MrSphay/odysseus
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix native tool-calling follow-up round on Gemini and Ollama (#867)

Browse Source 
The agent's multi-round (tool-result) follow-up request was rejected with
HTTP 400 on two providers, so tools ran but the agent never produced an answer:

- OpenAI-compatible streaming (Gemini 3) dropped the per-call thought_signature
  and collided parallel tool calls, which arrive with index=None: they all
  landed in slot 0, overwriting the first call's name and corrupting its
  arguments by concatenation, so the follow-up request 400'd. Capture and replay
  each call's extra_content (thought_signature), and give every parallel call
  its own accumulator slot (allocated above the max key, so sparse or mixed
  indices can't collide).
- Native Ollama /api/chat expects object tool-call arguments, but Odysseus
  carries them as a JSON string, which Ollama rejected ("Value looks like
  object, but can't find closing '}' symbol"). Convert them to objects in the
  Ollama payload builder.

Both compose with the no-prose null-content sanitize fix from #862.

Tested: python -m pytest tests/test_llm_core_streaming.py
tests/test_llm_core_ollama.py tests/test_agent_loop.py (53 pass), and
python -m py_compile src/llm_core.py src/agent_loop.py.
This commit is contained in:
James Arslan
2026-06-02 04:39:40 +02:00
committed by GitHub
parent 54ac4a74fb
commit a327df6936
5 changed files with 334 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
src/agent_loop.py
View File

@@ -1123,6 +1123,11 @@ def _append_tool_results(
"name": tc.get("name", ""),
"arguments": tc.get("arguments", "{}"),
},
# Gemini 3 requires the opaque thought_signature it returned with
# each function call to be echoed back on the follow-up turn, or
# the next request 400s. Replay it when present; other providers
# never emit it (their payload builders just ignore the field).
**({"extra_content": tc["extra_content"]} if tc.get("extra_content") else {}),
}
for j, tc in enumerate(native_tool_calls)
]

73
src/llm_core.py
View File

@@ -194,6 +194,43 @@ def _normalize_ollama_url(url: str) -> str:
return base.rstrip("/") + "/chat"
def _ollama_normalize_tool_messages(messages: List[Dict]) -> List[Dict]:
"""Adapt Odysseus' canonical OpenAI-style messages to native Ollama /api/chat.
Odysseus carries assistant tool calls in the OpenAI shape, where
`function.arguments` is a JSON *string*. Native Ollama expects it to be a
JSON *object*; given the string it fails the whole request with HTTP 400
"Value looks like object, but can't find closing '}' symbol", which aborts
every follow-up (tool-result) round. Parse the arguments back into an object
here, on a shallow copy, leaving non-tool messages untouched. The opaque
Gemini `extra_content` (thought_signature) is dropped — it is meaningless to
Ollama and only matters when the conversation is replayed to Gemini.
"""
out: List[Dict] = []
for m in messages or []:
tcs = m.get("tool_calls") if isinstance(m, dict) else None
if not tcs:
out.append(m)
continue
new_calls = []
for tc in tcs:
fn = tc.get("function") or {}
args = fn.get("arguments")
if isinstance(args, str):
try:
args = json.loads(args) if args.strip() else {}
except (json.JSONDecodeError, TypeError):
args = {}
call: Dict = {"function": {"name": fn.get("name", ""), "arguments": args or {}}}
if tc.get("id"):
call["id"] = tc["id"]
new_calls.append(call)
nm = dict(m)
nm["tool_calls"] = new_calls
out.append(nm)
return out
def _build_ollama_payload(
model: str,
messages: List[Dict],
@@ -204,7 +241,7 @@ def _build_ollama_payload(
) -> Dict:
payload: Dict = {
"model": model,
"messages": messages,
"messages": _ollama_normalize_tool_messages(messages),
"stream": stream,
}
options: Dict = {}
@@ -1040,6 +1077,7 @@ async def stream_llm(url: str, model: str, messages: List[Dict], temperature: fl
# ── OpenAI-compatible streaming ──
# Accumulate native tool_calls across streaming chunks
_tc_acc: Dict[int, Dict] = {} # index -> {id, name, arguments}
_tc_last_idx = [-1] # most-recently-touched slot, for providers that omit `index`
# For thinking models: prepend <think> to first content delta so frontend
# can detect thinking-in-progress (some models output </think> but no <think>)
_thinking_model = _supports_thinking(model)
@@ -1105,12 +1143,41 @@ async def stream_llm(url: str, model: str, messages: List[Dict], temperature: fl
yield f'data: {json.dumps({"delta": content})}\n\n'
# Native tool calls — accumulate across chunks
for tc in delta.get("tool_calls") or []:
idx = tc.get("index", 0)
func = tc.get("function") or {}
raw_idx = tc.get("index")
if raw_idx is None:
# Gemini's OpenAI-compat layer omits `index` on
# parallel tool calls (every delta arrives as
# index=None) and sends each call complete in one
# delta. Without this, all parallel calls collide
# into slot 0 — later calls overwrite the first's
# name and CORRUPT its arguments by concatenation,
# so only one malformed call survives and the
# follow-up round 400s. A function name marks the
# start of a new call → allocate a fresh slot;
# an arg-only continuation attaches to the last.
if func.get("name") or _tc_last_idx[0] < 0:
# Next free slot ABOVE any existing key (not
# len()), so a provider mixing integer indices
# with index=None can never collide.
idx = max(_tc_acc, default=-1) + 1
else:
idx = _tc_last_idx[0]
else:
idx = raw_idx
_tc_last_idx[0] = idx
if idx not in _tc_acc:
_tc_acc[idx] = {"id": "", "name": "", "arguments": ""}
if tc.get("id"):
_tc_acc[idx]["id"] = tc["id"]
func = tc.get("function") or {}
# Gemini 3 returns an opaque thought_signature in
# extra_content on the function-call delta. It MUST be
# echoed back on the assistant tool_call next round or the
# follow-up request 400s ("Function call is missing a
# thought_signature"). Preserve it verbatim; other
# providers never send it, so this is a no-op for them.
if tc.get("extra_content"):
_tc_acc[idx]["extra_content"] = tc["extra_content"]
if func.get("name"):
_tc_acc[idx]["name"] = func["name"]
if "arguments" in func:

