CortexDB 0.9.5 — the Code Intelligence Plane

The largest release since 0.9.0: CortexDB now has a code plane — a typed, versioned memory of your source code that recall and answers can draw on, built across four new crates and ~21 new /v1/code/* routes. Alongside it: the 10X performance campaign (boot in seconds instead of minutes, ~50 % artifact-compile cost cut, a two-tier LLM response cache), answer-free context preview, and a tl;dv connector bug-fix batch in cortexdb-connectors 0.2.15.

Compatibility: the /v1 API surface is purely additive (one new optional field on POST /v1/answer; everything else lives under the new /v1/code/* family). 0.9.x data directories load without a migration step — several stores switched to faster binary snapshot formats (FSN3, BTP2, LYS3) but all keep legacy read-compat. The code plane is off by default and free when off: with CORTEX_CODE_PLANE unset nothing is constructed and /v1/code/* answers 503 CODE_PLANE_DISABLED.

The Code Intelligence Plane (the headline)

CortexDB indexes registered source repositories into a content-addressed, generation-pinned code graph — definitions, imports, calls, framework/runtime resources — and serves it with honest precision labels. Every fact carries a precision tier (Syntax / Manifest / Compiler / History / Runtime) and a resolution status (Exact, ImportScoped, Ambiguous{candidates}, Unresolved, Dynamic), and both travel into the answer. There is no global name-match fallback anywhere — "never guess" is a permanent regression test, not a guideline. Reference: docs/CODE_PLANE.md; design docs/CODE_INTELLIGENCE_PLANE_DESIGN.md; operator quickstart in PRODUCTION_CONFIG.md.

What's in it:

• Storage + identity kernel: content-addressed store, gix git adapter (trees/diffs/renames/blame), an observation pipeline with an adaptive-coalescing file watcher, snapshot manifests with atomic head publication, and new WAL event arms so the plane replays like every other view.
• Extraction, 43+ languages: a sandboxed, per-file-budgeted tree-sitter pipeline — deep query packs for the launch languages plus a long-tail bundle, and a universal UTF-8 text fallback so docs/data files stay searchable. Pathological files are skipped with a recorded reason, never silently.
• Never-guess resolution + compiler tier: per-language resolvers (TS/JS tsconfig paths & export maps, Python shadowing/PEP 420, Rust module walking), then SCIP ingest (verified against scip-go and scip-typescript) where compiler facts supersede — never rewrite — syntax-tier facts.
• code_index posting engine: a new RocksDB store with delta segments, Roaring overlays, compaction-stable cursors, and hard query budgets — deep traversal without an explicit plan is refused with TRAVERSAL_REFUSED, never silently truncated. as_of time-travel materialization, freshness with hash-proof splicing, retention/GC with pins.
• Semantic code search (opt-in): one bounded structural representation per code-bearing file, embedded via CORTEX_CODE_EMBEDDING_* (OpenAI/Voyage/Cohere/Ollama), stored in a TQ2-compressed HNSW index with persistent-vector reranking and asymmetric query prefixes. Provider failure degrades to path/graph/BM25; a dimension mismatch fails instead of corrupting the index.
• Recall fusion: seven code question shapes behind a referent gate (CORTEX_CODE_SHAPES=shadow|active), a code-aware lexical namespace with an identifier tokenizer, and a server-side fusion stage doing shape-directed expansion + RRF with hard gating.
• Graph query & export: POST /v1/code/graph/query (bounded: ≤16 roots × 16 predicates, depth ≤6) and graph/export in nine formats — JSON, GraphML, Neo4j Cypher, self-contained vis/D3 HTML, SVG, Obsidian vault, Markdown wiki, DOT — all deterministic, no CDN.
• Evidence + certified answers: a code.context envelope with verbatim query-centered excerpts (exact line ranges in the citation, hard token ceiling, an uncertainty[] that names every gap), and POST /v1/code/certify for hash-verified list-implementations / count-call-sites answers that re-verify every input block and return incomplete reasons rather than overclaim.
• History & tests: hunk-precise diff→symbol anchoring with rename continuity, co-change over a bounded window, tests/for-change with static vs observed (lcov import) labeling.
• Everywhere it surfaces: a code-memory feature bundle + Admin UI page, cortexdb code CLI group (ten verbs), client.code.* in both SDKs, four read-only MCP tools (code_explore, code_impact, code_inventory, code_graph) with a compact --tool-profile code, admin-metrics code section (off-vs-empty is explicit), and hash-verified export/import index bundles. Tenant isolation is absolute: another tenant's repo is 404 and absent from listings.
• Benchmarked, honestly: the blind benchmarks/codeplane/ program records SWE-bench file localization at any@10 0.818 / all@10 0.754 with p50 18 ms queries (strongest no-LLM retrieval in our comparison; the write-up plainly notes LLM localizers still lead), 26/26 on the dynamic-multilang panel at 182 ms cold index, 54K LOC/s and ~22 bytes of index per LOC at production scale, and a 25M-LOC ingest in 93 s. Competitor arms are revision-pinned; manifests are append-only.

