CHANGELOG.md

Changelog

v0.14.0 — community PR hardening: NEO4J_DATABASE, NAMS cypher runtime, --ontology-file (2026-09-04)

Integrates five community PRs merged since v0.13.1 (#52, #56, #58, #59, #60), closes the gaps found while reviewing them, and adds test coverage at every level: +105 fast-suite tests, +3 Neo4j integration tests, a new bolt combo in the slow generated-suite runner, and 2 new tests inside every generated project (fast suite now 1,454 passing; the full CI suite with the connectors extra runs 1,866 passing / 1,881 collected). The new integration coverage immediately exposed — and this release fixes — a long-standing schema-DDL splitter bug that had been silently dropping five indexes/constraints from every seeded database.

Features (community PRs)

• NEO4J_DATABASE support end-to-end (#60, @henrardo). New --neo4j-database flag / NEO4J_DATABASE env var for the self-hosted backend, threaded through the CLI, wizard (new prompt on manual credential entry), Aura .env import (previously read and silently discarded), generated .env, the generated app's Settings, memory-layer MemorySettings, and the raw-driver session in execute_cypher(). Blank defers to the driver default (neo4j); set it for Aura instances provisioned via the Aura API/CLI, which commonly name the database after the instance id.
• Memory-layer failures now surface in /health (#60, @henrardo). store_message() failures are recorded into the same classified error state used at startup connect time; the bolt /health response gains memory, memory_error, and memory_error_detail fields, and the startup lifecycle checks the memory client instead of assuming success once the raw driver connects. Previously a wrong database name meant the app reported "ok" while every memory write failed with only a log line.
• execute_cypher() works on NAMS (#56, @benmyrgorod). New _execute_nams_cypher() routes through client.query.cypher with result-shape coercion (_coerce_nams_records), so domain agent tools and POST /cypher execute read queries on the hosted backend instead of dying on the never-connected bolt driver. _require_neo4j() now returns 503 when the NAMS client is missing, for every route.
• --ontology-file wired up (#58, @irene221b). The flag was documented (and tracked as #50) but had no code path. Scaffolds directly from a hand-written domain YAML — no LLM call — using the domain id declared in the file.
• Custom domains resolve by id (#52, @ecsricktorzynski, issue #30). load_domain() now searches ~/.create-context-graph/custom-domains/ (with a declared-domain.id fallback for renamed files), matching what list_available_domains() advertises. Bundled domains shadow same-id custom files.
• demo_scenarios optional in practice (#59, @irene221b). The generated Playwright spec indexed demo_scenarios[0].prompts[0] unconditionally, crashing scaffold generation for any domain that omits scenarios. Falls back to a generic prompt.

Bug Fixes (gaps found reviewing the PRs)

• --ontology-file scaffolds now include data/ontology.yaml. The renderer only wrote the ontology copy for --custom-domain (YAML string) or bundled domains (copy by id); hand-written ontologies matched neither branch, silently producing a scaffold without the file the docs promise. The raw YAML is now carried into the scaffold verbatim, and the flag also works when the wizard collects the remaining settings. --ontology-file + --custom-domain is now an explicit error instead of a silent precedence pick. (cli.py)
• NEO4J_DATABASE reaches scaffold-time ingest and the generated import script. #60 threaded the database through the generated app but not through ingest.py (_ingest_with_memory_client, _ingest_with_driver, reset_neo4j) or the scaffolded import_data.py bolt path — so --demo / --ingest / --reset-database / make import targeted the default database while the app read from the configured one, producing an empty graph on non-default-database instances. All bolt sessions now honor the setting; validate_connection() gains an optional database parameter. (ingest.py, neo4j_validator.py, templates/backend/connectors/import_data.py.j2)
• .env.example documents NEO4J_DATABASE. #60 updated dot_env.j2 only; the example file now carries the same commented block. (templates/base/dot_env_example.j2)
• Playwright spec survives a scenario with an empty prompts list. #59's guard handled a missing demo_scenarios but still crashed rendering when the first scenario had prompts: [] (legal per the Pydantic model). (templates/frontend/e2e/app.spec.ts.j2)
• Schema DDL splitter no longer eats statements or executes comment tails. Every consumer of generate_cypher_schema() output split on ; and skipped fragments starting with // — which (a) silently dropped the 5 real statements that sit behind comment headers (person_name, document_title, document_domain, document_name_unique and the local_file_fulltext index were never created by make seed/ingest), and (b) executed the tail of the "dimensions must match your embed model" comment as Cypher, raising CypherSyntaxError on every schema apply. New shared split_cypher_statements() strips comment lines before splitting; used by ingest.py (both bolt paths), the generated generate_data.py, and the integration suite. (ontology.py, ingest.py, templates/backend/shared/generate_data.py.j2)
• Test suite is hermetic against ~/.create-context-graph/custom-domains/. With #52, load_domain() joins list_available_domains() in scanning the user-local directory — so a contributor's saved custom domains leaked into every domain-iterating test (the "football-intelligence" failures reported while developing #60). An autouse conftest fixture isolates the path; tests that need custom domains patch it explicitly. (tests/conftest.py)

NAMS domain ontology activation

NAMS auto-binds every workspace to the generic nams-default ontology "until an explicit ontology is activated" — and pre-registers a server-side ontology for every bundled domain (a 1:1 catalog; the server's healthcare document is field-for-field our healthcare.yaml). Nothing ever activated one, so all data was stamped with (and extraction spoke) the default vocabulary. Now every NAMS touchpoint binds the workspace to the app's domain first:

• Generated memory.py: connect_memory() runs _ensure_nams_ontology() after the client connects — already active → no-op; catalog match by domain id → activate its latest version; unknown domain → create the ontology from the scaffold's new backend/app/ontology_document.json and activate it (the custom-domain path). Best-effort and logged: memory still works on nams-default if the ontology API is unavailable. Covers app startup and make seed (which connects through the same path).
• CLI ingest (run_nams_ingest) and the scaffolded import_data.py run the same ensure sequence as stage 0, pinned by the parity contract test (get_active → list → get → activate, keyword-called so the recorder can compare shapes).
• build_nams_ontology_document() (ontology.py) produces the server's OntologyDocument shape — {domain, entity_types, relationships}, excluding app-side sections (agent_tools, system_prompt, document_templates, …) — used by both the renderer (writes ontology_document.json into every scaffold) and the CLI ingest.

Verified against the production service: a fresh workspace on nams-default flips to healthcare on the generated app's first connect (reconnect is a no-op), and a custom test-domain scaffold creates and activates its ontology server-side with the full merged label set. Every stored entity is stamped with the active ontologyVersionId. Known limitation (server-side): the stored entity type remains {Person, Organization, Location, custom} regardless of the active ontology — activation governs the version stamp, validation mode, and extraction vocabulary, not the storage-type enum.

Live NAMS verification (and the bugs it caught)

The full flow — scaffold → ingest → boot → API — was exercised against the production NAMS service (memory.neo4jlabs.com) with a real API key and neo4j-agent-memory 0.5.0 (the version fresh scaffolds install). That surfaced five breaks the mocked suites couldn't see, all fixed and now pinned by tests; 19/19 live API checks pass afterward:

• Conversation memory silently failed on every message. The NAMS service only accepts messages addressed to conversation ids it minted at create time — client-chosen session ids 404 with "conversation not found," and MemoryIntegration (0.5.x) both posts the client id straight through and swallows the failure into an {"error": ...} return value. The generated memory.py now creates the conversation on first use per session and addresses the server-assigned id (_resolve_nams_conversation(), process-local cache — a restart starts a fresh conversation instead of erroring), and store_message() treats {"error": ...} returns as failures so they surface in /health rather than reading as success.
• Every document/body ingest write failed (role must be user, assistant, or system + the same conversation-404). All three NAMS ingest implementations (run_nams_ingest, the scaffolded import_data.py, and make seed's ingest_fixtures_nams) now create their message channels via create_conversation and send role="user" with a metadata.kind marker ("document" / "entity-body"). Before: 0/25 documents ingested; after: clean.
• /api/documents and /api/schema/visualization returned nothing. The live service coerces unknown entity_type values (OBJECT, EVENT) to custom, so the server-side OBJECT filter matched nothing — and it rejects empty search queries (query is required), so the schema view's enumerate-everything search 400'd. Both adapters are now cypher-first (using the scaffold's own _pole_type: OBJECT_ description marker for documents, and a type-count aggregation for the schema view), with the search-based flows kept as fallbacks.
• Graph expand / entity connections were dead. long_term.get_entity(id) doesn't exist in neo4j-agent-memory 0.5.x; expand_node_nams and get_entity_detail_nams now resolve id-addressed lookups and neighbor edges through the cypher API (which also surfaces the server-created SAME_AS resolution edges), keeping the old REST flow as a fallback.
• NAMS reset never deleted anything. long_term.delete_entity doesn't exist in 0.5.x (the old code swallowed the AttributeError and reported "0 entities removed"), the REST API has no delete endpoint, and the cypher API is read-only. --reset-database and the scaffold's make reset now say so honestly (with the current entity count) and point at the NAMS dashboard, instead of pretending. Docs updated to match.

Verified unchanged live: client.query.cypher read queries (the PR #56 dispatch) work and write queries are rejected; the reasoning trace ingest (start_trace/add_step/complete_trace) succeeds; add_relationship/add_fact/add_preference exist in 0.5.0 but raise NotSupportedError against NAMS, so the ccg-edges encoding remains the correct design; reasoning.list_traces is NotSupported client-side, so /api/traces degrades to an empty list on NAMS. Upstream issues worth filing against neo4j-agent-memory: MemoryIntegration should create conversations (or the service should honor client ids), add_message shouldn't silently target nonexistent conversations, and add_entity throws a client-side validation error when the server responds with a dedup/merge result.

