Self-hosted model endpoints for any role; Hermes v2026.9.24 with reasoning echo

[renmengye]

Operators want to run authors and judges on an open-weight model they serve themselves. vLLM exposes the OpenAI chat-completions, OpenAI Responses and Anthropic Messages APIs on one server behind a bearer key, but the kernel had no way to point a role at it.

Configuration

• Named endpoint profiles: OUTERLOOP_ENDPOINT_<NAME>_URL (the OpenAI-style base, e.g. https://llm.example.internal/v1), _KEY_FILE (absolute, owner-only), _MODEL (the served model), _API (anthropic, responses or chat).
• Roles select a profile with model[endpoint=name], or [endpoint=name] to use the profile's model: the author (OUTERLOOP_AUTHOR_ENDPOINT or the author model), panel lens specs, and standalone reviewers (REVIEW_ENDPOINT). The bracket syntax cannot collide with vendor ids that contain @, / or :.
• Preflight rejects an unknown profile, a missing or permissive key file, or a backend/API mismatch before any spend.

Per backend

• Claude Code: Anthropic API. The base URL drops a trailing /v1 (Claude Code appends /v1/messages), the auth token comes from the key file, and Vertex, Bedrock, Foundry and nonessential traffic are off, so a failing endpoint never falls back to a vendor.
• Hermes: a named custom provider with key_env and model.reasoning_echo: true, so each turn keeps its earlier reasoning (without it, custom endpoints lose it every turn).
• Codex: a custom model provider over the Responses API.
• Keys travel only through the environment (APPTAINERENV_* in containers), are redacted from verdict data, and keep the existing judge/author key separation.
• Run records now persist the author's route. Old records without it keep their native routing (legacy fixtures included).

Hermes pin: v2026.9.24. Audited against the new source: flags now use argparse (the quoting workaround is gone), reasoning_echo exists, and newer instruction-file names (AGENTS.override.md and aliases) are sanitized for judges. The trajectory format and custom-provider contract are unchanged.

Verified live against a self-hosted vLLM server, through the kernel's own build_harness/run_role, with a logging proxy:

• Hermes: verdict filed; 13 of 13 follow-up requests carried earlier reasoning; all requests to the local chat-completions endpoint.
• Claude Code: verdict filed; 14 of 14 carried earlier reasoning (thinking blocks); all to the local Messages endpoint. (The first live run found the doubled /v1, fixed here.)
• Codex over Responses: requests reached the local endpoint and returned 200, but the session filed no verdict; the served model's tool calls do not appear to come back through that server's Responses API. Documented as a known limitation; Claude Code and Hermes are the supported backends for self-hosted endpoints until this is resolved.

Gate: 2352 passed, 2 skipped; ruff, format, mypy clean.

🤖 Generated with Claude Code

[github-actions]

Round 1 — reviewed head 6d5aa001 — reviewer summarizer:hermes/gpt-5.6-terra over coverage+credentials+deployment+general+lifecycle+prose.

terra
Advisory findings from outerloop — the code owner decides. Reply to disagree; the outerloop:no-review label opts this PR out.

Verdict: nothing blocking — 1 advisory note.

Advisory (non-blocking):

• Endpoint reviewer settings never reach the shipped standalone reviewer workflows. [coverage+deployment] The endpoint resolver reads REVIEW_ENDPOINT and OUTERLOOP_ENDPOINT_<NAME>_URL, _KEY_FILE, _MODEL, and _API, but advisory-review-agent.yml exports none of them at this session environment and review-agent.yml neither defines inputs nor forwards them. As a result, hosted ubuntu-latest standalone reviews cannot select a profile and instead use the native path or skip. (.github/workflows/advisory-review-agent.yml:226; high confidence)

Merged one non-blocking configuration gap from the coverage and deployment lenses. Rejected findings: none; the credentials, general, lifecycle, and prose lenses reported no findings.

[renmengye]

On the advisory: intended for now. Endpoint profiles target deployments on the same network as the model server; a GitHub-hosted runner cannot reach a private self-hosted endpoint, so the hosted review workflows keep the native path. Forwarding the profile settings through the reusable workflows is a follow-up for self-hosted runners.