fix(agent): repair malformed tool_call arguments before API send by sirEven · Pull Request #12252 · NousResearch/hermes-agent

sirEven · 2026-04-18T19:16:09Z

What does this PR do?

Adds a JSON repair pipeline at the pre-send normalization point in run_agent.py (~L9179) to fix malformed tool_call.arguments before they reach the API. Currently, when json.loads() fails inside the except Exception: pass block, the invalid JSON is silently passed through to the API — which rejects it with HTTP 400 "invalid tool call arguments", crashing the entire session.

Related Issue

Death spiral: 400 invalid tool call arguments causes context compression loop instead of graceful abort #11001 — Death spiral: 400 invalid tool call arguments causes context compression loop
[Bug]: Malformed persisted tool calls can poison a session and cause repeated 400 errors on subsequent requests #4662 — Malformed persisted tool calls can poison a session
Hermes tool-calling pipeline can corrupt tool names and JSON arguments, causing generic tool-call failures #6841 — Tool-calling pipeline can corrupt JSON arguments

Type of Change

🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

run_agent.py ~L9179: Replaced except Exception: pass with a repair pipeline:
1. Empty/whitespace-only arguments → "{}"
2. Python None literal → "{}"
3. Strip trailing commas before } or ]
4. Auto-close unclosed { and [
5. Remove excess closing delimiters
6. Last resort: replace with "{}" (saves session, loses args — logged at WARNING)
All repairs logged at WARNING level for observability

How to Test

Use a model that produces malformed tool_call arguments (e.g., GLM-5.1 via Ollama proxy)
Without this fix: agent crashes with HTTP 400, session becomes permanently stuck
With this fix: agent logs the repair with WARNING, continues execution

Manual reproduction — in a session where the model generates truncated arguments:

Previously: except Exception: pass → 400 error → session death spiral
Now: repaired to valid JSON → session continues

Tested with live hermes chat session — zero 400 errors after patch. Confirmed Ollama proxy rejects broken args (400) and accepts repaired args (200).

How this differs from existing PRs

PR	Scope	Location	Approach
#11617	Compressor truncation	context_compressor.py	Changes Pass 3 truncation format
#11788	Compressor truncation	context_compressor.py	JSON-safe truncation + tests
#6691	Kimi models only	New kimi_json_sanitizer.py	Model-specific regex sanitizer
#5071	Persistence filter	Message persistence	Filter malformed calls on save
This PR	All models	run_agent.py pre-send	Model-agnostic repair at last line of defense

This PR is complementary, not competing — it catches what slips through upstream. It operates at the pre-send normalization point, which is the last chance to fix broken args before the API call goes out. Even if compressor fixes (#11617, #11788) land, models can still produce malformed args at inference time (truncation, Python None, etc.) — this catches those cases.

Checklist

I have read the Contributing Guide
My commit messages follow Conventional Commits
I searched for existing PRs — this is distinct from fix(compressor): keep truncated tool_call arguments as valid JSON #11617, fix(context_compressor): keep tool-call arguments JSON valid when shrinking #11788, Add JSON sanitizer for Kimi models (fix malformed tool call arguments) #6691, fix(agent): filter malformed tool calls before persistence and replay #5071
My PR contains only changes related to this fix
I have added tests for my changes — help wanted: guidance on test file location/conventions appreciated

Models like GLM-5.1 can produce invalid JSON in tool_call arguments (truncated strings, trailing commas, Python None, empty strings). The current code catches the exception from json.loads() but silently passes, allowing the broken JSON through to the API — which rejects it with HTTP 400 and crashes the entire session. This patch adds a repair pipeline at the pre-send normalization point (~L9179 in run_agent.py): 1. Empty/whitespace → '{}' 2. Python None literal → '{}' 3. Strip trailing commas before }/] 4. Auto-close unclosed { and [ 5. Remove excess closing delimiters 6. Last resort: replace with '{}' (saves session, loses args) All repairs are logged at WARNING level for observability. The fix is model-agnostic — it only activates on the except branch when json.loads() fails on tool_call arguments, so there is zero overhead for valid JSON. It acts as a last line of defense before the API request goes out, complementing upstream fixes that address specific sources of corruption (compressor truncation, per-model sanitizers, etc.). Tested: live session with hermes chat, no 400 errors after patch. Ollama proxy rejects broken args (400), accepts repaired args (200).