Testing

• tests/test_generated_client_runtime.py (new, 31 tests) — executes the rendered context_graph_client.py and memory.py against doubles: NAMS dispatch + result coercion + tool-event collection + not-connected error; bolt session database= threading; MemorySettings database pass-through; store_message() error recording/clearing/NotSupportedError handling; the _classify_memory_error buckets /health reports; the NAMS conversation-id translation (created once per session, server id targeted, bolt untouched, create-failure fallback) and the swallowed-{"error"} failure path.
• test_routes_integration.py (+6) — mounts the generated FastAPI app: POST /cypher on NAMS dispatches through execute_cypher (and maps errors to 400), API routes 503 when the NAMS client is missing while /health reports degraded, bolt /cypher still injects the $domain parameter, and a live store_message failure flips /health to degraded with classified fields — then a successful write clears it Three more pin the cypher-first adapters: documents enumerated by description marker (search fallback for custom-typed entities), schema visualization aggregated via cypher.
• test_cli.py (+12) — TestNeo4jDatabaseFlag (flag→.env, blank default, Aura import, flag-beats-file precedence, dry-run display) and TestOntologyFileFlag (scaffold, data/ontology.yaml copy, static demo data, invalid YAML exit 1, missing file exit 2, --custom-domain conflict, auto-slug).
• test_wizard.py (+5) — _parse_aura_env four-tuple contract: database read/absent/quoted, missing URI/password aborts.
• test_generated_project.py (+17) — template pins for the database threading (config/client/memory/import script/.env/.env.example, NAMS .env exclusion), memory-error surfacing in main.py, the NAMS cypher branch, and scenario-fallback rendering (none / empty-prompts / real prompts).
• test_ontology.py (+9) — split_cypher_statements unit tests (semicolon-in-comment, comment-header recovery, commented-out DDL dropped, all-domains executable-statement sweep, proof the old pattern dropped person_name) and custom-domain isolation meta-tests.
• test_custom_domain.py (+3) — resolution precedence: bundled shadows same-id custom, corrupt custom YAML skipped during the id scan, underscore files ignored.
• test_bolt_ingest_parity.py (+1) — the scaffolded bolt import session must target settings.neo4j_database.
• test_integration.py (+3, --integration) — explicit-database ingest lands data (via ProjectConfig.neo4j_database), validate_connection accepts a database name and rejects an unknown one. TestSchemaCreation now applies DDL through the shared splitter and asserts the previously-skipped indexes exist.
• test_generated_tests.py — the slow generated-suite runner now also scaffolds one bolt project (5 combos), and asserts the new backend-specific health tests actually ran. The generated test_routes.py gains two tests per backend: NAMS degraded-client reporting, bolt live-write-failure surfacing.
• scripts/e2e_smoke_test.py — asserts the /health contract shape per backend on startup (bolt: neo4j + memory fields, logging a warning with the classified error when degraded; NAMS: nams field).

CI

• The full test suite now runs on every PR. The full-suite job (formerly matrix, main-push-only) runs pytest --slow --functional on all pull requests and main pushes: the 176-combo domain × framework matrix, performance tests, generated-project venv suites, and the local-file vault functional tests (the connectors extra is installed, which both enables --functional and materializes ~165 connector-dependent tests that module-level importorskip guards silently excluded from dev-only CI — the job runs 1,866 tests). The secrets-dependent smoke-test job stays main-only — fork PRs can't access repository secrets.
• CI installs are lockfile-driven. Every job now uses uv sync --locked --extra ... + uv run --no-sync. The old recipe (uv pip install -e ".[dev]" followed by bare uv run pytest) let uv run's implicit lockfile sync downgrade locked base dependencies underneath freshly-installed latest extras — a new anyio release (requiring typing_extensions.sentinel, newer than the locked pin) made the skew fatal, failing test collection with ImportError: cannot import name 'sentinel' from 'typing_extensions' on every branch. --locked also means a pyproject.toml dependency change without a matching uv lock fails fast with a clear message instead of resolving to something untested. Makefile test/lint targets use uv run --extra dev ... so local runs resolve from the lock the same way.
• 14 accidentally-tracked .pyc files untracked. They were committed in the initial commit (before .gitignore applied) and have silently churned in diffs ever since; .gitignore already covers __pycache__/ going forward.
• Ruff's rule set and version are pinned. The lint job previously installed unpinned latest ruff with no project config, so ruff 0.16's expanded default rule set broke the build with 294 findings for rules this codebase never adopted. [tool.ruff.lint] now pins select = ["E4", "E7", "E9", "F"] (the set the codebase is written against — expand it deliberately, not via upstream default drift), and ruff>=0.16,<0.17 ships in the dev extra so CI and make lint run the same binary. Test-job matrix gains fail-fast: false so one Python version's failure no longer cancels the other's signal.

Docs

• reference/cli-options.md — --neo4j-database and --ontology-file rows, NEO4J_DATABASE in the env-var table.
• how-to/use-neo4j-aura.md — caution block on Aura API/CLI-provisioned database names and how the failure presents in /health.
• how-to/add-custom-domain.md — --ontology-file semantics: domain id from the file, data/ontology.yaml copy, mutual exclusion with --custom-domain.
• reference/generated-project-structure.md — .env listing includes NEO4J_DATABASE.

v0.13.1 — v0.13.0 feedback report triage (2026-05-20)

Addresses the May 20, 2026 v0.13.0 feedback report. The report mixed verified issues with claims that don't match the current codebase; each claim was verified before scoping work. This release closes every real issue, makes the generated app.models module load-bearing, and adds regression tests so the v0.12.0/v0.13.0 fixes can't silently come back.

Bug Fixes

• create-context-graph --dry-run no longer demands a NAMS API key. The non-interactive path validated MEMORY_API_KEY before reaching the --dry-run branch, so users couldn't preview a scaffold without first signing up. The credential gate is now scoped to non-dry-run flows; --dry-run skips it entirely. (cli.py:392)
• Dead template_id parameter removed from list_documents_nams. The NAMS branch of GET /documents already raises HTTP 501 when template_id is supplied, so the parameter could never reach the function meaningfully. Signature is now list_documents_nams(skip, limit). (templates/backend/shared/memory_adapter.py.j2, templates/backend/shared/routes.py.j2)
• DocumentBrowser entity badges use a stable composite key. The mentioned-entities map was the last key={\${e.name}-${i}`}(index-tainted) site in the frontend; switched tokey={`${selectedDoc.document.title}-${e.name}`} so badges don't collide when the user navigates back to the same document. (templates/frontend/components/DocumentBrowser.tsx.j2`)

Code Quality

• Generated models.py uses Field(...) for required fields. Previously, generate_pydantic_models emitted name: str = ... (bare Ellipsis literal). Valid Pydantic v2, but unfamiliar to contributors and harder to extend with constraints (Field(..., min_length=1)). Required fields now emit name: str = Field(...). (ontology.py:478)
• New GET /schema/models endpoint wires app.models into the runtime. Returns the JSON Schema for every Pydantic entity model generated from the domain ontology, useful for frontend codegen and OpenAPI clients. Makes the previously-unused app.models module load-bearing. (templates/backend/shared/routes.py.j2)

Testing

• TestCompositeKeyRegressions (tests/test_frontend_logic.py) — five new assertions pin the composite-key patterns introduced in v0.13.0 (ChatInterface entity/preference/tool-call badges, DecisionTracePanel step keys) plus the new DocumentBrowser fix.
• TestV0131ModelsPolish and TestV0131TemplateIdRemoval (tests/test_generated_project.py) — assert that the generated models.py uses Field(...) (and never bare = ...), that the /schema/models endpoint compiles, and that the list_documents_nams signature no longer takes template_id.
• Three new Playwright tests (templates/frontend/e2e/app.spec.ts.j2) — watch the browser console for React duplicate-key warnings across a multi-prompt chat sequence, verify the decision trace step list renders without page errors, and verify the document browser entity badges render without page errors.

Not Changed (and why)

The v0.13.0 feedback report flagged several issues that don't reflect the current code. Documenting here so contributors don't re-litigate:

• "Backend connectors removed from scaffolded projects" — false. templates/backend/connectors/ still exists and renderer.py:474-514 renders connector modules + backend/scripts/import_data.py whenever --connector is supplied. Runtime make import / make import-dry-run / make import-retry targets continue to work on every scaffold.
• "pyproject.toml.j2 bloats NAMS users with sentence-transformers" — false. The template already branches on backend mode at lines 16-20: NAMS scaffolds pin neo4j-agent-memory[litellm], self-hosted scaffolds pin neo4j-agent-memory[litellm,sentence-transformers,extraction,fuzzy]. NAMS users never pull PyTorch.
• "Restore media / insurance / supply-chain domains" — these never existed. git log --all -- src/create_context_graph/domains/*.yaml shows no history for these labels. v0.13.0's "restored domains" are legal, education, cybersecurity, government — total domain count remains 27.
• "Unused i in ContextGraphView line 210" — leave as-is. The unused index is in a fallback extractNodesAndRels parser, not a render hot path, and removing it would churn a path that hasn't drifted in months.

v0.13.0 — v0.12.0 feedback report fixes (2026-05-20)

Addresses the May 2026 v0.12.0 feedback report: one runtime bug on the --self-hosted ingest path, one React state bug in the streaming chat, four key={i} re-render hazards, dead/over-fetching code in the document adapter, four restored domains, and documentation for the ccg-edges encoding strategy.

Bug Fixes

• _ingest_via_bolt() now async (scaffolded import_data.py). app.context_graph_client.get_driver() returns an AsyncDriver, but the generated _ingest_via_bolt() was def and called with driver.session() as session: followed by sync session.run(...). Every self-hosted make import / make import-retry failed at runtime with a coroutine-not-iterable error. Function is now async def, uses async with driver, driver.session() as session: to own the driver lifecycle, and awaits every session.run(...). Both call sites (main(), _retry_deadletter()) wrap with asyncio.run(...) to mirror the NAMS branch.
• ChatInterface "done" handler no longer reads stale streaming state. The setMessages callback in the done SSE branch referenced streamingEntities/streamingPreferences from closure — these are useState-backed and could be one render behind the most recent entities_extracted/preferences_detected event. Accumulators are now useRef-backed (streamingEntitiesRef, streamingPreferencesRef); the done handler reads from .current and the two-step setMessages was collapsed into one.
• list_documents_nams() pushes the OBJECT filter to the server. Previously, the over-fetch buffer (skip + limit + 50) could be exhausted by unrelated POLE+O entities, silently dropping valid documents on busy NAMS instances. Now passes entity_type="OBJECT" to search_entities. Dead template_id filter at the same site removed.
• useEffect for externalInput now also depends on loading. An "Ask about this" click that landed mid-stream was silently dropped because the effect didn't re-fire when loading flipped back to false.

