# Changelog
Format based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
Project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
## [3.3.13] - 2026-05-21
### Changed
- **sqlitegraph 3.0.2 → 3.0.3** — Picks up AVX-512 SIMD performance improvements for graph traversals.
## [3.3.12] - 2026-05-21
### Fixed
- **`--root ./src --scan-initial` returned 0 files** — Auto-include logic added `src/` as a glob pattern, but FileFilter computed relative paths by stripping the root, so `src/lib.rs` became `lib.rs` and never matched `src/`. Now detects when root is already a source subdirectory and skips auto-include.
- **FTS5 index empty after bulk scan** — `scan_directory_with_filter` inserts directly into `graph_entities`, bypassing FTS5 triggers. Added automatic `rebuild_fts5()` after initial scan completes.
## [3.3.11] - 2026-05-19
### Changed
- **sqlitegraph 3.0.1 → 3.0.2** — Picks up V3Backend flush-error handling, PersistentHeaderV3 panic fixes, CliClient stack optimization, and HNSW docstring corrections.
## [3.3.9] - 2026-05-18
### Changed
- **sqlitegraph 2.2.5 → 3.0.1** — Upgraded core graph dependency. Brings built-in graph algorithms (SCC, cycles, topological sort), bulk insert APIs, and HNSW vector search. Internal `strongly_connected_components` implementation removed in favor of sqlitegraph's native `algo` module.
- **Schema compatibility auto-migration** — `db_compat.rs` updated for sqlitegraph 3.0.1 schema. Databases with older schemas are now auto-migrated on open. Newer schemas (forward-incompatible) are still rejected with `DB_COMPAT` markers.
### Fixed
- **`test_compute_delta_with_untracked`** — Removed CWD dependency by passing explicit `project_root` parameter. Test is now deterministic regardless of working directory.
- **`test_parse_frontmatter_float`** — Replaced `PI` comparison with exact `2.5` to avoid `f64::EPSILON` precision flakiness.
- **4 pre-existing clippy warnings** — Resolved in `indexer.rs`, `project_config.rs`, `watch_integration.rs`, and `source_inventory.rs`.
## [3.3.8] - 2026-05-12
### Added
- **`--detect-backend` flag** — New top-level flag that detects the backend format of a given database file. Usage: `magellan --detect-backend --db <path>`. Returns `sqlite` for valid SQLite databases, or exits with an error for missing or invalid files.
### Fixed
- **`status` on missing database** — Previously, `magellan status --db <missing>` silently created an empty database and reported 0 files/symbols. Now it fails with a clear "Database not found" error and exits non-zero.
## [3.3.7] - 2026-05-11
### Added
- **`#[cfg]` attribute extraction for CFG blocks** — During indexing, `#[cfg(...)]` attributes on functions are now parsed from the tree-sitter AST and stored in the `cfg_condition` column of `cfg_blocks`. All blocks within a cfg-gated function inherit the same condition. Supports `feature = "X"`, `all(...)`, `any(...)`, and `not(...)` forms.
- **`get_live_cfg_for_function()`** — Returns only CFG blocks whose `cfg_condition` evaluates true against the project's active features. Blocks without a condition are always included. This allows downstream tools (e.g. Mirage) to filter out dead code paths behind disabled feature flags.
- **`evaluate_cfg_condition()`** — Evaluates cfg condition strings against a set of active features. Supports `feature = "name"`, `all(...)`, `any(...)`, and `not(...)`. Unknown conditions conservatively return `true`.
### Changed
- **Schema v16** — Added `cfg_condition TEXT` column to `cfg_blocks`. Fresh databases include it; existing databases are upgraded automatically via migration.
## [3.3.6] - 2026-05-11
### Added
- **`magellan init` command** — Creates a `.magellan.toml` configuration file with sensible defaults. Detects the project root from `Cargo.toml` if present. Refuses to overwrite an existing config.
- **`.magellan.toml` configuration** — TOML-based project config with `[project]`, `[index]`, and `[watch]` sections. Supports `include`/`exclude` glob patterns for filtering what gets indexed.
