v0.1.10 — Shared workspace indexes, update priority and coverage

Overlapping workspace roots previously built separate copies of the same code data, and a first-time build could delay updates to an existing index. Version 0.1.10 shares immutable source text and embedding vectors, gives existing-workspace updates priority, and makes file coverage inspectable.

• Reuse indexed data across parent and child workspaces in either order. Each workspace retains its own file manifest, structural graph, and relative search paths; searches score only that workspace's units.
• Coordinate shared-cache writers across processes to avoid duplicate embedding requests. Workspace manifests and temporary build pins protect data during reclamation, including while a first-time build yields to updates.
• Prioritize updates to published indexes at both the global build limit and shared-cache writer lock. First-time builds finish their in-flight embedding requests before yielding, then resume from cached vectors.
• Add optional path, limit, and cursor arguments to index_status. Inspect a file or directory for indexed, pending, excluded, deleted, or missing entries, with exclusion reasons and bounded pagination. Coverage checks perform no model requests and do not wake idle indexing.
• Persist OCE_SHARED_INDEX_DIRECTORY for custom cache locations and tolerate Windows file-probe cleanup contention without leaking build capacity.

Upgrade:

npm install -g open-context-engine@0.1.10

Restart connected MCP clients after upgrading. Existing indexes remain on disk. The first synchronization builds canonical-path vectors once; subsequent overlapping workspaces reuse them. Shared pools are isolated by embedding provider, model, dimensions, and revision. Default data is stored in ~/.cache/opencontextengine/shared-index; custom state locations use .shared-index under their state parent unless OCE_SHARED_INDEX_DIRECTORY is configured.

Example index_status arguments:

{"directory_path":"/absolute/path/to/repository","path":"src/session.py"}

Use "path":"." to inspect the workspace and nextCursor for another page. Usage and configuration.

Validation: a clean checkout passed 76 Node tests and 110 Python tests on macOS (three platform-specific skips). Linux, macOS, and Windows CI passed for the release commit. The exact archive passed isolated installation, both CLI aliases, fresh Python setup and doctor, and real-model MCP checks for parent/child reuse, saved edits, coverage, pagination, environment-file exclusion, and restart restoration.

The verified npm archive and its SHA-256 checksum are attached.

v0.1.9 — Reduce idle resource use and reuse index shards

Idle MCP sessions could repeatedly scan large repositories, retain index workers, and keep cancelled retrieval work running. Version 0.1.9 bounds background resource use and makes saved-file updates reuse existing source and vector data.

• Pause background scans after 60 seconds without search activity; unload idle generations and release workspace leases after 120 seconds. Workers exit after their lease grace period when no requests or admitted builds remain.
• Reuse unchanged source text, parsed units, lexical features, and immutable vector shards. Existing embedding caches migrate without new model requests for cached vectors.
• Propagate cancellation and deadlines through retrieval queues, model requests, and parser subprocesses.
• Limit resident workspaces, concurrent builds, source/index size, and retained vector storage. Back off on repeated failures and reclaim abandoned generations.
• Use metadata-validated source hashes on macOS/Linux with periodic full reconciliation. Set OCE_STRICT_FRESHNESS=1 to reread content for every freshness check; Windows keeps strict checks by default.

Validation: local Node and Python suites, Linux/macOS/Windows platform CI, isolated archive installation and setup/doctor, and real-model MCP search before and after a saved-file edit. A 48,870-file CAD cache migrated 163,886 units without new embedding requests and restored successfully after unloading.

Upgrade with npm install -g open-context-engine@0.1.9, then restart connected MCP clients. Resource controls are documented in Quickstart. The release includes the verified npm archive and its SHA-256 checksum.

v0.1.8 — Configurable embedding concurrency

Indexing can now overlap embedding requests to remote model services with OCE_EMBEDDING_CONCURRENCY. The default remains 1; supported values are 1–8 per workspace. Start with 2 when the endpoint supports concurrent requests.

OCE_EMBEDDING_BATCH_SIZE=64
OCE_EMBEDDING_CONCURRENCY=2

The client bounds in-flight batches, matches out-of-order responses to their source documents, and commits successful batches to the vector cache. If a request fails, it stops submitting new batches and preserves successful in-flight results for retry. Changing concurrency or batch size does not invalidate cached vectors. Server-side GPU inference batch sizes are unchanged.

Save the setting in the shared user configuration (or through setup with environment variables), then restart the MCP client after active indexing finishes.