10X performance campaign

Boot, RAM, and LLM-cost work across the facts/artifact planes (docs/ENV_10X.md has the full env-var reference):

• Boot in seconds: layer snapshots + mutation journals (LYS3), binary fact-store (FSN3) and bitemporal (BTP2) snapshots, an HNSW journal-replay warm boot (full verify is now opt-in CORTEX_VECTOR_FULL_VERIFY=1), and all eleven store hydrations in parallel. Gate run: 26 s warm boot at 6.2 M events.
• Artifact lite profile (CORTEX_ARTIFACT_PIPELINE=lite): ~−50 % compile cost per session; deterministic triage turns signal-free turns into Complete bundles with no LLM call at all.
• Two-tier LLM response cache: RAM + persistent RocksDB, model-namespaced — a 48-call recompile replayed 46/48 from cache.
• Provider fleets + per-task routing: CORTEX_LLM_PROVIDERS JSON fleet with per-lane health rotation, CORTEX_LLM_ROUTE_<TASK> / CORTEX_LLM_TIMEOUT_SECS_<TASK> across 15 task kinds, per-provider cost accounting, and exponential retry cooling-off.
• RAM/WAL diet: embeddings moved out of WAL payloads and layer rows into an fp16 side column family; artifact store gained zstd + bloom filters; an O(n²) dedupe fixed.
• Ops: per-stage compiler LLM usage and enrichment call-family usage on admin metrics; config_lint now flags unknown/typo'd CORTEX_* env names at boot. New stress-rig gates (tools/stress-rig/) pin all of the above.

Answers & recall

• Answer-free context preview: POST /v1/answer accepts skip_answer_llm: true and returns the final quota-packed context_block + route/certification diagnostics with zero answer-stage LLM calls (the answer field fills only when a certified deterministic reducer succeeds).
• Certified deterministic reducers extended: seeded pairwise temporal deltas from ACL-checked sources, unique-explicit-day recovery that fails closed on ambiguity, artifact-route auditing in retrieval scorecards.
• Recall: focused lexical views run concurrently and fuse for noisy code tasks with primary results protected; direct and derived evidence channels interleave; source turns are recovered for untyped recall; derived memory is semantically retrievable at scale via a bounded vector backfill.

Connectors 0.2.15 — code-plane anchors + tl;dv bug fixes

• Code plane bridge: the GitHub connector emits code_anchors metadata (PR file lists, issue↔PR cross-links, push commit shas) that the server turns into discussed_in/changed_by bridge facts against file symbols — conversation memory linked to the code graph.
• tl;dv fixes from a real-data bug hunt (L1–L6): an expired API key no longer hard-crashes a one-shot sync (clean error, cursor held); solo/0-invitee recordings are labeled RESTRICTED instead of ORGANIZATION; a 24 h cursor-lookback window stops incremental sync from silently dropping late-transcribed meetings; TranscriptReady webhooks stamp the meeting's real happenedAt instead of the delivery time (and recover full meeting fidelity); episodes carry the real organizer name; JS-date parsing is locale-independent. Full detail in connectors/CHANGELOG.md.

Compatibility

• /v1 surface: additive only. New /v1/code/* family (503 when the bundle is off); new optional skip_answer_llm on POST /v1/answer.
• Data: no migration step. New binary snapshot formats write-forward with legacy read-compat; the code_index store is only opened when the code plane is enabled.
• One behavior note: CORTEX_ARTIFACT_WORKERS now gates the artifact compiler/projection workers directly (unset preserves historical behavior; =0 yields an immutable read-only artifact mode).
• New env vars are all registered with the boot-time config lint; the code-plane block is in .env.example, the 10X block in docs/ENV_10X.md.