Hard crash on GLM-4.6/5.2 (Fireworks): tool-call `arguments` rejected as 'invalid JSON', session dies

[fishloa]

Summary

Running an agentic coding session with a newer model on an OpenAI-compatible provider (GLM 5.2 via Fireworks AI), Crush hard-exits mid-session whenever the model emits an edit/write tool call, with:

ERROR  Agent processing failed: failed to start agent processing stream:
bad request: Invalid tool call in messages:
tool_calls[].function.arguments for function 'edit' must be a JSON object string (or an object), got invalid JSON.

The whole crush run process dies (exit 1). The model's accumulated progress in that turn is lost.

Environment

• Crush v0.76.0 (also nothing relevant in v0.77.0 changelog).
• Provider: custom fireworks (OpenAI-compatible, https://api.fireworks.ai/inference/v1).
• Model: accounts/fireworks/models/glm-5p2 (GLM 5.2). Headless crush run -m … -c … --quiet.

Key observation

It crashes only on tool calls with large/complex string arguments — edit and write (which carry code blobs with newlines/quotes/special chars). The same model's read/grep/reasoning tool calls in the same session work fine, and it produces excellent diagnoses. So GLM 5.2 is emitting the arguments field in a shape Crush's strict parser rejects (e.g. not a JSON-stringified object, or with unescaped content), specifically when the payload is big.

Other models on the same Fireworks provider (DeepSeek V4 Pro, Kimi K2.7) emit clean tool args and never hit this — resuming the same session on one of those lets the edit go through, which confirms it's a per-model tool-arg-serialization mismatch, not a session/config problem.

Impact

The model can't apply any file edit → unusable for coding with GLM 5.2, and a single bad tool-call kills the run rather than being recoverable/retried.

Suggested fix

Be lenient when normalizing tool_calls[].function.arguments from OpenAI-compatible providers:

• If arguments is a string that fails JSON.parse, attempt a repair pass (strip markdown fences, fix common escaping) before rejecting.
• Treat a malformed tool-call as a recoverable turn error (feed the parse error back to the model to retry) instead of a fatal stream error that exits the process.

Happy to provide a fuller repro (the GLM 5.2 raw tool-call payload) if useful.

[fishloa]

Root cause found — streaming tool-call delta assembly, not the provider/model

Reproduced against Fireworks directly; the fault is Crush's streaming assembly of multi-chunk tool_calls[].function.arguments, not GLM/Fireworks.

Non-streaming GLM 5.2 returns a single valid arguments JSON string — always parses.

Streaming (stream:true, which crush run uses) — counting delta.tool_calls[].function.arguments fragments for the SAME request:

model	# arguments delta chunks	concatenated → JSON.parse
glm-5p2 (Fireworks)	23	VALID
deepseek-v4-pro (Fireworks)	1	VALID

GLM 5.2 streams the arguments across ~23 small deltas; DeepSeek sends 1. Concatenated in order, GLM's 23 chunks form valid JSON — the provider stream is correct. The crash appears later as Invalid tool call in messages (a 400 when Crush replays the stored assistant tool-call next turn), i.e. Crush stored malformed/truncated arguments because its delta accumulator mishandles the many-fragment case (likely not appending every fragment keyed by tool_calls[].index).

Repro: stream a tool_choice:required request to accounts/fireworks/models/glm-5p2 with an edit-shaped function; observe 20+ argument delta fragments. Correct accumulation (append every delta.tool_calls[i].function.arguments by index) yields valid JSON. Models that stream args in many fragments (GLM 5.2) break; single-fragment models (DeepSeek) work. Fix is in the streaming accumulator.

[fishloa]

Workaround that works: disable streaming via extra_body

Confirmed fix without patching Crush: set provider-level extra_body: {"stream": false} on the openai-compatible provider. GLM 5.2 then applies edits cleanly — Crush handles the non-streamed response and the multi-fragment streaming accumulator is never exercised.

"providers": { "fireworks": { "type": "openai", "base_url": "...", "extra_body": { "stream": false }, "models": [...] } }

For headless crush run this is a free workaround (no live-token benefit lost). The underlying streaming-accumulator bug (many function.arguments fragments mis-assembled) still wants fixing for the streaming TUI path.