Validation: 70 Node tests passed on macOS (2 Windows-only tests skipped), 121 Python tests passed (1 platform-specific test skipped), and an isolated install passed setup, doctor, and a real-model MCP indexing/search smoke test with concurrency 2 and 2560-dimensional embeddings. Release artifacts include the npm archive and its SHA-256 checksum.

v0.1.7 — Fix Windows MCP console popups

Fix Windows console windows appearing when an agent calls OpenContextEngine MCP.

• Suppress console creation for runtime probes and Git, Node, Go, and Windows ACL helper processes.
• Bypass the Windows Python venv redirector while preserving the selected virtual environment and its installed packages.
• Preserve shared workers across MCP client exits.

Upgrade:

npm install -g open-context-engine@0.1.7

Restart all MCP clients after upgrading so existing workers exit and the new runtime is loaded.

Validation: Windows, macOS, and Linux platform CI passed, including native Windows console checks and virtual environment isolation. The release archive passed isolated installation, setup, doctor, and live MCP retrieval with saved-file updates and project switching.

Platform CI · Related issue #2

v0.1.6 — Indexing progress and automatic retry

Pending index synchronization now returns a normal updating status in MCP instead of appearing as a search failure. The response makes clear that search has not run yet and includes the available indexing progress.

• Report parsing, embedding, and finalization stages, with completed/total documents requiring new vectors during embedding.
• Retry a pending MCP search once, waiting up to 10 additional seconds (40 seconds of synchronization waiting by default). Set freshnessWaitMs: 0 to disable waiting and retry.
• Keep actual indexing failures and source changes during retrieval as explicit errors. Stale source is never substituted.

Update

npm install -g open-context-engine@0.1.6

Reconnect your MCP client after updating to load the new behavior. If indexing is still pending, check index_status for the same project and retry search_code once ready.

Validation

• All 68 Node tests and 61 Python tests passed locally, with no skipped tests.
• Linux, macOS, and Windows platform checks passed for release commit 014b3eaa5613f720dddf7aaa00dc22146ed03042.
• The exact archive passed isolated installation, both CLI aliases, fresh Python setup, doctor, and real-model MCP retrieval, including project switching, saved edits, file deletion, and restart recovery.
• npm reports latest=0.1.6; published integrity and a clean registry installation match the verified archive.

The exact npm archive and its SHA-256 checksum are attached.

v0.1.5 — 8k default retrieval budget

Searches that omit budget now return up to 8,000 tokens, previously 4,000. The default is consistent across the MCP tool, HTTP service, JavaScript client, and live retrieval engine, providing more room for related implementation evidence.

Pass budget: 4000 or another supported value when a smaller response is sufficient. The maximum remains 8,000 tokens. Existing model settings and indexes remain compatible.

Update

npm install -g open-context-engine@0.1.5

Reconnect your MCP client after updating so it loads the new default.

Validation

• 62 Node tests and 113 Python tests passed locally.
• Linux, macOS, and Windows platform checks passed for the release commit.
• The exact release archive passed isolated installation, both CLI aliases, setup, doctor, and MCP retrieval using configured remote models, including project switching, saved edits, file deletion, and reconnection.
• Explicit 4,000-token requests remain supported; omitted budgets were verified through MCP, HTTP, and the client.

The npm archive and its SHA-256 checksum are attached. Historical benchmark results retain their original evaluation budgets.

v0.1.4 — Shared workers across MCP sessions

OpenContextEngine 0.1.4 fixes index-lock conflicts when multiple MCP sessions access the same repository. Compatible MCP instances now reuse one worker for the same index instead of starting competing writers.

Changes

• Share repository workers across MCP instances, with authenticated discovery and compatibility checks.
• Keep a shared worker available while other clients are still using it, and shut it down after it becomes idle.
• Report index-lock conflicts clearly instead of only reporting worker exited (1).
• Preserve the lock-conflict diagnosis when a later startup retry times out on Windows.

Related issue: #1.

Upgrade

npm install -g open-context-engine@0.1.4

After upgrading, restart the MCP clients that use OCE so they load the new version and the old workers exit.

npm package

Validation

• Windows, macOS, and Linux CI all passed for the release commit: platform tests.
• All 61 Node tests passed in a clean checkout.
• Fresh npm installation, both CLI aliases, doctor, and real-model retrieval with two MCP clients passed, including saved-file updates and continued use after one client disconnects.