth status`
+  - `execute` now reads `gh_timeout` from config (default: 30s) instead of hardcoded fallback
+  - Follows ADR-022 pattern with `Ace::Review.get("defaults", "key")` for config access
+
+## [0.29.2] - 2025-12-30
+
+### Changed
+
+- Add ace-config dependency for configuration cascade management
+- Migrate from Ace::Core to Ace::Config.create() API (keep ace-support-core for ProcessTerminator)
+- Migrate from `resolve_for` to `resolve_namespace` for cleaner config loading
+
+## [0.29.1] - 2025-12-30
+
+### Changed
+
+- **Workflow presentation format**: Improved review results presentation in `review.wf.md` and `review-pr.wf.md`
+  - Separate "No Action Needed" section (no numbering) for INVALID and VERIFIED CORRECT items
+  - Numbered "Action Items" table with priority column
+  - New priority threshold selection: All, Medium+, High+, Critical only
+  - Clearer categorization step (Step 5) before presentation
+
+## [0.29.0] - 2025-12-30
+
+### Changed
+
+* Rename `.ace.example/` to `.ace-defaults/` for gem defaults directory
+
+
+## [0.28.0] - 2025-12-29
+
+### Changed
+- Migrate file system operations from `Ace::Core::Molecules` to `Ace::Support::Fs::Molecules` for direct ace-support-fs usage
+
+## [0.27.2] - 2025-12-28
+
+### Added
+- **Prioritize developer feedback in synthesis**: Human reviewer comments now receive special handling
+  - New "Developer Action Required" section appears before Consensus Findings
+  - Each unresolved comment gets its own subsection with exact text preserved
+  - Priority boosting ensures developer feedback is never ranked lower than Medium
+  - Added completeness requirements to prevent summarization of dev comments
+  - Added anti-patterns and verification checklist for developer feedback
+
+## [0.27.1] - 2025-12-28
+
+### Fixed
+- **Auto-discover repo for inline PR comments**: When running `ace-review --pr <number>` (local PR number), inline code comments were silently not fetched because GraphQL requires owner/repo format. Now automatically discovers repository via `gh repo view --json owner,name`
+- **Improved warning messages**: Upgraded warning messages from Debug to Warning level for better visibility when inline comments cannot be fetched
+
+## [0.27.0] - 2025-12-28
+
+### Added
+- **ADR-022 Configuration Pattern**: Migrate to gem defaults from `.ace.example/` with user override support
+  - Load defaults from `.ace.example/review/config.yml` at runtime
+  - Deep merge with user config via ace-core cascade
+  - Follows "gem defaults < user config" priority
+
+### Fixed
+- **Debug Check Consistency**: Standardized `debug?` method to use `== "1"` pattern (was `== "true"`)
+
+## [0.26.3] - 2025-12-27
+
+### Changed
+
+- **Add verification step to review workflows**: New Step 3 in `review.wf.md` and `review-pr.wf.md` requires verifying Critical/High priority action items before presenting to user
+  - Categorize items as VALID/INVALID/EDGE CASE/SUGGESTION
+  - Filter out false positives from LLM reviewers
+  - Prevents wasted investigation time on non-issues
+
+## [0.26.2] - 2025-12-26
+
+### Technical
+
+- Add timeout guidance for Claude Code agents in workflow instructions (10-minute timeout, inline mode)
+
+## [0.26.1] - 2025-12-26
+
+### Fixed
+
+- **Complete ace-git migration in SubjectExtractor**: Replace remaining ace-context references with ace-git equivalents
+  - `Ace::Context::Atoms::GitExtractor.tracking_branch` → `Ace::Git::Molecules::BranchReader.tracking_branch`
+  - `Ace::Context::Atoms::PrIdentifierParser.parse` → `Ace::Git::Atoms::PrIdentifierParser.parse`
+  - Fixes `uninitialized constant` errors when using ace-review after ace-context v0.16 migration
+
+## [0.26.0] - 2025-12-26
+
+### Changed
+
+- **Migrate to ace-git**: Replace duplicated Git/GitHub code with ace-git dependency
+  - Removed `Ace::Review::Molecules::GitBranchReader` - now uses `Ace::Git::Molecules::BranchReader`
+  - Removed `Ace::Review::Atoms::TaskAutoDetector` - now uses `Ace::Git::Atoms::TaskPatternExtractor`
+  - Removed `Ace::Review::Molecules::PrIdentifierParser` - now uses `Ace::Git::Atoms::PrIdentifierParser`
+  - Added `ace-git (~> 0.3)` as runtime dependency
+  - Eliminates code duplication and centralizes Git operations in ace-git package
+- **Fail fast on ace-git load**: Remove defensive begin/rescue around ace-git require since it's a hard dependency
+
+### Fixed
+
+- **GhCommentPoster fallback URL**: Fix URL construction when gh CLI output doesn't include URL
+  - Now correctly uses ace-git's combined `:repo` format (owner/repo) instead of legacy separate `:owner`/`:repo` keys
+
+### Removed
+
+- Deleted `lib/ace/review/molecules/git_branch_reader.rb`
+- Deleted `lib/ace/review/atoms/task_auto_detector.rb`
+- Deleted `lib/ace/review/molecules/pr_identifier_parser.rb`
+- Deleted corresponding test files
+- Removed `ace-git-diff` dependency (functionality migrated to ace-git)
+
+### Added
+
+- `mock_parse_result` test helper in test_helper.rb for consistent ParseResult mock creation
+- Tests for GhCommentPoster.extract_comment_url fallback path scenarios
+
+## [0.25.0] - 2025-12-17
+
+### Added
+
+* **Multiple `--subject` Flags**: Support combining multiple subject sources in a single review
+  * `ace-review --subject pr:77 --subject files:README.md --subject pr:79`
+  * Subjects are merged into unified ace-context config
