v1.1.0

Added

SCIP Index Ingestion

• Parses SCIP protobuf index files (.scip/index.scip, index.scip, dump.lsif) and creates HIGH-confidence edges (confidence=1.0, edge_source=scip)
• Supports indexes produced by scip-python, scip-typescript, rust-analyzer, scip-java, and scip-go
• SCIP edges win over tree-sitter edges on conflicts — deduplication keeps SCIP as primary, tree-sitter as secondary
• Additive, non-blocking pass after tree-sitter extraction; malformed SCIP files log a warning and fall back gracefully
• cortex status shows SCIP coverage percentage per language
• cortex index --generate-scip suggests the appropriate indexer for the detected primary language when no SCIP index is present

Confidence-Tagged Edges

• Every edge now carries edge_source (scip, framework_adapter, ast_direct, name_match) and confidence (1.0 / 0.8 / 0.5 / 0.3)
• All MCP tools include edge_source and confidence_tier in results so agents can filter by reliability
• Default query threshold: confidence >= MEDIUM (0.7) — filters out name-match heuristics by default
• Optional min_confidence parameter on trace_callers and blast_radius to lower or raise the threshold

Framework Adapters — six new adapters detect wiring that tree-sitter cannot see:

• FastAPI: Depends(X) → Injects edges; @app.get/post/router.get/post/put/delete → Routes edges; transitive dependency chains
• Express: app.use(middleware) / router.use(middleware) → Middleware edges; router.get/post/put/delete(path, handler) → Routes edges
• NestJS: @Controller() → Routes edges; @Injectable() constructor injection → Injects edges
• Spring: @Autowired / @Inject → Injects edges; @Component/@Service/@Repository/@Controller marks injectable; @Bean factory methods
• Django: urlpatterns path()/re_path() → Routes edges; @login_required/@permission_required → Middleware edges
• React: JSX component renders → Renders edges; useContext(SomeContext) → Injects edges
• Framework detection scans dependency manifests (package.json, requirements.txt, pyproject.toml, pom.xml, build.gradle, go.mod, Gemfile, composer.json) — only relevant adapters run
• Manual override via .cortex/config.toml frameworks = ["fastapi", "express"]

User-Configurable Pattern Rules

• .cortex/patterns.toml supports custom regex rules with named capture groups, edge kind, and confidence tier
• Enables framework-specific wiring for any framework not covered by built-in adapters

Evidence-Fusion Task Context (get_task_context overhaul)

• Multi-signal ranking: lexical match (BM25, 0.30) + embedding similarity (0.25) + SCIP reference distance (0.20) + git recency (0.15) + edge confidence (0.10) + file size penalty (−0.05)
• Weight redistributed to lexical (0.55) when embeddings are unavailable
• Greedy budget packing always includes top-1 result; each included file gets a one-line reason explaining selection
• Falls back to file-proximity heuristic (recently modified files matching keywords) when graph signals are weak
• Response includes confidence (0.0–1.0) and coverage_percent fields
• Non-empty results guaranteed on a healthy index

Ask Tool Overhaul

• Priority-ordered data source querying: SCIP edges first → framework adapter edges → AST-direct edges → grep fallback
• FallbackSuggestion returned when average confidence < 0.7 or zero results — includes grep commands and file-read suggestions derived from query terms
• Token cap: results capped at 1000 tokens when average confidence is below MEDIUM
• confidence_warning annotation on responses when results exist but confidence is low (never returns empty when results are available)
• Detects unindexed file paths in queries and returns targeted fallback suggestions

Index Health Gate

• Startup health check validates files_indexed > 0, node_count > 0, edge_count > 0
• All MCP tools return a structured HealthError (with reason, suggested action, and fallback grep commands) when the index is unhealthy — never serves confidently wrong data
• cortex_status tool is exempt from the health gate
• cortex index --repair rebuilds the index from scratch (clears all graph tables, resets FTS5, re-runs full pipeline) while preserving observations and ADRs

Token Savings Accounting (honest math)

• baseline_cost computed as (matching_file_count × avg_file_tokens) + grep_output_tokens using FTS5 to count matching files
• net_saved = baseline_cost − (cortex_response_tokens + query_overhead_tokens) — negative values reported without modification
• cortex status --savings displays: cumulative net saved, average per query, net-negative query count, baseline total vs actual total

