scillm

name	scillm
description	Internal Tau-owned LLM proxy on localhost:4001. Surfaces: chat/batch completions, scillm exec, OpenCode serve (coding delegate), OpenCode transport (DAG/SSE), standing Codex agents. Chutes, Gemini, Claude/Codex OAuth, OpenCode Go, Ollama. Auto-routes by model name. ZIP/PDF, JSON repair, batch pools. Project agents must not call this skill directly unless the human explicitly asks to operate SciLLM or the work is Tau/SciLLM maintenance.
allowed-tools	Bash, Read
triggers	["batch LLM calls","parallel completions","describe image","describe figure","describe table","VLM call","multimodal","extract JSON from","analyze image","LLM completion","preflight check","source grounding","grounding verification","verify grounded","call claude","call codex","call gemini","call glm","call opencode go","opencode serve","opencode transport","opencode agent","scillm-debugger","standing agent","scillm exec","call deepseek v4","call minimax","send zip to LLM","send PDF to LLM","coding delegate","patch agent"]
metadata	{"short-description":"scillm (LLM proxy — chat, exec, OpenCode serve, transport, standing agents)"}
provides	["llm-completion"]
composes	["task-monitor","create-evidence-case","analytics","create-figure","llm-eval-lab","memory","dogpile","debugger"]
taxonomy	["inference","llm"]

scillm — Internal Tau Provider Sidecar

Human onboarding: README.md · Repo contracts: docs/SCILLM_OPENCODE_SERVE.md, docs/SCILLM_OPENCODE_TRANSPORT_V1.md, docs/interactive-agents/

Critical Operating Rules

• Project-agent boundary: Do not invoke $scillm, /scillm, http://localhost:4001, /v1/chat/completions, /v1/scillm/exec, /v1/scillm/opencode/*, or /v1/scillm/agents/* directly during ordinary project work. SciLLM is a Tau-owned provider sidecar.
• Allowed direct use: Direct SciLLM calls are allowed only when the human explicitly asks to operate/debug SciLLM itself, when maintaining Tau/SciLLM internals, or inside a Tau-authored DAG/provider adapter that records the required receipts.
• Default routing: For model/provider work in another skill, express the need as a Tau DAG contract, Tau skill node, or the owning skill's Tau-mediated runtime. Do not paste SciLLM curl examples into project-agent workflows.
• Batch calls: httpx.AsyncClient + asyncio.create_task + asyncio.as_completed(tasks) unless the user explicitly requests asyncio.gather or strict input-order completion.
• No default gather for /scillm batches. Reorder by id / scillm_metadata after completion if needed.
• Batch metadata: Every batch item needs scillm_metadata.batch_id and scillm_metadata.item_id.
• Pick a surface first (below) only when you are operating inside the allowed direct-use boundary. Wrong surface = wrong tool loop or missing artifacts.

Setup (one-time per provider)

Provider	Setup	Model / surface
Claude	None if using Claude Code (~/.claude/.credentials.json)	claude-sonnet-4-6, claude-haiku-4-5
Codex	npm install -g @openai/codex && codex login	gpt-5.5
Gemini	GEMINI_API_KEY in .env	gemini-2.5-flash, text-gemini
GLM	GLM_API_Key in .env	text-glm
Chutes	CHUTES_API_KEY + CHUTES_API_BASE	chutes-deepseek, Org/Model
DeepSeek	DEEPSEEK_API in .env	text-deepseek
OpenCode Go	OPENCODE_GO_API_KEY in .env	opencode-go/deepseek-v4-pro, …
OpenCode serve	SCILLM_OPENCODE_SERVE_ENABLED=1; OPENCODE_SERVER_PASSWORD when starting serve	POST /v1/scillm/opencode/runs — agent profiles, not chat models
Ollama	ollama pull model:tag	Any model:tag

Rebuild: docker compose -p scillm -f deploy/docker/compose.scillm.core.yml up -d --build

Auth: GET /v1/scillm/auth with the configured local proxy key. For this repo's local proxy scripts, resolve it as ${SCILLM_MASTER_KEY:-${LITELLM_MASTER_KEY:-${SCILLM_PROXY_KEY:-sk-dev-proxy-123}}}; the dev default works only when the running proxy has not rotated SCILLM_MASTER_KEY/LITELLM_MASTER_KEY.

Invocation surfaces (Tau/SciLLM maintainers only)

Need	Use	Do not use
One-shot text/VLM	POST /v1/chat/completions	OpenCode serve for a paragraph
Pipeline gate / one headless CLI shot	scillm exec / POST /v1/scillm/exec	Product code authorship loops
Bounded repo investigate + optional patch	POST /v1/scillm/opencode/runs	Chat with opencode-go/* in a loop
DAG / debugger + SSE steer	POST /v1/scillm/opencode/transport/*	Blocking serve HTTP with no event tail
Multi-turn Codex in worktree	/v1/scillm/agents/*	OpenCode serve for standing lease loops

Why OpenCode serve sits between chat and exec

• Chat — one completion; no read/grep/skill loop.
• Exec — one bounded headless shot (codex exec, Pi, opencode run with skills/shell denied in config); for graph gates, not collaborative patching.
• OpenCode serve — bounded OpenCode session with an agent profile + optional skills[]. The project agent owns memory, validation, and merge; the worker returns evidence (assistant_text, events.jsonl, optional diff) — not auto-merged.
• Standing agents — multi-turn Codex with lease/handoff; see references/standing-agents.md.

