Changelog

All notable changes to the AURORA project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

[0.18.0] - 2026-04-14

Added

aur doctor store integrity checks — new StoreIntegrityChecks category runs alongside the existing six health-check categories. Detects silent corruption in the ACT-R store:
- FTS5 desync (missing/stale chunks_fts rows vs. chunks)
- Orphan code chunks (file_path no longer in file_index)
- Activation orphans (rows referencing deleted chunks, FK-bypass insurance)
- Reasoning-chunk growth warning above 10K (Record-phase unbounded caching signal)
- Retrieval roundtrip (end-to-end FTS sanity via highest-access seed chunk)
- Three mechanical failures (FTS desync, orphans, activation orphans) are auto-repaired under the existing aur doctor --fix flag.
Tiered access-history compaction behind AURORA_COMPACT_ACCESS_HISTORY=1 flag (default off). Bounds the unbounded-growth path in activations.access_history by collapsing older access records into time buckets for storage only. BLA ranking preserved to <0.001 because the decay formula (ln Σ t_j^(-d)) already treats old records as negligible.
- Four tiers: 0–7d verbatim, 7–30d hourly buckets, 30–180d daily buckets, 180d+ aggregate.
- AccessHistoryEntry gains an optional count: int = 1 field; BLA loop becomes count * t^(-d). Fully backward compatible — pre-change callers unchanged.
- Lazy trigger at length > 200; common-path record_access unaffected.
Design doc: docs/02-features/memory/STORE_HARDENING.md covering both changes, rationale, math, and rollout plan.

Changed

get_access_history() docstring documents the new optional count key and the compaction flag.

[0.17.6] - 2026-02-14

Added

aur spawn dependency-aware wave execution
- depends_on field in ParsedTask with - Depends: 1.0, 2.0 sub-bullet parsing
- Topological sort (Kahn's algorithm) producing parallel-safe waves
- Wave-based execution: sequential across waves, parallel within each wave
- Completed task outputs forwarded as context into dependent tasks
- Dry-run shows wave breakdown
aur spawn prompt-to-tasks decomposition
- decompose_prompt_to_tasks_md() via LLM for natural language → tasks.md
- Auto-detection of input type: file path vs prompt text
- Discovers available agents via AgentScanner/AgentParser
aur spawn output persistence
- SpawnRunStore class with .aurora/spawn/runs/<timestamp>/ structure
- Stores tasks.md, results/task-{id}.json, summary.json, meta.json
- Re-run detection via SHA-256 hash; skip-completed-tasks on re-runs
FEATURE_BACKLOG.md tracking parked features and future work

Fixed

Truncate long goals content before embedding to avoid 2048 char limit (full content kept for BM25/FTS5)

Changed

Replace HTML comment agent metadata with visible markdown sub-bullets in task output
Code formatting cleanup across 170+ files (black/isort)

[0.17.2] - 2026-02-14

Fixed

Suppress aurora_lsp.client warnings in aur mem search output (request_references failed noise)

Changed

Updated MEM_INDEXING.md with FTS5 pipeline architecture and chunk-type-aware weights
Updated TESTING_GUIDE.md with current counts (2,666 tests, 148 files)

Performance

Parallelize LSP deadcode analysis with concurrent futures (~3x faster on large codebases)
Add reference cache to LSP client to avoid redundant lookups

[0.17.1] - 2026-02-12

Added

FTS5 full-text search gate replacing activation gate in retrieval pipeline
- SQLite FTS5 virtual table for keyword filtering (Stage 1)
- Ensures rare but keyword-relevant chunks surface (activation gate starved infrequent content)
- Fallback to activation-based retrieval for old databases without FTS5
FTS5 store integration tests (246 tests) and FTS5 retrieval integration tests (149 tests)
LSP analysis/client tests (106 tests), verify lite tests (196 tests), early detection monitor tests (143 tests)

Fixed

Verify phase: reject empty descriptions, validate goals embed_chunk type
Python 3.14 async compatibility (anyio/asyncio event loop handling)

Changed

Updated black 24.1.1 → 26.1.0 and reformatted codebase
Removed stale BM25 persistence and lazy loading tests (479 tests replaced by FTS5 tests)

[0.17.0] - 2026-02-10

Added

Chunk-type-aware hybrid retrieval weights
- Code chunks: BM25 50% / ACT-R 30% / Semantic 20% (identifiers are exact tokens)
- KB chunks: BM25 30% / ACT-R 30% / Semantic 40% (prose benefits from embeddings)
- Hardcoded as _CODE_WEIGHTS and _KB_WEIGHTS in hybrid_retriever.py
- Replaces single global weight configuration

[0.16.0] - 2026-02-05

Removed

aur headless command — complete removal of headless mode and all related code
API key functionality — removed ANTHROPIC_API_KEY references and configuration

Changed

Major test cleanup (February 2026):
- Before: 5,500 tests, 314 files, 126k lines, unknown pass rate
- After: 2,608 tests, 144 files, ~48k lines, 100% pass rate, 57% coverage
- Removed: mock-heavy tests (~36k lines), failing/stale tests (~17k lines), import smoke tests (~12k lines)
- Consolidated all tests into packages/*/tests/ (removed root tests/ directory)
- Dead production code removed (~2,800 lines, LSP-confirmed 0 usages)
CI made self-sufficient — no .aurora/ dependency, use tmp_path or monkeypatch
Added P0 integration tests (92 tests across 5 areas, 21% → 56% coverage)
Added P1 batch 1 integration tests (87 tests: spawner recovery/observability, LSP languages/diagnostics, CLI escalation/health)
Added P1 batch 2 integration tests (100 tests: spawner circuit breaker/policies/heartbeat, LSP analysis/client, CLI agent search/doctor)

[0.15.0] - 2026-02-05

Added

Multi-language LSP support — expanded from Python-only to 5 languages:
- JavaScript/TypeScript: cross-file references via multilspy patches, didChangeWatchedFiles support
- Go: gopls integration (auto-discovers via go.mod)
- Java: jdtls integration
Go and Java tree-sitter indexing with language-specific parsers
LSP language configs with branch types, skip patterns, callback methods
Fixed TypeScript cross-file references in multilspy (wrong initialize_params.json)
Improved deadcode false positive handling for JS/Go/Java

Changed

Updated MEM_INDEXING.md with Go and Java language support
Updated GitHub username from amrhas82 to hamr0

[0.14.2] - 2026-02-05

Fixed

Aligned PlanStatus enum and template docs with actual usage

[0.14.1] - 2026-02-05

Added

Consolidated learned rules from friction pattern analysis into docs

[0.14.0] - 2026-02-04

Added

Pre-edit LSP check hook for Claude Code
- .claude/hooks/pre-edit-lsp-check.py — runs lsp check before every file edit
- Shows usage count and risk level (LOW/MED/HIGH) inline
- Text search fallback when LSP returns 0 refs
- Template at packages/cli/src/aurora_cli/templates/hooks/pre_edit_lsp_check.py
- Installed via aur init --tools=claude
Comprehensive memory indexing documentation (MEM_INDEXING.md)

[0.13.6] - 2026-02-04

Added

CLI Hybrid LSP Fallback:

Extended hybrid fallback to aur mem search CLI command
When LSP returns 0 refs, ripgrep text search provides file/ref counts
Shows ~Nf ~Mr format (tilde indicates approximate text-based counts)
Populates Files list from text search results
Updated documentation for hybrid fallback feature

[0.13.5] - 2026-02-04

Fixed

Dependencies:

Add missing nest-asyncio dependency required by aurora_lsp

[0.13.4] - 2026-02-04

Added

LSP Hybrid Fallback:

Text search fallback when LSP returns 0 references
Detects cross-package usage that LSP misses (lazy imports, installed packages)
Shows both LSP refs and text matches when they diverge
Risk level upgraded based on text matches

Fixed

MCP Server Configuration:

