fix(responses): HERMES_DISABLE_REASONING_INCLUDE env-var escape hatch

[nnnet]

Summary

The Responses API rejects include: [\"reasoning.encrypted_content\"] on non-reasoning models with:

HTTP 400: Encrypted content is not supported with this model.
param='include', code='model_not_found'

Users running Hermes against api.openai.com with a non-reasoning model (e.g. gpt-4o-mini) have no way to opt out — every request fails. This PR adds an env-var escape hatch.

Repro

1. ~/.hermes/.env: OPENAI_API_KEY=sk-...
2. ~/.hermes/config.yaml:

   model:
     default: gpt-4o-mini
     provider: auto
     base_url: https://api.openai.com/v1

3. Send any message via Telegram / API.
4. Observed: HTTP 400 on every call.
5. Workaround until this PR: switch to a reasoning model (gpt-5-nano etc.).

Root cause

agent/transports/codex.py:109,124 and agent/auxiliary_client.py:690 unconditionally set include=[\"reasoning.encrypted_content\"] whenever reasoning_enabled is true (the default). The parameter is only accepted by reasoning-capable models (o-series, gpt-5*); the Responses API returns 400 on every other model.

Fix

Add a single env var, HERMES_DISABLE_REASONING_INCLUDE, that suppresses both the include parameter and the reasoning block in the outgoing Responses-API kwargs. Default behavior (always send include) is preserved when the env var is unset or falsy.

Why an env var and not model-family detection?

1. ~30 tests assert include == [\"reasoning.encrypted_content\"] on every Responses-API path; teaching the transport about reasoning-capable model families would require revising every one of those test expectations. That's a larger refactor and a riskier PR.
2. The env var is opt-in and 100% backward compatible — nothing changes for existing users.
3. A follow-up PR can introduce agent.model_metadata.is_reasoning_model() to make this automatic; this PR unblocks users today.

Testing

• Docker image rebuilds (docker compose build gateway).
• docker compose up -d gateway — container healthy.
• API server (/v1/models, /v1/responses) responds normally after restart with the founder's gpt-5-nano default config (env var unset).
• In-container integration check:
  • _reasoning_include_disabled() returns False by default, True for HERMES_DISABLE_REASONING_INCLUDE in {1, true, yes, on}.
  • With env var unset: build_kwargs(\"gpt-4o-mini\", ...) includes [\"reasoning.encrypted_content\"] and reasoning={\"effort\":\"medium\",\"summary\":\"auto\"} (default behavior preserved).
  • With env var \"true\": same call returns kwargs with no include and no reasoning block.

Breaking changes

None. Default behavior is unchanged. The env var is opt-in.

Follow-up

A follow-up PR can introduce agent.model_metadata.is_reasoning_model(model_id) -> bool and switch the transport to use it automatically — at which point the env var can be deprecated in favour of automatic detection.

---

PR by Claude Code on behalf of @nnnet. Tracks internal bug tracker entry BUG-2.

[nnnet]

Closing — will reopen after founder approval of each fix individually (per request). Code unchanged, branch preserved on fork.