[Mffff4]

Summary

• Fixed integrity preservation of tool call arguments during streaming (SSE parsing, transformations, and response translation).

Related Issues

• Closes Repeated / duplicated characters in CLI tool-use output when using OmniRoute models #3701

Validation

• npm run lint
• npm run test:unit
• npm run test:coverage
• Coverage is still >= 60% for statements, lines, functions, and branches
• SonarQube PR analysis is green or any remaining issues are explicitly documented below

Tests Added Or Updated

• tests/e2e-tool-calls.sh (new)
• tests/unit/responses-transformer.test.ts
• tests/unit/sse-parser.test.ts
• tests/unit/sseTextTransform.test.ts
• tests/unit/streamingPiiTransform.test.ts
• tests/unit/translator-resp-openai-responses.test.ts

Coverage Notes

• Added unit and e2e tests covering SSE stream parsing, PII transform behavior with tool calls, and payload translation to ensure tool call arguments are preserved intact.

Reviewer Notes

• Added docs/fixes/TOOL_CALL_INTEGRITY.md detailing the integrity mechanism for streamed tool calls.

[gemini-code-assist]

Code Review

This pull request addresses a critical issue where streaming tool call arguments were corrupted or duplicated during SSE translation. It introduces a byte-preserving bypass for machine JSON fields (toolArgs and partialJson) in the text and PII transformation pipelines, and implements an overlap-aware delta assembler (appendToolCallArgumentDelta) to handle repeated snapshots. The changes are accompanied by comprehensive unit and E2E tests, as well as Dockerfile updates to support cloakbrowser. Feedback on the changes includes a high-severity fix for a fragile jq filter in the E2E test script that can cause false warnings, and a performance optimization for the overlap detection algorithm to prevent blocking the event loop.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[diegosouzapw]

Thanks @Mffff4 — this is a valuable fix and the diagnosis of #3701 is spot on. The most important part of your change is correct and lands as-is: routing tool-call arguments / partial_json around the PII and text-stream transforms (they're machine JSON, not human text) and emitting only the incremental slice (emittedDelta) instead of re-emitting the whole accumulated buffer. Those were the real sources of the duplicated-character corruption, and your tests for the PII bypass and the per-emit slicing are genuine regression guards.

During review I found one issue in the accumulator that I fixed in-place on your branch (pushed as a commit authored by you, with me as co-author) so the PR could merge cleanly:

open-sse/utils/toolCallArguments.ts — the fuzzy suffix/prefix-overlap dedup silently dropped bytes from legitimate incremental deltas. Providers like gpt-5.5 (Codex) and Kimi stream function.arguments as small incremental fragments, not full snapshots. When a fragment's leading byte(s) coincided with the buffer's trailing byte(s), the existing.endsWith(next) branch + the overlap loop collapsed them:

Deltas	Expected	Old heuristic
{"a":" x x "}	{"a":"xx"}	{"a":"x"}
{"cmd":"l l -l"}	{"cmd":"ll -l"}	{"cmd":"l -l"}

That turns a visible duplication bug into a silent truncation bug (ls -ll → ls -l) across all four streaming paths — including openai-to-claude.ts, which feeds the exact Claude Code / Droid clients in the issue. The fix restricts dedup to the unambiguous snapshot case only (an identical repeat, or a growing superset that still starts with the buffer) and appends every other fragment verbatim:

if (next === existing) return existing;     // identical snapshot repeat
if (next.startsWith(existing)) return next; // growing snapshot
return existing + next;                       // incremental delta — verbatim

I also added tests/unit/tool-call-arguments.test.ts (direct fragment-level coverage proving no byte loss) and an incremental-fragment test in responses-transformer.test.ts; both fail against the old heuristic and pass with the fix, and I strengthened your snapshot-dedup test to assert the streamed .delta events reconstruct the arguments exactly once.

Validated: 85/85 across the touched suites, typecheck:core clean, lint clean, no PII-redaction regression (content/reasoning redaction untouched; only machine tool-JSON bypasses it, by design). Merging into release/v3.8.24 for the upcoming release. Really appreciate the thorough write-up and the e2e harness. 🙌

[Mffff4]

Awesome, looking forward to the release so I can get back to using it properly)