- **Cargo.toml manifest parsing** — `magellan watch` reads `Cargo.toml` to extract feature flags, dependencies, and test/bench targets. Features are stored as JSON in `magellan_meta.project_metadata`. Dependencies are stored in `CargoManifest.dependencies`.
- **Config-driven indexing** — `magellan watch` loads `.magellan.toml` and respects `include`/`exclude` filters. Auto-detects `tests/` and `benches/` targets from `Cargo.toml` `[[test]]`/`[[bench]]` sections.
- **`magellan refresh` command** — Incrementally syncs the database with git working tree changes. Re-indexes modified files, removes deleted files, optionally includes untracked files. Supports `--dry-run` for preview.
### Fixed
- **`magellan init --path .` produced `name = "project"`** — Now canonicalizes the path before extracting the directory name, so `magellan init --path .` uses the actual project directory name.
## [3.3.5] - 2026-05-10
### Fixed
- **candidate-fact list SQL parameter mismatch** — `magellan candidate-fact list` (without --status) failed with "Wrong number of parameters passed to query". The no-status SQL branch used `LIMIT ?2` but only passed 1 parameter. Changed to `LIMIT ?1`.
- **query returns no symbols for indexed files** — `magellan query --file src/lib.rs` returned empty when the database stored paths with `./` segments (e.g., from `--root ./src`). `normalize_path_for_index` now strips `.` and resolves `..` path components so both indexing and querying produce matching paths.
## [3.3.4] - 2026-05-10
### Added
- **Crate-level docs for graph memory** — `lib.rs` now documents source-inventory and candidate-fact features in the crate root documentation visible on docs.rs
- **Public re-exports for candidate_fact API** — `CandidateFact`, `CandidateStatus`, `CandidateProperties`, `ValidationResult`, `ValidationError`, `ConflictSet`, `ConflictType`, `ResolutionStatus`, and key functions (`insert_candidate_fact`, `find_candidate_fact_by_id`, `list_candidate_facts_by_status`, `candidate_fact_review_queue`, `update_candidate_fact_status`, `validate_ontology`) re-exported from crate root for downstream crate users
### Changed
- **Docs: schema v12 → v14** — All internal `docs/` files updated from schema v12 to v14, including graph memory table documentation (`source_documents`, `candidate_facts`)
## [3.3.3] - 2026-05-10
### Fixed
- **Source-inventory frontmatter parser panic** — `magellan source-inventory --scan` panicked with `called Option::unwrap() on a None value` when scanning markdown files containing non-finite float values (NaN, inf) in YAML frontmatter. The `parse_frontmatter()` function used `serde_json::Number::from_f64(...).unwrap()` which returns `None` for non-finite floats.
- Replaced with safe `if let Some(n) = serde_json::Number::from_f64(fv)` check, falling through to string representation for non-finite values.
- Added 3 regression tests for negative numbers, floats, and non-finite float handling.
- **Candidate-fact submit silent errors** — `magellan candidate-fact submit` without `--candidate-id` sent an empty string as the candidate identifier, causing a UNIQUE constraint violation on the second insert that was obscured by a generic `.context()` error message.
- Auto-generates a UUID (`cf_<uuid>`) when `--candidate-id` is not provided.
- Error message now includes the candidate_id for diagnostics.
- Added regression test for duplicate candidate_id error reporting.
- **Watch mode database corruption** — `magellan watch --scan-initial` produced "database disk image is malformed" and "file is not a database" errors. Root cause: `CfgOps` opened a new SQLite connection per operation via `ChunkStore::connect()`, each without WAL mode or busy_timeout PRAGMAs, while 3 other connections were also writing to the same WAL file. Additionally, `insert_cfg_blocks()` ran `PRAGMA wal_checkpoint(TRUNCATE)` from one of these ephemeral connections, truncating the WAL while other connections had pending writes.
- Replaced all 7 `connect()` calls in `CfgOps` with `with_connection_mut`/`with_conn`, which use the shared `Arc<Mutex<Connection>>` instead of opening new connections.