+  * New `merge_typed_subject_configs()` public API for config merging
+
+### Fixed
+
+* **Nested Hash Merging**: Fixed `deep_merge_arrays` to recursively merge nested hashes
+  * Two typed subjects like `diff:HEAD~3` and `diff:HEAD` now correctly merge their nested `context.diffs` arrays
+  * Made merge operation immutable (no longer mutates input hashes)
+
+### Changed
+
+* **Simplified Subject Extraction**: Removed legacy content extraction paths
+  * Removed `extract(Array)` - arrays now handled via `merge_typed_subject_configs`
+  * Removed `extract_and_merge_multiple` - replaced by config passthrough
+  * Removed `subject-content.md` creation - all subjects use ace-context config passthrough
+  * Cleaner API: `parse_typed_subject_config` for single, `merge_typed_subject_configs` for multiple
+
+### Technical
+
+* Removed 17 tests for deprecated array extraction functionality
+* Added 8 tests for `merge_typed_subject_configs` merged config validation
+* Cleaned up `:subjects` intermediate key in CLI options
+
+## [0.24.2] - 2025-12-16
+
+### Fixed
+
+* **PR Subject Parsing**: Refined context diff detection and PR subject parsing for more reliable PR reviews
+  * Improved handling of PR references in subject configurations
+  * Better validation of PR references before fetching
+
+### Changed
+
+* **ContentChecker Extraction**: Refactored diff merging logic into dedicated ContentChecker component
+  * Cleaner architecture for content validation
+  * Improved maintainability of review subject processing
+
+### Technical
+
+* Added test coverage for PR and section processing features
+
+## [0.24.1] - 2025-12-16
+
+### Fixed
+- **pr: Array Consistency**: Changed `pr:` typed subject to return array format (`{"pr" => ["77"]}`) for consistency with `diffs:` and `files:` which are always arrays
+
+## [0.24.0] - 2025-12-16
+
+### Added
+- **Subprocess Timeout Protection**: Added 10-second timeout to `ace-taskflow` subprocess in `task:` subject resolution
+  - Prevents indefinite hangs when task lookup is stuck
+  - New `CommandTimeoutError` provides clear error message with command and timeout details
+  - Uses Ruby's `Timeout` module wrapping `Open3.capture3`
+- **Dual Extraction Paths Documentation**: Added class-level documentation to `SubjectExtractor` explaining the two code paths
+  - Direct extraction via `extract()` for immediate content
+  - Config passthrough via `parse_typed_subject_config()` for optimized ReviewManager flow
+
+### Fixed
+- **Comment Accuracy**: Updated misleading comment in `SubjectExtractor#use_ace_context`
+  - Comment now correctly reflects that ace-context supports both flat keys and nested `context:` structure
+
+## [0.23.2] - 2025-12-14
+
+### Fixed
+- **Upstream bug fixes**: ace-llm dependency fixes benefit ace-review users
+  - Zero-value generation parameters (`temperature: 0`) now preserved in MistralClient, AnthropicClient, GoogleClient
+  - All LLM clients standardized with GENERATION_KEYS pattern for consistency
+
+## [0.23.1] - 2025-12-14
+
+### Changed
+- **Workflow Simplification**: Simplified `review.wf.md` to match `review-pr.wf.md` pattern with full cycle workflow (review → plan → confirm → implement)
+
+## [0.23.0] - 2025-12-08
+
+### Added
+- **PR Comment Developer Feedback**: Extract developer feedback from PR comments and inline review threads
+  - New `--[no-]pr-comments` flag to control inclusion (default: true for `--pr` reviews)
+  - Creates `review-dev-feedback.md` report alongside LLM reviews
+  - Includes issue comments, review comments, and approval/change-request state
+  - Integrated into synthesis reports when multi-model review is used
+- **GhCommentResolver**: New molecule to reply to PRs and resolve review threads
+  - `reply` - Post comment with commit reference
+  - `resolve_thread` - Resolve review threads via GitHub GraphQL API
+  - `reply_and_resolve` - Combined operation for workflow automation
+- **Empty-Body Review Support**: Approvals and change-requests without body text now included with placeholder
+
+### Fixed
+- **Thread ID Validation**: Added format validation (`PRRT_xxx` pattern) to prevent GraphQL injection
+- **Markdown Table Safety**: Escape pipe characters in table preview to prevent broken rendering
+
+### Changed
+- **Pagination Warnings**: Warn when GraphQL results are truncated (>100 threads or >50 comments per thread)
+- **Comment Fetch Failure Logging**: Log warning instead of silent failure when PR comment fetch fails
+- **Table Readability**: Wrap IDs in backticks for improved readability in markdown tables
+
+### Documentation
+- **README Updated**: Document `--[no-]pr-comments` flag and developer feedback feature
+- **CLI Reference**: Added `--[no-]pr-comments` to GitHub PR options section
+
+## [0.22.0] - 2025-12-03
+
+### Added
+- **Auto-Save Feature**: Automatically save reviews to task directories based on git branch name
+  - Enable with `auto_save: true` in config
+  - Configurable branch patterns via `auto_save_branch_patterns`
+  - Release directory fallback via `auto_save_release_fallback`
+  - Disable per-command with `--no-auto-save` flag
+- **Multi-Model Auto-Save**: Individual model reports now saved to task directory (not just synthesis)
+- **Auto-Save Integration Tests**: Comprehensive test coverage for branch detection and task resolution
+
+### Fixed