Breaking Changes

• --framework maf alias removed. Use --framework anthropic-tools instead. Click now rejects maf with the standard "Invalid value for '--framework'" error. The FRAMEWORK_ALIASES dict and the resolved_framework property on ProjectConfig are gone — call sites use config.framework directly.

Domains

• Restored 4 domains as YAML definitions: legal, education, cybersecurity, government. Each ships with the full schema (entity_types, relationships, document_templates, decision_traces, demo_scenarios, agent_tools, system_prompt, visualization) plus a deterministic static-fallback fixture. Domain count: 23 → 27.

Code Quality

• key={i} replaced with stable identifiers in 4 locations. ChatInterface.tsx: suggested-prompt buttons keyed by prompt text; entity/preference badges keyed by ${type}-${name}-${i} and ${category}-${preference}-${i}. DecisionTracePanel.tsx: trace steps keyed by step-${i}-${action.slice(0, 32)}.

Documentation

• ccg-edges encoding strategy documented. New Docusaurus page (docs/docs/explanation/ccg-edges.md) explains the fenced YAML block inside entity descriptions, fragility (what happens if descriptions are edited or truncated), the migration playbook once NAMS adds native add_relationship, and the parity test that pins the format. Scaffolded README gains a short "How NAMS stores relationships" section linking to the page.
• Custom domain generation walkthrough. New Docusaurus page (docs/docs/how-to/custom-domains.md) covers create-context-graph --custom-domain "..." end-to-end, including the few-shot prompting strategy, the validate-retry loop, and how to convert a generated domain into a permanent contribution.

Testing

• Bolt ingest parity test. New tests/test_bolt_ingest_parity.py mirrors test_nams_ingest_parity.py — renders the scaffolded template, exec's it against a mock AsyncDriver, and asserts the Cypher statement + parameter sequence matches a frozen golden file.
• AST async-shape check. tests/test_generated_project.py now parses generated import_data.py and asserts _ingest_via_bolt is AsyncFunctionDef, every session.run call is inside an Await, and both call sites wrap with asyncio.run. backend/scripts/import_data.py was added to PYTHON_FILES so it gets py_compile-checked on every scaffold.
• Streaming ref-accumulation test. tests/test_frontend_logic.py now AST-checks that the done handler in ChatInterface.tsx.j2 reads from streamingEntitiesRef.current/streamingPreferencesRef.current, not the useState values.
• E2E coverage for new panels. Playwright tests for DecisionTracePanel (tab-switch + step rendering) and DocumentBrowser (pagination + markdown detail view).

v0.12.0 — NAMS-native connector ingest + ccg-edges relationship encoding (2026-05-19)

The big shift in v0.12.0 is that connector ingest (make import in a scaffolded project) now writes through NAMS REST on NAMS scaffolds — previously the generated import_data.py only spoke bolt Cypher, so the SaaS connectors didn't actually work on the default backend. The two ingest paths (CLI demo-fixture seeding and connector-driven imports) now share a single write shape pinned by a contract test.

New Features

• ccg-edges relationship encoding for NAMS. NAMS REST has no add_relationship endpoint yet, so each entity's outbound edges are now encoded into the source entity's description as a fenced ccg-edges YAML block (deterministically sorted by type then target). The frontend graph view recognizes this marker and renders edges from it; the agent reads them out naturally as part of the description. A one-shot migration replays them as native edges once NAMS gains add_relationship — the seam is _build_ccg_edges_block() in both src/create_context_graph/ingest.py and src/create_context_graph/templates/backend/connectors/import_data.py.j2.
• Dual-tracked documents on NAMS. Documents now land as both long_term.add_entity(name=title, type=OBJECT, description=...) AND short_term.add_message(role="document", content=...). The long-term entity is the queryable source of truth (matches the bolt :Document shape and powers /documents); the short-term message is extraction fuel for the NAMS server-side extractor. The document browser now reads from long-term entities and strips the ccg-edges / _pole_type: markers for a clean preview.
• Per-connector BODY_FIELDS mapping. Each connector declares BODY_FIELDS: dict[str, str] mapping entity_label → property_name for the prose body. The ingestor pipes that body through short_term.add_message so the NAMS extractor can mine it for secondary entities. Wired up for Comment.body (Linear), Message.content (Claude AI / ChatGPT / Claude Code), DecisionThread.content / Reply.content (Google Workspace), and Document.description / Section.description (local-file). Pure-metadata entities (Person, Project, Label) opt out by omission.
• Idempotent connector imports. Generated import_data.py now tracks per-connector watermarks in .context-graph/watermarks.json so re-runs only fetch deltas. Failed records (one at a time) are appended to .context-graph/deadletter.jsonl and the watermark only advances on a clean run. NAMS add_entity is trusted to merge by name for entity-level idempotency.
• --dry-run and --retry modes for import_data.py. --dry-run fetches connectors and writes data/fixtures.json without touching the memory backend (useful for sanity-checking a large pull before paying the write cost). --retry drains .context-graph/deadletter.jsonl back through the writer.
• New Make targets. make import-dry-run (fetch-only, writes fixtures.json), make import-retry (drain deadletter). The legacy make import-and-seed target collapsed into make import — it now both fetches and ingests in one step on either backend.
• run_nams_ingest() extracted as a reusable async function in ingest.py. Takes an already-open MemoryClient plus the fixture dict, ontology, optional body_fields, and optional on_event callback; returns counts and a list of failure records. Used by both the CLI's Rich-progress path and the scaffolded import_data.py. Pinned to the same call sequence as the generated importer via the new tests/test_nams_ingest_parity.py contract test (~526 LOC, 14 assertions).

Breaking Changes

• make import-and-seed removed from generated Makefiles. Replaced by the now-idempotent make import, which both fetches from connectors and ingests on either backend. Existing scaffolds with the old target keep working until regenerated; the generated import_data.py is the source of truth.
• /documents?template_id=... returns HTTP 501 on NAMS. Template-based filtering relied on the MENTIONS graph edges that NAMS doesn't yet have. The bolt path is unchanged; un-filtered GET /documents works on both backends.

Bug Fixes

• ingest_data() legacy bolt signature restored. v0.11.0 had changed the signature to take a ProjectConfig exclusively, breaking library callers that were on the older (neo4j_uri, neo4j_username, neo4j_password) triple. ingest_data() now accepts either: a ProjectConfig instance OR the legacy three positional bolt args; _coerce_ingest_config() synthesizes a bolt ProjectConfig from the legacy triple. CLI usage is unaffected.
• Cypher identifier injection guards on bolt fallback. _ingest_with_driver and _ingest_with_memory_client now validate relationship types, source labels, and target labels against [A-Za-z_][A-Za-z0-9_]* before string-interpolating them into a Cypher template. Unsafe identifiers log a warning and skip the record instead of executing. Closes the CodeQL py/code-injection-shaped finding flagged on the bolt ingest path. (NAMS path was never affected — it never builds Cypher.)
• Labeled MATCH on bolt relationship ingest. The bolt connector path was falling back to a label-less MATCH (a {name: $name}) for relationship endpoints, which would match across labels and silently mis-merge. Now uses MATCH (a:SourceLabel ...) / MATCH (b:TargetLabel ...) when the connector supplies both labels, with the existing sanitization guard.
• Deadletter retry on partial-failure imports. Records that failed during a connector run (rate limit, transient 5xx) are appended to .context-graph/deadletter.jsonl with their original payload; make import-retry (or python scripts/import_data.py --retry) replays them through the same write path. Records whose failure category isn't retryable (e.g. a permanently unsupported entity shape) are logged but kept in the deadletter for inspection rather than dropped silently.

Documentation

• docs/docs/explanation/memory-backends.md — replaced the "best-effort B-partial port" framing with the actual hybrid write shape: ccg-edges encoding, dual-tracked documents, BODY_FIELDS extractor channel. Added a note about the test_nams_ingest_parity.py contract test that pins the two ingest paths together.
• docs/docs/how-to/use-nams.md — rewrote the "Seeding a relationship-rich graph" section to reflect that NAMS scaffolds now do encode relationships (just inside descriptions) rather than dropping them, with a ccg-edges example block. Kept the --self-hosted --demo recommendation for native-edge workflows (expand_node, GDS, arbitrary Cypher).
• src/create_context_graph/ingest.py — module docstring rewritten end-to-end documenting the new NAMS write shape, the BODY_FIELDS extension point, and why the two ingestors are duplicated by design (rendered template vs. runtime CLI consumer).

Internal / CI

• New tests/test_nams_ingest_parity.py contract test — drives run_nams_ingest and the rendered import_data.py against a shared fixture and asserts identical call sequences against a mock MemoryClient. Pins entity → ccg-edges → body → document → trace ordering, fenced-block format, body-field routing, and decision-trace shape.
• Expanded NAMS test coverage. ~1,500 new test lines across test_ingest_nams.py, test_memory_adapter.py, test_nams_ingest_parity.py, test_routes_integration.py, and test_doc_snippets.py — covering ccg-edges round-trips, dual-tracked document reads, BODY_FIELDS per connector, deadletter retry behavior, watermark persistence, the 501 guard on template-filtered /documents, and the bolt Cypher identifier sanitizers.

v0.11.3 — Streaming for CrewAI/Strands + NAMS hardening (2026-05-19)

Rolls up streaming chat for the last two non-streaming frameworks, a much better NAMS failure-mode UX (classified errors surfaced in /health, memory-backend auto-detection from .env), lighter NAMS dependencies, custom-domain robustness fixes, and silent-failure cleanup across all 8 agent templates.