Handle 'already exists' gracefully when configuring aurora MCP server
Use explicit --scope local for both remove and add-json commands

ML Model Loading:

Suppress verbose 'Loading weights' progress bar during model loading
Disable transformers progress bar via logging API
Disable huggingface_hub progress bars

Changed

Moved stale test test_conflict_detection_resolution.py to archive

[0.13.2] - 2026-02-03

Added

CLAUDE.md Template:

MCP tools documentation now included in managed block template
aur init creates CLAUDE.md with MCP tools section
aur init --config refreshes managed block with MCP information
Managed block markers (AURORA:START/END) for automatic updates

Changed

CLAUDE.md restructured with managed block for auto-updates
Template now documents lsp and mem_search MCP tools with usage examples

[0.13.1] - 2026-02-03

Fixed

Package Structure:

Bundle aurora_lsp into aurora-actr instead of separate package
Remove aurora-lsp PyPI dependency (now included directly)
Add multilspy>=0.0.15 as direct dependency

Deprecated

aurora-lsp package on PyPI (use aurora-actr>=0.13.1 instead)

[0.13.0] - 2026-02-03

Added

MCP Tools:

lsp MCP tool with 3 actions (deadcode, impact, check)
mem_search MCP tool with LSP enrichment (used_by, called_by, calling)
CODE_QUALITY_REPORT.md generation with severity levels (CRITICAL, HIGH, MEDIUM, LOW)
MCP configurator integration for Claude, Cursor, Cline, Continue
aurora-lsp package as mandatory dependency (10 languages supported)
36 comprehensive TDD tests for MCP tools

CLI Enhancements:

aur doctor now shows MCP tools status
aur init configures MCP servers automatically

Documentation:

docs/02-features/mcp/MCP.md - Complete MCP reference
docs/02-features/lsp/LSP.md - LSP integration guide
Updated CLAUDE.md with MCP tools section
Updated AGENTS.md with MCP tool documentation
Updated README.md with MCP overview

Removed

Deprecated Slash Commands:

/aur:search - Replaced by mem_search MCP tool
/aur:get - Replaced by mem_search MCP tool

Changed

Slash commands reduced from 6 to 4 (plan, tasks, implement, archive)
Memory search now includes LSP relationship data and git metadata

[0.12.2] - 2026-02-02

Changed

Simplified aur --help output (removed verbose examples, cleaner format)

[0.12.1] - 2026-02-02

Fixed

Include defaults.json in package distribution (was missing from wheel)
Added [tool.setuptools.package-data] configuration for data files

Changed

README: Removed headless references, updated doc paths to new structure, updated version

[0.12.0] - 2026-02-02

Removed

Headless Command (Complete Removal):

aur headless command and all related code (~1,400 lines)
packages/cli/src/aurora_cli/commands/headless.py
packages/cli/src/aurora_cli/templates/headless.py
packages/cli/src/aurora_cli/concurrent_executor.py
22+ headless test files

Tool Providers (Complete Removal):

Entire packages/cli/src/aurora_cli/tool_providers/ directory
Provider implementations: claude, cursor, codex, gemini, opencode, generic
Orchestrator and registry (~2,000 lines)

API Key References:

Removed ANTHROPIC_API_KEY from all documentation
Removed API key troubleshooting sections
Removed API key configuration options

Documentation Cleanup (300+ files):

Removed docs/archive/ (~180 files)
Removed docs/analysis/ (~10 files)
Removed docs/arch/ (~3 files)
Removed docs/commands/ (~4 files)
Removed docs/development/ (~25 files)
Removed docs/examples/ (~3 files)
Removed docs/features/ (~2 files)
Removed docs/fixes/ (~15 files)
Removed docs/guides/ (~20 files)
Removed docs/performance/ (~10 files)
Removed docs/reference/ (~12 files)
Removed docs/reports/ (~15 files)
Removed docs/workflows/ (~2 files)

Added

New Documentation Structure:

docs/00-context/ - Architecture and context
docs/01-product/ - Product documentation
docs/02-features/ - Feature docs (cli/, agents/, soar/, memory/)
docs/03-logs/ - Session logs
docs/04-process/ - Process docs (getting-started/, troubleshooting/, reference/, development/)

Changed

CLAUDE.md:

Reduced from ~150 lines to ~85 lines (43% reduction)
Removed verbose explanations
Condensed to essential commands and patterns

Init Command:

Removed headless references from next steps
Updated help text

[0.11.2] - 2026-02-02

Removed

CLI Cleanup - Deprecated Commands:

aur health command (dead code, never registered)
aur query command (MCP-dependent, deprecated)
aur plan init and aur plan create subcommands (use aur goals instead)
aur verify command (merged into aur doctor)

MCP Deprecation:

MCPFunctionalChecks class from health checks (~360 lines)
MCP configuration functions from init helpers
All MCP-related test files (~20 files, ~5,000 lines)

Added

Doctor Command Enhancements:

New InstallationChecks class verifying:
- Python version (>= 3.10)
- Core package imports (aurora_core, aurora_context_code, aurora_soar, aurora_reasoning, aurora_cli)
Installation checks merged from removed aur verify command

Init Command Enhancements:

create_default_config() function creates .aurora/config.json with sensible defaults
Config file created automatically during aur init

Changed

Spawn Command:

--verbose is now the default (use --quiet or -q to suppress output)
Changed from --verbose flag to --verbose/--quiet boolean pair

Fixed

Test Suite:

Fixed Subgoal model field mismatch in plan command tests (recommended_agent → assigned_agent)
Fixed timezone-naive vs timezone-aware datetime issues in archive tests
Removed broken MCP patches from init unified tests

[0.10.1] - 2026-01-30

Added

SOAR Pipeline Enhancements:

Vague word detection in complexity assessment
- Detects ambiguous terms: "it", "this", "that", "something", "stuff", "thing", etc.
- Increases complexity score for queries needing clarification
Complexity-based subgoal limits (2-4-6 by complexity tier)
- SIMPLE: max 2 subgoals
- MEDIUM/MODERATE: max 4 subgoals
- COMPLEX/CRITICAL: max 6 subgoals
- Reduces over-decomposition and improves response times
log_path field in aur soar response metadata
- Returns path to SOAR reasoning log for debugging
SOAR cache circuit breaker and failed log cleanup
- Prevents repeated cache failures from blocking queries
- Automatic cleanup of corrupted cache entries

Fixed

SOAR Pipeline:

Normalize dependency format in collect phase (sg-N → int conversion)
Handle both int and string dependency formats in Phase 8 respond
Display Phase 3 when using cached decomposition for better UX
Show cached decomposition indicator in aur soar Phase 3
Correct phase_callback attribute name (removed underscore prefix)
Add SOAR complexity value mapping (MEDIUM → MODERATE for consistency)
Use SOAR Phase 1 complexity assessment instead of reassessing locally
Fail explicitly when SOAR verification fails instead of degrading to fallback
Normalize dependency format in verify_lite (sg-N → int)

Agent Discovery:

Enable agent discovery for all 20 tools via dynamic path resolution
Fixed agent path resolution for tools outside standard directories

Memory Search:

Improve search UX with better help message and silent model loading
Suppress verbose embedding model loading messages

Goals Command:

Resolve aur goals failures (examples file path + enum value)

CI/CD:

Add timeouts to CI workflow to prevent runaway jobs

Changed

Documentation:

Update subgoal limits to 2-4-6 in aur-soar and aur-goals command docs
Update aur-soar and aur-goals with v0.10.0 optimization details
Clarify upgrade path from 0.9.x to 0.10.0 in README
Add pip uninstall command to README

Code Organization:

Organize documentation and scripts into proper directories
ML dependency improvements and cleanup

[0.10.0] - 2026-01-26

Added

Epic 1: Memory Search Performance (Foundation Caching)

