omp-saia-plugin

SAIA (Academic Cloud Hessen) provider for the OMP coding agent

An OMP plugin that auto-registers all SAIA Academic Cloud models as a provider — no manual configuration needed.

This is the OMP port of pi-saia-plugin. It is part of a broader agent memory ecosystem; see the agent-memory-research survey [arXiv 2512.13564] for the research foundation driving our memory architecture design.

Features

• Dynamic discovery — the provider is registered via OMP's runtime provider API (pi.registerProvider() with fetchDynamicModels); OMP queries the SAIA /v1/models endpoint through the same model-cache pipeline as built-in providers (24 h TTL, omp models refresh to force)
• Zero config — API key from $SAIA_API_KEY (OMP loads .env files automatically) or a config-sourced key; no reload needed once the key is set — omp models re-runs discovery
• Skill included — /skill:saia-models documents available models and usage
• OpenAI-compatible — uses standard openai-completions API (chat-ai.academiccloud.de/v1)

Installation

# From a local checkout (recommended for development)
omp plugin link /path/to/omp-saia-plugin

# From git
omp plugin install github:koalajoe23/omp-saia-plugin

The plugin is enabled automatically on install. Restart omp if it was running.

Available Models

Models are dynamically discovered from the SAIA API. The table below lists currently known models with context windows sourced from official SAIA docs.

Model ID	Name	Context	Reasoning
saia/glm-4.7	GLM 4.7	200K	✅
saia/glm-5.3-flash	GLM 5.3 Flash	1.25M	✅
saia/qwen3.5-397b-a17b	Qwen 3.5 397B	256K	✅
saia/qwen3.5-122b-a10b	Qwen 3.5 122B	256K	✅
saia/qwen3.8-27b	Qwen 3.8 27B	262K	✅
saia/devstral-2-123b-instruct-2512	DevStral 2 123B	256K	❌
saia/openai-gpt-oss-120b	GPT-OSS 120B	128K	✅
saia/qwen3.6-35b-a3b	Qwen 3.6 35B	262K	✅
saia/deepseek-v4-flash-0731	DeepSeek V4 Flash 0731	1M	✅
saia/mistral-medium-3.5-128b	Mistral Medium 3.5	256K	✅
saia/gemma-4-31b-it	Gemma 4 31B	256K	✅
saia/qwen3-30b-a3b-instruct-2507	Qwen 3 30B	256K	❌
saia/qwen3-coder-next	Qwen 3 Coder Next	256K	❌
saia/qwen3-omni-30b-a3b-instruct	Qwen 3 Omni 30B	256K	❌
saia/apertus-70b-instruct-2509	Apertus 70B	64K	❌
saia/meta-llama-3.1-8b-instruct	Llama 3.1 8B	128K	❌

Reasoning status is verified against the live API with reasoning_effort probes (2026-09-10); the ✅ set can change as SAIA updates model capabilities. Date-stamped variants (e.g. saia/deepseek-v4-flash-0731) inherit capability and context from their base id. The actual model set may vary — run omp models | grep ^saia to see what's currently available. Most reasoning models expose the full minimal…max effort ladder; mistral-medium-3.5-128b (only none/high), openai-gpt-oss-120b (low/medium/high), and qwen3.8-27b (low/medium/xhigh) are restricted to their backend's accepted values.

Date-stamped variants: SAIA sometimes serves date-stamped variants of a model (e.g. saia/deepseek-v4-flash-0731 alongside the base deepseek-v4-flash). Capability lookups strip a trailing -NNNN stamp, so variants inherit the base model's reasoning support and context window instead of falling back to defaults.