New Features

• Streaming chat for CrewAI and Strands. Both frameworks now implement handle_message_stream(), so the /chat/stream SSE endpoint streams text deltas as the model produces them — previously these two frameworks emitted tool events live but text only arrived at the end of the run. CrewAI subscribes to LLMStreamChunkEvent on the crew event bus; Strands iterates agent.stream_async(). Both include a 60s timeout, partial-text fallback assembly, and emit entities_extracted / preferences_detected events. This closes out the streaming matrix — all 8 frameworks now stream text.
• --import-preview CLI flag. Parses a chat export file (--import-file … + --import-type …) and prints a sanity-check summary (entity counts, conversation date range, sample titles) without scaffolding or ingesting. Useful before committing to a long import of a 1 GB+ ChatGPT/Claude AI export. Implemented in cli.py::_run_import_preview().
• NAMS error classification surfaced in /health. New _classify_memory_error() in memory.py.j2 buckets NAMS init failures into auth / rate_limit / network / config / unknown, with human-readable messages mapped per category (e.g. "NAMS authentication failed — verify MEMORY_API_KEY at https://memory.neo4jlabs.com"). The /health endpoint now returns nams_error, nams_error_message, nams_error_detail, and nams_dashboard so the frontend can show a useful diagnostic instead of a generic "memory unavailable". Exposed via get_error_category() / get_error_message() / get_error_detail() for the FastAPI startup banner.
• Memory-backend auto-detection from .env. New @model_validator in config.py.j2 reconciles memory_backend with the credentials actually present in .env: flips nams → bolt if MEMORY_API_KEY is blank but NEO4J_URI is set (and vice-versa), printing a warning. An explicit MEMORY_BACKEND env var still wins. Default Neo4j credentials in generated .env are now empty rather than baked-in placeholder passwords.
• Lighter NAMS dependency footprint. Generated pyproject.toml for NAMS scaffolds drops the sentence-transformers extra (NAMS does embeddings server-side) — extras shrink from [litellm,sentence-transformers] to [litellm]. _resolve_embedding_model() short-circuits to None on NAMS so the generated venv no longer pulls torch for a backend that doesn't use local embeddings.

Bug Fixes

• Custom-domain renderer crash. _get_domains_path() was imported inside a narrow try: scope in renderer.py, raising UnboundLocalError in the success path after partially writing ontology.yaml. Import hoisted to module scope.
• Custom-domain generation produced silently-truncated ontologies. custom_domain.py now checks the LLM response stop_reason for truncation, validates with Pydantic, and asserts completeness (non-empty system_prompt / visualization / agent_tools) before accepting. Provides actionable retry messages and clearer errors when the Anthropic/OpenAI SDK isn't installed.
• Agent-template degradations across all 8 frameworks. Jinja conditionals that checked param.default truthiness were rewritten to param.default is defined, and a silent fallback that swallowed Jinja syntax errors was removed so render failures now surface instead of degrading to stub code. Affects every framework template (anthropic_tools, claude_agent_sdk, crewai, google_adk, langgraph, openai_agents, pydanticai, strands).
• Partial streamed text discarded on agent errors. Strands and CrewAI now accumulate emitted text deltas and return the partial response when a generator raises mid-stream, instead of throwing away accumulated output. Also drops a redundant RuntimeError branch from the memory error classifier.
• NAMS Docker builds crashed on spacy download. The v0.11.2 fix for the generated Makefile is now mirrored in Dockerfile.backend.j2 via {% if not is_nams %} — NAMS images no longer fail at build time on a download command for a package they don't depend on.
• Frontend "Ask about" button rendered on non-string entity names. ContextGraphView.tsx.j2 adds a typeof === "string" guard before reading .properties.name.
• E2E selector regex out of date. e2e/app.spec.ts.j2 regexes updated from /try a demo scenario/i to /try these/i to match the current welcome card label.

Documentation

• docs/docs/how-to/use-nams.md — new "Seeding a relationship-rich graph" section. Documents NAMS's current lack of add_relationship REST support and shows two working patterns: (Option A) scaffold with --self-hosted --demo for the rich dev experience, flip MEMORY_BACKEND=nams for production reads; (Option B) seed bolt first then migrate. Includes a TODO(nams-relationships) pointer for the future server-side API.
• Generated README (base/README.md.j2) — minor wording updates for the NAMS sign-up flow and the --import-preview workflow.

Internal / CI

• CodeQL py/incomplete-url-substring-sanitization cleanup. Test assertions in test_nams_adapter.py switched from URL substring checks ("memory.neo4jlabs.com" in env) to full-URL or content-phrase matches.
• Frontend devDeps pinned (package.json.j2). Added @types/react-dom ^19.0.0; overrides section pins lodash ^4.17.24 and postcss ^8.5.10 to dodge known vulnerable transitive versions.
• New test coverage: ~600 new test lines spread across test_nams_adapter.py (NAMS error classification, backend auto-detection, Dockerfile spacy guard), test_renderer.py (template-degradation guards, custom-domain regression), test_custom_domain.py (truncation/completeness validation), test_generated_project.py (CrewAI/Strands streaming surface), test_chat_import.py (--import-preview), test_cli.py, test_doc_snippets.py, and test_routes_integration.py.

v0.11.2 — Post-release stabilization + pre-release smoke-render target (2026-05-19)

Rolls up three follow-up fixes surfaced by running v0.11.0/v0.11.1 end-to-end on a fresh machine, plus a durable safeguard against the same class of bugs.

Bug Fixes

• Full matrix + performance test suites broken on CI. test_matrix.py (184 combos) and test_performance.py (23 domains) defined their own local runner = CliRunner() fixtures that bypassed the auto---self-hosted shim added to test_cli.py in v0.11.0. With NAMS as the new default, every matrix/perf invocation hit the "NAMS API key required for non-interactive mode" guard and failed. 207 of 1,398 slow-suite tests failed on the v0.11.1 tag. Fix: moved _AutoSelfHostedRunner and the runner / nams_runner fixtures to tests/conftest.py so every test file inherits the auto-self-hosted behavior. Removed the now-duplicate fixtures from test_cli.py, test_matrix.py, and test_performance.py.

• make install and Docker builds crashed on NAMS scaffolds with No module named spacy. Both the generated Makefile's install-backend target and the generated Dockerfile.backend ran python -m spacy download en_core_web_sm unconditionally, but spacy is only present in the [extraction] extra which NAMS scaffolds correctly omit (entity extraction happens server-side on NAMS). Fix: wrapped the spacy download line with {% if not is_nams %} in both Makefile.j2 and Dockerfile.backend.j2. On bolt scaffolds the Makefile path is additionally guarded by an import spacy check so it stays robust even if the user uninstalls the extraction extras post-scaffold. Four regression tests in test_nams_adapter.py::TestBoltRenderedTemplates: test_makefile_skips_spacy_download_on_nams, test_makefile_guards_spacy_download_on_bolt, test_dockerfile_skips_spacy_download_on_nams, test_dockerfile_includes_spacy_download_on_bolt.

• make test in generated projects crashed with No module named pytest. The generated pyproject.toml didn't declare pytest or httpx anywhere, so uv sync never installed them — the generated tests/test_routes.py scaffold couldn't run. Fix: added [project.optional-dependencies] dev = ["pytest>=8.0", "httpx>=0.27"] to pyproject.toml.j2, and changed the generated Makefile's install-backend target from uv sync → uv sync --extra dev. Generated projects can now run make test out of the box.

New Tooling

• Root make smoke-render target — full scaffold → install → import-check → run-generated-tests sweep for both backends, in <1 min, no Neo4j / NAMS / LLM keys required. Catches the class of breakage the mocked unit suite can't see:
  • dep-resolution failures (uv sync conflicts)
  • install-time crashes (e.g. spacy download on NAMS)
  • import-time failures in generated app.main (e.g. questionary default validation, framework SDKs that validate API keys at module-load time)
  • generated test-scaffold regressions
• Sub-targets make smoke-render-nams and make smoke-render-bolt for per-backend runs. make smoke-render-clean removes the scratch directory (/tmp/ccg-smoke-render by default).
• Verified passing locally:
  • NAMS: render → install (no spacy) → import-check → 2 generated tests pass
  • Bolt: render → install (with guarded spacy download) → import-check → 2 generated tests pass

Process Note

The three issues bundled here all surfaced from running the actual product end-to-end on a fresh machine after the v0.11.0/v0.11.1 tags. Each was a class of issue the mocked unit suite couldn't catch by design (CLI fixture bypass, install-time shell commands, generated-project deps). make smoke-render is the durable answer — run it before tagging future releases.

v0.11.1 — Wizard framework-prompt fix (2026-05-19)

Bug Fixes

• Interactive wizard crashed at the framework prompt with ValueError: Invalid 'default' value passed. The value ('Strands') does not exist in the set of choices. — questionary.select(default=...) validates the default against Choice.title, not Choice.value. The wizard was passing the display label ("Strands") but choices used the framework key ("strands") as their value. Fixed by removing the default= argument entirely and reordering the choices list so DEFAULT_FRAMEWORK is first (questionary highlights the first row on entry). Regression test TestQuestionaryConstruction added to tests/test_wizard.py — exercises the real questionary.select constructor (only .ask is stubbed) so any bad default= argument fails at construction time. Total test count: 1,177 → 1,179.

v0.11.0 — NAMS by Default + LiteLLM Provider Injection (2026-05-19)

Breaking Changes

• Default memory backend flipped from self-hosted Neo4j to NAMS — create-context-graph my-app now scaffolds against the hosted Neo4j Agent Memory Service by default. The wizard collects a NAMS API key as its memory step. Use --self-hosted (or any explicit --neo4j-* flag) to opt into the legacy bolt path. Existing scaffolded projects are unaffected; only newly generated projects pick up the new default.
• Generated pyproject.toml pins neo4j-agent-memory>=0.4.0,<0.6.0 (was >=0.1.0). Extras conditional on backend: NAMS scaffolds get [litellm,sentence-transformers]; self-hosted scaffolds additionally get [extraction,fuzzy] for local entity extraction.
• ingest_data() library signature changed — now takes a ProjectConfig instead of separate Neo4j credentials. CLI users see no change; programmatic users of the create_context_graph package must update callers.
• MEMORY_API_KEY env var added to generated .env files. When set, the library auto-routes to NAMS even if MEMORY_BACKEND is unspecified.