36
tests/test_agent_loop.py
View File

@@ -301,3 +301,39 @@ class TestAppendToolResultsNativeContent:
assert messages[0]["content"] == "thinking..."
assert messages[1]["role"] == "user"
assert "tool output" in messages[1]["content"]
class TestAppendToolResultsThoughtSignature:
"""Gemini 3 returns an opaque thought_signature (in extra_content) with each
function call and rejects the follow-up turn with HTTP 400 unless it is
echoed back on the assistant tool_call. _append_tool_results must replay it
when present, and omit the field entirely otherwise (other providers never
send it)."""
def test_extra_content_is_replayed_when_present(self):
native = [{
"id": "call_g",
"name": "app_api",
"arguments": '{"action": "get_memory"}',
"extra_content": {"google": {"thought_signature": "EuIDCt8DAQ=="}},
}]
messages = []
_append_tool_results(
messages, "", native, [{}], ["mem"],
used_native=True, round_num=1,
)
tc = messages[0]["tool_calls"][0]
assert tc["extra_content"] == {"google": {"thought_signature": "EuIDCt8DAQ=="}}
# function payload is still well-formed alongside it
assert tc["function"]["name"] == "app_api"
assert tc["id"] == "call_g"
def test_no_extra_content_key_when_absent(self):
native = [{"id": "call_o", "name": "app_api", "arguments": "{}"}]
messages = []
_append_tool_results(
messages, "", native, [{}], ["r"],
used_native=True, round_num=1,
)
# No empty/None extra_content leaks onto non-Gemini tool calls.
assert "extra_content" not in messages[0]["tool_calls"][0]