- Removed the rogue WAL checkpoint from `insert_cfg_blocks()`. Only the watch loop checkpoints, from a single coordinated connection.
- Added `batch_mode: bool` toggle to `CodeGraph`, `SymbolOps`, `ReferenceOps`, and `CallOps`. Watch mode sets `batch_mode = false` before any file processing, falling back to individual per-insert auto-commit mode to avoid `BEGIN IMMEDIATE` contention.
- CfgOps now shares the same ChunkStore connection as CodeGraph instead of creating its own.
- Verified: two consecutive `magellan watch --scan-initial` runs complete with zero corruption, `PRAGMA integrity_check` passes, 146 files indexed cleanly.
## [3.3.2] - 2026-05-09
### Fixed
- **Missing SQLite transactions during indexing** — Fixed the root cause of ~27x throughput degradation when indexing larger codebases. Each file's symbol/reference/call inserts were executing in auto-commit mode, creating one WAL frame per INSERT. With ~20 symbols + ~20 references + ~30 calls per file, 146 files generated ~24,000 individual SQL executes. At 3.3x scale (483 files) this exploded to ~79,000 executes, causing super-linear WAL checkpoint overhead.
- `SymbolOps` now uses `sqlitegraph::bulk_insert_entities` + `bulk_insert_edges` wrapped in `TransactionGuard` (BEGIN IMMEDIATE...COMMIT) for all symbol nodes and DEFINES edges per file.
- `ReferenceOps` now batches reference node inserts and REFERENCES edges via the same bulk APIs.
- `CallOps` now batches call node inserts, CALLER edges, and CALLS edges via the same bulk APIs.
- All three modules receive an `sqlite_backend: Option<Arc<SqliteGraphBackend>>` field during `CodeGraph::open`, with graceful fallback to individual inserts for non-SQLite backends.
- Verified: `magellan watch --scan-initial` over `./src` (146 files, 2,853 symbols, 3,196 references, 4,265 calls) completes cleanly; DB integrity check passes; `magellan doctor` reports all OK.
## [3.3.1] - 2026-05-06
### Added
## [3.3.0] - 2026-05-06
### Added
- **`SymbolLookup` cache** — `graph/symbol_lookup.rs` provides an in-memory name→ID lookup table, eliminating repeated O(N) DB scans during reference resolution and query operations.
### Fixed
- **Watch-mode database corruption** — Fixed intermittent database corruption during `magellan watch --scan-initial` caused by uncoordinated SQLite connections writing to the same WAL:
- Reduced sqlitegraph connection pool from 2 → 1 (eliminates inter-pool WAL races)
- `rebuild_fts5()` now uses the shared `side_conn` instead of opening a new connection during flush
- `delete_file_facts` now cleans up `file_metrics` rows (both normal delete path and orphan cleanup)
- Added `PRAGMA integrity_check` verification after initial scan flush
- Added stress tests (5 sequential watch cycles, `#[ignore]`'d for CI speed)
- **FTS5 index rebuild on `magellan refresh`** — The `refresh` command now automatically rebuilds the `symbol_fts` virtual table after applying changes. Previously, `llmgrep` queries returned stale or empty results until a manual `INSERT INTO symbol_fts(symbol_fts) VALUES('rebuild')`.
- **`context` command `--path` alias** — `context symbol`, `context impact`, and `context affected` now accept `--path` as an alias for `--file`, consistent with `find` and `context file` commands.