Details: references/opencode-serve.md · Repo: docs/SCILLM_OPENCODE_SERVE.md

Transport (DAG): references/opencode-transport.md · docs/SCILLM_OPENCODE_TRANSPORT_V1.md

Exec profiles: references/exec-workers.md · docs/SCILLM_EXEC.md

How to call

These examples are for Tau provider adapters and SciLLM maintenance. Project agents should route through Tau instead of copying these calls into task workflows.

Chat: POST http://localhost:4001/v1/chat/completions — OpenAI format. Auth: configured proxy bearer, X-Caller-Skill: <project>.

SCILLM_PROXY_KEY="${SCILLM_MASTER_KEY:-${LITELLM_MASTER_KEY:-${SCILLM_PROXY_KEY:-sk-dev-proxy-123}}}"
curl -s http://localhost:4001/v1/chat/completions \
  -H "Authorization: Bearer $SCILLM_PROXY_KEY" \
  -H "X-Caller-Skill: my-project" \
  -H "Content-Type: application/json" \
  -d '{"model":"chutes-deepseek","messages":[{"role":"user","content":"What is 2+2?"}]}'

OpenCode serve (multi-step coding delegate):

SCILLM_PROXY_KEY="${SCILLM_MASTER_KEY:-${LITELLM_MASTER_KEY:-${SCILLM_PROXY_KEY:-sk-dev-proxy-123}}}"
curl -s -X POST http://localhost:4001/v1/scillm/opencode/runs \
  -H "Authorization: Bearer $SCILLM_PROXY_KEY" \
  -H "X-Caller-Skill: my-project" \
  -H "Content-Type: application/json" \
  -d '{"prompt":"Inspect tests/test_foo.py; do not edit.","agent":"build","skills":["memory","scillm"],"timeout_s":600}'

Never put opencode-go/kimi-k2.6 in "agent" — that is a chat model. Never call raw http://127.0.0.1:4096 from product code.

Verify serve: bash scripts/sanity_opencode_serve.sh (from scillm repo root).

Slash wrapper: /scillm "…" · /scillm --model moonshot-text "…" for explicit SciLLM operations only.

Project-Agent Routing Boundary

Project agents should not choose SciLLM surfaces directly. Use Tau as the orchestration and receipt boundary for provider/model work.

1. /memory recall --brief --q "<task>"
2. /dogpile … (if novel/ambiguous) → paste into prompt
3. /debugger (if stuck or hidden runtime state) → breakpoint proof before patch
4. Create or run a Tau DAG / Tau skill node for provider work
5. Tau provider adapter may call SciLLM and must return receipts
6. Validate artifacts; project agent merges or fork-retries
7. /memory store lesson (after verified fix)

Skill	Who runs	Connection
/memory	Project agent (+ optional skills[])	Ground prompt; store after success
/dogpile	Project agent first	Paste synthesis into prompt
/debugger	Project agent when stuck	Proof before asking serve to patch
/scillm	Tau provider adapter or SciLLM maintainer	Sidecar localhost:4001 only, with receipts

Models and routing (summary)

Use model names directly (claude-*, gpt-*, gemini-*, opencode-go/*, Org/Model, model:tag). Discover: GET /v1/scillm/providers, GET /v1/scillm/opencode-go/models?refresh=true.

Avoid deprecated broad alias text for QRA/corpus repair. For quota-sensitive VLM prefer gpt-5.5 or vlm-chutes over generic vlm.

Full tables, Chutes cold-start, OpenCode Go caveats: references/models-and-routing.md

Reference map (load on demand)

Topic	File
Chat, JSON, VLM, message shapes	references/chat-calls.md
Batch, pools, as_completed, OpenCode Go batches	references/batch-calls.md
Source grounding	references/grounding-and-hedged.md
ZIP/PDF/images/files	references/files-multimodal.md
Claude / Codex OAuth	references/oauth-claude-codex.md
scillm exec profiles	references/exec-workers.md
OpenCode serve parameters, fork, skills	references/opencode-serve.md
OpenCode transport SSE	references/opencode-transport.md
Standing /v1/scillm/agents/*	references/standing-agents.md
Middleware, cascade, retry, cache	references/proxy-internals.md
Ops endpoints	references/ops-endpoints.md
Paved path contract	docs/SCILLM_PAVED_PATH_CONTRACT.md

Ops (quick)

Endpoint	Purpose
GET /health/liveliness	Proxy alive
GET /v1/scillm/health	Groups, fallbacks, concurrency
GET /v1/scillm/auth	OAuth token health
POST /v1/scillm/batch/completions	Server-side model_pool batches
POST /v1/scillm/opencode/runs	OpenCode serve run
POST /v1/scillm/opencode/transport/runs	Transport run
GET /v1/scillm/agents/registry	Standing workers

Full table: references/ops-endpoints.md

Composable skills

Skill	Integration
/memory	Recall before work; optional "memory" in OpenCode skills[]
/dogpile	Research before hard problems; optional "dogpile" in skills[]
/debugger	Breakpoint proof before patch; /opencode/serve/debugger/run
/task-monitor	Long-run monitoring
/create-evidence-case, /analytics, /create-figure, /llm-eval-lab	Tau-mediated provider lane, skill-local deterministic path, or explicit SciLLM maintenance only

Composable skills must not instruct project agents to call http://localhost:4001 directly. If a provider/model call is needed, route it through Tau or keep it inside the skill's owned runtime with explicit proof boundaries.