Repo Brief (get_repo_brief — new MCP tool)

• Zero-parameter cold-start summary under 400 tokens: detected languages and frameworks, top 5 entry points, top 10 hotspots (complexity × churn), auth/security patterns, test coverage shape, index health
• Cached in repo_brief_cache table; invalidated when files_indexed, node_count, or edge_count changes
• Returns partial info even when the index is unhealthy

Tool Surface Management

• Default tools (always-on, 10): get_repo_brief, get_task_context, ask, trace_callers, blast_radius, get_complexity_hotspots, get_git_hotspots, search_symbols, write_observation, read_observations
• Experimental tools (opt-in via .cortex/config.toml experimental_tools = true, 7): find_taint_paths, check_dependencies, decompose_boundaries, generate_steering, find_dead_code, generate_sbom, find_similar_functions
• Smart-tools mode (cortex serve --smart-tools, 5): get_repo_brief, ask, get_task_context, write_observation, read_observations
• All tool descriptions kept under 100 tokens
• semantic_search only appears in the tool manifest when embeddings are built

Hybrid Semantic Retrieval

• Incremental embedding: only re-embeds functions whose content hash changed since the last run; removes stale embeddings for deleted nodes
• cortex semantic enable builds the initial embedding index (Ollama nomic-embed-code or bundled ONNX model)
• Cosine similarity integrated as a 0.25-weight signal in evidence-fusion ranking
• cortex status shows embedding count and a degradation warning when running in BM25-only mode

Correctness Benchmarks

• cortex benchmark command runs JSON-based test suites with ground-truth answers for trace_callers, blast_radius, get_task_context, and ask
• Reports pass rate, per-tool accuracy, and average token savings vs grep baseline
• 24-case self-referential benchmark suite ships in benchmark/cortex_self_benchmark.json
• CI gate: cortex benchmark runs on every release and exits with code 1 if pass rate drops below 70%
• cortex status displays a warning when the last benchmark run was below the 70% threshold

Schema Migrations

• 0009_confidence_edges.sql: adds edge_source and confidence columns to edges; adds new edge kinds (Injects, Middleware, Routes, Renders); migrates existing edges to ast_direct / 0.5
• 0010_scip_coverage.sql: scip_coverage table (per-file SCIP tracking) and index_health singleton table
• 0011_enhanced_savings.sql: adds baseline_cost, net_saved, query_terms columns to token_savings
• 0012_repo_brief_cache.sql: repo_brief_cache table for cached repo briefs

Changed

• trace_callers now traverses Injects edges (framework DI injection sites) and follows Implements edges to include callers of interface/trait methods; annotates each result with edge_source and confidence_tier
• blast_radius now traverses all edge types (Calls, Imports, Injects, Implements, Middleware, Routes, Renders) with confidence filtering
• get_task_context replaced with evidence-fusion ranking (see above); response format extended with confidence, coverage_percent, and reasons fields
• cortex status extended with SCIP coverage %, active framework adapters, per-language file breakdown, semantic search status, and benchmark pass rate warning
• Indexing pipeline now runs framework detection before adapters, SCIP ingestion after tree-sitter, and updates index_health at the end of every index run
• Universal exclusions hardened: generation markers (// Code generated, # AUTO-GENERATED, @generated) detected in first 2KB; *.map source maps excluded; .cortexignore support verified

Fixed

• get_task_context no longer returns empty on a healthy index — file-proximity fallback guarantees at least one result
• ask no longer returns confidently wrong answers from a corrupt or empty index — health gate blocks all tools when unhealthy
• trace_callers missed framework-wired connections (FastAPI Depends, Spring @Autowired, NestJS constructor injection) — now included via Injects edge traversal
• Low-confidence name-match edges no longer pollute results by default — filtered at confidence >= 0.7
• User-defined pattern rules (.cortex/patterns.toml) were incorrectly stored with edge_source=name_match (confidence 0.3) instead of edge_source=framework_adapter (confidence 0.8) — pattern rule edges now pass the default min_confidence >= 0.7 filter and appear correctly in trace_callers, blast_radius, and ask results