New Features

• NAMS hosted backend support — Generated app/memory.py now constructs MemoryClient(MemorySettings(backend="nams", nams=NamsConfig(api_key=...))) on the NAMS path. Sign-up panel printed in the wizard with the https://memory.neo4jlabs.com landing URL.
• --self-hosted CLI flag — Preserves the legacy bolt-Neo4j path with full demo fixtures, schema DDL, and relationship-rich graph view. Recommended for workshops, screen recordings, demos, and air-gapped use.
• LiteLLM provider injection — Generated memory layer reads MEMORY_LLM and MEMORY_EMBEDDING env vars (LiteLLM-style provider strings, e.g. anthropic/claude-haiku-4-5, bedrock/anthropic.claude-3-haiku-20240307-v1:0, vertex_ai/gemini-1.5-flash, ollama/llama3). Native adapters resolve first (Anthropic, OpenAI, Bedrock, Vertex AI, SentenceTransformers); everything else routes through LiteLLM. Default fallback: sentence-transformers/all-MiniLM-L6-v2 for embeddings, anthropic/claude-haiku-4-5 (or openai/gpt-4o-mini) for entity extraction.
• Streamlined 6-prompt wizard — Collapsed from 11 prompts. Domain picker switched to questionary.autocomplete. Inline Anthropic-key prompt removed (deferred to post-scaffold .env editing with a prominent reminder panel). Advanced settings (MCP toggle, extraction toggles, extra API keys) gated behind a single Y/N prompt. Median wizard run is now ~6 prompts vs ~11.
• Default agent framework: AWS Strands — was previously unselected; the wizard now suggests strands as the default. All 8 frameworks remain supported.
• Backend-aware route adapters — Generated app/routes.py dispatches /expand, /documents, /traces, /schema/visualization, /entities/{name}, /search, /cypher to the NAMS REST adapter (app/memory_adapter.py) or bolt Cypher path based on MEMORY_BACKEND. /gds/* returns 501 on NAMS.
• Backend-aware MCP config — claude_desktop_config.json ships in NAMS or bolt shape depending on the scaffold. NAMS forces mcp_profile=core because extended-profile tools rely on unsupported endpoints (preferences/facts).
• Backend-aware make reset — On NAMS, enumerates entities via REST and deletes one-by-one (slow but correct, with a printed warning). On bolt, retains today's MATCH (n) DETACH DELETE n.
• NAMS-aware make seed — Generated generate_data.py branches on settings.memory_backend. On NAMS, delegates to memory_adapter.ingest_fixtures_nams() which does the B-partial port (see below). On bolt, applies schema + ingests via Cypher.
• Health endpoint backend-aware — /health returns {"memory_backend": "nams", "nams": <bool>} on NAMS or {"memory_backend": "bolt", "neo4j": <bool>} on self-hosted.

NAMS Write-Path Caveats (v0.4)

The NAMS REST API exposes a narrower write surface than bolt Cypher. The CLI does best-effort B-partial ingest with these documented gaps:

• Relationships are dropped on NAMS — add_relationship is not yet exposed by NAMS REST. The CLI logs a single warning per ingest run. The graph view shows entities but no edges.
• Entity properties collapse into description — NAMS REST accepts only {name, type, description} per entity. All other properties (status, severity, blood_type, etc.) are serialized into a markdown block inside description. The frontend property panel renders this markdown so the data remains readable.
• Preferences and facts are unsupported — auto_preferences=True is forced off on NAMS via ProjectConfig.effective_auto_preferences. auto_extract=True still runs but extracted relationships are silently dropped.
• Schema DDL skipped on NAMS — NAMS owns its schema. CREATE CONSTRAINT/CREATE INDEX statements from generate_cypher_schema() are no-ops on the NAMS path.

For the full relationship-rich demo experience, scaffold with --self-hosted --demo.

New CLI Flags

Flag	Purpose
--self-hosted	Use self-hosted Neo4j instead of NAMS (the v0.10 default behavior)
--nams-api-key	NAMS API key (also reads MEMORY_API_KEY env)
--nams-endpoint	Override NAMS endpoint URL (defaults to https://memory.neo4jlabs.com/v1)
--memory-llm	LiteLLM provider string for memory entity extraction
--memory-embedding	LiteLLM provider string for memory embeddings

New Docs Pages

• Use NAMS — sign-up, API key, switching between NAMS and self-hosted, troubleshooting
• Configure Memory Providers — LiteLLM provider strings, native adapters, default fallback behavior, per-provider auth examples
• Memory Backends — conceptual NAMS vs self-hosted comparison, choosing per-project, frontend dispatch architecture

Tests

• 1,177 passing fast tests (was 1,102 in v0.10). 50 new tests across 4 files:
  • test_ingest_nams.py (15) — _ingest_with_nams dispatch, entity serialization, document/trace ingestion, relationship-skip warning, missing-API-key error path, reset_memory_store for both backends.
  • test_wizard.py (7) — drives the interactive wizard via patched questionary; covers NAMS happy path, NAMS+advanced, self-hosted Docker, self-hosted existing-Neo4j, and edge cases.
  • test_memory_adapter.py (18) — renders a project, imports the generated memory_adapter.py via importlib, exercises every adapter function with AsyncMock MemoryClient.
  • test_routes_integration.py (10, gated by pytest.importorskip("fastapi")) — renders a NAMS or bolt project, mounts the generated FastAPI app via TestClient, asserts correct dispatch on /health, /documents, /search, /schema/visualization, /expand, /traces, /gds/*.
  • test_nams_adapter.py extended with 5 runtime dispatch tests.
• scripts/e2e_smoke_test.py — added --backend {bolt,nams} flag. bolt default preserves existing flow; nams exercises the hosted-memory scaffold path (requires MEMORY_API_KEY env).
• [dev] extras — now include fastapi, httpx, pydantic-settings so route integration tests run in CI.

Implementation Notes

• MemorySettings / MemoryClient / MemoryIntegration construction split in memory.py.j2 — explicit MemoryClient(settings) then MemoryIntegration(client=client, ...) (instead of having MemoryIntegration build the client implicitly). Enables NAMS backend + LiteLLM provider injection in one place.
• CodeQL false-positive fix — test_nams_adapter.py assertion changed from substring host check ("memory.neo4jlabs.com" in env_example) to full URL match ("https://memory.neo4jlabs.com/v1" in env_example) to satisfy py/incomplete-url-substring-sanitization.
• Generated test scaffold (backend/tests/test_routes.py) — fixture patches both bolt and NAMS connect/close paths so the generated test suite works regardless of backend.

v0.10.0 — Local-File Document Connector (2026-05-18)

New Connector

• local-file connector — Deterministic ingestion of local Markdown, PDF, HTML, AsciiDoc, and Word documents into :Document → :Section hierarchies, with LINKS_TO edges between sections and documents. No LLMs, no embeddings, no randomness. URI-keyed nodes integrate with the existing MERGE-on-(name, domain) pipeline without changes to ingest.py. Section URIs use GitHub/Pandoc slug rules (NFKD-normalize, ASCII-lower, [a-z0-9_] runs collapse to -); duplicate-heading collisions disambiguate per-parent (-1, -2, …); skipped heading levels (e.g. H1 → H3) preserve the original level on the child node rather than synthesizing intermediate H2s. Parser strategy per format: markdown-it-py with GFM extensions (Markdown); three-tier fallback pypdf outline → /StructTreeRoot → pdfplumber font-size heuristic (PDF); BeautifulSoup + lxml (HTML); pure-Python regex with block-delimiter state tracking (AsciiDoc); python-docx (Word). Adds 8 optional dependencies to the connectors extra: markdown-it-py, mdit-py-plugins, pdf-oxide, pypdf, pdfplumber, beautifulsoup4, lxml, python-docx. Implementation in connectors/local_file_connector.py and connectors/_local_file/ subpackage (parser, mapper, slug, link-resolver). 1,721-line test_local_file_connector.py plus test_local_file_vault.py functional test (make test-functional) round-trips a 14-file fixture vault.

Bug Fixes

• reset_database() connection lifecycle — Generated context_graph_client.py previously assumed a driver was already open. It now opens its own driver when _driver is None and closes it via try/finally, preserving any pre-existing connection. TestResetDatabase regression coverage added in tests/test_generated_project.py.

Tests

• 1,102 passing fast tests (1,321 collected including slow/integration/functional).

v0.9.5 — Options Intelligence Domain & GitHub Connector Enhancements (2026-04-29)

New Domain

• options-intelligence (23rd domain) — Options market intelligence covering 0DTE analysis, dealer positioning (GEX/DEX/VEX/CHEX), gamma regime classification, key levels, and trading strategies. 8 entity types (Underlying, OptionsContract, ExposureLevel, Regime, KeyLevel, Trade, MarketEvent, Strategy), 17 relationships (HAS_OPTION, EXPOSURE_AT, IN_REGIME, TRIGGERED_BY, FLIPPED_TO, PRECEDED_BY, etc.), 10 agent tools (get_regime, get_key_levels, get_exposure_by_strike, get_trades_by_strategy, etc.), 5 document templates (market briefs, trade journals, regime analysis), 9 decision traces (trade entry, regime flip, level breach, VIX spike, etc.), 4 demo scenarios. Pre-generated fixture: 65 entities, 125 relationships, 25 documents. 17 new property clamp ranges in generator.py for options-specific values (delta [-1, 1], gamma [0, 0.15], IV [0.05, 2.0], strike, GEX, etc.). 8 label-specific name pools and ID prefixes added in name_pools.py. Tickers (SPX, SPY, QQQ, IWM, AAPL) added to the global _TICKER_POOL.

