fix(translator): coerce submit_pr_review functionalChanges/findings to arrays

[NomenAK]

Problem

Some strict-schema downstream tools (e.g. submit_pr_review with required functionalChanges/findings array fields) hard-reject when those fields arrive as null, plain object, empty string, or stringified array — malformation modes that can emerge from OpenAI-shape upstream models routed through the streaming openai-to-claude translator.

When the downstream tool rejects, the caller often enters a retry loop it cannot recover from on its own.

Root cause

The streaming OpenAI → Anthropic translator at open-sse/translator/response/openai-to-claude.ts passes tool-call deltas through verbatim. When the upstream model emits a malformed required-array field, the assembled tool args land downstream in the rejected shape.

A related side effect: the strip-empty heuristic from #1852 — correct in the general case — strips legitimate empty [] from these tools' args, so the field appears omitted to the strict-schema validator.

Fix

Adds a small TOOL_SHIMS registry under open-sse/translator/helpers/toolCallShim.ts. The SSE translator applies it on the streaming OpenAI → Anthropic tool_use path. For a registered tool the shim:

1. Buffers raw tool args per shimmed tool (rather than passing partial deltas through).
2. Suppresses the passthrough input_json_delta events for those tools while buffering.
3. Emits one corrective input_json_delta carrying the fully-patched JSON just before content_block_stop.

Handles five malformation modes per field: omitted, null, plain object, empty string, stringified array. Zero overhead for non-shimmed tools.

Adding another strict-schema tool to the registry is one entry.

Test plan

• 20/20 unit + streaming integration tests in tests/unit/translator-tool-call-shim.test.ts (new, 263 lines)
• 4 adjacent translator suites stay green: translator-resp-openai-to-claude, openai-to-claude-strip-empty, empty-tool-name-loop, context-pinning-tool-calls (23 tests combined)
• Typecheck passes against tsconfig.typecheck-core.json
• Verified stable for ~24h on a downstream deploy under review-harness traffic (no recurring rejection observed since deploy)

Scope and trade-offs

Intentionally surgical:

• Does not revise the strip-empty heuristic from [BUG] WebSearch returns empty result blocks on v3.7.7 with Claude Code #1852 — that fix is correct for the 99%+ case. The shim provides the per-tool defense where the heuristic is wrong, without changing the general path.
• Registry is hard-coded (one entry, submit_pr_review). Open to making it config-driven if you'd prefer.
• Covers the SSE streaming path only — that's where the malformations surfaced in flight. Happy to extend to the non-streaming branch if you think it's worth it.

Happy to split into smaller commits, revise, or scope down further.

[gemini-code-assist]

Code Review

This pull request implements a shimming layer for tool calls to handle cases where upstream models emit malformed data, specifically ensuring that fields in the submit_pr_review tool are coerced into arrays. The implementation involves buffering tool arguments and re-emitting a corrected JSON delta at the end of the stream. A review comment correctly identified a potential memory performance issue where arguments are buffered for all tool calls regardless of whether they require shimming; the reviewer suggested limiting buffering to shimmed tools and removing a redundant string check.

[gemini-code-assist]

The current implementation buffers tool arguments for all tool calls, which introduces significant memory overhead for non-shimmed tools. For large tool outputs (e.g., file contents or large data sets), this could lead to excessive memory consumption or OOM errors in high-traffic environments. Buffering should be restricted to only those tools that require shimming.

Additionally, since argBuffer is initialized to "" at line 175, the (toolInfo.argBuffer || "") check is redundant.

          if (toolInfo.shimmed) {
            // Buffer the raw stream for shimmed tools to re-emit corrected JSON at stop time.
            toolInfo.argBuffer += tc.function.arguments;
            // Suppress passthrough; we emit one corrective delta at finish.
            continue;
          }

[diegosouzapw]

Thanks @NomenAK — squash-merged into release/v3.8.0 as 52285d8a. The shim approach is elegant: zero overhead for non-shimmed tools, single corrective input_json_delta before content_block_stop, and the 20 unit + streaming tests cover all five malformation modes per field plus the buffering contract. Easy to add new strict-schema tools to TOOL_SHIMS later. Appreciated.