Repository navigation
v2.1.0
·
127 commits
to master
since this release
Immutable
release. Only release title and notes can be modified.
Synapse 2.1.0
Answers in the shape you declare: a model step's answer and a tool's answer can hold groups and lists of groups, and every step that reads them is checked against that shape, on every path a run can take. Every model picker now says what each model is and what it costs, and Synapse AI offers first the models that build well. A loop that ran zero times says so, older runs are a click away, and a run's path draws each step by what it does.
New
- Answers with a shape. A model step's answer fields can now be a group (fields of its own) or a list of groups, as well as lists of numbers or yes-or-nos, to four levels, and any field can be marked "may be empty". In the step editor you can build the fields by hand, paste an example answer, or paste a JSON Schema; "What the model is asked for" shows the schema the run sends. Each provider gets the shape in the strongest form it supports: a schema where it holds one, a required tool call where it doesn't, or plain words with one repair. The
claudeCLI's answer is read in the declared shape too. A problem in an answer is named by its path, for exampleleads[2].email is missing. - A tool's answer has a shape too. The custom tool editor has a new "What it answers" section, in HTTP and Python mode, with the same fields, example and schema paste. When a Tool step's answer lacks a field its tool declares, the step warns and names the path, and the run carries on. A step that later fails on that missing field says which tool left it out. An MCP tool's declared answer is read the same way, as its server's word.
- One check of every read, on every path. When Synapse AI builds or checks a workflow, and as you edit one in the editor, each value a step reads is checked against the steps that can write it on the way there: groups and lists to their full depth, a Tool step's answer against what its tool declares, through branches, loops and parallels. A field that only some branches write is flagged where a condition or code reads it. An Ingest step reached before its records are set is refused, with the fix.
- Pickers that say what each model is. Every model picker shows each model's context (small, standard or large), a cost tier and its price per 1M tokens. Synapse AI's own picker shows first the models that build well and leaves out the ones that aren't for building, small local models included; its Default row says why when the model it would use isn't one. An agent's Model and a step's Model override say how to choose for the job. When an automatic pick can only choose a small model, its reason says so.
- Older runs. The Runs page has "Show older", so every run can be reached however many came after it. A workflow's Recent Runs now shows that workflow's own runs and pages back the same way, where it used to show only those among everyone's newest hundred.
Fixed
- A loop that ran zero times says so. A Loop over an empty list used to hand the next steps an empty result as if it had run, and a body step's value from earlier in the run was passed on as this run's. Now the next steps read that the loop ran zero times, and why.
- A Parallel's join gets every branch's values. A step after a Parallel could be refused for reading a value one of the branches writes. The check now sees what every branch brings to the join.
- A run draws each step by what it does. A Tool step whose arguments are set runs without asking a model, but the run screens drew it as a model call. A workflow's path, the canvas, its minimaps and a step's result card now draw each step by what its settings make it do, before any run and after.
- A custom tool's Save keeps its output schema. Saving a Python tool in the custom tool editor erased an output schema set through the API or an import. The new "What it answers" section edits and keeps it.
- The run page reads a run as it moves. An open run page read the whole run every 3 seconds while it also followed the run's live events. It now reads it when the events say something changed, with a slow safety read every 30 seconds, and goes back to polling if the stream ends or breaks.
- Synapse AI's local-model tip suggests gpt-oss:20b. Local models of 9B or fewer are no longer offered for building, so the tip for running Synapse AI on Ollama now names
gpt-oss:20b: about a 14 GB download, it runs in 16 GB of memory and reads 128K. On a narrow screen the pull command stays on one line, so you can copy it whole. - The panel says why a message was refused. When the builder refused a message, for example one with an image for a model that can't see images, the panel said only "The builder could not be reached (422)". It now says why, in the builder's own words, and what to do instead.
- A new workflow that runs one you name is built new. Asked to make a workflow that runs, calls or starts after one you name, Synapse AI could plan it as a change to that workflow when it wasn't open. It now builds a new workflow, and a change you ask for by name stays a change.
- Synapse AI recommends only models you can use. Asked which model to pick, it could suggest a model CLI that wasn't signed in here and say you could use it straight away. It now leaves out a CLI that isn't signed in and says how to sign in to it.
synapse status --json now also says how Synapse was started: by you or by autostart, and with which options.
Install
macOS and Linux
curl -LsSf https://synapseorch.com/install.sh | bashWindows (PowerShell)
irm https://synapseorch.com/install.ps1 | iexnpm, pip or Docker
npm install -g synapse-orch-ai && synapse
pip install synapse-orch-ai && synapse # Python 3.11–3.14 and Node.js 20.9+For Docker, see Docker: the image is synapseorchai/synapse-ai:2.1.0, also :latest.
Upgrading
- From 2.0.x: run
synapse upgrade. Your data stays where it is, and the database is backed up first. Workflows you've already saved run as before: a flat answer's schema is unchanged, the new read checks run where a workflow is built or edited, never when it runs, and a tool answer that lacks a declared field is a warning, not a failure. One change you may notice: a Loop that runs zero times now leaves its steps' results empty and says so, so a value written to the same key before the Loop is no longer passed on as the Loop's result. If a workflow relied on that as a fallback, set the fallback after the Loop. - From 1.x: run
synapse upgradeon your 1.x install. It moves you to 2.0 the same way 2.0.0 did; read the 2.0.0 notes for what changes.
Full changelog: v2.0.20...v2.1.0 · Documentation