fix(api-server): preserve tool_call_id in conversation_history

[yzxuyang]

Problem

When sending follow-up messages via the Hermes Web UI through the API Server, DeepSeek (and other strict OpenAI-format providers) returns a 400 error:

messages[3]: missing field `tool_call_id`

Root Cause

_handle_runs() in gateway/platforms/api_server.py reconstructs conversation_history messages with only role and content fields, dropping tool_call_id, tool_calls, and other metadata.

This affects two code paths:

1. Explicit conversation_history from request body (web UI) [line 2409]
2. Multi-message input array mode [line 2438]

Fix

Copy all fields from the incoming message dict instead of creating a new dict with only role + content. This preserves tool_call_id and other metadata while still ensuring role/content values are stringified.

Related

Fixes tool-calling conversations via Hermes Web UI when using providers that enforce the OpenAI message schema strictly (DeepSeek, OpenAI, etc.).

[alt-glitch]

Likely duplicate of #17422 — same root cause: api_server.py strips tool_call_id when reconstructing conversation_history messages.

[alt-glitch]

Likely duplicate of #17422 — same root cause: api_server.py strips tool_call_id when reconstructing conversation_history messages.

[teknium1]

Thanks for isolating the two /v1/runs history projections. The underlying defect is still present on current main: explicit history is reduced to {role, content} at gateway/platforms/api_server.py:4217, and multi-message input is reduced the same way at gateway/platforms/api_server.py:4243. Those histories are passed to AIAgent.run_conversation() at gateway/platforms/api_server.py:4348-4351, so dropping tool_calls and tool_call_id can break the required assistant→tool pairing.

Problems

• There is no regression test for either preservation path. tests/gateway/test_api_server_runs.py:156-165 only covers an invalid history shape.

Suggested changes

• Add tests for explicit conversation_history and multi-message input that assert the mock agent receives both assistant tool_calls and tool tool_call_id unchanged.

Automated hermes-sweeper review.

[GottZ]

This was generated by AI during triage.

Summary

Three PRs address the reported history-normalization defect. #17422 preserves tool metadata in /v1/responses and /v1/runs; #17632 preserves metadata in the two /v1/runs projections; #19415 broadens the fix across chat completions, responses, and runs with a role-aware allowlist plus regression tests for tool and reasoning continuity.

Related pull requests

• fix(gateway): preserve tool_call_id in api_server conversation history #17422 [closed] duplicate — (+30/-3) — n/a: Reconstructs messages in gateway/platforms/api_server.py while preserving tool_call_id, tool_calls, and name at the /v1/responses and /v1/runs history paths, directly addressing the reported metadata loss. Closed, but remains relevant as an overlapping reference implementation for the same root cause.
• fix(api-server): preserve tool_call_id in conversation_history #17632 related — (+8/-2) — n/a: Copies incoming message dictionaries in the explicit-history and multi-message-input /v1/runs paths, preserving tool_call_id, tool_calls, and other metadata instead of retaining only role and content. The [contributor:184 commits] discussion identifies it as likely duplicate of fix(gateway): preserve tool_call_id in api_server conversation history #17422, while the MAINTAINER-BOT VERDICT says keep_open with high salvageability because no regression tests were added.
• fix(gateway): preserve tool and reasoning continuity #19415 duplicate — (+566/-23) — n/a: Adds a shared role-aware history normalizer across /v1/chat/completions, /v1/responses, and /v1/runs, preserving tool and Codex/reasoning continuity fields while dropping unrelated metadata, with regression tests covering the affected paths. The [contributor:184 commits] discussion calls it likely duplicate of fix(api-server): preserve tool_call_id in conversation_history #17632, but the diff shows a broader allowlisted implementation covering the same defect plus reasoning-history replay.

Duplicates

#17422 and #17632 substantially overlap on preserving tool metadata in api_server.py; #19415 covers that same root cause as a broader allowlisted implementation. The duplicate chain is #17422 → #17632 → #19415, with #19415 retaining the broader salvageable scope.

Suggested consolidation

Keep open #19415 with a salvage path: retain its shared role-aware allowlist and regression tests, and verify the full gateway history-normalization behavior before further triage. Close #17632 as duplicate of #19415 despite the MAINTAINER-BOT VERDICT recommending keep_open with high salvageability, because #19415's diff contains the same /v1/runs coverage plus responses/chat paths and tests; #17422 is already closed and remains the earlier overlapping reference.

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 Dup17422 ["PRs duplicating each other"]
        P17422["PR #17422 (closed)"]
        P17632["PR #17632 (open)"]
        P19415["PR #19415 (open)"]
    end
    class P17422 closed
    class P17632 open
    class P19415 open
    class P17632 target
    click P17422 "https://github.com/NousResearch/hermes-agent/pull/17422"
    click P17632 "https://github.com/NousResearch/hermes-agent/pull/17632"
    click P19415 "https://github.com/NousResearch/hermes-agent/pull/19415"

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 (state tag in the node label).

Cross-PR triage: Reviewed 3 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 39 kB of PR diffs, 7 kB of issue/PR text, 2 kB of discussion (7 comments), 0 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.