Governance

Codendum controls who can use the model: the organization’s networks, one API key per user, limits per user and overall, and access logs keyed by user. It does not govern what goes through the model. It does not inspect prompts or answers, apply content policies, protect personal data, guard agents or keep a tamper-evident audit trail. Classrooms and companies often need some of these, for example to protect students’ or employees’ personal data, or to meet internal policies and the EU AI Act.

Today

Area	What Codendum provides
Access	Network allowlist, TLS, one API key per user, per-user and global limits
Exposure	Only /v1/chat/completions and /v1/models reach vLLM; everything else stays on the host
Accountability	nginx access logs with time, user id, path, status and durations, never keys or bodies
Privacy	vLLM does not log prompts; no persistence beyond in-memory caches
Content, personal data, agents, audit	Not covered

Possible future integration with Admina

Admina is an open-source framework for building AI applications that are governed by design. It is licensed under Apache-2.0, self-hosted, and maintained by the author of Codendum (GitHub). Its OpenAI-compatible gateway supports vLLM as an engine. It is a natural candidate to add the missing governance layer, placed between the proxy and the model:

 User workstations ──HTTPS──▶ nginx container    network allowlist, TLS, per-user keys and limits
                                  │  authenticated user id (set by nginx, never by the client)
                                  ▼
                              Admina gateway     policies, personal-data redaction, injection firewall,
                                  │              agent loop breaker, audit log
                                  ▼
                              vLLM container     127.0.0.1:8000

The integration would add one more container to the Docker deployment. For each Admina capability, this is what it would bring to a shared coding-agent service:

Admina capability	Use in Codendum
Personal-data redaction on requests and streamed responses	Keep personal data in prompts and code out of the model, including OpenCode’s streamed answers
Prompt-injection firewall	Screen instructions hidden in repositories, issues or web pages that agents read
Agent loop breaker	Stop runaway agent loops. In the measurements, a coding request took a median of 15 model calls in the simulation and 33 in a real OpenCode session.
MCP tool-call allowlist	Restrict which MCP servers and tools agents may reach, when OpenCode is configured with MCP
Forensic audit log with hash chains	Tamper-evident record of requests per user, with retention decided by the organization
EU AI Act classification	Support the organization’s assessment of its obligations, for example in education

Open points to design and measure before the integration:

• how nginx passes the authenticated user id to Admina, so that policies and audit records are per user;

• which policies suit a classroom and which suit a company team;

• where the audit log is stored, and for how long;

• the effect on time to first token and on throughput. The effect would be measured with scripts/bench.sh and scripts/bench-classroom.sh against the results in Benchmarks.

Status. This is a possible future integration. It is not implemented or tested in Codendum yet, and no configuration in this repository depends on Admina.