fix(agent): JSON-serialize non-string tool results to prevent API 400

[hermesagent26]

Two-layer fix for HTTP 400: invalid message content type: map[string]interface{}.

Root cause: MCP tools and memory helpers return Python dicts/lists as tool results. The OpenAI SDK expects tool role content to be a string or content-parts list, not raw objects.

Changes:

1. run_agent.py:_tool_result_content_for_active_model — serializes non-str results before appending to messages.
2. agent_runtime_helpers.py:sanitize_api_messages — coerces tool role content to JSON string as a safety-net before every API call.

Fixes the 'model provider failed after retries' loop caused by a single bad tool result poisoning the entire message history.

[teknium1]

Thanks for targeting a real strict-provider failure: current run_agent.py:5086-5087 still returns a plain non-multimodal dict unchanged, and agent/agent_runtime_helpers.py:2462-2657 has no content-type repair despite being used by both outbound paths.

Problems

• agent/agent_runtime_helpers.py:1678 JSON-encodes every non-string tool content value. This also encodes valid content-part lists: run_agent.py:5089-5117 deliberately returns such lists for multimodal tool results, so the added safety net would discard their structured vision payload.
• run_agent.py:3665-3668 changes Kimi/Ollama reasoning detection but is unrelated to tool-result serialization. Current run_agent.py:5505-5516 documents host-driven Kimi detection to avoid model-name routing; please keep this PR scoped to the serialization fix.
• The PR changes only runtime files; add regression tests for a fresh dict tool result and persisted dict tool content through sanitize_api_messages.

Suggested changes

• Preserve strings and valid content-part lists; serialize only invalid tool-content shapes, with explicit None behavior.
• Cover both the construction and pre-API sanitizer paths with focused tests.

Automated hermes-sweeper review.

[teknium1]

This blanket check also serializes valid OpenAI content-part lists. _tool_result_content_for_active_model intentionally returns such a list for multimodal results, so the final sanitizer would turn a vision payload into JSON text. Preserve valid lists and coerce only invalid tool-content shapes.

[teknium1]

This Kimi/Ollama reasoning change is unrelated to the stated serialization fix. Please split it out: current main documents Kimi detection as host-driven specifically to avoid model-name routing (run_agent.py:5505-5516).

[GottZ]

This was generated by AI during triage.

Summary

Two PRs address the strict-provider HTTP 400 caused by non-string tool-result content: #29920 applies the fix at result creation and outbound sanitization, while #29937 implements similar serialization but bundles several unrelated provider, context-length, and skills-sync changes. Both sanitizers currently JSON-encode all non-string tool content, including valid structured content-part lists, so neither diff is merge-ready as written.

Related pull requests

• #29920 related — (+19/-0) — keep open and revise: The focused two-layer fix addresses fresh and persisted dict/list tool results, but the contributor keep_open review correctly identifies that the sanitizer also destroys valid multimodal content-part lists; remove the unrelated Kimi/Ollama routing change and add regression tests for fresh dict results, persisted invalid content, and preserved valid content-part lists.
• #29937 [closed] duplicate — (+106/-14) — keep closed as a superseded duplicate: Its core serialization and sanitizer changes overlap #29920 and have the same valid-content-part-list defect, while the diff additionally bundles unrelated llama.cpp context-length, system-message, custom-provider, and skills-sync changes; it remains relevant as an alternative implementation and evidence that the fix should stay narrowly scoped.

Duplicates

#29920 and #29937 substantially duplicate the same two-layer tool-result serialization change; #29937 is the broader, closed duplicate.

Suggested consolidation

Revise and then merge #29920 as the focused consolidation target: preserve strings and valid content-part lists, serialize only invalid tool-content shapes, remove the unrelated Kimi/Ollama change, and add the requested regression coverage. Keep #29937 closed as the duplicate because its relevant fix is covered by #29920 and its remaining changes belong in separate PRs.

Complex graph

flowchart LR
    classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a
    classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151
    classDef best stroke-width:3px,stroke:#b45309
    classDef target stroke-width:3px,stroke:#4338ca
    subgraph Dup29920 ["PRs duplicating each other"]
        P29920["PR #29920 (open)"]
        P29937["PR #29937 (closed)"]
    end
    class P29920 open
    class P29937 closed
    class P29920 target
    click P29920 "https://github.com/NousResearch/hermes-agent/pull/29920"
    click P29937 "https://github.com/NousResearch/hermes-agent/pull/29937"

Loading

Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed or no verify verdict yet (state tag in the node label).

Cross-PR triage: Reviewed 2 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 15 kB of PR diffs, 2 kB of issue/PR text, 2 kB of discussion (3 comments), 1 verify verdict. verdicts reflect diff content, not PR titles. Part of an automated triage batch.