+- **Multi-Model Auto-Save**: All model reports now saved to task directory, matching explicit `--task` behavior
+
+### Technical
+- Remove unused `project_root` variable in TaskReportSaver
+- Stabilize GitBranchReader tests with Open3 mocking and real detached HEAD test
+
+## [0.21.0] - 2025-12-03
+
+### Added
+- **Multi-Model Report Synthesis**: Automatically synthesize reviews from multiple LLM models
+  - New `ace-review synthesize --session <dir>` standalone command
+  - Auto-triggered after multi-model execution when 2+ models succeed
+  - Identifies consensus findings, strong recommendations, unique insights, and conflicting views
+  - Produces prioritized action items combining all model feedback
+  - Configurable synthesis model via `--synthesis-model` or `synthesis.model` config
+  - Disable with `--no-synthesize` flag or `synthesis.enabled: false` config
+- **ReportSynthesizer Molecule**: LLM-powered report consolidation with structured prompt
+- **Synthesis Prompt Template**: `handbook/prompts/synthesis-review-reports.system.md`
+- **E2E Integration Test**: Full test coverage for multi-model auto-synthesis flow
+
+### Changed
+- **Default Synthesis Model**: `claude:sonnet` (was `google:gemini-2.5-flash`)
+
+### Documentation
+- **Configuration Defaults**: Clarify gem defaults in README
+  - Default preset is `code` (basic single-model review)
+  - Default `auto_execute` is `false` (prompts for confirmation)
+  - Projects can override in their `.ace/review/config.yml`
+
+## [0.20.6] - 2025-12-02
+
+### Fixed
+- **SlugGenerator**: Remove trailing hyphen after max_length truncation
+
+### Documentation
+- **Multi-Model Reviews**: Add section documenting CLI usage, preset config, and output structure
+- **Preset Resolution Chain**: Document preset lookup order (project → gem defaults)
+
+## [0.20.5] - 2025-12-02
+
+### Technical
+- Update documentation to use `code-pr` preset instead of deprecated `pr` preset
+
+## [0.20.4] - 2025-12-02
+
+### Added
+- **LLM Timeout**: Configurable timeout (default: 300s) to prevent indefinite hangs
+  - Set via `defaults.llm_timeout` in config
+- **Model Name Validation**: CLI validates model names contain only safe characters
+
+### Fixed
+- **Example Config**: Default preset changed from 'pr' to 'code' (existing preset)
+
+## [0.20.3] - 2025-12-02
+
+### Fixed
+- **Config Loading**: Use correct `Ace::Core.get` API to load `.ace/review/config.yml`
+  - Was using `Ace::Core.config.get("ace", "review")` which returned nil
+  - Now `defaults.preset` and other settings are properly read from config
+
+## [0.20.2] - 2025-12-02
+
+### Changed
+- **Config-based Preset Default**: `defaults.preset` in config now used instead of hardcoded "pr" fallback
+  - CLI `--preset NAME` overrides config default
+  - Helpful error message when no preset specified and no config default set
+
+## [0.20.1] - 2025-12-02
+
+### Added
+- **Config-based Settings**: Move runtime options from ENV to config file
+  - `max_concurrent_models` - configurable in `defaults` section
+  - `auto_execute` - skip confirmation prompt when set to `true`
+- **Improved CLI Output**: Task directory output now shows directory once, then lists filenames
+
+### Changed
+- **Preset Consolidation**: Replace duplicated `pr.yml` with DRY `code-pr.yml` extending `code`
+- **Concurrency Guard**: Clamp `max_concurrent_models` to minimum 1 to prevent crashes
+- **Input Validation**: Filter blank entries from comma-separated model lists
+
+## [0.20.0] - 2025-12-02
+
+### Added
+- **Multi-Model Concurrent Execution**: Run code reviews against multiple LLM models simultaneously
+  - New `--model` flag accepts comma-separated models or multiple flags (e.g., `--model claude:opus,gpro`)
+  - Configurable concurrency via `ACE_REVIEW_MAX_CONCURRENT_MODELS` environment variable (default: 3)
+  - Thread-safe parallel execution with progress indicators
+  - Individual model failures don't stop other executions
+  - Preset support via `models:` array in YAML configuration
+
+### Fixed
+- **Output File Handling**: Pass `output_file` parameter to LlmExecutor for correct file creation
+- **Effective Model Logic**: `effective_model` now respects `models` array for metadata and filenames
+- **Task Report Filenames**: Use full model slug in filenames to prevent overwrites when using same-provider models
+- **Task Path Propagation**: Fix result key from `:task_path` to `:path` so saved locations surface in CLI output
+
+## [0.19.2] - 2025-12-01
+
+### Fixed
+- **Task Integration Fixes**: Multiple fixes to ensure `--task` flag works correctly
+  - Add missing `require` for `ace/taskflow/organisms/task_manager` in TaskResolver
+  - Pass actual review file path to TaskReportSaver instead of session directory
+  - Add defensive guard for nil/empty `task[:path]` to prevent crashes
+
+### Changed
+- **Test Consistency**: Refactored TaskResolver tests to use `Minitest::Mock` consistently
+
+## [0.19.1] - 2025-11-29
+
+### Fixed
+- Fix PR diff generation to use actual PR content instead of origin...HEAD when using `--pr` flag with presets
+- Remove problematic default subject from `code-pr.yml` preset that contained `origin...HEAD`
+- Add comprehensive integration tests for PR diff generation behavior
+
+## [0.19.0] - 2025-11-27
+
+### Added
+- **Specification Review Focus**: New `scope/spec` focus for reviewing specifications and proposals
