[wahyuzero]

Summary

Fixes a streaming protocol bug in the Kiro executor that caused random Unterminated string JSON parse errors in downstream consumers (opencode, claude-code, etc.) for tool calls with substantial argument content.

Root cause

open-sse/executors/kiro.js reads toolUseEvent.input and emits its JSON.stringify as an OpenAI tool_calls[].function.arguments delta on every event. Kiro/AWS CodeWhisperer streams input as partial objects that grow across events:

event 1: { command: "cat /home" }
event 2: { command: "cat /home/wxsys" }
event 3: { command: "cat /home/wxsys/Project/naskin/.impeccable.md" }

Stringifying each one produces overlapping prefixes:

{"command":"cat /home"}
{"command":"cat /home/wxsys"}
{"command":"cat /home/wxsys/Project/naskin/.impeccable.md"}

The OpenAI streaming protocol requires arguments deltas to be concatenable string fragments that join into one valid JSON. Concatenating the three above produces:

{"command":"cat /home"}{"command":"cat /home/wxsys"}{"command":"cat /home/wxsys/Project/naskin/.impeccable.md"}

→ JSON.parse fails with Unexpected non-whitespace character after JSON at position N or Unterminated string depending on where the parser bails out.

The truncation position is non-deterministic because it tracks the prefix length of whichever partial object Kiro happened to send. This makes the bug present as random "the network truncated my tool call" failures — {"comm, {"comman, {"command": — when in fact the protocol contract was violated upstream.

Reproduction

Any Kiro session that streams a tool whose arguments grow over multiple toolUseEvents will hit this. Common triggers:

• File paths longer than a few words
• Multi-line content parameters in write tools
• Shell commands that get progressively expanded

I encountered it in opencode while writing a 1500-byte markdown file via the write tool routed through 9router → Kiro → Claude Opus 4.7. The write call failed with Unterminated string at variable byte offsets across retries.

Fix

1. String-form input payloads are already concatenable deltas — emit them incrementally as before, tracking accumulated text per toolCallId in state.toolArgsEmitted.

2. Object-form input payloads are buffered into state.toolArgsBuffered keyed by toolCallId. The latest canonical (full JSON of the most recent partial object) overwrites earlier buffered values — earlier partials are noise once a more complete one arrives.

3. New helper flushBufferedToolArgs() is invoked at all three finish boundaries to emit each buffered canonical exactly once as a complete arguments string:

   • messageStopEvent
   • The dual-event finish path (hasMeteringEvent && hasContextUsage)
   • The TransformStream flush() callback

The result: downstream consumers receive arguments deltas that concatenate into valid JSON regardless of whether Kiro chose to send string fragments or growing partial objects.

Tests

New regression coverage (tests/unit/kiro-tool-args-streaming.test.js, 7 tests):

• Single buffered tool call → valid flush
• 4-event partial-object stream (the exact reported repro) → 1 valid flush
• Multiple concurrent tool calls → indices preserved
• Idempotent flush (canonical equals already-emitted) → no re-emit
• Empty buffer → no-op
• Contract test: re-stringifying partials is unparseable
• Contract test: emitting once is parseable

No regression: all 14 existing kiro tests (openai-to-kiro.test.js, rtkKiro.test.js) continue to pass.

End-to-end verified: built the patched CLI tarball locally, reinstalled 9router@0.4.66, restarted the proxy, and reproduced the original failure case (substantial write tool call ~1500 bytes) — now succeeds. Multiple test writes confirmed.

Notes

• The fix is executor-only — translator/response/kiro-to-openai.js has the same shape of bug but is a separate (non-streaming-executor) code path. If this PR lands, that file deserves a follow-up; happy to include it here if preferred.
• state initialization adds two new Maps (toolArgsEmitted, toolArgsBuffered) — no change to existing field types.
• flushBufferedToolArgs is exported so the unit tests can target it directly without standing up a full TransformStream pipeline.

[tonydwb]

Code Review Summary

Verdict: Approved — PR #1573

Reviewed via Hermes Agent batch. Changes are well-scoped for a PR of this size. No blocking issues identified in this pass.

---

Reviewed by Hermes Agent