72
tests/test_llm_core_ollama.py
View File

@@ -41,3 +41,75 @@ def test_llm_call_posts_native_ollama_payload(monkeypatch):
assert seen["headers"]["Authorization"] == "Bearer ollama-key"
assert seen["json"]["stream"] is False
assert seen["json"]["options"] == {"temperature": 0.2, "num_predict": 7}
# ---------------------------------------------------------------------------
# Tool-call argument serialization for native Ollama
#
# Odysseus carries assistant tool calls in the OpenAI shape, where
# `function.arguments` is a JSON *string*. Native Ollama /api/chat expects a
# JSON *object* and rejects the string form with HTTP 400 ("Value looks like
# object, but can't find closing '}' symbol"), aborting every follow-up
# (tool-result) round. _build_ollama_payload must parse it back to an object.
# ---------------------------------------------------------------------------
def _assistant_tool_call_msgs():
"""A canonical OpenAI-style assistant tool call + tool result, as produced by
agent_loop._append_tool_results (arguments are a JSON string)."""
return [
{"role": "user", "content": "what do you know about me?"},
{
"role": "assistant",
"content": None,
"tool_calls": [
{
"id": "call_0",
"type": "function",
"function": {"name": "app_api", "arguments": '{"action": "get_memory"}'},
}
],
},
{"role": "tool", "tool_call_id": "call_0", "content": "Memory: user is James."},
]
def test_ollama_payload_parses_string_arguments_to_object():
payload = llm_core._build_ollama_payload(
"gpt-oss:120b", _assistant_tool_call_msgs(), temperature=0.0, max_tokens=0,
)
asst = payload["messages"][1]
args = asst["tool_calls"][0]["function"]["arguments"]
# The whole point: arguments must be a dict, not the JSON string.
assert args == {"action": "get_memory"}
assert not isinstance(args, str)
assert asst["tool_calls"][0]["function"]["name"] == "app_api"
assert asst["tool_calls"][0]["id"] == "call_0"
def test_ollama_payload_drops_gemini_thought_signature():
"""A cross-provider fallback can hand Ollama a tool call that still carries
Gemini's opaque extra_content; it is meaningless to Ollama and must not leak."""
msgs = _assistant_tool_call_msgs()
msgs[1]["tool_calls"][0]["extra_content"] = {"google": {"thought_signature": "AAAA"}}
payload = llm_core._build_ollama_payload(
"gpt-oss:120b", msgs, temperature=0.0, max_tokens=0,
)
tc = payload["messages"][1]["tool_calls"][0]
assert "extra_content" not in tc
assert tc["function"]["arguments"] == {"action": "get_memory"}
def test_ollama_payload_leaves_plain_messages_untouched():
msgs = [{"role": "user", "content": "hello"}]
payload = llm_core._build_ollama_payload("m", msgs, temperature=0.0, max_tokens=0)
assert payload["messages"][0] == {"role": "user", "content": "hello"}
def test_ollama_payload_tolerates_malformed_arguments():
msgs = [{
"role": "assistant",
"tool_calls": [{"function": {"name": "x", "arguments": "{not json"}}],
}]
payload = llm_core._build_ollama_payload("m", msgs, temperature=0.0, max_tokens=0)
# Falls back to an empty object rather than raising.
assert payload["messages"][0]["tool_calls"][0]["function"]["arguments"] == {}