Module-level HybridRetriever instance caching with LRU policy
- Cache keyed by (db_path, config_hash) with configurable size (default: 10)
- Reduces cold search time by 30-40% (15-19s → 10-12s)
- Configurable via AURORA_RETRIEVER_CACHE_SIZE and AURORA_RETRIEVER_CACHE_TTL
ActivationEngine singleton caching per database path
- Reduces warm search time by 40-50% (4-5s → 2-3s)
- Automatically shared across all retrievers using same database
Shared QueryEmbeddingCache with LRU policy
- Cache hit rate typically >60% in SOAR multi-phase operations
- Reduces embedding computation on repeated queries
Enhanced BM25 index persistence validation
- Disk-cached at .aurora/indexes/bm25_index.pkl
- Load time <100ms (vs 9.7s rebuild)
Cache invalidation support via aur mem index and config changes
Comprehensive caching guide in documentation

Epic 2: Lazy BM25 Loading + Dual-Hybrid Fallback

Lazy BM25 index loading deferred to first retrieve() call
- Creation time: 150-250ms → 0.0ms (99.9% improvement)
- Thread-safe double-checked locking ensures single load
- No impact on search performance (index loaded once and reused)
Dual-hybrid fallback mode (BM25 + Activation) when embeddings unavailable
- Automatically triggers when embedding model not installed or AURORA_EMBEDDING_PROVIDER=none
- Quality testing showed 85-100% overlap with tri-hybrid (BM25 + Activation + Semantic)
- Graceful degradation from tri-hybrid → dual-hybrid → activation-only
Performance benchmarks and qualitative validation

Planning System Refactor (Complete R1-R9)

New /aur:tasks slash command for regenerating tasks from PRD
- Available in all 20 tool configurators (6 commands total: search, get, plan, tasks, implement, archive)
- Reads prd.md, goals.json, agents.json and regenerates tasks.md
- Includes same TDD hints and format as /aur:plan
source_file field added to Subgoal and SubgoalData models
- Tracks primary source file for each subgoal
- Extracted from LLM responses with validation
- Supports both code files and markdown documentation
- Backward compatible (optional field, defaults to None)
- Integrated into both aur goals and aur soar decomposition
agents.json schema documentation in PLAN_REFERENCES
- Complete template with required and optional fields
- Reference to official schema file
TDD hints in tasks.md template
- Format: tdd: yes|no and verify: command
- Guidelines for when to use TDD (models, APIs, bugs) vs when to skip (docs, config)
- Matches 2-generate-tasks agent pattern
3-SIMPLE-STEPS.md comprehensive planning guide created
- Step-by-step workflow documentation
- Code-aware (with goals.json) vs prompt-based planning paths
- Artifact generation order and dependencies

Task Execution Improvements

Wave-based execution with context passing for dependent tasks
Topological sorting for dependency-aware parallel execution
Invalid dependency reference validation

Changed

Planning System

Plan folders now use slug-only format (e.g., improve-search instead of 0001-improve-search)
- Backward compatible: numbered plans (NNNN-slug) still work
- Folder override: Re-running same goal replaces previous attempt
- Plan ID validators accept both formats
Artifact generation order now explicitly documented
- Order: plan.md → prd.md → design.md → agents.json → tasks.md
- tasks.md generated LAST as it depends on PRD content
Plan templates emphasize goals.json as recommended but optional
- Code-aware planning (with goals.json) preferred for production
- Prompt-based planning (skip goals.json) valid for simpler workflows

Code Quality

Applied 4,425 safe formatting fixes across codebase
Replaced all datetime.utcnow() with datetime.now(timezone.utc) (Python 3.12+ compatibility)
Fixed all ARG001-ARG005 ruff violations (104 unused argument fixes)
Prefixed unused parameters with underscore for clarity

Fixed

Planning System

Fixed aur plan view schema mismatch
- Changed sg.recommended_agent → sg.assigned_agent
- Removed non-existent sg.agent_exists checks
- Added ideal_agent gap detection
Removed non-existent command references from templates
- Removed aur plan validate references
- Removed --specs flag from aur plan list
- Replaced show --json --deltas-only with view --format json
Fixed configurator method signature incompatibilities (codex.py, kilocode.py)

Memory Search

Fixed embedding model wait logic in mem search
- Changed from BM25-only fallback to proper embedding wait
Fixed dual-hybrid fallback test configuration (caplog logger)

SOAR & Spawner

Fixed fallback when SOAR decomposition returns empty subgoals
Fixed retry_feedback parameter passing to decompose_query
Fixed _verbose parameter name in _display_goals_results call
Corrected parameter names in function calls (Phase 2B Task 15.1)

Type Checking & Linting

Resolved all Pyright type-checking errors
- Fixed Config attribute access with type: ignore comments
- Fixed Subgoal model attribute references
- Fixed method signatures to match base classes
Fixed datetime deprecation warnings throughout codebase
Removed 12 unused test variables in collect tests
Removed commented code in packages/cli

Removed

Slash Commands

Removed aur:checkpoint slash command from all tool configurations
- All 20 tool configurators now generate 6 commands: search, get, plan, tasks, implement, archive
- Removed checkpoint template from templates/slash_commands.py
- Removed checkpoint command from templates/commands.py
- Existing .aurora/checkpoints/ directories are preserved (not deleted)
- Config keys related to checkpoints are silently ignored for backward compatibility

CLI Implementation

Removed checkpoint-related CLI options from aur spawn command
- Removed --resume, --list-checkpoints, --clean-checkpoints, --no-checkpoint flags
- Deleted packages/cli/src/aurora_cli/planning/checkpoint.py
- Deleted packages/cli/src/aurora_cli/execution/checkpoint.py
- Removed checkpoint helper functions from spawn_helpers.py

Planning System

Removed spec generation from planning system
- Plan creation now generates 5 files instead of 8 (plan.md, prd.md, design.md, agents.json, tasks.md)
- Removed 4 capability spec files: planning, commands, validation, schemas specs
- Updated all references from "8 files" to "5 files"
- Removed spec validation from validate_plan_structure()
- Removed spec entries from show_plan() file status

Performance

Startup Improvements

Lazy BM25 index loading: 99.9% improvement (150-250ms → 0ms)
Module-level retriever caching: 30-40% faster cold searches (15-19s → 10-12s)
Activation engine singleton: 40-50% faster warm searches (4-5s → 2-3s)
BM25 index persistence: <100ms load vs 9.7s rebuild
Query embedding caching: >60% cache hit rate in multi-phase operations

Memory Efficiency

Thread-safe double-checked locking for lazy initialization
LRU caching prevents unbounded memory growth
Shared instances reduce memory footprint

Documentation

New Guides

Created docs/guides/3-SIMPLE-STEPS.md - Comprehensive planning workflow guide
Added Epic 1 caching guide with configuration examples
Added Epic 2 documentation (lazy loading + dual-hybrid fallback)
Added comprehensive Phase 2 lessons learned

Updated References

Updated PLAN_REFERENCES with agents.json schema and examples
Updated PLAN_STEPS with artifact generation order
Updated README.md with real command output examples
Updated CODE_QUALITY_REPORT with Phase 2A and 2B results
Updated all documentation to reflect 6-command structure (tasks added)
Removed checkpoint references from all documentation
Removed spec file references from all documentation

Testing

New Tests

Added performance tests for caching (Epic 1 Task 6.0)
Added integration tests for memory search caching (Epic 1 Task 5.0)
Added wave execution and performance regression tests
Added 18 planning system tests for spec removal and enhancements
Added dual-hybrid fallback tests

Test Improvements

Fixed test fixtures to use assigned_agent instead of recommended_agent
Added None checks for optional fields in ShowResult tests
Configured proper caplog logging for pytest

[0.9.4] - 2026-01-19

Fixed

Code Quality:

Fixed lint errors in multiple files
- Removed unused variables in packages/cli/src/aurora_cli/commands/memory.py
- Removed unused variables in packages/cli/src/aurora_cli/concurrent_executor.py
- Removed unused variables in packages/cli/src/aurora_cli/file_change_aggregator.py
- Removed unused variables in packages/spawner/src/aurora_spawner/spawner.py
- Fixed import ordering across all packages (243 auto-fixes)

Changed

Project Organization:

Moved development documentation from root to docs/development/
- ADHOC_SPAWNING_VALIDATION.md
- EARLY_DETECTION_ENHANCEMENT_ANALYSIS.md
- GOALS_SOAR_REFACTOR_PLAN.md
- PARALLEL_SPAWN_FAILURE_ANALYSIS.md
- STARTUP_OPTIMIZATIONS.md
Moved utility scripts from root to tools/
- benchmark_startup.py
- profile_indexing.py
- profile_indexing_detailed.py
Removed duplicate CLAUDE (copy).md file
Root directory now contains only essential files: README, CHANGELOG, CLAUDE.md, OPENCODE.md, conftest.py

[0.9.3] - 2026-01-19

Fixed

Spawner: Rate Limit Error Handling:

Rate limit errors now fail fast without wasteful retries
- Detects patterns: "rate limit", "429", "quota exceeded", "too many requests"
- Prevents 2-4 retry attempts that would hit the same quota limit
- Retry policy now skips rate limits with clear message: "Rate limit exceeded - quota exhausted, retries would fail"
Circuit breaker no longer opens on rate limit errors
- Rate limits are external API constraints, not agent failures
- Agents remain healthy in circuit breaker even when quota exhausted
- Early exit in record_failure() prevents tracking as circuit breaker failure
New FailureReason.RATE_LIMIT metric for separate tracking
- Distinct from inference failures in observability system
- Allows monitoring of quota-related issues separately
Files changed:
- packages/spawner/src/aurora_spawner/observability.py - Added RATE_LIMIT enum
- packages/spawner/src/aurora_spawner/spawner.py - Added rate limit detection
- packages/spawner/src/aurora_spawner/timeout_policy.py - Prevent rate limit retries
- packages/spawner/src/aurora_spawner/circuit_breaker.py - Skip circuit breaker for rate limits

Testing

Comprehensive test coverage for rate limit handling:
- 19 new unit tests in tests/unit/spawner/test_rate_limit_handling.py
- Tests cover detection, retry prevention, circuit breaker isolation, metrics, and edge cases
- All existing tests continue to pass

Documentation

Removed "Rate Limit Error Handling" from FEATURES_BACKLOG.md (now implemented)
Updated backlog last modified date to January 19, 2026

[0.9.1] - 2026-01-16

Performance

Startup Optimization - 40-60% Faster Initialization:

Lazy embedding model imports: Deferred torch/sentence-transformers loading eliminates 20-30s startup delay
- Only loads ML dependencies when semantic search is actually used
- Commands like aur mem index, aur soar, aur goals now start instantly
- Embedding model loads on-demand during first semantic search
SQLite connection pooling: Reuses validated connections across Store instances
- Eliminates redundant schema checks (60-90ms saved per invocation)
- Thread-safe per-database pool locks
- Skips repeated PRAGMA execution for pooled connections
Deferred schema initialization: Schema validation deferred until first database access
- Store creation now takes <1ms (was 30-50ms)
- Schema checks only run once per connection pool
Combined health monitoring: Merged two config methods into one
- Single config lookup pass (saves 10-15ms)
- Reduced logging verbosity
Lazy agent loading: Discovery adapter loads agents on-demand
- Eliminates 25-40ms of manifest parsing at startup
- Agents still available when needed, just loaded lazily

Total Impact:

Cold start: 150-250ms → 60-100ms (40-60% faster)
Warm start with pooled connections: 150-250ms → 30-50ms (70-80% faster)
Commands without semantic search: Nearly instant startup (no ML loading)

Affected Commands:

aur soar - Orchestrator startup optimized
aur goals - Goal decomposition starts faster
aur mem index - Indexing starts immediately (model loads during processing)
aur mem search - Search command starts instantly (model loads on first query)

Changed

Database Layer:

New ConnectionPool class in packages/core/src/aurora_core/store/connection_pool.py
SQLiteStore now uses deferred schema initialization
Connection pooling enabled by default for better performance

Orchestrator:

Combined _configure_proactive_health_checks() and _configure_early_detection()
Health monitoring configuration now single-pass

CLI Commands:

Removed eager agent registry population from aur soar
Discovery adapter handles agent loading on-demand

Testing

New benchmark script: benchmark_startup.py
Performance test suite in tests/performance/
Documentation in docs/PERFORMANCE_TESTING.md

[0.9.0] - 2026-01-16

Added

Spawner: Unified Parallel Execution Infrastructure:

New spawn_parallel_tracked() function as single source of truth for mature spawning
- Stagger delays (default 5s) between agent starts to avoid API burst limits
- Per-task heartbeat monitoring with HeartbeatEmitter + HeartbeatMonitor
- Global timeout calculation based on waves and policy (num_waves * policy_max + stagger + buffer)
- Circuit breaker pre-checks for fast-fail on known broken agents
- Retry with exponential backoff + LLM fallback
- Execution metadata collection (early_terminations, fallback_count, retried_tasks, etc.)
- Spinner display during execution showing active task count and elapsed time
Both aur spawn and SOAR collect phase now use the same infrastructure
New CLI options for aur spawn:
- --stagger-delay (default: 5.0s)
- --policy (default: patient)
- --max-concurrent (default: 4)
- --no-fallback to disable LLM fallback

New spawner configuration in defaults.json:

"spawner": {
  "max_concurrent": 4,
  "stagger_delay": 5.0,
  "default_policy": "patient"
}

display_name field in SpawnTask for progress display even when agent is None

Changed

Goals-SOAR Integration Refactor (Unified UX):

Removed brittle adapter layer between SOAR and goals command
Pydantic validators now coerce instead of reject:
- Agent names: automatically add '@' prefix if missing
- Subgoal IDs: automatically add 'sg-' prefix if numeric
- Dependencies: automatically convert numeric IDs to 'sg-N' format
Removed normalization functions (normalize_agent_name, normalize_dependency_id)
aur goals and aur soar now have unified UX until divergence point
SOAR phases 1-5 display naturally in both commands
Goals command shows agent assignments table after SOAR phases complete

CLI Output Improvements:

Ad-hoc spawned agents now display correct agent name instead of "llm"
Spinner shows during Phase 5 Collect: ⠧ Working... 3 active (125s)
Suppressed verbose HuggingFace Hub warnings during embedding model loading
- No more WARNING:huggingface_hub.utils._http noise during aur soar
- Retries still work, just not logged to console
Observability failure logs changed from WARNING/ERROR to DEBUG level
- Progress callbacks handle user-facing feedback
- Clean console output during normal operation
- Use --verbose or --debug to see detailed logs

SOAR Collect Phase:

Refactored to use spawn_parallel_tracked() instead of inline tracked_spawn()
Sets display_name=agent.id when creating SpawnTasks for proper progress display
Collect phase now inherits all spawner improvements (stagger, heartbeat, global timeout)

Fixed

Spinner regression in Phase 5 - spinner now properly displays and cleans up
Ad-hoc agent display name showing "llm" instead of actual agent name
Missing finally block in spawn_parallel_tracked() to stop spinner
HuggingFace warning spam during model loading

Performance

Stagger delays prevent API burst limits and improve success rates
Global timeout prevents hanging on stuck agents
Heartbeat monitoring detects and handles stalled agents

Testing

40 unit tests passing for planning models (Goals-SOAR refactor)
All spawner functionality tested via existing test suite
Clean output verified with manual aur soar and aur spawn testing

[0.8.0] - 2026-01-15

Added

Spawner: Early Failure Detection & Circuit Breaker:

Circuit breaker pattern with CLOSED/OPEN/HALF_OPEN states (2 failures → skip agent for 120s)
Early error detection with 11 regex patterns (rate limit, 429, connection errors, auth failures)
Per-attempt failure tracking for faster circuit opening within same query
Soft error handling: timeouts/rate-limits don't stop execution of other agents
Heartbeat emitter support for real-time progress tracking
All spawner logs changed to DEBUG level for clean CLI output
New file: packages/spawner/src/aurora_spawner/circuit_breaker.py (~100 lines)

SOAR UI Enhancements:

Plan Decomposition table display after Phase 4 verification
- Shows subgoal number, description, and assigned agent
- Spawned agents marked with asterisk (*) indicator
Summary panel with intelligent gap detection
- Format: "X subgoals • Y assigned • Z spawned"
- Lists missing agents that will fallback to LLM
- Proper singular/plural grammar (1 subgoal vs 2 subgoals)
is_spawn flag propagation from verify_lite through orchestrator to UI

Changed

SOAR Phase Improvements:

Collect phase now distinguishes soft vs hard errors
- Soft errors (timeout, rate limit, quota) continue execution
- Hard errors (auth failures, code errors) stop immediately
Timeouts don't block other agents from running in parallel
Better error messages and user feedback throughout pipeline

Memory:

aur mem search now records access and updates activation scores
ACT-R decay formula applied after each search (frequency + recency)
Chunks retrieved more often get higher activation in future searches

Fixed

Summary panel grammar: "1 subgoals" → "1 subgoal"
Gap detection using is_spawn flag instead of unreliable string matching
Non-friendly WARNING logs changed to DEBUG level
Clean CLI output with no verbose error messages during normal operation

Performance

Circuit breaker prevents wasting time on broken agents (skip after 2 failures)
Error patterns kill processes within 5s instead of waiting for 300s timeout
Parallel agent execution continues even when some agents fail

[0.6.7] - 2026-01-12

Improved

CLI Output Enhancements:

Added "For more details: aur mem stats" hint after indexing issues summary
Added log file path display: "Skipped files logged to: .aurora/logs/index.log"
Improved user guidance when indexing encounters warnings or errors

Documentation Updates:

Updated aur agents list documentation with all subcommands (--all, --category, search, refresh)
Updated aur init documentation with 3-step flow, --tools flag, and agent refresh option
Added agent discovery explanation and re-run options to init docs
Updated Quick Reference table with new agent commands

[0.6.6] - 2026-01-12

Fixed

Synchronized all package versions to 0.6.6
Pre-commit lint fixes (MD5→SHA256, docstrings, f-strings)

[0.6.5] - 2026-01-12

Added

TypeScript/JavaScript Parser Support:

Added TypeScriptParser for .ts and .tsx files
- Extracts classes, functions, arrow functions, methods, interfaces, and type aliases
- JSDoc comment extraction
- Handles React/JSX syntax
Added JavaScriptParser for .js, .jsx, .mjs, .cjs files
- Extracts classes, functions, arrow functions, function expressions, and methods
- JSDoc comment extraction
New dependencies: tree-sitter-typescript>=0.20.0, tree-sitter-javascript>=0.20.0
Auto-registered in parser registry

Memory Index Improvements:

File-level git blame caching for 336x speedup on subsequent function lookups
Batch embedding using native sentence-transformers batching
Two-line progress display with overall progress and phase detail
Files-by-language breakdown in aur mem stats output
Success rate now calculated from parseable files only (not all discovered)
Detailed index log written to .aurora/logs/index.log with:
- Summary statistics
- Files by language breakdown
- Failed files with error messages
- Skipped files (parseable but no extractable elements)

Fixed

aur init tool detection and duplicate display issues
Progress bar showing grey at 100% completion
Total Files count in stats now matches language breakdown sum
Warning help text now points to aur mem stats

[0.6.2] - 2026-01-10

Fixed

Build Configuration:

Removed non-existent headless/py.typed reference from aurora-soar package build config
Fixes FileNotFoundError during package installation in CI environments
Note: The headless directory was removed in previous refactoring (commit 5299248), but pyproject.toml reference was left behind

[0.6.1] - 2026-01-10

Fixed

Critical Dependency Fixes:

Fixed aurora-implement package name (was incorrectly named "implement")
Added missing dependencies to aurora-cli package:
- aurora-reasoning>=0.1.0
- aurora-planning>=0.1.0
- aurora-spawner>=0.1.0
- aurora-implement>=0.1.0
Updated all sub-package versions to 0.6.0 for consistency
Fixes ModuleNotFoundError: No module named 'aurora_spawner' when installing from PyPI

CI/CD Fixes:

Updated Makefile install targets to include all packages
Updated GitHub Actions workflow to install spawner, implement, and planning packages
Fixes 26 import errors in CI test suite

What this means for users:

pip install aurora-actr now correctly installs all required packages
Sprint 4 features (aur goals, aur spawn) now work out of the box
No manual package installation required

[0.6.0] - 2026-01-10

Added - Planning Flow & Execution (PRD-0026: All 4 Sprints)

aur goals Command:

New command for decomposing high-level goals into actionable subgoals with agent assignments
SOAR-based goal decomposition with 2-7 subgoals per goal
Automatic agent matching using keyword patterns and LLM fallback
Memory context integration with ACT-R activation-based search
Generates goals.json in .aurora/plans/NNNN-slug/ for /plan skill integration
Interactive user review flow with $EDITOR integration
Validation: 10-500 character goals, CLI tool existence checking
Options:
- --tool, -t: CLI tool selection (claude/cursor/windsurf)
- --model, -m: Model selection (sonnet/opus)
- --context, -c: Context files for informed decomposition
- --no-decompose: Skip decomposition for simple single-task goals
- --format, -f: Output format (rich/json)
- --yes, -y: Skip confirmation prompts
- --verbose, -v: Show detailed progress

Planning Flow Workflow:

aur goals "feature description" - Decompose goal → creates goals.json
/plan skill in Claude Code - Generate PRD and tasks from goals.json
aur implement or aur spawn - Execute tasks sequentially or in parallel

Documentation:

docs/commands/aur-goals.md - Comprehensive command reference (200+ lines)
docs/workflows/planning-flow.md - Complete workflow guide (400+ lines)
examples/goals/goals-example.json - Reference OAuth2 authentication example
Updated README.md with planning flow quick start
Updated COMMANDS.md with goals command and goals.json format specification

Test Coverage:

24 unit tests for goals command (100% coverage of core functionality)
4 E2E tests for command workflow
5 E2E tests for goals → /plan integration
Total: 33 new tests passing

Related Files:

Core implementation: packages/cli/src/aurora_cli/commands/goals.py
Planning models: Extended packages/cli/src/aurora_cli/planning/models.py
Planning core: Extended packages/cli/src/aurora_cli/planning/core.py
Tests: packages/cli/tests/test_commands/test_goals.py
E2E tests: tests/e2e/test_goals_command.py, tests/e2e/test_goals_plan_flow.py

Configuration:

Environment variables: AURORA_GOALS_TOOL, AURORA_GOALS_MODEL
Config file support: [goals] section with default_tool and default_model
Resolution order: CLI flag → env var → config file → default

aur spawn Command (Sprint 3):

New command for parallel task execution from terminal
Reads tasks.md with agent metadata comments ()
Parallel spawns CLI tools (claude, cursor, aider, etc.) per task
Reports completion status with rich progress display
Lightweight ~100-line spawning primitive (90% reduction vs complex orchestration)

Enhanced aur soar Command (Sprint 3):

CLI-agnostic execution using CLIPipeLLMClient
Parallel spawning for research sub-questions
Terminal-based orchestrator for complex queries
Tool/model resolution: CLI flag → env var → config → default

Infrastructure (Sprints 1-2):

