llms.txt
Compare with other sites' llms.txt files →# OpenSOP
> Process as Infrastructure for agentic processes — declare, version, fork, run, and audit a process as a file in your repo.
OpenSOP makes agentic processes reproducible *as procedures*. A naive prompt re-derives the procedure on every invocation — different shape, different fields, invented items, no audit trail. An OpenSOP process pins the steps, validates step outputs against declared required/type/enum constraints (retrying LLM steps on mismatch), and writes an append-only receipt to disk: the same procedure, the same output contract, and a full audit trail on every run. (Step *values* still depend on models, commands, inputs, and external services — what OpenSOP fixes is the procedure, the schema, and the receipt, not the laws of nondeterminism.)
**Terraform is to cloud resources what OpenSOP is to agentic processes.**
---
## What a process is
A process is a `.sop.json` (or `.sop.yaml`) file declaring typed inputs, an ordered list of steps, and typed outputs. It lives in your repo, reviews in your PRs, ships in your commits.
Step types executable by the local CLI today: `automated`, `shell`, `noop`, `form`, `approval`, `wait`, `llm`, `webhook`, `subprocess`. Additional types — `loop`, `judgment`, `notification` — are defined in the spec but run server-side (see https://github.com/Chosen9115/opensop/blob/main/SPEC.md §3 capability matrix for what runs where). Each step declares what it requires and what it returns. When a required upstream output is missing, execution stops — the runtime does not ask an LLM to fill the gap.
Spec version: **0.7**. Full grammar in https://github.com/Chosen9115/opensop/blob/main/SPEC.md.
---
## Why it helps an agent
- **Reproducibility of procedure, not luck.** The process file is the source of truth: the same steps, the same declared output schema, and the same audit format on every run. Where a naive prompt drifts run-to-run — different fields, invented items, unstructured prose — the process holds the contract: output that conforms to the schema every time (step *values* still depend on models and inputs).
- **Scope discipline.** The prompt is part of the file. Output fields are declared. The LLM step validates its response against `expected_output_schema` and retries on mismatch — it cannot silently return the wrong shape.
- **Auditability.** Every run writes `manifest.json`, `audit.jsonl`, and `context.json` under `$OPENSOP_LOCAL_HOME/runs/<run_id>/` — the active cell's `.opensop/runs/` when run inside a cell, otherwise `~/.opensop-local/runs/`. Inspect any run after the fact.
- **Discovery before invention.** Before re-deriving a procedure, search for an existing process: `opensop search <keyword>` or `opensop suggest "<task description>"`. Write new only when no match clears a confidence threshold.
- **Evolution without copy-paste.** Fork a process, adapt it, record lineage. `opensop fork`, `opensop lineage`, `opensop diff` — processes evolve like code.
---
## Quick start
Install (deps: bash 4+, jq):
```bash
curl -fsSL https://raw.githubusercontent.com/Chosen9115/opensop/main/cli/bin/opensop \
-o /usr/local/bin/opensop && chmod +x /usr/local/bin/opensop
```
Discover existing processes before writing a new one:
```bash
opensop search lead # ranked keyword search (local)
opensop suggest "qualify an inbound lead" # intent-based match
opensop list # enumerate all processes in cell chain
```
Run a process:
```bash
opensop run ./extract-action-items.sop.json --input notes="$(cat meeting.txt)"
opensop status <run_id>
opensop show <run_id> --json # full receipt
```
Machine-readable command index (for agent consumption):
```bash
opensop help # full command reference
opensop help --json # JSON index: command, summary, usage, category, backend
opensop help --json | jq '.[] | select(.category=="discovery")'
```
Validate a process file without running it:
```bash
opensop dry-run ./my-process.sop.json --input key=value
```
---
## Key documents
| Document | Purpose |
|---|---|
| https://github.com/Chosen9115/opensop/blob/main/SPEC.md | Authoritative format grammar, step-type semantics, and `/sop/*` HTTP API contract (v0.7) |
| https://github.com/Chosen9115/opensop/blob/main/MANIFESTO.md | The thesis: why processes are infrastructure |
| https://github.com/Chosen9115/opensop/blob/main/docs/AGENTS.md | Agent guide: Discover -> Run -> Build -> openSOP-ize -> Iterate & Evolve |
| https://github.com/Chosen9115/opensop/blob/main/EVOLUTION.md | Mineralization tiers (m0-m6), fork lineage, and evolution policy (experimental) |
| https://github.com/Chosen9115/opensop/blob/main/cli/README.md | Full CLI command reference, step type matrix, pause/resume protocol |
| https://github.com/Chosen9115/opensop/blob/main/README.md | Project overview, quick start, step type table, optional server |
---
## Optional
The CLI runs processes locally with no server. For shared orchestration, a team audit log, or a monitoring UI, point the CLI at any server that implements the `/sop/*` API contract in https://github.com/Chosen9115/opensop/blob/main/SPEC.md §4.2.
Reference server implementation: https://github.com/Chosen9115/opensop-rails — a Rails 8 server adding team visibility, webhook-triggered starts (HMAC-authenticated), and a shared audit log. Use it with:
```bash
opensop config set url https://your-server.example.com
opensop config set token <your-api-token>
opensop --remote run lead-qualification --input lead_name="Alice"
opensop --remote search lead
```
The server is optional infrastructure, not the core. `opensop run` never requires it.