Limits
Every limit of the agent: analyses a day, analyses per problem, pull requests a month on Sypnia Cloud, sizes, timeouts, and what happens at each limit.
Engine Cloud The agent has limits so that a bad day cannot turn into a large bill, yours or ours. At every limit Sypnia says so in plain sight; it never analyses less quietly.
Analyses
| Limit | Self-hosted engine | Sypnia Cloud |
|---|---|---|
| Analyses a day, per project | 100 by default. Yours to set with SYPNIA_AI_MAX_ANALYSES_PER_DAY; 0 is no limit | Free 20 · Builder 100 · Team 500 |
| Analyses a day, per problem | 3 by default (SYPNIA_AI_MAX_ANALYSES_PER_PROBLEM_PER_DAY) | 3 |
Asking by API (POST …/analyze) | 10 a minute per token | 10 a minute per token |
A day is a calendar day in UTC. Every run counts: the first analysis of a problem, a new one because it grew or came back, and one you asked for.
At the daily limit, a new problem gets analysis.status limit_reached. It keeps its technical title, still alerts, and its prompt carries the evidence. The next day, problems that are still open and were skipped are analysed first, those that hit the most people first. An usage.limit alert is created at 80 % and at 100 %.
GET /v1/usage returns analyses.used_today and analyses.limit_per_day.
Pull requests Cloud
| Plan | Pull requests Sypnia opens a month |
|---|---|
| Free | 5 |
| Builder | 50 |
| Team | 300 |
Counted per organization, per calendar month in UTC, when a pull request is opened. A pull request that Sypnia could not write does not count. See Pull requests. The engine has no pull requests and so no such limit.
Sizes
These are the same everywhere. They keep one analysis small, fast and cheap.
| What | Limit |
|---|---|
| Frames of your own code sent, with source lines | 8 frames, 5 lines on each side |
| Breadcrumbs sent | 30 |
| Spans of the trace sent | 40, the failing ones and their parents first |
| Changed files of the release sent with their diff | 10 files, 20,000 characters of diff in all |
| The whole input | About 24,000 characters. Longer context is cut from the least useful end and the analysis notes it |
| The model’s answer | 2,000 tokens by default (SYPNIA_AI_MAX_OUTPUT_TOKENS) |
| Evidence items in an analysis | 1 to 6 |
| The ready prompt | 12,000 characters |
One analysis is therefore a few thousand tokens in and under two thousand out. What that costs depends on the model you chose; multiply by your provider’s prices.
Time
| What | Limit |
|---|---|
| One model call | 60 seconds by default (SYPNIA_AI_TIMEOUT_SECONDS) |
| Retries | Twice more, after 5 and 30 seconds. Then failed, and tried again when the problem grows |
| From first failure to an analysis | Usually under 3 minutes: up to 2 for the problem to appear, then the model call |
Weekly suggestions
| What | Limit |
|---|---|
| Runs | One a week per project and environment |
| Model calls in a run | At most 25: one per page with findings, the pages with most views first |
| Suggestions kept from a run | At most 20 |
| Counted as analyses? | No. They have their own cap, above |
The suggestions that come from fixed rules have no limit and need no model.
Change the limits
On a self-hosted engine every number marked with a setting is yours to change, in the configuration reference. On Sypnia Cloud the limits are the plan’s; see Plans and limits.