151
tests/test_llm_core_streaming.py Normal file
View File

@@ -0,0 +1,151 @@
"""Streaming tool-call accumulation tests for the OpenAI-compatible path.
Regression for Gemini's OpenAI-compat layer, which (a) attaches an opaque
thought_signature in `extra_content` on the function-call delta and (b) omits
`index` on PARALLEL tool calls — every parallel delta arrives as index=None.
The accumulator must give each parallel call its own slot (otherwise they
collide into slot 0, overwriting the first call's name and concatenating —
corrupting — its arguments) and must preserve extra_content per call.
"""
import json
import asyncio
from src import llm_core
class _FakeResp:
def __init__(self, lines):
self._lines = lines
self.status_code = 200
async def aiter_lines(self):
for ln in self._lines:
yield ln
async def aread(self):
return b""
class _FakeStreamCtx:
def __init__(self, lines):
self._lines = lines
async def __aenter__(self):
return _FakeResp(self._lines)
async def __aexit__(self, *a):
return False
class _FakeClient:
def __init__(self, lines):
self._lines = lines
def stream(self, method, url, **kw):
return _FakeStreamCtx(self._lines)
def _drive(monkeypatch, lines, model="gemini-3.1-pro-preview-customtools"):
"""Run stream_llm against a canned SSE line list; return parsed events."""
monkeypatch.setattr(llm_core, "_get_http_client", lambda: _FakeClient(lines))
monkeypatch.setattr(llm_core, "_is_host_dead", lambda u: False)
monkeypatch.setattr(llm_core, "note_model_activity", lambda *a, **k: None)
monkeypatch.setattr(llm_core, "_clear_host_dead", lambda *a, **k: None)
async def run():
events = []
async for chunk in llm_core.stream_llm(
"https://generativelanguage.googleapis.com/v1beta/openai/chat/completions",
model,
[{"role": "user", "content": "hi"}],
headers={"Authorization": "Bearer k"},
tools=[{"type": "function", "function": {"name": "x", "parameters": {}}}],
):
for ln in chunk.split("\n"):
ln = ln.strip()
if ln.startswith("data: ") and ln[6:] != "[DONE]":
try:
events.append(json.loads(ln[6:]))
except ValueError:
pass
return events
return asyncio.run(run())
def _sse(delta):
return "data: " + json.dumps({"choices": [{"delta": delta}]})
def test_parallel_calls_with_null_index_do_not_collide(monkeypatch):
# Two parallel calls, each complete in one delta, both with index=None
# (exactly what Gemini's OpenAI-compat layer emits). Only the first carries
# a thought_signature.
lines = [
_sse({"tool_calls": [{
"index": None, "id": "call_a", "type": "function",
"function": {"name": "get_memory", "arguments": "{}"},
"extra_content": {"google": {"thought_signature": "SIG0"}},
}]}),
_sse({"tool_calls": [{
"index": None, "id": "call_b", "type": "function",
"function": {"name": "bash", "arguments": '{"command":"echo hi"}'},
}]}),
"data: [DONE]",
]
events = _drive(monkeypatch, lines)
calls = next(e["calls"] for e in events if e.get("type") == "tool_calls")
assert len(calls) == 2, f"parallel calls collided: {calls}"
by_name = {c["name"]: c for c in calls}
assert set(by_name) == {"get_memory", "bash"}
# arguments are NOT corrupted by concatenation
assert by_name["get_memory"]["arguments"] == "{}"
assert by_name["bash"]["arguments"] == '{"command":"echo hi"}'
# signature preserved on the first call only, exactly as received
assert by_name["get_memory"]["extra_content"] == {"google": {"thought_signature": "SIG0"}}
assert "extra_content" not in by_name["bash"]
def test_single_call_chunked_arguments_still_accumulate(monkeypatch):
# Conformant OpenAI style: index present, arguments streamed in pieces.
lines = [
_sse({"tool_calls": [{"index": 0, "id": "c", "type": "function",
"function": {"name": "search", "arguments": '{"q":"'}}]}),
_sse({"tool_calls": [{"index": 0, "function": {"arguments": 'cats"}'}}]}),
"data: [DONE]",
]
events = _drive(monkeypatch, lines, model="gpt-4o-test")
calls = next(e["calls"] for e in events if e.get("type") == "tool_calls")
assert len(calls) == 1
assert calls[0]["name"] == "search"
assert calls[0]["arguments"] == '{"q":"cats"}'
def test_null_index_chunked_arguments_attach_to_last_call(monkeypatch):
# index=None where the name arrives first, then an arg-only continuation:
# the continuation must attach to the just-started call, not open a new one.
lines = [
_sse({"tool_calls": [{"index": None, "id": "c", "type": "function",
"function": {"name": "search", "arguments": '{"q":'}}]}),
_sse({"tool_calls": [{"index": None, "function": {"arguments": '"dogs"}'}}]}),
"data: [DONE]",
]
events = _drive(monkeypatch, lines)
calls = next(e["calls"] for e in events if e.get("type") == "tool_calls")
assert len(calls) == 1, f"continuation opened a spurious call: {calls}"
assert calls[0]["arguments"] == '{"q":"dogs"}'
def test_sparse_integer_indices_then_null_do_not_collide(monkeypatch):
# Hardening: a provider that uses sparse integer indices (0 and 2) and then
# a null-index call must allocate ABOVE the max key, not at len()==2 (which
# would overwrite slot 2). Three distinct calls must survive.
lines = [
_sse({"tool_calls": [{"index": 0, "id": "a", "function": {"name": "f0", "arguments": "{}"}}]}),
_sse({"tool_calls": [{"index": 2, "id": "b", "function": {"name": "f2", "arguments": "{}"}}]}),
_sse({"tool_calls": [{"index": None, "id": "c", "function": {"name": "fn", "arguments": "{}"}}]}),
"data: [DONE]",
]
events = _drive(monkeypatch, lines)
calls = next(e["calls"] for e in events if e.get("type") == "tool_calls")
assert sorted(c["name"] for c in calls) == ["f0", "f2", "fn"], f"collision: {calls}"