Ollama's Silent Tool-Call Failures: Conversation-History Round-Trip, Dropped Calls and Retyped Arguments

An agent loop calls a local Ollama server, the HTTP status is 200, the stream ends cleanly, and the assistant message contains nothing: no content, no tool_calls, done_reason: "stop". The loop treats that as “the model is done” and stops. Nothing was done. No error was raised.

That is the failure this article is about. It is not one bug with one cause. Four separate, currently open reports against Ollama between versions 0.34.1 and 0.40.8 describe four different ways a tool call disappears or comes back mutated, and the source of tag v0.40.12 shows which of those paths are still present in the newest release at the time of writing.

The evidence boundary matters, so state it up front: four public issue reports written by other people, plus reads of the tagged source. No Ollama server was started for this article, so nothing here is a benchmark, and nothing here is our own reproduction. Where a number is quoted, it is the reporter’s own count on their own harness, at the version they named.

The signature to recognise before anything else

A dropped tool call and a genuinely finished turn look identical in most clients, because both are done_reason: "stop" with an assistant message. The distinguishing field is eval_count. A dropped call still generated tokens: the model wrote a call, the server failed to return it as tool_calls, and the tokens are counted. The reporter on qwen3:8b saw eval_count between 20 and 160 on replies whose content and tool_calls were both empty.

So the operational rule is narrow and testable: an assistant turn with empty content, no tool_calls, done_reason of stop and a non-zero eval_count is a server-side loss, not a completed task. A turn with eval_count: 0 may genuinely be empty output. Most agent frameworks collapse both into “finished” and move on.

The rest of this article is the four paths that produce the first case, then the guard.

The four paths in one table

Two of the four are parser or renderer defects. One is a type loss in the API decode that only becomes visible through a renderer. One is a model- and build-specific suppression that has not been shown to exist on newer versions. Treating them as a single “Ollama drops tool calls” defect would be wrong, and the scope of each is the reason the fixes are tracked as separate pull requests.

Path 1: history replay retypes an integer

The first report (issue #19191) is precise enough to read as a specification. With flint:40b-a2b-qat-int4 on Ollama 0.40.8 (f580b54), a request whose history contains an earlier tool call carrying timeout_ms: 30000 returns the next call with {"command":"pwd","timeout_ms":"30000.0"}. Without that earlier call in history, the same model returns proper integers on nine of nine runs; with it, the wrong form appears on three of three.

Two layers produce that, and separating them matters if you are deciding what to patch.

The decode layer loses the integer type before any model sees the prompt. In api/types.go at tag v0.40.12, ToolCallFunctionArguments.UnmarshalJSON rebuilds the arguments with json.Unmarshal into an ordered map, which decodes every JSON number as float64. Message.UnmarshalJSON wraps its alias type, so this path runs for every history message that carries tool calls. This is type loss, not value loss: float64 holds integers exactly far beyond any plausible timeout value, but the fact that the value was an integer is gone by the time a renderer formats it.

The render layer decides whether the type loss becomes visible text. In model/renderers/flint_json.go at the same tag, flintJSONValue handles floats itself: it formats with strconv.FormatFloat(..., 'g', -1, ...), appends .0 whenever the result contains no ., e or E, and returns the string as a json.Number, so a float64 of 30000 is turned into the literal 30000.0 before the shared writer sees it. flint.go sends every non-string tool-call argument through that conversion on the way into <arg_value>, which is how 30000 becomes <arg_value>30000.0</arg_value> in the Flint prompt. Flint then copies the form it was shown, and returns it as a string.

That .0 is this renderer’s, not a rule of Ollama’s JSON handling in general. pythonScalar in model/renderers/json.go at the same tag re-reads the marshalled value with UseNumber and returns a number literal that contains no ., e or E unchanged; only a literal that does contain one is parsed to float64, formatted with 'f', and given a trailing .0 when that result has no dot. A raw float64 of 30000 marshals to 30000 and comes out as 30000 there.

