Skip to content

v0.6.0

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 01 Aug 15:11
· 1 commit to main since this release
Immutable release. Only release title and notes can be modified.
v0.6.0
acd558c

Security

  • Upgraded modelcontextprotocol/go-sdk from v1.4.0 to v1.7.0, which clears
    two HIGH advisories affecting v1.4.0: GHSA-89xv-2j6f-qhc8 (CVE-2026-33252,
    cross-site tool execution against Streamable HTTP servers without inbound
    auth) and GHSA-q382-vc8q-7jhj (null-Unicode JSON parsing).

  • HTTP transport now verifies request origin explicitly. The SDK applied
    cross-origin protection by default in v1.4.1–v1.5.0 but made it opt-in again
    in v1.6.0, so the handler now passes its own http.CrossOriginProtection.
    Browser cross-origin POST requests to /mcp are rejected with 403.
    Non-browser clients (no Origin/Sec-Fetch-Site header) are unaffected.

  • Cursors are always signed in HTTP mode. When OPENGROK_MCP_CURSOR_SECRET
    is unset and the transport is HTTP, the server generates a random 32-byte
    process-local HMAC key at startup instead of leaving cursors unsigned. Over
    HTTP a cursor is an attacker-reachable blob, so unsigned was the wrong
    default. Cursors signed with a generated key are invalidated on restart and
    are not valid across replicas — set OPENGROK_MCP_CURSOR_SECRET explicitly
    for either. Stdio behavior is unchanged (unsigned, with the existing startup
    warning). Implements ROADMAP item 001, Option A.

Fixed

  • Search paging no longer drops results. OpenGrok pages by file while
    page_size caps returned lines, so line matches beyond the cap inside an
    already-fetched file were discarded and unreachable by any cursor — a
    page_size of 2 over a query with 85 line matches in the first two files
    surfaced 2 and lost 83. Cursors now carry a line offset that resumes inside
    the file window, so a full walk yields every matching line exactly once
    (verified live: 910 results, 910 unique, 0 duplicates, matching the corpus).
    Previously minted cursors decode with a zero line offset and stay valid.
    total_pages counts file pages and is therefore a lower bound on a full
    walk — stop on a null next_cursor, not on total_pages.
  • Identical searches now return identical results. The OpenGrok response was
    decoded into a map and ranged over directly, so Go's randomized map iteration
    reordered results on every call; combined with page truncation, the same query
    returned different hits run to run. The decoder now records and follows
    OpenGrok's emission order, preserving relevance ranking.
  • PAGE_SIZE_TRUNCATED no longer reports data loss. Same code, same
    trigger, but it now states the surplus lines are reachable via next_cursor
    instead of implying they were dropped.
  • expand_context is no longer silently ignored. Expansion was skipped
    whenever response_mode resolved to compact, which is the default under the
    economy profile — so an explicit expand_context: true had no effect on a
    stock server. An explicitly passed expand_context now outranks the compact
    default. Omitting it still keeps economy responses lean.
  • Search modes the instance cannot serve report a capability gap, not a parse
    error.     OpenGrok answers 400 both for an unparseable query and for a mode it
    does not serve. When startup probing already found the mode unavailable, the
    new SEARCH_MODE_UNSUPPORTED error says so instead of offering query-syntax
    advice that cannot help. Genuine parse failures still return
    QUERY_PARSER_FAILED.
  • file_type documents that it takes an analyzer name, and warns when it
    matches nothing.     file_type maps to OpenGrok's type parameter, which wants
    the analyzer name (golang), not the extension (go). A wrong value is not an
    upstream error — it silently returns zero hits. The schema now says so, and a
    new FILE_TYPE_NO_MATCH warning fires when a file_type filter yields no hits.
  • Snippets no longer leak OpenGrok's <html> marker. For struct-field
    definition hits OpenGrok returns a synthesized symbol+scope string prefixed
    with <html> rather than the source line. The marker is stripped; <b> match
    highlighting is preserved.
  • Tool errors now carry their remediation in the text content. The
    suggestion field (which valid fields to use, what to change) was only
    present in structuredContent. Many MCP clients never forward that to the
    model, so an agent hitting e.g. UNKNOWN_FIELD saw only Unknown field "max_results" for operation "code". with no indication that page_size was
    the field it wanted. suggestion is now appended to the text content.
    structuredContent is unchanged.

Added

  • results_on_page on every paginated response. total_hits is OpenGrok's
    count of matching files, but code-search results are one per matching
    line, so len(results) could exceed total_hits with no way to tell from
    the response. results_on_page always equals len(results), making the two
    units reconcilable. OpenGrok reports no global line count (resultCount is
    constant regardless of how much is fetched), so there is deliberately no
    global counterpart. total_hits semantics are unchanged — this is additive.
  • Compact schemas mark fields that only some operations accept. The
    top-level property bag is the union of every operation's fields (so strict
    clients stop stripping them), while each oneOf branch rejects fields it does
    not own. Six such fields — max_results, lines_before, lines_after,
    expand_context, sort, max_hits_per_file — now carry an
    (operation=... only) note, so agents stop discovering ownership by rejection.
    Costs ~1,800 bytes on tools/list.

Changed

  • page_size documents that paging is file-based while it caps lines.
    OpenGrok pages by file; page_size caps returned line results. Line matches
    beyond page_size inside an already-fetched file are dropped and cannot be
    reached by paging — raise page_size instead. Behavior is unchanged; this
    documents a pre-existing defect now tracked in docs/limitations.md.
  • tools/list now emits "idempotentHint": false on every tool. The SDK
    dropped omitempty from ToolAnnotations.IdempotentHint in v1.7.0. Read-only
    annotations are otherwise unchanged.
  • List results carry SDK-emitted ttlMs and cacheScope fields (SEP-2549
    cache hints, ttlMs: 0 / cacheScope: "public"). These are protocol-level
    fields written by the SDK, not part of this server's tool contract.
Assets 16
Loading