fix(server): gemma4 bare call: tool extract for OpenAI chat

[bakon11]

Summary

Gemma-4 free-form tool use (no structural tags / specials stripped at detokenize) often emits bare call:NAME{ARGS} instead of wrapped <|tool_call>… bodies. The engine-backed gemma4 parser needs special-token IDs; OpenAI /v1/chat/completions then returned plain content and clients (e.g. Hermes) never executed tools.

Changes

• Text-seam Gemma4ToolParser: extract bare call:NAME{ARGS} (stream + non-stream), still supports wrapped blocks
• OpenAI chat: for --tool-call-parser gemma4, prefer text-seam MakeToolParser over engine path so detokenized/free-form bodies become real tool_calls
• USAGE: document the gemma4 dual surface

Test plan

• Lab: non-stream + stream tools=[{terminal}] → finish_reason=tool_calls, args {"command":"date"}
• CI gates
• Regression: wrapped <|tool_call>call:… still extracts

Notes

Small serve/parser fix; independent of ROCm #317. Does not change structural-tag defaults.

FOLLOWING_AGENTS_PROTOCOL

[localai-org-maint-bot]

@bakon11 Please add automated stream and non-stream regression coverage before this parser/dispatch change merges. In particular, exercise wrapped calls, bare calls, ordinary prose containing call:, nested braces/string delimiters, incomplete calls, and a marker split across streaming chunks. The last case matters because the new no-marker branch immediately emits all current text; a partial ca/call: prefix can otherwise leak as content before the next chunk turns it into a tool call. The PR test plan currently leaves wrapped regression unchecked and the diff changes parsing behavior without a checked-in test.

[bakon11]

@localai-org-maint-bot Addressed CHANGES_REQUESTED:

1. Bugfix — streaming no-marker branch no longer emits trailing partial <|tool_call> / bare call: prefixes (the CI failure gemma4 stream: start tag split across chunks buffers cleanly). Incomplete bare calls without { are buffered, not leaked as content.
2. Regressions in tests/.../test_gemma4.cpp:
   • non-stream: wrapped, bare complete, prose prefix, callback: prose (not a tool), nested braces + string delims, incomplete bare
   • stream: bare complete, call: split across chunks, prose-then-bare, incomplete call buffer, start-tag split (extra coverage)

Local: test_tool_parser_gemma4 69/69 (was 58 + new cases; previously 1 fail on tip).

[localai-bot]

Landed on main (as one commit authored to you — the branch's two commits had to be squashed because check-doc-checkpoint is a per-commit obligation and the second moved the parser without its docs/USAGE.md line; Don Mirror's marker-buffering work is credited in the commit body). The bare-call: scanner itself is a good catch and went in as written.

One repair applied on landing — worth knowing about because it is silent and easy to reintroduce:

StripTrailingToolMarkerPrefixes was widened from <|tool_call> prefixes to call: prefixes unconditionally. Since kCall is "call:", that buffers a trailing c, ca, cal or call out of streamed content. This seam has no end-of-stream flush — extract_tool_calls_streaming emits content.substr(streamed_content_len_) and nothing ever calls back to release the tail — so anything held when the stream ends is simply lost.

Concretely: a reply ending in "topic", "music", "specific" or "logic" loses its last letter. It was <|tool_call> before, which starts with < and effectively never ends a message, so widening it changed the risk class.

The fix reuses your own rule. ScanBareCalls already refuses to treat call: as a marker when the preceding character is alphanumeric, _ or /; the buffer now applies the same test, so a partial inside a word streams through while a genuine split marker after whitespace is still buffered:

const std::size_t at = content_out.size() - k;
if (at > 0) {
  const char prev = content_out[at - 1];
  if (std::isalnum(static_cast<unsigned char>(prev)) || prev == '_' || prev == '/') {
    break;  // part of a word, not the start of a bare call
  }
}
content_out.erase(at);

Three regressions added, test_tool_parser_gemma4 72/72. They are mutation-checked: removing the boundary guard fails two of them on the swallowed trailing letter.

Left open, recorded rather than fixed: bare-call detection is inherently heuristic — prose of the form ... call:name{...} becomes a tool call, and the args text is consumed out of content. That is the premise of the feature, not a defect in this PR, and it is only reachable with --tool-call-parser gemma4 and tools enabled. If you want it tightened later, matching the parsed name against request.tools would remove nearly all false positives.

Reviewed and landed with Claude Code.

[localai-bot]

Landed on main as bc6bd88c456d, authored to you.

Closing manually rather than by merge because the branch's commits were squashed into one, so the head SHA here never becomes reachable from main and GitHub cannot auto-close it. The squash was forced by check-doc-checkpoint, which runs per non-merge commit (git rev-list --no-merges): this branch split code from the public document that had to change with it, so at least one commit owed a docs/ line it did not carry, and the range gate went red. Squashing with --author keeps your attribution on the landing commit.

The commit message records what changed on landing and why. Nothing was dropped.