GitHub Connector Enhancements

• Configurable per-resource limits — GITHUB_LIMIT=20 sets the default cap for issues, PRs, and commits. Override individually with GITHUB_ISSUES_LIMIT, GITHUB_PRS_LIMIT, GITHUB_COMMITS_LIMIT. Pagination now uses itertools.islice to stream results instead of materializing full pages.
• Issue/PR body import toggle — GITHUB_IMPORT_BODY=true (default) controls whether issue and PR bodies are imported as Document nodes. Issue/PR bodies are now labeled "Body" rather than "Document" to better reflect their source.
• Cross-link issues, PRs, and commits — GITHUB_LINK_ISSUES_PRS=true (default) creates CLOSES and REFERENCES edges between commits, issues, and PRs. Regex matches (close|closes|closed|fix|fixes|fixed|resolve|resolves|resolved) #N → CLOSES; bare #N → REFERENCES. GraphQL closingIssuesReferences provides authoritative PR closures. GITHUB_LINK_SOURCE=both selects regex / graphql / both (falls back to both on invalid values). CLOSES takes precedence over REFERENCES for the same pair, regex and GraphQL results are deduped, and references to numbers outside the fetched set are silently skipped. GraphQL failures log a warning and return empty rather than hard-failing.

Bug Fixes

• Connector relationship field names — All 7 connectors (GitHub, Notion, Jira, Slack, Gmail, Google Calendar, Salesforce) now emit source_name/target_name in relationship dicts (was source/target), aligning with the ingest schema.
• Generated config.py Settings fields — Added env-var fields for all 7 connectors (GitHub, Notion, Jira, Slack, Salesforce, Linear, etc.) so credentials and toggles are honored. Previously silently dropped by extra: "ignore".
• Generated pyproject.toml — Added connector package dependencies (PyGithub, notion-client, atlassian-python-api, slack-sdk, etc.) so generated projects install required SDKs.
• memory.py.j2 boolean rendering — Switched | tojson to | capitalize so Jinja-emitted booleans render as Python True/False rather than JSON true/false.
• CLI command/comment alignment — Fixed inconsistent spacing in CLI help output so commands and comments line up.
• Test patch target for is_connected — Tests now patch app.main.is_connected (where it's used after from … import is_connected) rather than app.context_graph_client.is_connected (where it's defined). Generated test_routes.py mock fixture also sets dummy API keys (ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY) before importing app.main, since PydanticAI's Agent(...) validates the key at module import time. Added regression guard test_test_file_mocks_is_connected in tests/test_generated_project.py so accidental removal is caught in the fast unit suite.
• options-intelligence fixture quality — Rewrote fixture with consistent per-underlying scoping. Cross-underlying contamination eliminated (each Underlying now has scoped OptionsContract, ExposureLevel, KeyLevel). Strike scales realistic per underlying (SPX ~5800, SPY ~580, QQQ ~500, IWM ~212, AAPL ~230). Regime timestamps chronologically consistent with FLIPPED_TO/PRECEDED_BY edges. Document content references title entities. Decision-trace placeholders replaced with concrete values from fixture entities. Greeks normalized (delta -1..1, gamma 0..0.15, proper put delta signs).

Improvements

• make schema target — Added to the generated Makefile. Comments on make seed and make import-and-seed clarified.
• Tests — 12 new tests for GitHub linking (regex/GraphQL/dedupe/precedence). 89 new tests for body-import toggle and configurable limits. 54 additional renderer tests covering connector field renames, memory boolean rendering, config Settings fields, and the new Makefile target.

v0.9.4 — GitHub Actions release workflow bugfixes (2026-04-17)

Bug Fixes

• Upgrade npm cli for trusted publishing - project currently uses npm 20 but we need the latest npm cli in order to perform the trusted publishing OIDC workflow.

v0.9.3 — GitHub Actions release workflow bugfixes (2026-04-17)

Bug Fixes

• Target npm-wrapper directory for node pkg - previously workflow was looing for package.json in top-level directory, but it lives in npm-wrapper.

v0.9.2 — Batch Seeding Performance (2026-04-15)

Bug Fixes

• Batch entity seeding with UNWIND — make seed previously executed one MERGE query per entity. With large Claude Code imports (27k+ entities), this caused seeding to appear to hang. Entity, relationship, and document creation now use UNWIND $batch with batches of 500, reducing ~27,000 round-trips to ~55.
• Batch ingestion in import_data.py — The --ingest path in generated import scripts also uses UNWIND batching.

v0.9.1 — Claude Code Connector Fixes & Docs (2026-04-15)

Bug Fixes

• Fix make import crash — The scaffolded Claude Code connector crashed with int(None) when max_sessions was passed as None from import_data.py. All credential reads now use the or idiom for None-safety.
• Fix missing Settings fields — Added claude_code_scope, claude_code_since, claude_code_max_sessions, claude_code_content_mode, claude_code_base_path to the generated config.py Settings class so .env variables are honored (previously silently dropped by extra: "ignore"). Also added google_client_id, google_client_secret, gws_folder_id for the Google Workspace connector.
• Fix dict vs attribute access — All template connectors return dict from fetch(), but import_data.py used attribute access (data.entities). Changed to data["entities"] across the board.
• Don't write empty fixtures.json on crash — The import script now skips writing fixtures.json when no data was collected, instead of silently overwriting it with empty lists.
• Fix make test-connection — Combined two sequential asyncio.run() calls into one, fixing the "Event loop is closed" error that made it print both "successful" and "failed".

Improvements

• Expanded scaffolded connector (268 → 457 lines) — The template connector now extracts 9 entity types (added GitBranch, Error, Decision, Preference, Alternative) with 14 relationship types (added ON_BRANCH, ENCOUNTERED_ERROR, MADE_DECISION, CHOSE, REJECTED, NEXT, PRECEDED_BY, USED_TOOL, EXPRESSES_PREFERENCE). Includes secret redaction, language detection from file extensions, file path validation, and [rerun: bN] suffix stripping.
• Claude-Code-specific demo scenarios — When --connector claude-code is active, the "Try these" prompts now show relevant questions ("What files have I modified?", "Show me decisions", "What are my coding preferences?") instead of generic software-engineering prompts about PRs and incidents.

Documentation

• Fixed dark-mode announcement bar (was white on dark background)
• Added explicit @easyops-cn/docusaurus-search-local plugin configuration
• Marked ANTHROPIC_API_KEY as required for the chat agent in the Claude Code tutorial
• Replaced "copy .env.example to .env" with "edit the generated .env" in the tutorial
• Added CLI flag literals next to connector display names in the intro page
• Added "Focus" column with one-line disambiguators to the domain catalog
• Custom 404 page with links to Introduction, Quick Start, and "Report Broken Link"
• "See all 22 domains →" link on the homepage carousel
• Tighter scroll transitions for the memory-type explainer
• Higher-contrast step numbers on the "How it works" section

Tests

• 7 new tests (1060 total): Settings fields, dict access, entity types, redaction, scenario override

v0.9.0 — Claude Code Connector, Google Workspace & Security (2026-04-02)

New Connectors

• Claude Code connector — Reads local session JSONL files from ~/.claude/projects/ with no authentication required. Parses user/assistant messages, tool_use/tool_result blocks, and progress entries. Extracts 7 entity types (Project, Session, Message, ToolCall, File, GitBranch, Error) with 10 relationship types. Includes heuristic decision extraction (user corrections, deliberation markers, error-resolution cycles, dependency changes) and preference extraction (explicit statements, package frequency). Secret redaction (API keys, tokens, passwords, connection strings) applied by default. 8 session intelligence agent tools injected via the renderer. 5 CLI flags (--claude-code-scope, --claude-code-project, --claude-code-since, --claude-code-max-sessions, --claude-code-content). Implementation split into connectors/claude_code_connector.py and connectors/_claude_code/ subpackage (parser, redactor, decision_extractor, preference_extractor).
• Google Workspace connector — Imports from 6 Google APIs (Drive Files, Comments, Revisions, Activity, Calendar, Gmail) with OAuth2 authentication and dynamic scope building. Extracts decision traces from resolved comment threads in Google Docs (question, deliberation, resolution, participants). 10 decision-focused agent tools (find_decisions, decision_context, who_decided, document_timeline, open_questions, meeting_decisions, knowledge_contributors, trace_decision_to_source, stale_documents, cross_reference). Cross-connector linking detects Linear issue references in comment bodies, doc names, email subjects, and meeting descriptions. 9 CLI flags for scoping imports. Rate limiting (950 queries/100s with exponential backoff).

Security

• Replaced weak cryptographic hashing — Switched from MD5/SHA1 to SHA-256 for content hashing in connectors (code scanning alerts #8 and #9).

Bug Fixes

• Session collision fix — Session.name now uses session_id as the unique MERGE key to avoid cross-session collisions when importing Claude Code data.
• Linear connector hardening — Don't fail on blank Linear team key; improved robustness of Linear import with better error handling.

Improvements

• Google Workspace connector template improvements and renderer integration
• Updated import_data.py template to handle new connectors
• 10 total SaaS connectors (GitHub, Notion, Jira, Slack, Gmail, Google Calendar, Salesforce, Linear, Google Workspace, Claude Code)
• 955 passing tests (1,165 collected including slow/integration)

v0.8.2 — Linear Connector (2026-04-02)

New Features

• Linear SaaS connector — GraphQL-based connector with cursor-based pagination and rate limiting against https://api.linear.app/graphql. Maps 12 entity types (Issue, Project, Cycle, Team, Person, Label, WorkflowState, Comment, ProjectUpdate, ProjectMilestone, Initiative, Attachment) to the POLE+O entity model with 26 relationship types. Imports issue relations, threaded comments with resolution tracking, project updates with health status, milestones, initiatives, attachments, and Linear Docs. Issue history entries are transformed into decision traces capturing state transitions, assignment changes, and priority changes with actor attribution. Uses stdlib only (urllib.request).
• Linear connector hardening — Named constants, structured logging, URLError/JSONDecodeError/429 handling with retry, pagination safety limits (MAX_PAGES), null-safe field access, team key validation during authenticate(), incremental sync via updated_after.