@sirEven

Cherry-picked from PR #12252 by @sirEven. Models like GLM-5.1 via Ollama can produce malformed tool_call arguments (truncated JSON, trailing commas, Python None). The existing except Exception: pass silently passes broken args to the API, which rejects them with HTTP 400, crashing the session. Adds a multi-stage repair pipeline at the pre-send normalization point: 1. Empty/whitespace-only → {} 2. Python None literal → {} 3. Strip trailing commas 4. Auto-close unclosed brackets 5. Remove excess closing delimiters 6. Last resort: replace with {} (logged at WARNING)

Follow-up for PR #12252 salvage: - Extract 75-line inline repair block to _repair_tool_call_arguments() module-level helper for testability and readability - Remove redundant 'import re as _re' (re already imported at line 33) - Bound the while-True excess-delimiter removal loop to 50 iterations - Add 17 tests covering all 6 repair stages - Add sirEven to AUTHOR_MAP in release.py

@sirEven

Cherry-picked from PR #12252 by @sirEven. Models like GLM-5.1 via Ollama can produce malformed tool_call arguments (truncated JSON, trailing commas, Python None). The existing except Exception: pass silently passes broken args to the API, which rejects them with HTTP 400, crashing the session. Adds a multi-stage repair pipeline at the pre-send normalization point: 1. Empty/whitespace-only → {} 2. Python None literal → {} 3. Strip trailing commas 4. Auto-close unclosed brackets 5. Remove excess closing delimiters 6. Last resort: replace with {} (logged at WARNING)

Follow-up for PR #12252 salvage: - Extract 75-line inline repair block to _repair_tool_call_arguments() module-level helper for testability and readability - Remove redundant 'import re as _re' (re already imported at line 33) - Bound the while-True excess-delimiter removal loop to 50 iterations - Add 17 tests covering all 6 repair stages - Add sirEven to AUTHOR_MAP in release.py

teknium1 · 2026-04-20T12:13:13Z

Merged via PR #13005 (#13005). Your commit was cherry-picked onto current main with your authorship preserved in git log. Follow-up: extracted to a testable helper, bounded the loop, added 17 tests. Thanks @sirEven!

@sirEven

Cherry-picked from PR NousResearch#12252 by @sirEven. Models like GLM-5.1 via Ollama can produce malformed tool_call arguments (truncated JSON, trailing commas, Python None). The existing except Exception: pass silently passes broken args to the API, which rejects them with HTTP 400, crashing the session. Adds a multi-stage repair pipeline at the pre-send normalization point: 1. Empty/whitespace-only → {} 2. Python None literal → {} 3. Strip trailing commas 4. Auto-close unclosed brackets 5. Remove excess closing delimiters 6. Last resort: replace with {} (logged at WARNING)

Follow-up for PR NousResearch#12252 salvage: - Extract 75-line inline repair block to _repair_tool_call_arguments() module-level helper for testability and readability - Remove redundant 'import re as _re' (re already imported at line 33) - Bound the while-True excess-delimiter removal loop to 50 iterations - Add 17 tests covering all 6 repair stages - Add sirEven to AUTHOR_MAP in release.py

@sirEven

Cherry-picked from PR NousResearch#12252 by @sirEven. Models like GLM-5.1 via Ollama can produce malformed tool_call arguments (truncated JSON, trailing commas, Python None). The existing except Exception: pass silently passes broken args to the API, which rejects them with HTTP 400, crashing the session. Adds a multi-stage repair pipeline at the pre-send normalization point: 1. Empty/whitespace-only → {} 2. Python None literal → {} 3. Strip trailing commas 4. Auto-close unclosed brackets 5. Remove excess closing delimiters 6. Last resort: replace with {} (logged at WARNING)

Follow-up for PR NousResearch#12252 salvage: - Extract 75-line inline repair block to _repair_tool_call_arguments() module-level helper for testability and readability - Remove redundant 'import re as _re' (re already imported at line 33) - Bound the while-True excess-delimiter removal loop to 50 iterations - Add 17 tests covering all 6 repair stages - Add sirEven to AUTHOR_MAP in release.py

@sirEven

Cherry-picked from PR NousResearch#12252 by @sirEven. Models like GLM-5.1 via Ollama can produce malformed tool_call arguments (truncated JSON, trailing commas, Python None). The existing except Exception: pass silently passes broken args to the API, which rejects them with HTTP 400, crashing the session. Adds a multi-stage repair pipeline at the pre-send normalization point: 1. Empty/whitespace-only → {} 2. Python None literal → {} 3. Strip trailing commas 4. Auto-close unclosed brackets 5. Remove excess closing delimiters 6. Last resort: replace with {} (logged at WARNING)

