Truncated tool-call arguments can be silently written to disk (stale partial-parse nativeArgs reused on finalization failure)

[canblmz1]

Summary

When a streamed tool call's arguments are cut off mid-value (e.g. the model hits max_tokens while still writing a string argument), the value that was only ever meant for live streaming-progress display can end up used as the final, executed value with no error or diagnostic indicating it was incomplete.

Where

Observed at commit abaf732bb7a96848a7c6f5dd2ce485b86bff086d.

• src/core/assistant-message/NativeToolCallParser.ts: processStreamingChunk() uses the partial-json package (^0.1.7 in src/package.json) to extract a best-effort value from the JSON accumulated so far, purely to show progress during streaming - createPartialToolUse()'s write_to_file case builds nativeArgs.content directly from that lenient parse.
• When the stream ends, finalizeStreamingToolCall() calls the strict parseToolCall() (JSON.parse on the full accumulated text). A genuinely truncated payload correctly throws and the function returns null.
• src/core/task/Task.ts (search for finalizeStreamingToolCall returned null): when finalization returns null, the code reuses the existing tool-use object - the one built during streaming from the partial-json output - and just sets .partial = false. It does not clear nativeArgs.
• src/core/assistant-message/presentAssistantMessage.ts: the guard here is !block.nativeArgs. Since nativeArgs was already populated during streaming, this guard does not trip.
• src/core/tools/WriteToFileTool.ts: the only content check is newContent === undefined. A truncated-but-present string passes this too.

Repro / PoC

Given a write_to_file call whose content argument is cut off mid-value:

{"path":"config/database.yml","content":"production:\n  host: db.prod.internal\n  password: correct-horse-battery-sta

Running the actual partial-json package (v0.1.7, the version pinned in src/package.json) on this string:

parse(text) -> { path: "config/database.yml",
                  content: "production:\n  host: db.prod.internal\n  password: correct-horse-battery-sta" }

No error. content is present and looks like a normal, complete string - exactly what nativeArgs.content would still hold when finalization fails and the stale streaming-phase object is reused, so WriteToFileTool would write this content to disk exactly as shown, silently.

For contrast, running the same input through prefix-safe-json (an incremental JSON parser I maintain, built specifically around never exposing a value until it's unambiguously complete):

outcome: "truncated", executable: false
stableValue: { path: "config/database.yml" }   // content is absent, not guessed

Impact

Any tool whose schema includes a free-form string argument (file content, shell command, etc.) can have that argument silently truncated-but-executed when the model's response is cut off mid-argument - a data-integrity issue for any tool with a real side effect.

Context

This codebase shares this exact code path (including comments) with RooCodeInc/Roo-Code, which this appears to be derived from - I reported the same issue there as well, in case it's useful for cross-referencing a fix.

Suggested direction (not prescriptive)

When finalizeStreamingToolCall() returns null, clear/discard the reused tool-use object's nativeArgs (and ideally params) rather than only marking it non-partial - falling back to the existing !block.nativeArgs guard at that point instead of leaving stale partial-parse data in place.

I found this by cloning the repository and reading the real source (paths/lines above), then verified the specific claim about partial-json's output by running the actual package version you depend on - not speculation. Happy to answer questions or share the small repro script if useful.