+  - Goal clarity validation (single objective, no ambiguous terms, clear success criteria)
+  - Usage expectations analysis (target audience, scenarios, inputs/outputs)
+  - Test strategy evaluation (testable criteria, edge cases, validation approach)
+  - Completeness checking (required sections, dependencies, assumptions)
+  - Implementation feasibility assessment (achievable requirements, realistic estimates)
+  - Consistency and traceability verification
+- **Spec Preset**: New `spec.yml` preset for specification reviews
+  - Default subject: `origin/main...HEAD` filtered to `**/*.s.md` (task specs)
+  - Combines spec focus with standard format and tone guidelines
+
+## [0.18.0] - 2025-11-17
+
+### Added
+- **GitHub Pull Request Review Mode**: Support for reviewing GitHub pull requests directly
+  - New `--pr` option to specify pull request identifier (number, URL, or owner/repo#number)
+  - `--post-comment` option to automatically post review as GitHub comment (requires `gh` CLI)
+  - `--dry-run` option to preview comment without posting
+  - Multiple PR identifier formats supported: `123`, `https://github.com/owner/repo/pull/123`, `owner/repo#123`
+  - Automatic repository detection from git remote when using PR number only
+  - Comprehensive error handling for authentication, network issues, and PR state
+  - Retry logic with exponential backoff (capped at 32 seconds) for network resilience
+  - PR state validation (prevents posting to closed/merged PRs)
+  - Rich PR metadata included in review context (title, author, branch names, state)
+  - Secure comment posting via tempfiles (prevents command injection)
+  - Review cache saved to `.cache/ace-review/sessions/review-{timestamp}/`
+- **New Molecules**:
+  - `GhCliExecutor`: Safe execution of GitHub CLI commands with timeout and error handling
+  - `PrIdentifierParser`: Parse and normalize PR identifiers to owner/repo/number format
+  - `GhPrFetcher`: Fetch PR diffs and metadata with retry logic
+  - `GhCommentPoster`: Post review comments to GitHub with dry-run support
+- **New Atoms**:
+  - `RetryWithBackoff`: Reusable retry logic with exponential backoff for operations with transient failures
+- **New Error Classes**: Specific errors for GitHub integration (`GhCliNotInstalledError`, `GhAuthenticationError`, `PrNotFoundError`, `PrStateError`, `GhNetworkError`)
+- **CLI Options**:
+  - `--gh-timeout <seconds>`: Configure timeout for GitHub CLI operations (default: 30 seconds)
+- **Markdown Sanitization**: LLM review output is now sanitized and wrapped in collapsible `<details>` tags
+  - Automatically closes unclosed code fences to prevent broken GitHub comment rendering
+  - Wraps review content in expandable section for better PR comment readability
+- **README Documentation**: Comprehensive guide for GitHub PR review mode with examples, timeout configuration, and troubleshooting
+
+### Changed
+- **Default Timeout**: Reduced default timeout for GitHub CLI operations from 600 seconds (10 minutes) to 30 seconds
+  - Provides faster failure feedback for network issues
+  - Users can override with `--gh-timeout` option for large PRs or slow connections
+- **Retry Logic Architecture**: Extracted retry logic from `GhPrFetcher` into reusable `RetryWithBackoff` atom
+  - Improves testability and code reusability
+  - Cleaner separation of concerns following ATOM architecture
+
+### Fixed
+- **Architectural Compliance**: Moved `GhCliExecutor` from `atoms/` to `molecules/` to properly reflect its side effects (shell command execution)
+- **Test Coverage**: Uncommented and fixed previously failing tests in `gh_pr_fetcher_test.rb`
+  - Fixed complex mocking chain issues by extracting retry logic into testable atom
+  - All failure path and retry exhaustion tests now pass
+
+## [0.17.0] - 2025-11-17
+
+### Added
+- **Task Integration**: New `--task` flag to save review reports to task directories
+  - Accepts task references in multiple formats: `114`, `task.114`, `v.0.9.0+114`
+  - Reports saved to `<task-dir>/reviews/` with timestamped filenames
+  - Filename format: `YYYYMMDD-HHMMSS-{provider}-{preset}-review.md`
+  - Graceful degradation when ace-taskflow unavailable or task not found
+  - Created `TaskResolver` molecule for task reference resolution
+  - Created `TaskReportSaver` molecule for report persistence
+  - Updated `ReviewManager` organism to orchestrate task-aware saving
+  - Added ace-taskflow ~> 0.19 as runtime dependency
+
+## [0.16.1] - 2025-11-15
+
+### Fixed
+- **Git Worktree Cache Path Resolution**: Fixed cache directory creation to use project root instead of current working directory
+  - Resolves issue where caches were created in deeply nested, incorrect paths in git worktrees (e.g., `/path/.ace-wt/task.094/ace-context/.cache/ace-review/sessions/`)
+  - Modified `ReviewManager#create_cache_directory` to use `Ace::Core::Molecules::ProjectRootFinder.find_or_current`
+  - Added `require "ace/core/molecules/project_root_finder"` to review_manager.rb
+  - Each worktree now maintains its own cache at `.cache/ace-review/sessions/` relative to worktree root
+  - Updated tests to expect cache at project root location
+  - All 161 ace-review tests pass with no breaking changes to main repo usage
+  - Transparent fix - tool "just works" in worktrees without user configuration
+
+### Changed