New aurora-spawner package: Core spawning primitives
New implement package: In-Claude execution for /implement skill
SOAR collect phase now uses real agent spawning (replaced _mock_agent_execution())
Deprecated AgentRegistry, unified on agent_discovery module
Agent gap detection with LLM fallback for capability matching

Changed

Documentation Reorganization:

Consolidated 280+ scattered files into organized structure
Created 4 main directories: reference/, guides/, development/, archive/
Added 3 specialized directories: commands/, workflows/, examples/
Created README.md index for each section
Updated all internal documentation cross-references
Moved root-level completion summaries to archive
Comprehensive Sprint 4 documentation (1,695 lines COMMANDS.md, 1,954 lines TOOLS_GUIDE.md)

Removed

Test Artifacts:

Removed accidentally committed MagicMock/ test fixtures from repository root
Added to .gitignore to prevent future commits

[0.5.0] - 2026-01-06

Removed

Deprecated MCP tools (aurora_query, aurora_search, aurora_get) - replaced by slash commands and CLI equivalents
MCP configuration from aur init - now requires --enable-mcp flag for testing/development
MCP checks from aur doctor - MCP FUNCTIONAL section no longer displayed

Added

/aur:implement placeholder command - Future Aurora native workflow command for plan-based implementation
--enable-mcp flag - Opt-in flag for MCP server configuration during aur init
Comprehensive MCP deprecation documentation:
- docs/MCP_DEPRECATION.md - Architecture rationale and re-enablement guide
- docs/MIGRATION.md - Tool replacement mapping and behavior changes
- docs/ROLLBACK.md - Complete rollback procedures with 3 options

Changed

Slash commands now recommended - /aur:search and /aur:get are preferred over MCP tools
MCP infrastructure preserved - All 20+ MCP configurators and session cache kept dormant for future use
Default behavior - Fresh installations skip MCP configuration by default

Migration Guide

Tool Replacements:

Deprecated MCP Tool	CLI Replacement	Slash Command	Notes
`aurora_query`	`aur soar "query"`	N/A	Full SOAR pipeline execution
`aurora_search`	`aur mem search "query"`	`/aur:search "query"`	Formatted table output
`aurora_get`	N/A	`/aur:get N`	Session cache retrieval

Breaking Changes: None - deprecated tools were already replaced by superior alternatives.

Re-enablement: Use aur init --enable-mcp --tools=claude to configure MCP for testing/development.

Rollback Options:

Git tag: git checkout mcp-deprecation-baseline (complete revert)
Feature flag: Set "mcp": {"enabled": true} in config (fastest, no code changes)
Revert commits: See docs/ROLLBACK.md for detailed instructions

Reference

PRD: docs/prd/PRD-0024-mcp-tool-deprecation.md
Git Tag: mcp-deprecation-baseline (rollback baseline)
Documentation: docs/MCP_DEPRECATION.md, docs/MIGRATION.md, docs/ROLLBACK.md

[0.4.0] - 2026-01-05

BREAKING CHANGES

Package Structure Simplified:

All sub-packages (aurora-core, aurora-cli, etc.) now bundled directly into aurora-actr
Single package installation: pip install aurora-actr includes everything
No more separate package dependencies on PyPI
Package size: ~528KB (still tiny!)

What This Means:

✅ Users: Just run pip install aurora-actr - everything works
✅ Simpler installation, no verbose sub-package output
✅ All features included by default (CLI, MCP, memory, reasoning, planning, SOAR)
⚠️ If you had aurora-cli or other sub-packages installed separately, uninstall them first

Optional Dependencies:

[ml]: Machine learning features (sentence-transformers, torch) for semantic embeddings
[dev]: Development tools for contributors (pytest, mypy, ruff, etc.)

Installation:

# Full installation (recommended)
pip install aurora-actr

# With ML features
pip install aurora-actr[ml]

# Development (contributors only)
pip install aurora-actr[dev]

[0.3.1] - 2026-01-05

[0.3.0] - 2026-01-05

Added - Developer Experience & Configuration

Project-Local Configuration (Breaking Change):

All Aurora artifacts now stored in project-local ./.aurora/ directory (except budget tracker)
Plans created in ./.aurora/plans/ with sequential numbering (0001, 0002, etc.)
Config loading detects project mode when ./.aurora/ exists
Database, plans, and metadata now project-scoped for better isolation

Enhanced aur mem stats:

Added Last Indexed timestamp display (e.g., "2 hours ago")
Added Success Rate percentage when < 100%
Displays failed files list (up to 10) with error details
Displays warnings list (up to 10) for parse issues
Metadata persisted in .aurora/.indexing_metadata.json

Improved aur doctor Health Checks:

New TOOL INTEGRATION section:
- CLI tools detection (checks for claude, cursor, aider, cline)
- Slash commands status (configured/not configured)
- MCP servers status with tool names listed
Enhanced CODE ANALYSIS section:
- Tree-sitter language parsers detection and listing
- Shows available languages (python, javascript, typescript, etc.)
Removed duplicate/unnecessary checks:
- Removed redundant MCP check from CONFIGURATION section
- Removed API key check (no longer needed)

MCP Server Configuration:

MCP servers now configured automatically during aur init Step 3
Supports Claude, Cursor, Cline, Continue
Shows "(+ MCP server)" notation in init output

Fixed

Fixed CLAUDE.md stub overwriting existing content (now prepends stub)
Fixed error messages to reference aur mem stats instead of aur doctor for indexing issues
Fixed plan numbering to use project-local sequence instead of global
Fixed init command to show "Check with aur mem stats" for indexing errors

Changed

Updated all package versions from 0.1.0 → 0.2.0
Updated main version from 0.2.0 → 0.3.0
Doctor output now shows binary status (configured/not) instead of fractions

Added - BM25 Tri-Hybrid Search

Tri-Hybrid Retrieval Architecture:

BM25 Keyword Matching (30% weight): Code-aware tokenization for exact identifier matches
- Splits camelCase: getUserData → ["get", "user", "data", "getuserdata"]
- Splits snake_case: user_manager → ["user", "manager", "user_manager"]
- Preserves acronyms: HTTPRequest → ["HTTP", "Request", "httprequest"]
- Handles dot notation: auth.oauth → ["auth", "oauth"]
Staged Retrieval Pipeline:
- Stage 1: BM25 filtering (top-k=100 candidates from activation-ranked chunks)
- Stage 2: Tri-hybrid re-ranking (30% BM25 + 40% Semantic + 30% Activation)
Backward Compatibility: Dual-hybrid mode (set bm25_weight=0.0 for original 60/40 activation/semantic)

Configuration:

New HybridConfig parameters: bm25_weight, stage1_top_k, use_staged_retrieval
Config validation ensures weights sum to 1.0
Supports config loading from aurora_config.get("context.code.hybrid_weights")

Implementation:

BM25Scorer class with Okapi BM25 algorithm (k1=1.5, b=0.75)
Index persistence: save/load BM25 indexes
Result format now includes bm25_score field alongside activation_score and semantic_score
Backup of v1 hybrid retriever saved as hybrid_retriever_v1_backup.py

CLI Features:

--show-scores flag: Display detailed score breakdown with intelligent explanations in rich box-drawing format
- BM25 explanations: exact keyword match, strong term overlap, partial match
- Semantic explanations: very high/high/moderate/low conceptual relevance
- Activation explanations: access count, commit count, last used time
- Box-drawing format with Unicode characters for visual clarity
- Git metadata display: commit count and last modified time
--type filter: Search specific element types (function, class, method, knowledge, document)
Type abbreviations: Search results display abbreviated types (func, meth, class, code, reas, know, doc) for improved readability
Knowledge chunk support: Index and search markdown documentation files

Knowledge & Reasoning Chunk Support:

KnowledgeParser: Parse markdown files into searchable chunks (section-based splitting)
MarkdownParser: Registered in global parser registry for automatic .md file handling
ReasoningChunk: Support for SOAR reasoning patterns (created during pipeline execution)
CodeChunk validation expanded to support "knowledge" and "document" chunk types

Testing & Quality:

Unit Tests: 52 tests (15 BM25 + 5 staged + 6 knowledge + 4 reasoning + 4 type abbreviations + 7 box drawing + 6 BM25 explanations + 5 semantic explanations + 5 activation explanations)
Shell Tests: 20 acceptance tests covering exact match, CamelCase, staged retrieval, knowledge indexing, type abbreviations, box-drawing format, score explanations
Integration Tests:
- test_e2e_search_quality.py: MRR validation (target ≥0.85)
- test_index_rebuild.py: Index rebuild and BM25 IDF recalculation
- test_performance_benchmarks.py: Query latency and memory usage benchmarks
Type Safety: MyPy strict mode, all BM25 code type-checked
Lint: Ruff clean (code formatting and import organization)

Performance:

Simple queries: <2s latency (exact identifier matches)
Complex queries: <10s latency (semantic concept search)
Memory usage: <100MB for 10K chunks
Indexing throughput: >2 files/sec, >10 chunks/sec

Documentation:

Updated docs/cli/CLI_USAGE_GUIDE.md with --show-scores and --type examples
Added BM25 architecture section to docs/KNOWLEDGE_BASE.md
Comprehensive docstrings in BM25Scorer and HybridRetriever

[0.2.1] - 2025-12-29

Fixed - Phase 1: Core Restoration

Critical Bug Fixes (P0 Priority):

Issue #2 - Database Path Unification: All commands now use ~/.aurora/memory.db as single source of truth. No more local aurora.db files created. Config-based path resolution ensures consistency across all operations.
Issue #4 - Activation Tracking: Fixed record_access() not being called during search. Activation scores now properly update with each chunk access, using ACT-R decay formula. Scores vary across chunks instead of remaining static.
Issue #15 - Query Retrieval Integration: Direct LLM queries now retrieve context from indexed codebase before generating responses. Memory store properly initialized and passed to all query execution modes.
Issue #16 - Git-Based BLA Initialization: Chunks now initialized with activation scores based on Git commit history at FUNCTION level using git blame -L <start>,<end>. Frequently-edited functions have higher initial activation than rarely-touched functions in the same file. Gracefully falls back to base_level=0.5 for non-Git directories.

Critical Bug Fixes (Found During Manual E2E Testing):

Bug #1: Fixed retrieve_by_activation filtering out negative BLA values (sqlite.py:355) - Changed filter to use -inf when min_activation==0.0 to include all chunks
Bug #2: Fixed Query command not passing memory_store to Direct LLM (main.py:340) - Memory store now initialized and passed for all query modes
Bug #3: Fixed HybridRetriever not extracting CodeChunk content properly (hybrid_retriever.py:236) - Now extracts full function signature and docstring with proper line ranges
Bug #4: Fixed Config not respecting AURORA_HOME environment variable (config.py:226,359, init.py:151) - Added _get_aurora_home() helper function, tests now use isolated config/database

Feature Improvements (P1 Priority):

Issue #6 - Complexity Assessment: Added 16 domain-specific keywords (soar, actr, activation, agentic, marketplace, research, analyze, etc.). Multi-question queries (2+ ?) get +0.3 complexity boost. Domain queries now correctly classified as MEDIUM/COMPLEX.
Issue #9 - Auto-Escalation Logic: Low confidence queries (<0.6) now trigger escalation to SOAR pipeline. In non-interactive mode, auto-escalates automatically. In interactive mode, prompts user with clear choice.
Issue #10 - Budget Management: Implemented aur budget command group with 4 subcommands: show, set <amount>, reset, history. Budget checked before LLM calls, queries blocked with helpful error if limit exceeded.
Issue #11 - Error Handling: Added user-friendly error messages for authentication, API, network, configuration, and budget errors. Stack traces hidden by default, shown with --debug flag. Each error includes actionable solutions.

Changed

Database Management:

Default database location changed from ./aurora.db to ~/.aurora/memory.db
Migration tool added to aur init for existing local databases
Config file now specifies database path for all operations

Activation Scores:

Initial BLA calculated from Git commit history at function level (not file level)
Negative BLA values are valid (ACT-R log-odds representation)
Each function has individual activation based on its edit history
Access tracking now updates base_level, access_count, and last_access_time

Query Execution:

Direct LLM mode now retrieves context from memory store (not just SOAR mode)
Context includes file paths, line ranges, and full chunk content
Queries about indexed code return accurate answers from codebase

Complexity Assessment:

Threshold adjusted to confidence < 0.6 for auto-escalation
Domain keywords expanded from 8 to 16 terms
Multi-question detection added (2+ ? triggers complexity boost)

Documentation

Updated Documentation:

CHANGELOG.md - Added v0.2.1 release notes with all fixes and improvements
CLI_USAGE_GUIDE.md - Added budget commands section with examples
TROUBLESHOOTING.md - Added budget exceeded errors, configuration errors with --debug usage
PRD-0010 - Complete Phase 1 implementation documented

New Features Documented:

FUNCTION-level Git tracking via git blame -L <start>,<end>
Base-level activation (BLA) initialization from commit history
Negative BLA values are valid (ACT-R log-odds)
Budget management workflow with enforcement and history

Reference

Issues Fixed: #2, #4, #6, #9, #10, #11, #15, #16 (8 issues total)

Test Coverage: 12 E2E/integration tests added, 62 unit tests added, 4 critical bugs fixed during manual testing

Test Results: 31 budget tests passing, 10 integration tests passing, all 4 bugs verified fixed. E2E tests have environment setup issues (unrelated to bug fixes).

[Unreleased]

[0.3.1] - 2026-01-05

Added

Retrieval Quality Handling (TD-P2-016):

3-tier retrieval quality system: no match / weak match / good match detection
Interactive prompts for weak matches in CLI (groundedness < 0.7 OR < 3 high-quality chunks)
Automatic groundedness scoring to prevent hallucination on weak context
Activation threshold filtering (≥0.3) to exclude low-relevance chunks
--non-interactive flag for CI/CD and automation workflows
Production deployment guide for tuning activation and groundedness thresholds
MCP tools remain non-interactive (unaffected by retrieval quality handling)
Comprehensive test coverage: 7 integration tests + 18 edge case tests
Documentation in CLI_USAGE_GUIDE.md, TROUBLESHOOTING.md, and SOAR_ARCHITECTURE.md

Decision Matrix:

Scenario	Chunks	Groundedness	CLI Interactive	CLI Non-Interactive	MCP/Headless
No match	0	N/A	Auto-proceed + note	Auto-proceed	Auto-proceed
Weak match	>0	<0.7 OR <3 high-quality	Prompt user (3 options)	Auto-continue	Auto-continue
Good match	>0	≥0.7 AND ≥3 high-quality	Auto-proceed	Auto-proceed	Auto-proceed

Changed

Test Suite Systematic Cleanup (Phases 1-5):

Removed 20 low-value tests (constructor tests, implementation detail tests)
Archived 7 performance benchmarks for manual execution
Converted 79 @patch decorators to Dependency Injection pattern
Added 26 integration tests for CLI, memory manager, and auto-escalation
Improved test pyramid from 95/4/1 to 76/21/3 (unit/integration/E2E)
Increased coverage from 74.95% to 81.06% (+6.11 percentage points)
Created comprehensive testing documentation (TESTING.md, TEST_REFERENCE.md)
Marked 86 tests with pytest markers (critical, core, integration, e2e)
Fixed Python 3.11/3.12 compatibility (from 28+27 failures to 0)

Testing Documentation:

Added docs/development/TESTING.md - Testing principles and best practices
Added docs/development/TEST_REFERENCE.md - Test categorization and markers
Added docs/development/TESTING_TECHNICAL_DEBT.md - Technical debt tracking