## [3.2.0] - 2026-05-06
### Added
- **Context Analysis Commands** (Cross-Project Symbol Queries):
- `magellan context build` — Build context index
- `magellan context summary` — Show project summary
- `magellan context list` — Paginated symbol listing (multi-DB)
- `magellan context symbol` — Symbol detail with callers/callees/source
- `magellan context file` — File-level symbol context
- `magellan context impact` — Blast radius analysis (transitive callers)
- `magellan context affected` — Dependency reach analysis (transitive callees)
- Multi-DB support: pass a directory to `--db` to query all `.magellan/*.db` files
- JSON envelope output with `schema_version`, `execution_id`, `data`
### Fixed
- `--db` flag now appends instead of overwriting when specified multiple times
### Changed
- **FTS5 Full-Text Search Integration** (Schema v12):
- `symbol_fts` FTS5 virtual table for fast prefix searches
- 2.5× speedup on prefix queries (0.005s → 0.002s)
- Automatic index rebuild after batch indexing
- Migration: `magellan migrate --db <path>`
## [3.1.9] - 2026-05-03
### Added
- **FTS5 Integration Complete**:
- Schema v12 with `symbol_fts` FTS5 table
- `CodeGraph::rebuild_fts5_index()` for automatic sync
- Performance: 60% faster prefix searches
- Documentation: `FTS5_INTEGRATION_COMPLETE.md`
### Changed
- Version bump: 3.1.8 → 3.1.9
## [3.1.8] - 2026-05-01
### Added
- **Validation Infrastructure** (Phase 0):
- `scripts/validate-completion.sh` - 5-gate validation pipeline (stubs, check, test, clippy, db)
- `.claude/hooks/stub-check.fish` - Pre-commit hook blocking TODO/unimplemented/panic in non-test code
- `.claude/hooks/build-check.fish` - Pre-commit hook running cargo check/test/clippy
- `docs/superpowers/skills/no-stubs-enforcement.md` - Skill for zero-tolerance stub policy
- `docs/superpowers/skills/verification-before-completion.md` - Skill for completion gates
- `docs/superpowers/MASTER_PLAN.md` - Master plan for toolchain improvement
- **Phase 3: Cross-Project Federation**:
- `magellan registry scan --root <dir>` - Discovers .magellan/*.db files recursively
- `magellan registry list --root <dir>` - Lists discovered databases with stats
- `src/registry_cmd.rs` - New command module for database discovery
- **Phase 4: Editor Integration** (config file parsing):
- `src/config.rs` - Configuration management module (~/.config/magellan/config.toml)
- `magellan config show` - Display current configuration
- `magellan config init` - Create default config file
- Supports editor provider settings (ollama, openai, anthropic, custom)
- Supports registry settings (auto_scan, scan_roots)
- Added `toml` crate dependency for config parsing
- **Phase 5: Validation Pipeline** (verification infrastructure):
- `scripts/validate-completion.sh` - Pre-completion 5-gate validation pipeline
- `magellan verify --root <dir> --db <db>` - Path integrity verification (filesystem vs DB)
- `splice verify --before <snap> --after <snap>` - Snapshot comparison tool
- `.claude/hooks/` - Pre-commit hooks (stub-check.fish, build-check.fish)
- `src/checksum.rs` - File/span checksum computation for audit trails
- `src/code_validator.rs` - Rust-analyzer validation for LSP-verified mutations
### Fixed
- **Path resolution in query functions** (`symbol_nodes_in_file`, `symbol_nodes_in_file_with_ids`):
- Relative paths (e.g., `src/main.rs`) were not matching paths in the database (which stores absolute paths)
- Added `resolve_query_path()` helper that canonicalizes relative paths using current working directory
- `rebuild_file_index()` now stores absolute canonical paths in the in-memory index
- Now all query operations work consistently regardless of path format
- **Legacy database schema compatibility**: `ensure_cfg_schema()` and `ensure_coverage_schema()` now detect existing cfg_edges table with legacy schema before attempting schema modifications that would fail on legacy databases
- **splice dead-code / splice reachable**: Fixed by using local magellan path in splice's Cargo.toml
- Stale installed binary causing `backfill` command to fail with "Direct SQLite connection not available for shared backend" - rebuilt and reinstalled to `~/.local/bin/magellan`
- **mirage hotspots**: Fixed `mirage hotspots --entry main` returning "0 functions" (needed `mut db` for `conn_mut()` call)