+- **Dependencies**: Updated to use ace-support-core ~> 0.10.1 for worktree support
+
+## [0.16.0] - 2025-11-13
+
+### Added
+- **Preset Composition**: Support for composing review presets from reusable base configurations
+  - New `presets:` array at root level for preset composition
+  - Recursive preset loading with circular dependency detection (max depth: 10)
+  - Smart merging strategies: arrays concatenate+deduplicate, hashes deep merge, scalars last-wins
+  - **Composition order**: Base presets are loaded first, then the composing preset (last wins for scalars)
+  - Full backward compatibility - existing presets without `presets:` key continue to work unchanged
+  - New `PresetValidator` atom for validation and circular dependency detection
+  - Preset name validation (prevents path traversal, enforces length limits)
+  - Enhanced `PresetManager` molecule with `load_preset_with_composition` method
+  - Comprehensive test coverage: 23 validator tests, 22 manager composition tests, 11 integration tests
+- **Example Preset Refactoring**: New DRY preset structure
+  - `code.yml` base preset with common review instructions
+  - `code-pr.yml` composed preset for pull request reviews
+  - `code-wip.yml` composed preset for work-in-progress reviews
+
+### Changed
+- **PresetManager**: Enhanced to support preset composition while maintaining backward compatibility
+  - Recursive loading with visited set tracking
+  - Deep merge support for nested hash structures
+  - Array deduplication during composition
+  - Intermediate caching prevents redundant composition (particularly beneficial for deeply nested presets)
+  - Standardized internal metadata format (string keys for consistency)
+  - Deep metadata stripping from nested structures
+  - Added `strip_composition_metadata` helper method for DRY code
+  - Performance instrumentation with debug mode support
+
+### Technical
+- Integrated test suite performance optimizations from v0.15.1
+- Updated test patterns to match new test helper structure
+- All 56 tests passing (23 validator + 22 manager + 11 integration)
+
+## [0.15.1] - 2025-11-11
+
+### Technical
+- Optimize test suite performance with mocking (2.2x faster, 2.03s → 0.93s)
+  - Add `Ace::Context.load_auto()` mocking in test_helper
+  - Add `GitExtractor` mocking (staged_diff, working_diff, tracking_branch)
+  - Remove real git operations from integration tests
+  - Fix test issues (super calls, initialization timing, assertions)
+  - All 108 tests passing (16 atoms + 53 molecules + 29 organisms + 10 integration)
+
+## [0.15.0] - 2025-11-10
+
+### Added
+- **Section-Based Content Organization**: Support for `instructions.context.sections` format
+  - Integration with ace-context v0.17.5+ section-based content organization
+  - Structured organization of review content into semantic sections (focus, style, diff, etc.)
+  - All built-in presets (pr, code, security, docs, performance, ruby-atom, agents, test) now use sections
+  - PresetManager enhanced to preserve `instructions` field through resolution chain
+  - New format detection helper for automatic backward compatibility
+
+### Changed
+- **ReviewManager**: Enhanced to support both legacy `system_prompt` and new `instructions` formats
+  - Automatic format detection ensures seamless migration
+  - New `create_system_context_file_with_instructions()` method for section-based contexts
+  - Full backward compatibility maintained for existing user presets
+- **CLI Output**: Updated to properly display system and user prompt file paths
+  - Shows both `system.prompt.md` and `user.prompt.md` file paths
+  - Provides correct `ace-llm query` command with `--file` and `--context` parameters
+  - Maintains backward compatibility with legacy `prompt_file` format
+
+### Documentation
+- **README.md**: Added comprehensive documentation for new section-based format
+  - Examples of `instructions` format with section organization
+  - Legacy format documentation for backward compatibility
+  - Migration guidance and best practices
+
+### Testing
+- **Comprehensive Test Coverage**: Added 6 new test methods for section-based functionality
+  - Format detection validation
+  - Section-based context file creation
+  - Integration testing for both formats
+  - Backward compatibility verification
+  - All tests passing with 100% success rate
+
+## [0.14.0] - 2025-11-05
+
+### 🚀 **Major Performance Upgrade - Ruby API Migration & Context Fix**
+
+### Changed
+- **Architecture**: Complete migration from CLI subprocess to Ruby API calls
+  - Replaced `ace-llm-query` subprocess calls with direct `Ace::LLM::QueryInterface.query()` calls
+  - Eliminated all temp file creation and subprocess overhead
+  - Achieved 98-99% reduction in LLM call latency (70-135ms → 1-2ms)
+
+### Added
+- **Ruby API Integration**: Direct ace-llm Ruby library usage
+  - No more temp file management for prompts
+  - Rich response metadata (usage stats, model info, provider details)
+  - Structured exception-based error handling
+  - Enhanced session files with `llm_metadata.yml`
+
+- **Performance Benefits**:
+  - Eliminated process spawning overhead
+  - Removed shell interpretation delays
+  - Native Ruby object handling (no JSON parsing)
+  - Direct method calls with immediate response
+
+- **Enhanced Error Handling**:
+  - Specific exception types (`Ace::LLM::Error`, `Ace::LLM::ProviderError`, etc.)