The downstream damage is durable rather than transient. Once a call carrying the string form is appended to history, every later call that sets the same argument is rendered from the same mutated history and fails the same validation, so an agent that survived turn two dies on turn three and stays dead for the session. The reporter’s client rejected it with invalid type: string "30000.0", expected u64, and their shell tool could not run any command for the rest of the session.

Both halves are addressed in open pull requests, neither merged into v0.40.12: api: preserve integer types in decoded tool call arguments (PR #19195) switches the decode to preserve number representation, and Fix tool calls rendering integer arguments as decimals (PR #19252) stops the renderer from rewriting the numeric form. If your client validates argument types, check your server version against both before assuming a fix is present.

Path 2: the call names a tool that is not in tools

The second path is the one that most often hides inside a working agent, because the trigger is a design pattern people adopt deliberately: a tool offered on demand, named inside another tool’s description, rather than listed in the request’s tools array.

The report is issue #19187. On Ollama 0.40.7 with qwen3:8b, /api/chat streaming, num_ctx 32768, a call to such a name returned no content and no tool_calls, with eval_count between 20 and 160. Replaying sixteen captured requests, the reporter got that empty reply on 28 of 32 runs. Generating raw text for the same requests with the model’s own template produced a well-formed <tool_call> to the unlisted name every time, and adding the missing names to tools with empty schemas restored the call on 30 of 32. The model was calling; the server was dropping the call on the way out.

An independent reporter reproduced it on 0.40.8 (Linux, CPU) with qwen3:0.6b, and their counts are cleaner: with the target tool absent from tools, eight of eight seeds returned no content and no tool_calls with eval_count 25; with it listed, eight of eight returned the call. They saw the same empty reply through /v1/chat/completions, with finish_reason stop, so switching endpoints does not route around it.

The mechanism proposed in PR #19240 is that the parser buffers a tagged call for a name it does not recognise, and the content accessor returns an empty string for non-JSON tags, so the API emits a successful response with neither content nor tool calls. That is the contributor’s account, not a path confirmed by reading v0.40.12, so treat it as the best available explanation rather than a verified one. The PR is open.

The practical consequence: if your agent exposes tools dynamically by mentioning them in a description, the request that “should” work will silently produce nothing, and the retry loop will keep producing nothing. Listing the tool, even with an empty parameter schema, is the difference between 28 of 32 empty and 30 of 32 working.

Path 3: the repair closes braces, never brackets

The third path is in gemma4 argument parsing (issue #19273), and it is the clearest case of a one-line log message that looks harmless. The reporter’s build is Ollama 0.40.6 on Windows with CUDA, and they note that model/parsers/gemma4.go has not changed since commit a9d8953 (2026-09-22), so their observation extends to 0.40.11.

A gemma4 call that leaves the edits list unclosed still ends with its closing tag, so the model finished the call. The parser’s repair candidates close a string delimiter, and they add a missing } only when another repair already changed the text. At v0.40.12, in model/parsers/gemma4.go, the object-close repair returns immediately when the argument text already ends in }, and no repair step inserts a ] at all. So a call of the shape call:edit{edits:[{newText:S,oldText:S,path:S} has nothing in the repair path that can rescue it. The parser logs gemma4 tool call parsing failed and returns a reply with no content, no tool_calls, done_reason: "stop", and no error field.

The reporter’s counts come from five days of agent evals with gemma4:e4b through OpenClaw: 30 dropped edit calls in the server log, every one leaving the edits list unclosed, across 16 failed runs in 28 rounds. Their four observed call shapes split on what is missing — path inside the item with both ] and the outer } absent (20 occurrences), ] missing before ,path (5), ] and outer } missing (4), and one case with two items where only the ] was missing. Those are one reporter’s task mix, not a benchmark, but the shape distribution is the useful part: a single missing bracket, not a mangled call.

PR #19274 adds a repair that closes the lists a finished gemma4 call leaves open. It is open. Until it lands, a gemma4 agent should treat the log line gemma4 tool call parsing failed as an incident marker and re-ask, not as noise.

Path 4: the announce-then-call pattern on an MLX build

The fourth path is the narrowest and the one most likely to be a version artifact, so it is stated with its environment attached.