Usage

# List available models
omp models | grep ^saia

# Switch to a model
/model saia/glm-4.7

# With thinking level
/model saia/qwen3.5-397b-a17b:high

# Force a refresh of the model list (bypass the 24 h cache)
omp models refresh

# Load the skill for documentation
/skill:saia-models

Automatic capability reconciliation

The plugin keeps model capabilities (reasoning, vision, context windows) fresh in the background: a deferred cycle at startup (when the store is stale), then every 6 h while omp runs. Each cycle refreshes the model list, probes unknown models for reasoning (tiny ~2-token calls, ≤2 per cycle), re-scrapes SAIA's docs page weekly for context windows, and stores the result in ~/.omp/agent/saia-models.json. omp picks the fresh data up at its next model discovery (omp models / omp models refresh).

Trigger a reconcile manually: /saia-refresh — then omp models refresh to surface the result immediately.

All timings are env-configurable (invalid values fall back to defaults):

Variable	Default	Meaning
SAIA_RECONCILE_STARTUP_DELAY_MS	5000	defer reconcile after session start
SAIA_RECONCILE_INTERVAL_MS	21600000	background cycle (6 h)
SAIA_RECONCILE_STALE_AFTER_MS	43200000	catch-up at startup if store older
SAIA_RECONCILE_PROBES_PER_CYCLE	2	probe budget per cycle
SAIA_RECONCILE_PROBE_TIMEOUT_MS	20000	per-probe timeout
SAIA_RECONCILE_REVERIFY_MS	604800000	per-model re-verify (7 d)
SAIA_RECONCILE_SCRAPE_INTERVAL_MS	604800000	docs scrape cadence (7 d)
SAIA_RECONCILE_SCRAPE_TIMEOUT_MS	15000	scrape fetch timeout
SAIA_RECONCILE_STORE_PATH	~/.omp/agent/saia-models.json	store location
SAIA_RECONCILE_DISABLED	unset	set to any truthy value to disable

Failures degrade silently to the last known state or the static tables; nothing blocks startup or discovery.

API Key

The API key is resolved from $SAIA_API_KEY (OMP also loads .env files from the project, ~/.omp/agent/, and ~/.omp/):

export SAIA_API_KEY=your-key
omp models | grep ^saia

Without a key the saia provider is registered but carries no models; set the variable and run omp models to trigger discovery — no reload needed.

Package Structure

omp-saia-plugin/
├── package.json              # OMP plugin manifest (omp.extensions)
├── extensions/
│   ├── index.ts              # Provider registration (pi.registerProvider + fetchDynamicModels)
│   ├── discovery.ts          # API client: fetchModels(), buildModelDefs()
│   ├── config.ts             # ModelDef → ProviderModelConfig transformer
│   ├── constants.ts          # Provider ID, base URL, context window table
│   └── types.ts              # TypeScript interfaces for API responses
├── skills/
│   └── saia-models/
│       └── SKILL.md          # Model documentation skill
└── test/
    └── extensions.test.ts    # bun test for discovery/config logic

Development

git clone https://github.com/koalajoe23/omp-saia-plugin
cd omp-saia-plugin
bun install
bun check    # tsc --noEmit
bun test

Test locally:

omp plugin link /path/to/omp-saia-plugin
omp models | grep ^saia

Porting Notes (pi → OMP)

pi	OMP
pi.registerProvider(id, { models }) — one-shot fetch at startup	pi.registerProvider(id, { fetchDynamicModels }) — OMP's runtime discovery pipeline (SQLite model cache, 24 h TTL, omp models refresh)
thinkingLevelMap (pi thinking levels → reasoning_effort)	thinking: { mode: "effort", efforts: [...] } — OMP levels map 1:1 to vLLM values; compat.supportsReasoningEffort: true
compat.supportsThinkingTokenBudget	not exposed by OMP's compat schema (dropped)
apiKey: "$SAIA_API_KEY" (deferred env resolution)	resolved at registration; env is re-read inside the discovery callback so the key can appear later
pi install / /reload	omp plugin link / omp plugin install; no reload needed for model changes

Acknowledgements

This plugin is based on pi-saia-plugin by Tobias Weiss: the model discovery logic, context-window tables, reasoning verification data, and skill content are derived from that project, ported from the pi extension API to OMP's. Thanks to the original author and contributors.

License

MIT