Skip to content

v2.12.0

Choose a tag to compare

@github-actions github-actions released this 12 Oct 17:10
3813e84

Changed

  • The gno agents block (v5), the gno skill and the MCP gno_search description narrow the exact-text rung: gno search is for a term, identifier, quote or error you already have verbatim, and a question, or a name or topic taken from it, goes to gno query --fast. In headless agent sessions given only the v4 block, one agent often ran gno search on a word or name lifted from a natural-language question and never reached semantic search; with the v5 wording its first retrieval call fit the question in 14 of 16 runs instead of 6 of 16, with no change for exact-phrase questions. Run gno agents update and gno skill install --force to pick it up.
  • The Web UI's Fast depth runs the same hybrid fast search as gno query --fast (keyword and semantic, no expansion or reranking) instead of keyword search only; it falls back to keyword search when the server reports vector search unavailable.
  • MCP gno_context returns the gno-context-agent-v2 text to the model: each evidence passage also carries its content date and, for a superseded memory fact, the URI of the fact that replaced it, so agents no longer cite an outdated fact as current. Promotion receipts recorded with v1 still validate.

Fixed

  • gno search answers long natural-language questions in about 0.2-0.4 s instead of 2-11 s on a 2,300-document index: highlighted snippets are built only for the returned results instead of for every match. Results and scores are unchanged. Snippets and their line ranges now come from the part of the document that matches the query instead of always its first chunk.
  • Memory facts carry their record date (memory.createdAt) as their content date in Context Capsule evidence, ask sources, recency sorting and date filters. The next gno update reprocesses existing documents once to pick this up; embeddings are reused.
  • Opening an index with a config that does not list all of its collections no longer deletes them. Every command, the SDK, gno serve, gno daemon and gno mcp synced the config into the index on open and deleted each collection the config omitted together with its documents, so gno --config other.yml search ..., or an SDK client given a test config but no dbPath, silently emptied the normal index. Such an open now fails with an error that names the collections and how to resolve it (add them back, use a separate index, or run gno cleanup), and changes nothing; a running gno serve, gno daemon, gno mcp or SDK client refuses requests the same way (REST 409 STALE_COLLECTIONS) if another client indexes a collection its config omits. gno status lists them under Collections not in this config (JSON staleCollections). Deleting a collection's documents now takes an explicit removal: gno collection remove / rename (and their MCP, REST and SDK forms) or gno cleanup, which now names each collection on stderr before deleting it and accepts --json. A collection removed from the config by hand, also while gno serve or gno daemon runs, keeps its documents and is reported until you run gno cleanup or add it back.
  • gno status, gno doctor, MCP gno_status, SDK status(), GET /api/status and the Web UI now name a collection whose folder is missing or unreadable (an unmounted drive, an offline sync folder) with the error, instead of only counting a recent error and reporting the index healthy. The collection gets rootError ({ code, message }), health turns degraded and gno doctor adds a collection-roots warning; its documents stay indexed, and the report clears as soon as the folder is back, without waiting for a sync.
  • When the local model worker cannot start, its error now says why instead of only exited.
  • gno agents update only refreshes files that already contain the GNO block. A file without one is reported absent with a hint to run gno agents install and is left untouched; update no longer appends a block or creates files such as ~/.hermes/SOUL.md.
  • gno skill install, uninstall and paths at user scope use CLAUDE_CONFIG_DIR for Claude and CODEX_HOME for Codex when set, as gno agents already did, so Claude Code and Codex profile directories receive the skill. A blank value counts as unset and surrounding whitespace is ignored (also for gno agents).

Development

  • bun run eval:memory runs again (it hung since 2.8.3): indexing and the model worker now find their own files when loaded outside Bun's module loader, and the pinned eval runner is patched so an eval that throws exits non-zero instead of hanging. The ask eval's sources carry the hashes grounded answers require, so it measures generated answers again, and the agentic eval's write lane checks embeddings through the index's own status.

Security

  • Updated @modelcontextprotocol/client and @modelcontextprotocol/server to 2.3.1 (the 2.0 client's OAuth flow could send credentials to an authorization server chosen by the MCP server), sharp to 0.35.5 (librsvg CVE-2026-96889) and the development dependency vitest to 4.1.11 (path traversal through the mocker redirect).