The report is issue #19353. On qwen3.8:27b-mlx with Ollama 0.34.1 (macOS 26.5.2, Apple M4 Pro, 48 GB, qwen3_5 architecture, nvfp4), a system prompt that asks the model to say something in natural language before calling the tool produces prose only: tool_calls is null and done_reason is stop. The same request with a tool-directive prompt returns parsed calls, and a control model, qwen3:4b, returns finish_reason: tool_calls with a non-empty preamble on three of four runs of the identical prompt. Only the system prompt differs between the two cases.

This is the shape of an announce-then-call instruction, which is a common agent pattern, so the report matters even though the version does not overlap the 0.40.x series. What it does not support is any claim about v0.40.12. Nothing here shows the behaviour reproduced on a newer build, and no source shows it fixed. If you run a preamble-style prompt on an MLX build, check this first; if you run it on a current CPU or CUDA build, you have no evidence either way from this report.

The guard: an empty turn is an error, not a completion

Everything above is detectable client-side without patching the server, because the failure is a specific shape on the wire rather than an exception. The whole check is small enough to paste into the response handler:

def lost_tool_call(msg, done_reason, eval_count):
    return (done_reason == "stop"
            and not (msg.get("content") or "").strip()
            and not msg.get("tool_calls")
            and (eval_count or 0) > 0)

When that predicate is true, treat the turn as a transport-level failure for the loop’s purposes: do not append it to history as an assistant completion, do not fall through to a final answer, and either re-ask with the tool registered in tools or surface an error to the caller. Appending the empty message is the specific mistake that turns a recoverable drop into a session-ending stop, because the next request then replays a history in which the model appears to have said nothing.

Three supporting behaviours make the guard usable in practice:

• Log done_reason and eval_count on every turn, not only on errors. Without eval_count you cannot tell a dropped call from a genuinely empty reply, and you will spend the debugging session on the model.
• Validate argument types against the tool schema on the way in, before you append a call to history. The integer-to-string mutation survives forever once replayed, so the cheapest place to catch it is the moment it arrives.
• Before concluding a fix is missing or present, read the server version and compare it with the open pull requests listed above. The v0.40.12 release was published on 2026-10-27T00:07:12Z, and none of the four pull requests referenced here is merged into it.

Where to look first when a turn comes back empty

Order the checks by cost, cheapest first, and stop at the first one that explains the observation:

1. History serialisation. Does the request you sent contain the argument types you intended, or did a previous response already hand you "30000.0"? If the latter, the server-side retype happened one turn earlier.
2. The tools array. Is every function name the model could plausibly call actually listed, including the ones you only mention inside descriptions?
3. The server log. One line, gemma4 tool call parsing failed, names the parser path outright; its absence does not clear the other paths, which log nothing.
4. The model and build. Preamble suppression and MLX render paths are model- and build-specific, and are the last thing to blame, not the first.

For the surrounding mechanics, the Ollama API integration guide covers the /api/chat request shape these probes use, and the gemma4 setup guide covers installation and the errors that are not tool-call related. If a model looks worse rather than silent, why a local LLM feels dumber covers quantisation, context and sampling — a different failure class from state corruption.

What this article does not claim

It does not claim that all Ollama parsers or renderers share one defect: each of the four paths is reported against one model-parser or model-renderer pair, and the code observations are specific files at one tag. It does not claim the paths were reproduced here; no server was run. It does not claim the reports are current for v0.40.12 in either direction — none of them shows the behaviour on that tag, and none of the four pull requests is merged. The counts quoted, including 28 of 32, 30 of 32, eight of eight, three of three, 30 dropped calls and 16 of 28 runs, are the reporters’ own on their own harnesses, at the versions they named, and are not normalised for hardware. No search-volume, keyword or traffic figure is claimed anywhere; the demand signal for this topic was repository issue incidence, which measures reported pain, not query volume.

The source records, fetched page bodies and code files behind this article are a private local author check kept beside the draft, not a downloadable dataset: the issue and pull-request payloads are third-party content, and republishing them is not something this article does. Every claim above links to the public page it came from, and those pages are the record to open.