+  - Structured error responses with error categorization
+  - Better debugging information with error types
+  - Graceful handling of API vs CLI availability
+
+- **Rich Metadata**:
+  - Token usage information (`usage` field)
+  - Model information and provider details
+  - Response timing and metadata
+  - Session persistence of LLM interaction data
+
+### Technical
+- **Dependency**: Added `ace-llm (~> 0.1)` runtime dependency
+- **API Compatibility**: Maintains identical external interface
+- **Backward Compatibility**: All existing CLI options and workflows unchanged
+- **Error Recovery**: Enhanced error messages and recovery paths
+
+### Fixed
+- **Context Generation Bug**: Fixed empty user.context.md files that had no subject configuration
+  - Updated `create_user_context_file` method to properly handle subject configuration fallbacks
+  - Eliminated redundant subject.md file creation (subject now handled via ace-context workflow)
+  - Enhanced configuration flow: explicit config → preset config → default "staged" configuration
+  - Improved handling of file paths, preset shortcuts, and structured configurations
+
+### Performance
+- **98-99% faster** LLM calls
+- **Zero temp file overhead**
+- **Direct Ruby object responses**
+- **Immediate availability** of results and metadata
+
+### Technical
+- **Streamlined Session Structure**: Removed subject.md and context.md files, now using ace-context workflow
+- **Enhanced Configuration Handling**: Better fallback logic for subject configuration processing
+- **Updated Tests**: Modified test expectations to match new v0.14.0 session structure
+
+## [0.13.1] - 2025-11-05
+
+### Fixed
+- **Implementation Gap**: Actually completed the v0.13.0 architectural changes that were documented but not fully implemented
+- **Removed Legacy Code**: Eliminated all prompt splitting logic and fallback methods as claimed in v0.13.0 CHANGELOG
+- **Updated Tests**: Fixed test expectations to match new architecture and removed tests for removed methods
+- **File Structure**: Corrected session file naming to use `system.prompt.md` and `user.prompt.md`
+- **CLI Integration**: Fixed `undefined method 'subject_config'` error in ReviewManager parameter naming
+- **ace-llm-query Interface**: Updated to use correct `--system` and `--prompt` flags instead of non-existent `--user` flag
+
+### Technical
+- **Code Cleanup**: Removed 214 lines of legacy code while maintaining functionality
+- **Syntax Validation**: All Ruby files now pass syntax validation
+- **Architecture Alignment**: Implementation now matches documented CHANGELOG claims
+
+## [0.13.0] - 2025-11-05
+
+### 🎯 **Major Architecture Fix - System/User Prompt Separation**
+
+### Changed
+- **Architecture**: Complete overhaul of prompt generation and processing
+  - Removed arbitrary prompt splitting (`split_and_save_prompts` method)
+  - Removed combined prompt generation (`build_review_prompt` method)
+  - Implemented proper ace-context integration throughout
+  - Fixed fundamental misunderstanding of system vs user prompts
+
+### Added
+- **System Prompt Generation**:
+  - Creates `system.context.md` with YAML frontmatter containing prompt:// references
+  - Integrates context configuration (e.g., "project" → presets: ["project"])
+  - Uses ace-context to generate `system.prompt.md`
+  - Proper base system instructions included after frontmatter
+
+- **User Prompt Generation**:
+  - Creates `user.context.md` with subject configuration
+  - Supports commands, files, diffs, and inline content from presets
+  - Uses ace-context to generate `user.prompt.md`
+  - Handles all subject types from preset configurations
+
+- **LLM Integration**:
+  - LlmExecutor requires separate system and user prompts
+  - New format: `--system-prompt` and `--user-prompt` flags via ace-llm-query
+  - Removed legacy single prompt support for cleaner architecture
+
+- **Session Structure**:
+  ```
+  session/
+  ├── system.context.md   # ace-context input (system prompt config)
+  ├── system.prompt.md    # ace-context output (generated system prompt)
+  ├── user.context.md     # ace-context input (user prompt config)
+  ├── user.prompt.md      # ace-context output (generated user prompt)
+  ├── subject.md          # Extracted subject content
+  ├── context.md          # Legacy context content
+  ├── metadata.yml        # Session metadata
+  └── review.md           # LLM output
+  ```
+
+### Fixed
+- **Configuration Structure**: Renamed `prompt_composition` → `system_prompt` in all preset configs
+- **Preset Parsing**: Updated all preset parsing logic to use new structure
+- **Backward Compatibility**: All existing preset configurations continue to work unchanged
+- **Ruby Syntax**: All syntax errors resolved and code validated
+
+### Technical
+- **YAML Frontmatter**: Follows ace-context patterns exactly with proper context structure
+- **Error Handling**: Comprehensive fallback mechanisms for ace-context failures
+- **Cache Management**: Enhanced cache-first storage model with proper file organization
+- **Token Optimization**: Potential to reduce token usage through ace-context processing
+
+### Breaking Changes
+- **LlmExecutor API**: Removed legacy single prompt support, now requires system_prompt and user_prompt parameters
+- **ReviewManager**: Removed fallback prompt generation methods, ace-context is now required
+- **Session Structure**: Updated file naming from legacy `prompt.md` to `system.prompt.md` and `user.prompt.md`
+
+## [0.12.0] - 2025-11-05
+
+### Added
+- **Context.md Pattern**: Adopt ace-docs context.md pattern for improved reproducibility
+  - ContextComposer molecule generates context.md with YAML frontmatter
+  - ContextExtractor delegates to ContextComposer for ace-context integration
+  - Cache-first storage with `.cache/ace-review/sessions/` directory
+  - Context.md files saved with embedded files and ace-context configuration
+
+### Changed
+- **Storage Model**: Implement cache-first storage approach
+  - Working files stored in cache directory instead of release folder
+  - Final reports copied to release folder `.ace-taskflow/v.*/reviews/`
+  - Removed `.tmp` extensions from all session files