Bug Fixes

• Fix traces silent failure — Decision trace ingestion no longer silently fails on malformed data.

Documentation

• New tutorial: linear-context-graph.md — end-to-end guide for importing Linear project data
• Updated CLI options reference and SaaS data import guide

v0.8.1 — Schema Node Fix & Responsive Docs (2026-03-31)

Bug Fixes

• Click on schema node works again — The Button component was added to the template for the "Ask about [entity]" feature but was never added to the @chakra-ui/react import. Clicking a schema node crashed the React component, preventing double-click expand from working.
• Python 3.14 boundary — Added <3.14 to requires-python for forward compatibility.

Improvements

• Improved responsive design for Docusaurus landing page

v0.8.0 — Embedding Fix, Data Quality & Documentation (2026-03-28)

Critical Fixes

• neo4j-agent-memory no longer requires OpenAI API key — Removed [openai] extra from the generated pyproject.toml dependency. Conversation memory now uses local sentence-transformers (all-MiniLM-L6-v2, 384 dims) by default. If OPENAI_API_KEY is set in the environment, automatically upgrades to OpenAI text-embedding-3-small (1536 dims). Added sentence-transformers>=2.0 as an explicit dependency so local embeddings work out of the box with zero API keys.
• openai-agents framework warns about missing API key — CLI now displays a clear warning when --framework openai-agents is selected without --openai-api-key. The interactive wizard prompt text changes to indicate the key is "required" (not optional) for this framework.

Data Quality

• 67 new entity name pools — Added domain-appropriate names for every entity label across all 22 domains. LABEL_NAMES now has 118 entries (up from 51), eliminating all "Label 1" / "Label 2" fallback names. Covers agent-memory (Conversation, Memory, Session, ToolCall), digital-twin (Alert, Asset, Sensor, Reading, MaintenanceRecord, System), golf-sports (Round, Handicap, Hole, Course, Tournament), hospitality (Room, Reservation, Guest, Staff), oil-gas (Well, Equipment, Reservoir, Formation, Permit), personal-knowledge (JournalEntry, Note, Bookmark, Contact, Topic, Project), retail-ecommerce (Order, Product, Customer, Campaign, Category), vacation-industry (Booking, Package, Resort, Season), wildlife-management (Sighting, Camera, Habitat, Individual, Threat), conservation (Stakeholder), data-journalism (Correction), GIS (Boundary, Coordinate, Feature, Layer, MapProject, Survey), GenAI/LLM-Ops (Model, Prompt, Evaluation, Experiment), product-management (Epic, Metric, Objective, Release, Feedback, UserPersona), and scientific-research (Paper, Researcher, Grant, Institution).
• Post-generation value clamping — LLM-generated entities are now post-processed by _validate_and_clamp() in generator.py. Clamps 28 property types to domain-reasonable ranges (e.g., price_per_night: $30–$2,000; duration_hours: 0.25–24; rating: 1–5; latitude: -90–90). Also corrects taxonomy class mismatches (e.g., Bengal Tiger → "mammalia", not "aves").
• Richer entity descriptions — Added _LOCATION_LABELS, _EVENT_LABELS, and _OBJECT_LABELS sets (parallel to existing _PERSON_LABELS/_ORGANIZATION_LABELS) for POLE-type-aware descriptions. Added 7 label-specific description overrides for Medication, Permit, Sensor, Equipment, Paper, Model, and Species. Fallback descriptions no longer say "record tracked in the knowledge graph".
• digital-twin fixture fix — Fixed label casing in digital-twin.json (UPPERCASE → PascalCase) to match the domain YAML schema.
• Domain-scoped entity MERGE keys — Changed entity MERGE from {name: $name} to {name: $name, domain: $domain} in both generate_data.py.j2 and ingest.py. Prevents constraint violation warnings when multiple domains share a single Neo4j instance.

Framework Fixes

• google-adk AttributeError guard — Added try/except AttributeError around runner.run_async() in both handle_message and handle_message_stream to gracefully handle the google-genai SDK's BaseApiClient cleanup error when _async_httpx_client was never initialized.

Documentation

• Quick-Start page — New docs/quick-start.md with a 5-step guide: scaffold → Neo4j setup → configure → seed → start.
• use-neo4j-local guide — New docs/how-to/use-neo4j-local.md covering @johnymontana/neo4j-local (npx), Neo4j Desktop, and Docker standalone with troubleshooting tips.
• Domain catalog — New docs/reference/domain-catalog.md listing all 22 domains with entity types, agent tool counts, sample questions, and scaffold commands. Auto-generated from domain YAML files.
• Architecture diagram — Mermaid flowchart added to the Introduction page showing CLI → Template Engine → Backend/Frontend → Neo4j data flow. Added @docusaurus/theme-mermaid for rendering.
• switch-frameworks 404 fix — Added slug: switch-frameworks to frontmatter so /docs/how-to/switch-frameworks resolves correctly.
• Updated navigation — Sidebar now includes quick-start, use-neo4j-local, and domain-catalog pages.

Frontend UX

• Larger status indicator — Backend health dot enlarged from 8px to 12px with a text label ("Connected" / "Degraded" / "Offline").
• Health check retry on initial load — First page load now retries the health check 3 times with exponential backoff (1s, 2s, 4s) before showing "Offline". Prevents the transient "Internal Server Error" on initial Next.js compilation.
• Improved empty graph state — Empty knowledge graph panel now shows a link icon, "Your knowledge graph will appear here" heading, and actionable guidance text instead of a minimal "No graph data to display" message.

Testing

• 691 passing tests (89 new), up from 602
• New tests/test_fixtures.py (88 tests) — Cross-validates all 22 domains:
  • Schema alignment: fixture entities have all required YAML properties
  • Agent tool property references: Cypher queries only reference properties that exist in schema or fixtures
  • Label coverage: fixtures include entities for every YAML-defined label
  • Data quality: numeric property values fall within reasonable ranges

v0.7.0 — Documentation Site Redesign & CI (2026-03-28)

New Features

• Docusaurus landing page redesign — New animated terminal hero section with domain-specific demo commands, improved hero animation timing, and terminal width/height fixes.
• Mobile navigation — Responsive nav bar with mobile layout improvements and design polish.
• CI matrix job — Full test suite (including domain × framework matrix, perf, and generated project tests) now runs in the matrix CI job on push to main.

Documentation

• 4 new docs pages — "Use Neo4j Aura", "Use Docker", "Why Context Graphs?", "Framework Comparison"
• Updated sidebars with all new pages

Bug Fixes & Data Quality

• Bug fixes and data quality improvements across domains
• Updated docs and test coverage

v0.6.1 — Stability, Data Quality & Tool Coverage (2026-03-28)

Critical Bug Fixes

• CrewAI dependency fix — Changed crewai>=0.1 to crewai[anthropic]>=0.1 in framework dependencies. The crewai agent template uses llm="anthropic/claude-sonnet-4-20250514" which requires the anthropic extra. Without it, the generated project crashes on startup with ImportError: Anthropic native provider not available.
• CLI non-interactive mode fix — The CLI no longer requires a positional PROJECT_NAME argument when all flags (--domain, --framework) are provided. Auto-generates a slug like healthcare-pydanticai-app. Also added TTY detection with helpful error messages for CI/CD environments.

Data Quality Improvements

• Document Markdown rendering — Static document content now uses Markdown headings (##) instead of RST-style ===/--- separators. The DocumentBrowser component renders content with ReactMarkdown.
• Entity-derived document titles — Document titles now reference primary entities: "Discharge Summary: Maria Elena Gonzalez" instead of generic "Discharge Summary #1".
• Realistic entity descriptions — Replaced generic "Comprehensive patient profile for..." with POLE-type-aware descriptions using domain roles and industries (e.g., "Dr. Sarah Chen, attending physician specializing in healthcare").
• Domain-aware Organization.industry — Added DOMAIN_INDUSTRY_POOL for all 22 domains. Healthcare organizations get "Hospital Systems" instead of "Technology".
• Realistic decision trace observations — Observations now reference actual entity names: "Verified Dr. Sarah Chen against healthcare standards" instead of generic "Found 7 relevant records".
• Improved thinking text filter — Added continuation patterns to catch multi-sentence agent thinking blocks between tool calls.

New Agent Tools (All 22 Domains)

• list_* tools — Every domain now has a list tool for its primary entity type (e.g., list_patients, list_players, list_accounts) with sort and limit parameters.
• get_*_by_id tools — Every domain now has a direct ID lookup tool that returns the entity with all connections (e.g., get_patient_by_id, get_player_by_id).
• Gaming-specific — Added get_top_players tool (sort by level) for the gaming domain.

Frontend Improvements

• "Ask about this" button — Clicking a node in the Knowledge Graph shows an "Ask about [entity]" button that sends a query to the chat.
• Node hover tooltips — Graph nodes show full name, labels, and key properties on hover.
• Health polling optimization — Reduced polling frequency from 30s to 60s.
• Responsive hint text — Keyboard shortcut hint hidden on small screens to prevent overlap.
• Suggested question max width — Pill buttons capped at 320px to prevent layout stretching.
• Scrollable label badges — Label filter badges in the graph panel scroll when they overflow.
• Seed constraint fix — Entity seeding now uses ON CREATE SET / ON MATCH SET to avoid constraint violations on re-seed.

Documentation

• 4 new docs pages — "Use Neo4j Aura", "Use Docker", "Why Context Graphs?", "Framework Comparison"
• Updated sidebars — All new pages added to Docusaurus navigation

Testing

• 602 passing tests (57 new), up from 545

v0.6.0 — Comprehensive Testing Feedback (2026-03-28)

Framework Fixes

• CrewAI no longer hangs — Added explicit llm="anthropic/claude-sonnet-4-20250514" to prevent defaulting to OpenAI. Added request-level logging and reduced timeout to 60s.
• Strands serialization fix — Added _extract_text() helper that robustly extracts text from agent results, handling ParsedTextBlock serialization issues from newer Anthropic SDK versions.
• Google ADK API key support — Added --google-api-key CLI flag (GOOGLE_API_KEY env), wizard prompt when google-adk is selected, and GOOGLE_API_KEY in generated .env/.env.example templates.

