fix(agent): coordinate truncated tool call argument repair

[konsisumer]

What does this PR do?

Coordinates truncated tool-call argument handling around one repair contract: repairable malformed JSON continues normally, while unrepairable truncated JSON is reported back to the model as a tool error instead of executing with {} or aborting prematurely.

Related Issue

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Security fix
• Documentation update
• Tests (adding or improving test coverage)
• Refactor (no behavior change)
• New skill (bundled or hub)

Shared root cause

• agent/message_sanitization.py returned only repaired argument text, so callers could not distinguish a legitimate empty {} from the last-resort {} used after failed repair.
• agent/conversation_loop.py had a separate invalid-JSON truncation branch that rejected router-mislabeled truncated tool calls before trying the shared repair routine.
• Streaming and pre-request transcript sanitation therefore made inconsistent decisions for the same broken tool-call arguments.

How this fixes each issue

• 【Bug】Truncated tool_call arguments silently succeed instead of returning error #35151: unrepairable truncated tool-call arguments now produce a tool-result error telling the model to shorten or split the content; Hermes no longer executes the tool with silent empty arguments on that path.
• [RFC] Attempt JSON repair on truncated tool call args before aborting #35574: router-mislabeled truncated tool-call arguments are passed through the existing repair routine before being treated as invalid, so simple missing braces/brackets and dangling commas are repaired and continue normally.

Changes Made

• Added repair_tool_call_arguments_with_status() so callers can tell whether repair succeeded even when the returned JSON is {}.
• Reused that status-bearing repair in the conversation loop, streaming assembly, and pre-request tool-call argument sanitizer.
• Converted unrepairable truncated live tool calls into assistant/tool recovery messages instead of a partial abort.
• Added regression coverage for repair success/failure status, pre-request sanitizer behavior, and router-mislabeled truncated tool calls.

How to Test

1. HERMES_HOME=/private/tmp/hermes-test-home /opt/homebrew/bin/timeout -k 30 480 sh -c 'pytest tests/run_agent/test_repair_tool_call_arguments.py tests/run_agent/test_tool_call_args_sanitizer.py tests/run_agent/test_run_agent.py -q -x --timeout=60'
2. BASE=$(git merge-base origin/main HEAD); { git diff --name-only --diff-filter=d "$BASE"; git ls-files --others --exclude-standard; } | grep -E '\.pyi?$' | sort -u | xargs ruff check
3. python scripts/check-windows-footguns.py run_agent.py agent/message_sanitization.py agent/agent_runtime_helpers.py agent/chat_completion_helpers.py agent/conversation_loop.py tests/run_agent/test_repair_tool_call_arguments.py tests/run_agent/test_tool_call_args_sanitizer.py tests/run_agent/test_run_agent.py
4. Broad suite attempted with /opt/homebrew/bin/timeout -k 30 480 sh -c 'pytest tests/ -q -x --timeout=60 "$@"' sh; it aborted during collection because fastapi/uvicorn are not installed and this Homebrew Python is externally managed, so the lazy dependency installer cannot install them.

What platforms tested on

• macOS on darwin-arm64 (local)

Checklist

Code

• I've read the Contributing Guide / AGENTS instructions
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched the provided issue cluster context for duplicate/member PRs
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass; broad collection is blocked locally by missing dashboard dependencies as noted above
• I've added tests for my changes
• I've tested on my platform: macOS darwin-arm64

Documentation & Housekeeping

• Documentation update N/A
• cli-config.yaml.example update N/A
• CONTRIBUTING.md / AGENTS.md update N/A
• Cross-platform impact considered; changed-file Windows footgun scan passed
• Tool descriptions/schemas update N/A

Screenshots / Logs

• Focused tests: 446 passed
• Ruff changed files: passed (with a pre-existing warning about run_agent.py:107 # noqa directive format)
• Windows footgun scan: passed

Refs #35151
Refs #35574

[tonydwb]

Code Review Summary

Verdict: LGTM

Improves tool call argument repair by distinguishing successful normalization from unrepairable failures. Truncated args now return tool errors instead of empty-object execution.

Looks Good

• ToolCallArgumentsRepair dataclass cleanly separates repair status
• Truncated args now return model-visible errors instead of silent empty execution
• Repairable truncations still get one attempt before fallback
• Comprehensive tests covering repair success, failure, and truncation paths

---

Reviewed by Hermes Agent

[teknium1]

Thanks for the careful investigation and focused tests.

This automated hermes-sweeper review is closing the PR as not planned under the standing model-output-repair policy.

• The PR adds new reconstruction/repair behavior for malformed and truncated model-emitted tool-call JSON, including a new trailing-comma repair and routing repairable truncations through execution.
• Current main instead fail-closes at the execution boundary: agent/tool_executor.py:78-94 accepts only a valid JSON object and produces a model-visible Invalid tool arguments result stating that the tool was not executed.
• Commit 5e50f18b (fix(agent): reject malformed tool call arguments (#61784)) is an ancestor of current main; tests/run_agent/test_malformed_tool_arguments.py:50-98 covers truncated arguments and verifies they are not dispatched while valid siblings still run.
• The linked issue [RFC] Attempt JSON repair on truncated tool call args before aborting #35574 was already closed under this same policy. As its maintainer comment notes, a focused change that preserves invalid-argument rejection while improving a model-visible recovery/error path, without reconstructing malformed content, can be proposed separately.

This is a design-direction decision, not a judgment on the quality of the contribution. Thanks again for the clear root-cause analysis and regression coverage.

---

Automated hermes-sweeper review.

---

Closed as not-planned per standing maintainer policy (model-output-repair). This is a design-direction decision, not a code-quality judgment — see the Contribution Rubric in AGENTS.md for what the project is looking for. If you believe this policy was misapplied to your change, comment here and a maintainer will take a look.