### Changed
- Hook configuration updated to run stub-check and build-check via proper git pre-commit hook (not Claude Code hooks)
- Clippy validation now checks `--lib --bins` only (not tests) to avoid pre-existing test issues
- splice now depends on magellan from crates.io (version 3.1.7) for external users
- All Phase 1 P0 commands verified working: magellan backfill, magellan dead-code, splice dead-code, splice reachable, mirage hotspots, mirage unreachable
- Removed AI terminology from public documentation — this is a code intelligence toolchain, not an AI product
### Phase 2 Investigation (2026-05-01)
Investigated consistency issues from MASTER_PLAN Phase 2:
- **`llmgrep search --query "fn"`**: Returns 0 results is **expected behavior** - "fn" is a Rust keyword, not a symbol name. Use `--kind fn` to filter by function type (e.g., `llmgrep search --query test --kind fn` returns 729 results)
- **`splice query --label` vs `magellan query --file`**: Different semantics (label-based vs file-based query) - not a bug, design difference
- **`mirage hotpaths`**: Command exists and works correctly (discovered during investigation)
### Known Issues
- `mirage hotspots`: Verified working after `mut db` fix in mirage repo
- `mirage unreachable`: Verified working - `--within-functions` flag name is correct
- `tests/stress_concurrent_edits.rs::stress_database_integrity` can still deadlock (pre-existing)
- `tests/call_graph_tests.rs::test_cross_file_call_resolution` may fail (pre-existing)
## [3.1.7] - 2026-04-27
### Added
- `backfill` command to recompute metrics and derived data.
- `index` command to index a single source file.
- `delete` command to remove one file from the index.
- `cross-file-refs` command to report references to an FQN from other files.
- LCOV coverage ingestion with `ingest-coverage`.
- Stable `status` JSON coverage shape:
- `coverage.available`
- `coverage.covered_blocks`
- `coverage.covered_edges`
- optional `coverage.source`, `coverage.revision`, `coverage.ingested_at`
- CFG coverage side tables:
- `cfg_block_coverage`
- `cfg_edge_coverage`
- `cfg_coverage_meta`
- Optional `external-tools-cfg` feature for C/C++ and Java CFG extraction through
installed external tools.
### Changed
- Public documentation now describes the supported SQLite `.db` workflow as the
normal source-of-truth storage model.
- Multi-language single-file indexing now dispatches all supported language
parsers through the parser pool, including Python, JavaScript, and TypeScript.
- Compatibility preflight now distinguishes non-SQLite files, missing
`graph_meta`, missing `graph_meta.id = 1`, and schema mismatches before any
database mutation.
### Fixed
- Python files indexed through `CodeGraph::index_file()` now produce symbol
facts.
- JavaScript and TypeScript files indexed through `CodeGraph::index_file()` now
produce symbol facts.
- Incompatible database errors now use deterministic `DB_COMPAT` markers.
- Conflict cleanup around coverage/status wiring and command module staging.
### Known Issues
- `tests/stress_concurrent_edits.rs::stress_database_integrity` can still
deadlock and fail after its timeout. Treat this as unresolved until fixed in a
current verification run.
## [3.1.6]
### Added
- Schema version 11.
- CFG block hashes, statement snippets, and 4D coordinate columns in SQLite CFG
storage.
- CFG edge extraction improvements for short-circuit operators, `?` error paths,
and match guards.
- Status reporting for file, symbol, reference, call, code chunk, and coverage
counts.
### Changed
- SQLite is documented as the supported public database workflow.
- JSON output continues to use schema version `1.0.0`.
## [3.1.5] - 2026-03-20
### Added
- Symbol lookup indexes for faster symbol resolution.
### Fixed
- Rescan event handling issues in watch/index workflows.
## [3.1.4] - 2026-03-19
### Fixed
- CLI query issues in `status`, `query`, `find`, and `refs`.
- Unsafe algorithm implementation paths replaced with graph traversal helpers.
## [3.1.2] - 2026-03-15
### Added
- `refresh` command for git-aware database synchronization.
### Fixed
- Stale file index handling when files are deleted.
## [3.1.1] - 2026-03-15
### Added
- Improved symbol ranking and ambiguity reporting.
- `GraphStats` API.
### Fixed
- Ambiguous `find` command behavior.