+  - Split prompts into `prompt-system.md` and `prompt-user.md` files
+
+### Enhanced
+- **Ace-Context Integration**: Full integration with ace-context via `load_file_as_preset()`
+  - Follows ace-docs pattern exactly for consistency
+  - Support for presets, files, diffs, and commands in YAML frontmatter
+  - Fail-fast error handling with clear error messages
+  - Backward compatible CLI interface
+
+### Technical
+- **ContextComposer**: New molecule for context.md generation
+- **ReviewManager**: Updated with cache-first storage and prompt splitting
+- **ContextExtractor**: Refactored to delegate to ContextComposer
+- **Comprehensive Tests**: Added test coverage for all new functionality
+- **Backward Compatibility**: All existing presets work without modification
+
+## [0.11.2] - 2025-11-01
+
+### Changed
+
+- **Dependency Migration**: Updated to use renamed infrastructure gems
+  - Changed dependency from `ace-core` to `ace-support-core`
+  - Changed dependency from `ace-test-support`
+  - Part of ecosystem-wide naming convention alignment for infrastructure gems
+
+## [0.11.1]
+ - 2025-10-24
+
+### Fixed
+- Address PR #3 review issues for ace-git-diff integration
+
+### Technical
+- Standardize diff/diffs API documentation to ace-git-diff format
+- Add comprehensive integration and unit tests
+- Update all presets and workflow instructions with standardized diff format
+
+## [0.11.0] - 2025-10-23
+
+### Changed
+- Integrated with ace-git-diff for unified diff operations
+- SubjectExtractor now handles new diff: format with ranges
+- All example presets updated to use diff: key instead of commands:
+- Added ace-git-diff (~> 0.1.0) as runtime dependency
+- Delegates to ace-context which now uses ace-git-diff
+
+### Technical
+- Maintains backward compatibility with old diff: string format
+- Supports both diff: { ranges: [...] } and legacy diff: "range" formats
+
+## [0.10.0] - 2025-10-14
+
+### Added
+- Standardize Rakefile test commands and add CI fallback
+
+### Technical
+- Add proper frontmatter with git dates to all managed documents
+
+## [0.9.9] - 2025-10-08
+
+### Changed
+
+- **Test Structure Migration**: Migrated to flat ATOM structure
+  - From: `test/ace/review/molecules/`
+  - To: `test/molecules/`
+  - Aligns with standardized test organization across all ACE packages
+
+## [0.9.8] - 2025-10-07
+
+### Changed
+
+- **Workflow documentation updated** for ace-context v0.9.6+ integration
+  - Updated `review.wf.md` to reflect unified YAML schema
+  - Removed references to deprecated `patterns:` key
+  - Added comprehensive configuration schema section documenting `files:`, `diffs:`, `commands:`, `presets:`
+  - Enhanced common scenarios with file-based review examples
+  - Added "Review Specific Files" and "Compose Multiple Sources" scenarios
+  - Improved troubleshooting section with "No code to review" → use `files:` guidance
+  - Updated preset file examples to show proper YAML structure
+  - Added simple string shortcuts documentation (staged, working, pr)
+  - Updated all command examples to use correct YAML keys
+
+## [0.9.7] - 2025-10-07
+
+### Fixed
+
+- **ace-llm-query integration**: Fixed command construction to work with updated ace-llm-query API
+  - Replaced non-existent `--file` flag with new `--prompt` flag (added in ace-llm v0.9.1)
+  - Added proper PROVIDER:MODEL positional argument as first parameter
+  - Added `--output` flag to save review reports directly to session directory
+  - Added `--timeout 600` for 10-minute timeout on long reviews
+  - Added `--format markdown` for consistent markdown output
+  - Output filename now uses model short name: `review-report-{model-short}.md`
+  - Example: `review-report-gemini-2.5-flash.md` for model `google:gemini-2.5-flash`
+
+### Changed
+
+- **LlmExecutor API**: Updated `execute` method to require `session_dir:` parameter
+  - Enables direct file output to session directory
+  - Returns `output_file` path in result hash
+- **ReviewManager**: Updated to pass `session_dir` to LlmExecutor
+  - Simplified result handling (no longer needs to save output separately)
+
+## [0.9.6] - 2025-10-06
+
+### Changed
+
+- **ace-context integration**: Refactored to use ace-context for unified content aggregation
+  - `SubjectExtractor` now delegates to `Ace::Context.load_auto` for all content extraction
+  - `ContextExtractor` now delegates to `Ace::Context.load_auto` for all content extraction
+  - Preserved special behaviors (staged/working/pr keywords, project context defaults)
+  - Eliminated duplicated file reading, command execution, and git extraction logic
+
+### Added
+
+- **ace-context dependency**: Added `ace-context ~> 0.9` as runtime dependency
+- **Enhanced composition**: Can now combine files + commands + diffs + presets in unified configs
+- **Preset support in context**: `presets:` key now functional in context configuration
+- **Diffs support**: New `diffs:` key supported in subject and context configs
+
+### Removed
+
+- **Redundant atoms**: Deleted `git_extractor.rb` and `file_reader.rb` (now in ace-context)
+- **Duplicated logic**: All content extraction now centralized in ace-context
+
+## [0.9.5] - 2025-10-06
+
+### Changed
+
+- **Workflow command renamed**: `review-code.wf.md` → `review.wf.md`
+  - Claude command changed from `/ace:review-code` to `/ace:review` for simplicity
+  - Updated workflow invocation from `wfi://review-code` to `wfi://review`
+
+### Fixed
+
+- **Storage path detection**: Removed hardcoded storage defaults that prevented smart detection
+  - `storage_config` now only checks user config, not module defaults
+  - Properly implements 3-tier priority: user config → ace-taskflow → cache directory
+  - Fallback path changed from `./reviews` to `.cache/ace-review/sessions/`
+- **LLM command execution**: Fixed remaining `ace-llm` command reference in `llm_executor.rb`
+  - Changed from `ace-llm query` to direct `ace-llm-query` invocation
+  - Renamed method from `execute_ace_llm` to `execute_ace_llm_query` for clarity
+- **Configuration comments**: Updated config file comments to reflect correct detection order
+
+## [0.9.4] - 2025-10-05
+