Document & Trace Ingestion Fix

• --ingest now creates proper Document and DecisionTrace nodes — Both ingestion paths now create :Document and :DecisionTrace/:TraceStep nodes using direct Cypher, matching the generate_data.py pattern that the frontend expects. Previously, Documents and Decision Traces panels appeared empty after --ingest.
• Entity MERGE fix — Direct driver ingestion now uses MERGE (n:Label {name: $name}) SET ... instead of MERGE (n:Label {all_props}), preventing duplicate nodes.

Data Quality

• Domain-aware base entities — Person, Organization, Location, Event, and Object entities now use domain-specific names and roles (doctors for healthcare, traders for finance, game designers for gaming, etc.).
• Fixed templated property values — Properties like "Metformin 500mg - Contraindications" now replaced with realistic values. Added pools for contraindications, dosage_form, allergies, sector, lead_reporter, manufacturer, mechanism_of_action, population_trend, and habitat.

Frontend UI Improvements

• Redesigned chat input — Bordered container with focus highlight and keyboard shortcut hint (Chakra UI Pro inspired).
• Suggested questions redesign — Pill-shaped buttons with full text (no 60-char truncation), "Try these" label with Sparkles icon.
• Message avatars — User and assistant messages now have Circle avatars with User/Bot icons.
• Tool progress counter — Shows "Running tool N of M..." during tool execution.

CLI

• --demo convenience flag — Shortcut for --reset-database --demo-data --ingest
• --google-api-key flag — New CLI flag with GOOGLE_API_KEY env variable support

Testing

• 545 passing tests (35 new), up from 510

v0.5.3 — Agent Loop Breakout Fix (2026-03-27)

Bug Fixes

• Prevent agents from returning pre-tool text as the final answer — PydanticAI's run_stream fires a FinalResultEvent the moment any TextPart begins streaming. When Claude emits "I'll search for..." alongside a tool call, run_stream treats that text as the final output and exits before tool results are incorporated. Replaced agent.run_stream() + stream_text() with agent.run() in handle_message_stream, which completes the full agent loop before emitting text.
• Ruff lint fixes — Resolved lint errors across connectors, tests, and scripts.
• Test assertion fix — Directory conflict test now asserts on "not empty" instead of "already exists" for better cross-platform compatibility.

v0.5.2 — Agent Framework Refinements (2026-03-26)

• Improved Anthropic Tools and Claude Agent SDK agent templates
• Enhanced context_graph_client event handling and error recovery
• ChatInterface component improvements
• Better error handling in API routes
• generate_data.py improvements for data quality
• 74 new ontology validation tests

v0.5.1 — UX Improvements & Bug Fixes (2026-03-25)

Bug Fixes

• SSR hydration fix in frontend components
• PydanticAI tool serialization fix — agent tools now return JSON string types correctly
• Google ADK hyphenated domain name sanitization
• HuggingFace warning suppression in agent templates

Improvements

• Retry button on chat errors
• Agent thinking text collapsible filter — reasoning steps render in a collapsible "Show reasoning" section
• Strands max_tokens configuration support
• Cypher query validation tests across all 22 domains

v0.5.0 — Data Quality & Domain Completeness (2026-03-24)

New Features

• 22 complete domain ontologies with pre-generated LLM fixture data shipped for all domains
• Domain-specific static name pools — 200+ realistic names across 50+ entity labels (medical diagnoses, financial instruments, software repos, etc.)
• Label-aware ID prefixes (PAT- for Patient, ACT- for Account, etc.)
• 12+ domain-specific property pools (currency codes, ticker symbols, drug classes, medical specialties, severities)
• domain property on all ingested entities for cross-domain isolation when sharing a Neo4j instance
• Structured document templates for static fallback data generation

Bug Fixes

• Fixed missing SSE event messages in chat streaming
• Float value clamping for confidence/rating/efficiency fields

Testing

• 510 passing tests (145 new), up from 365

v0.4.6 — Conversation Fetching Fix (2026-03-24)

• Fixed conversation history fetching in context_graph_client for multi-turn sessions

v0.4.5 — Framework & Build Fixes (2026-03-24)

• Fixed pyproject.toml build configuration bug
• Strands framework default changed from Bedrock to AnthropicModel
• Agent template improvements across multiple frameworks
• Domain YAML fixes for gaming, genai-llm-ops, healthcare, personal-knowledge, product-management, retail-ecommerce, software-engineering, and trip-planning

v0.4.4 — Strands Model Default (2026-03-24)

• Changed Strands agent framework default from AWS Bedrock to Anthropic native model (AnthropicModel)

v0.4.3 — API Keys & Docker Support (2026-03-23)

• Fixed API key handling and validation across agent frameworks
• Added Dockerfile.backend template for Docker builds
• Makefile improvements for containerized deployments

v0.4.2 — E2E Testing & Bug Fixes (2026-03-23)

• Bug fixes across agent templates
• Playwright e2e test scaffolding for generated projects (app.spec.ts, playwright.config.ts)
• Improved e2e smoke testing infrastructure

v0.4.1 — Streaming & E2E Infrastructure (2026-03-23)

• Server-Sent Events (SSE) streaming for real-time chat responses and tool call visualization
• POST /chat/stream endpoint with asyncio.Queue-based event streaming
• Token-by-token text streaming for PydanticAI, Anthropic Tools, Claude Agent SDK, OpenAI Agents, LangGraph
• Real-time tool call events with Timeline/Spinner/Collapsible UI components
• Text delta batching (~50ms) to optimize React re-renders
• E2E smoke testing infrastructure (scripts/e2e_smoke_test.py)
• Documentation updates

v0.4.0 — Hardening, Security & DX (2026-03-23)

Bug Fixes

• Critical: Enum identifier sanitization — special characters (A+, A-, 3d_model) in domain ontology enum values now generate valid Python identifiers with value aliases
• Critical: Graceful degradation when Neo4j is unavailable — backend starts in degraded mode, /health endpoint reports connectivity status
• High: Cypher injection prevention in GDS client — label parameters validated against entity type whitelist
• High: CrewAI async/sync deadlock resolved — replaced bare asyncio.run() with nest_asyncio-compatible helper, crew execution moved to thread
• High: Claude Agent SDK model version now configurable via ANTHROPIC_MODEL environment variable
• Medium: Silent exception swallowing replaced with structured warning messages in ingest.py and vector_client.py
• Medium: JSON parsing errors in agent tool calls now return helpful error messages instead of crashing
• Medium: Input validation (max_length) added to chat and search request models
• Low: CLI validates empty project names before entering wizard
• Low: Healthcare YAML blood type enums properly quoted

New Features

• --dry-run CLI flag — preview what would be generated without creating files
• --verbose CLI flag — enable debug logging during generation
• /health endpoint in generated projects — returns Neo4j connectivity status and app version
• CORS origins configurable via CORS_ORIGINS environment variable
• constants.py module in generated projects — centralizes magic strings (index names, graph projections, embedding dimensions)
• Document browser pagination (page size 20, prev/next controls)
• Semantic HTML landmarks (<main>, <section>, <aside>) and ARIA labels in frontend
• Actionable error messages in chat interface — distinguishes backend errors, network failures, and Neo4j unavailability

Security

• Query timeouts (30s default) on all Neo4j operations
• Credential warnings in generated .env.example
• CORS production configuration guidance

Testing

• 365 passing tests (51 new), up from 314
• New: enum identifier sanitization edge cases
• New: models.py compilation across all 22 domains (prevents enum regression)
• New: v0.4.0 feature validation (health endpoint, constants, graceful degradation, input validation, CORS, pagination)
• New: CLI validation and flag tests

v0.3.0 — Memory Integration, Multi-Turn & Graph Visualization (2026-03-23)

• neo4j-agent-memory integration for multi-turn conversations
• Interactive NVL graph visualization (schema view, double-click expand, drag/zoom, property panel)
• LLM-generated demo data (80-90 entities, 25+ documents, 3-5 decision traces per domain)
• Markdown rendering in chat with tool call visualization
• Document browser and entity detail panel
• Improved graph visualization and frontend styling
• Docusaurus documentation site setup and deployment
• Improved domain fixture data quality
• 314 passing tests

v0.2.0 — Connectors & Custom Domains (2026-03-22)

New Features

• 7 SaaS data connectors — GitHub, Notion, Jira, Slack, Gmail, Google Calendar, Salesforce
• Each connector implements BaseConnector ABC with authenticate(), fetch(), and get_credential_prompts()
• Gmail/Google Calendar prefer gws CLI with Python OAuth2 fallback
• Custom domain generation — generate complete domain ontology YAMLs from natural language descriptions using LLM (Anthropic/OpenAI)
• Custom domains saved to ~/.create-context-graph/custom-domains/ for reuse
• Neo4j Aura .env import and neo4j-local support in wizard
• Documentation site (Docusaurus) with deployment configuration

v0.1.1 — Initial Bug Fixes (2026-03-22)

• Bug fixes for CLI and template rendering
• Test improvements and expanded coverage
• Added .gitignore to generated projects

v0.1.0 — Initial Release (2026-03-22)

• Interactive CLI scaffolding tool (create-context-graph) invoked via uvx or npx
• 7-step interactive wizard with Questionary prompts
• 8 agent frameworks: PydanticAI, Claude Agent SDK, OpenAI Agents SDK, LangGraph, CrewAI, Strands, Google ADK, Anthropic Tools
• Domain ontology system with YAML definitions and two-layer inheritance (_base.yaml)
• Jinja2 template engine generating full-stack projects (FastAPI backend, Next.js + Chakra UI v3 frontend)
• Neo4j schema generation (constraints + GDS projections)
• Static and LLM-powered synthetic data generation pipeline
• Neo4j data ingestion via neo4j-agent-memory or direct driver fallback
• Domain-specific agent tools with Cypher queries
• NVL graph visualization component