Fixed

Python 3.11/3.12 test failures caused by @patch decorators (79 instances)
Import organization issues (19 auto-fixes)
Test fragility across Python versions

[0.2.0] - 2025-01-24

Added

MCP Server Integration:

Added FastMCP-based server with 5 tools for Claude Desktop integration
aurora_search: Semantic search across indexed codebase with hybrid scoring
aurora_index: Index directory with configurable file patterns
aurora_stats: Database statistics and metrics
aurora_context: Retrieve full file or specific function content
aurora_related: Find related chunks using ACT-R spreading activation
MCP configuration schema in CLI config with always_on, log_file, max_results
aurora-mcp control script: start, stop, status commands
Performance logging decorator for all MCP tools with latency tracking
Cross-platform MCP setup for Windows, macOS, and Linux

Package Consolidation:

Meta-package installation: pip install aurora installs all 6 components
Post-install hook displaying component-level feedback (✓ Core, ✓ CLI, etc.)
Optional dependencies: [ml] for sentence-transformers, [mcp] for FastMCP, [all] for everything
Automated import path migration script: aurora_* → aurora.* namespace
aurora-uninstall helper script with --keep-config flag

CLI Enhancements:

aur --verify command: Installation health check, dependency verification, diagnostics
Flexible --headless global flag: Both aur --headless file.md and aur headless file.md work
Enhanced help text with multiple examples for all commands
Improved error messages with actionable recovery steps for common failures
Better Path handling in aur init command

Documentation:

Comprehensive MCP Setup Guide (docs/MCP_SETUP.md) with 9 sections
Troubleshooting Guide (docs/TROUBLESHOOTING.md) with platform-specific notes
Updated README for v0.2.0 with MCP integration as primary workflow
Installation verification documentation

Testing:

Comprehensive MCP test suite (120+ tests) via Python client
Integration tests for all 5 MCP tools
MCP server startup/shutdown tests
Error handling and edge case tests
Performance and logging tests
Platform compatibility tests (Linux, with Windows/macOS support)
Smoke test suite with 11 end-to-end CLI tests

Fixed

Bug #1: aur init crash due to duplicate from pathlib import Path import (line 88 shadowing)
Bug #2: aur mem index API mismatch - changed add_chunk() to save_chunk(), fixed embeddings serialization
Bug #3: Dry-run import error - corrected import path to aurora.context_code.semantic.hybrid_retriever
Fixed 39 import statements across packages (aurora.* -> aurora_* for backward compatibility)
Resolved memory leak in MemoryStore.close() for Python 3.12
Fixed deprecated datetime.utcnow() for Python 3.12 compatibility

Changed

BREAKING: Import paths migrated to aurora.* namespace (e.g., from aurora.core.store import SQLiteStore)
- Old aurora_core.* imports still work but deprecated
- Use migration script: python scripts/migrate_imports.py
Version bumped from 0.1.0 to 0.2.0 in pyproject.toml
Project description updated to include MCP integration
README reorganized: MCP integration as primary workflow, standalone CLI as secondary
Updated feature list with Windows support and cross-platform compatibility

Deprecated

Old import paths (aurora_core.*, aurora_context_code.*, etc.) - use aurora.* namespace instead
Individual package installation (pip install -e packages/core) - use meta-package (pip install aurora)

Migration Guide (Breaking Changes)

Import Path Migration:

Old (v0.1.0):

from aurora_core.store import SQLiteStore
from aurora_context_code import PythonParser
from aurora_soar import SOAROrchestrator

New (v0.2.0):

from aurora.core.store import SQLiteStore
from aurora.context_code import PythonParser
from aurora.soar import SOAROrchestrator

Automated Migration:

python scripts/migrate_imports.py --dry-run  # Preview changes
python scripts/migrate_imports.py           # Apply changes

Installation Migration:

Old (v0.1.0):

pip install -e packages/core
pip install -e packages/context-code
pip install -e packages/soar
pip install -e packages/cli

New (v0.2.0):

pip install aurora[all]  # Single command

Backward Compatibility:

Old import paths still work in v0.2.0 but will be removed in v0.3.0
Individual package installation still works but not recommended
Migrate before v0.3.0 release to avoid breakage

[0.1.0] - 2025-01-15

Added

Phase 1 Foundation:

Core storage layer with SQLite and in-memory implementations
Chunk types: CodeChunk and ReasoningChunk with validation
Python code parser using tree-sitter
Context providers with scoring and caching
Configuration system with JSON schema validation
Agent registry for capability-based selection
Comprehensive testing framework (243 tests, 88% coverage)

Phase 2 SOAR Pipeline:

9-phase SOAR orchestrator: Assess → Retrieve → Decompose → Verify → Route → Collect → Synthesize → Record → Respond
Multi-provider LLM integration (Anthropic Claude, OpenAI, Ollama)
Cost tracking with soft/hard budget limits
Conversation logging with markdown format
ReasoningChunk pattern caching with ACT-R activation
60-70% query optimization via keyword-based assessment

Phase 3 Advanced Memory:

ACT-R activation engine with BLA, spreading activation, context boost
Semantic embeddings with sentence-transformers
Hybrid retrieval: 60% activation + 40% semantic similarity
Multi-tier caching: hot cache + persistent cache + activation cache
Query optimization: type filtering, threshold filtering, batch calculation
Performance: <500ms for 10K chunks, 30%+ cache hit rate

CLI Implementation (v1.1.0):

Complete CLI with aur command
Auto-escalation between direct LLM and full AURORA pipeline
Memory commands: aur mem index/search/stats
Configuration management with environment variable overrides
Headless reasoning mode for autonomous experiments
Dry-run mode for testing without API costs

Performance

Simple query latency: 0.002s (target: <2s) - 1000x faster
Complex query latency: <10s (met target)
Memory usage: 39.48 MB for 10K chunks (target: <100MB)
Chunk storage: ~15ms (target: <50ms)
Chunk retrieval: ~8ms (target: <50ms)
Throughput: >100 queries/second (target: >10)

Release Links

v0.2.0 GitHub Release (pending)
v0.1.0 GitHub Release (pending)

Version History

v0.2.0 (2025-01-24): MCP integration, package consolidation, CLI enhancements
v0.1.0 (2025-01-15): Initial release with SOAR pipeline and advanced memory

Last Updated: 2026-02-14

FilesExpand file tree

CHANGELOG.md

Latest commit

History

CHANGELOG.md

File metadata and controls

Changelog

[Unreleased]

[0.18.0] - 2026-04-14

Added

Changed

[0.17.6] - 2026-02-14

Added

Fixed

Changed

[0.17.2] - 2026-02-14

Fixed

Changed

Performance

[0.17.1] - 2026-02-12

Added

Fixed

Changed

[0.17.0] - 2026-02-10

Added

[0.16.0] - 2026-02-05

Removed

Changed

[0.15.0] - 2026-02-05

Added

Changed

[0.14.2] - 2026-02-05

Fixed

[0.14.1] - 2026-02-05

Added

[0.14.0] - 2026-02-04

Added

[0.13.6] - 2026-02-04

Added

[0.13.5] - 2026-02-04

Fixed

[0.13.4] - 2026-02-04

Added

Fixed

Changed

[0.13.2] - 2026-02-03

Added

Changed

[0.13.1] - 2026-02-03

Fixed

Deprecated

[0.13.0] - 2026-02-03

Added

Removed

Changed

[0.12.2] - 2026-02-02

Changed

[0.12.1] - 2026-02-02

Fixed

Changed

[0.12.0] - 2026-02-02

Removed

Added

Changed

[0.11.2] - 2026-02-02

Removed

Added

Changed

Fixed

[0.10.1] - 2026-01-30

Added

Fixed

Changed

[0.10.0] - 2026-01-26

Added

Changed

Fixed

Removed

Performance

Documentation