+### Changed
+
+- **Dynamic storage path**: Storage now defaults to `$(ace-taskflow release --path reviews)`
+  - Falls back to `./reviews` if ace-taskflow not available
+  - Config `storage.base_path` commented out by default, uses smart detection
+  - User can still override by uncommenting and setting custom path
+- **Review file organization**: All review files now stored together with `.tmp` pattern
+  - Session files in `{release_path}/reviews/review-{timestamp}/`
+  - Temporary files use `.tmp` extension: `prompt.md.tmp`, `subject.md.tmp`, `context.md.tmp`
+  - Committable files: `metadata.yml`, `review.md`
+  - Gitignore pattern changed from `.ace-review-sessions/` to `**/*.tmp`
+- **Command detection**: Binary check updated from `ace-llm` to `ace-llm-query`
+  - Error message now correctly references `ace-llm-query`
+
+### Fixed
+
+- Review sessions no longer create separate `.ace-review-sessions` directory
+- All review artifacts now properly organized in release-specific folders
+- Temporary working files automatically gitignored via `.tmp` extension
+
+## [0.9.3] - 2025-10-05
+
+### Changed
+
+- **Configuration file renamed**: `code.yml` → `config.yml` for consistency with ace-* naming conventions
+  - Updated all references in code, tests, and documentation
+  - Both `.ace.example/review/config.yml` and `.ace/review/config.yml` now use new name
+- **Preset organization improved**: All presets now stored as individual files
+  - Extracted 7 presets from main config to separate `.yml` files in `review/presets/`
+  - Main `config.yml` now contains only defaults and storage settings
+  - Presets: pr, code, docs, security, performance, test, agents, ruby-atom
+- **Configuration cascade integration**: Removed hardcoded paths in favor of ace-core
+  - `PresetManager` now uses `Ace::Core::Molecules::ConfigFinder` for all file discovery
+  - Automatic cascade resolution across `./.ace → ~/.ace` without hardcoded paths
+  - Preset files discovered automatically across entire configuration cascade
+  - Maintains backward compatibility with fallback for environments without ace-core
+
+### Fixed
+
+- Configuration system now properly respects ace-core's configuration cascade
+- Preset loading works correctly from both local and user config directories
+
+## [0.9.2] - 2025-10-05
+
+### Fixed
+
+- **Prompt resolution** now works correctly via ace-nav integration
+  - Fixed custom `PromptResolver` that wasn't working properly
+  - Replaced with `NavPromptResolver` using ace-nav's universal resolution
+  - Registered ace-review prompts with ace-nav protocol for proper discovery
+- **Critical command injection vulnerability** in `GitExtractor`
+  - Fixed unsafe string interpolation in git commands
+  - Now uses array arguments with `Open3.capture3(*command_parts)`
+- **Code organization issues**
+  - Fixed overly complex `ReviewManager#execute_review` method
+  - Replaced hash options with proper `ReviewOptions` class
+  - Improved separation of concerns throughout
+
+### Changed
+
+- Refactored `ReviewManager` into clearer, testable steps
+- Dependencies now include `ace-nav ~> 0.9` for proper prompt resolution
+
+## [0.9.1] - 2025-10-05
+
+### Fixed
+
+- Replaced Zeitwerk with explicit requires following ace-gems conventions
+- Fixed all require_relative paths and namespace references
+- Removed unnecessary dependencies (zeitwerk, tty-*, rainbow, dry-cli)
+- Replaced dry-cli with OptionParser for consistency with other ace gems
+- Simplified output formatting to use plain text without external libraries
+
+### Changed
+
+- Minimal dependencies - now only requires ace-core (~> 0.9)
+- CLI implementation now follows standard ace-gems patterns
+
+## [0.9.0] - 2025-10-05
+
+### Changed
+
+- **BREAKING**: Simplified CLI interface from `ace-review code` to just `ace-review`
+- Tool is now more universal - presets determine what type of review (code, docs, security, etc.)
+- Cleaner, more intuitive command structure
+- Migration from v0.8 legacy code-review system
+
+### Migration
+
+Update all commands from:
+```bash
+ace-review code --preset pr
+```
+
+To:
+```bash
+ace-review --preset pr
+```
+
+## [0.1.0] - 2025-10-05
+
+### Added
+
+- Initial release of ace-review gem
+- Migrated from dev-tools code-review implementation
+- ATOM architecture with atoms, molecules, organisms, and models
+- Preset-based review configuration system
+- Prompt composition with base, format, focus, and guidelines modules
+- Prompt cascade resolution (project → user → gem)
+- prompt:// URI protocol for prompt references
+- Support for direct file path references in prompts
+- Multiple focus module composition
+- Integration with ace-taskflow for release-based storage
+- CLI command: `ace-review code` with various options
+- Built-in presets: pr, code, docs, security, performance, test, agents
+- Example configuration files in .ace.example/
+- Comprehensive prompt library migrated from dev-handbook
+- LLM execution via ace-llm integration
+- Session management for dry-run mode
+- List commands for presets and prompts
+
+### Changed
+
+- **BREAKING**: Replaced `code-review` command with `ace-review code`
+- **BREAKING**: Removed `code-review-synthesize` CLI (use `wfi://synthesize-reviews` workflow)
+- **BREAKING**: Configuration moved from `.coding-agent/code-review.yml` to `.ace/review/config.yml`
+- **BREAKING**: Storage location now defaults to `.ace-taskflow/<release>/reviews/`
+- Preset files now support separate directory at `.ace/review/presets/`
+- Improved preset override system with `--add-focus` option
+- Enhanced prompt resolution with multiple lookup strategies
+
+### Migration Notes
+
+To migrate from the old code-review system:
+
+1. Install ace-review gem
+2. Copy `.coding-agent/code-review.yml` to `.ace/review/config.yml`
+3. Update workflow files to use `ace-review code` instead of `code-review`
+4. Synthesis is now handled via workflow instructions only (no CLI command)
+
+
+## [0.40.4] - 2026-02-22
+
+### Fixed
+- Hidden deprecated 'feedback skip' subcommand from help (use 'verify --skip' instead)
+- Standardized quiet, verbose, debug option descriptions to canonical strings