Follow-up for PR NousResearch#12252 salvage: - Extract 75-line inline repair block to _repair_tool_call_arguments() module-level helper for testability and readability - Remove redundant 'import re as _re' (re already imported at line 33) - Bound the while-True excess-delimiter removal loop to 50 iterations - Add 17 tests covering all 6 repair stages - Add sirEven to AUTHOR_MAP in release.py

@sirEven

Cherry-picked from PR NousResearch#12252 by @sirEven. Models like GLM-5.1 via Ollama can produce malformed tool_call arguments (truncated JSON, trailing commas, Python None). The existing except Exception: pass silently passes broken args to the API, which rejects them with HTTP 400, crashing the session. Adds a multi-stage repair pipeline at the pre-send normalization point: 1. Empty/whitespace-only → {} 2. Python None literal → {} 3. Strip trailing commas 4. Auto-close unclosed brackets 5. Remove excess closing delimiters 6. Last resort: replace with {} (logged at WARNING)

Follow-up for PR NousResearch#12252 salvage: - Extract 75-line inline repair block to _repair_tool_call_arguments() module-level helper for testability and readability - Remove redundant 'import re as _re' (re already imported at line 33) - Bound the while-True excess-delimiter removal loop to 50 iterations - Add 17 tests covering all 6 repair stages - Add sirEven to AUTHOR_MAP in release.py

@sirEven

Cherry-picked from PR NousResearch#12252 by @sirEven. Models like GLM-5.1 via Ollama can produce malformed tool_call arguments (truncated JSON, trailing commas, Python None). The existing except Exception: pass silently passes broken args to the API, which rejects them with HTTP 400, crashing the session. Adds a multi-stage repair pipeline at the pre-send normalization point: 1. Empty/whitespace-only → {} 2. Python None literal → {} 3. Strip trailing commas 4. Auto-close unclosed brackets 5. Remove excess closing delimiters 6. Last resort: replace with {} (logged at WARNING)

Follow-up for PR NousResearch#12252 salvage: - Extract 75-line inline repair block to _repair_tool_call_arguments() module-level helper for testability and readability - Remove redundant 'import re as _re' (re already imported at line 33) - Bound the while-True excess-delimiter removal loop to 50 iterations - Add 17 tests covering all 6 repair stages - Add sirEven to AUTHOR_MAP in release.py

@sirEven

Cherry-picked from PR NousResearch#12252 by @sirEven. Models like GLM-5.1 via Ollama can produce malformed tool_call arguments (truncated JSON, trailing commas, Python None). The existing except Exception: pass silently passes broken args to the API, which rejects them with HTTP 400, crashing the session. Adds a multi-stage repair pipeline at the pre-send normalization point: 1. Empty/whitespace-only → {} 2. Python None literal → {} 3. Strip trailing commas 4. Auto-close unclosed brackets 5. Remove excess closing delimiters 6. Last resort: replace with {} (logged at WARNING)

Follow-up for PR NousResearch#12252 salvage: - Extract 75-line inline repair block to _repair_tool_call_arguments() module-level helper for testability and readability - Remove redundant 'import re as _re' (re already imported at line 33) - Bound the while-True excess-delimiter removal loop to 50 iterations - Add 17 tests covering all 6 repair stages - Add sirEven to AUTHOR_MAP in release.py

teknium1 mentioned this pull request Apr 20, 2026

fix(agent): repair malformed tool_call arguments before API send #13005

Merged

teknium1 closed this in #13005 Apr 20, 2026

alt-glitch mentioned this pull request Apr 24, 2026

Compressed sessions with corrupted tool_calls.arguments JSON brick chats with HTTP 400 (invalid_tool_call_format) #15236

Closed

teknium1 mentioned this pull request Apr 27, 2026

Add JSON sanitizer for Kimi models (fix malformed tool call arguments) #6691

Closed

teknium1 mentioned this pull request Jun 10, 2026

Bug: Ollama cloud model glm-5.1 produces malformed JSON tool calls in long contexts #13042

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix(agent): repair malformed tool_call arguments before API send#12252

fix(agent): repair malformed tool_call arguments before API send#12252
sirEven wants to merge 1 commit into
NousResearch:mainfrom
sirEven:fix/tool-call-arg-repair

sirEven commented Apr 18, 2026

Uh oh!

teknium1 commented Apr 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

sirEven commented Apr 18, 2026

What does this PR do?

Related Issue

Type of Change

Changes Made

How to Test

How this differs from existing PRs

Checklist

Uh oh!

teknium1 commented Apr 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants