Environment


    Repository:  https://github.com/NousResearch/hermes-agent.git
    Upstream HEAD: 627abbb1e (chore(release): map davidvv in AUTHOR_MAP)
    Local changes: gateway/run.py (2-line fix), run_agent.py (gap-filler)
    Provider:      DeepSeek (custom, base_url: https://api.deepseek.com/v1)
    Model:         deepseek-v4-pro


    Error Output

    From ~/.hermes/logs/errors.log — multiple sessions, same crash pattern:


    2026-04-25 10:23:30,754 ERROR [20260425_064607_5cbb93af] root:
      Non-retryable client error: Error code: 400 -
      {'error': {'message': 'The reasoning_content in the thinking mode
       must be passed back to the API.', 'type': 'invalid_request_error',
       'param': None, 'code': 'invalid_request_error'}}

    2026-04-25 10:27:22,828 ERROR [20260425_064607_5cbb93af] root:
      (same session, retry — same 400)

    2026-04-25 10:30:09,120 ERROR [20260425_102739_d592f4f1] root:
      (new session, second turn — same 400)

    2026-04-27 08:05:38,043 WARNING [20260427_080149_e689a3] root:
      Failed to get summary response: Error code: 400 -
      'The reasoning_content in the thinking mode must be passed back to the API.'

    2026-04-28 15:08:09,337 WARNING [20260428_142857_d1ae44] root:
      Failed to get summary response: Error code: 400 -
      'The reasoning_content in the thinking mode must be passed back to the API.'


    Error hits both the main agent loop AND the session summarizer — because both go through _run_agent → agent_history rebuild.

    The Broken Pipeline (Step by Step)


    TURN 1:
      DeepSeek API → returns reasoning_content: "Let me think..."
      run_agent.py → stores in msg["reasoning_content"]
      session.py    → saves to SQLite ✅ (line 1151 — reasoning_content IS in schema)
      session.py    → writes to JSONL transcript ✅

    TURN 2:
      session.py    → reads history from SQLite ✅ (line 1183 — reasoning_content IS loaded)
      run.py:9940   → rebuilds agent_history from loaded messages
                      ❌ DROPS reasoning_content — not in the preserved-fields tuple
      agent_history → [{role:"user",...}, {role:"assistant", content:"...", reasoning:...}]
                      ^ reasoning_content is ABSENT
      run_agent.py  → _copy_reasoning_content_for_api() gap-filler fires
                      BUT: it copies from msg["reasoning_content"] (which is missing)
                      → nothing to copy, gap remains
      DeepSeek API  → sees: prior assistant turn WITHOUT reasoning_content
                      current turn WITH reasoning_content
                      → HTTP 400: "reasoning_content must be passed back"

    ALL SUBSEQUENT TURNS: same 400 — session is permanently broken


    What Each File Does (and Doesn't Do)

    | File | Saves reasoning_content? | Loads reasoning_content? | Rebuilds history with it? |
    |---|---|---|---|
    | gateway/session.py | ✅ line 1151 | ✅ line 1183 | N/A |
    | gateway/run.py | ❌ line 9940 | N/A | ❌ dropped |
    | run_agent.py | ✅ stores in msg dict | ✅ gap-filler at 7259 | ✅ but operates on already-broken agent_history |

    The problem is in gateway/run.py alone. Session I/O is fine. The gap-filler in run_agent.py is correct but helpless — it receives history that was already stripped during the DB→agent_history rebuild.

    The Fix (Unified Diff)

    diff
    --- a/gateway/run.py
    +++ b/gateway/run.py
    @@ -9937,10 +9937,11 @@
                             if role == "assistant":
    -                            for _rkey in ("reasoning", "reasoning_details",
    +                            for _rkey in ("reasoning", "reasoning_content",
    +                                          "reasoning_details",
                                               "codex_reasoning_items"):
                                     _rval = msg.get(_rkey)
    -                                if _rval:
    +                                if _rval is not None:
                                         entry[_rkey] = _rval


    Change 1: Add "reasoning_content" — this is DeepSeek's native reasoning field, stored correctly by session.py but dropped during history rebuild.

    Change 2: if _rval: → if _rval is not None: — the gap-filler in run_agent.py sets reasoning_content to "" on tool-call turns that lack it. if _rval evaluates "" as falsy and drops it again, causing the same 400 on the next turn. is not None preserves empty strings through the DB round-trip.

    Reproducibility

    100% reproducible. Any multi-turn DeepSeek session on unpatched main will crash on turn 2. The first message always succeeds (no prior history to violate), then the session is dead. Affects all DeepSeek reasoning models (V4-Pro, R1, Reasoner) via both /v1 and /anthropic endpoints.