Limits and errors

Asking questions

Question timeout: 5 minutes. A question that takes longer returns an error and closes the chat turn.
Rows per query: Cassis fetches at most 10,000 rows per query. A larger result shows as “10,000+ rows” in the app, and the answer says it was cut and suggests aggregating, filtering, or downloading the CSV.
Result rows: Up to 100 rows inline over MCP, with truncated marking a capped result set. For more, narrow the question or ask for aggregates.
CSV download: Download CSV in the app runs the query again and fetches up to 200,000 rows. A file cut at that limit says so.
Zero rows: Not an error. The query ran and matched nothing.
Null results: The SQL was not executed. On a schema-only project that is always the case: see Run without a warehouse connection.
Not answerable: The question needs information, context or data that Cassis cannot supply. Read the explanation, then provide the missing information, narrow the question or repair its source before retrying. See ask_question.
Model overloaded: When the model is briefly overloaded or rate-limited, the tool returns a plain-language message rather than a raw error. Retry in a moment.

Transport: Stateless Streamable HTTP: each POST has its own transport. Transport state is not kept between requests. GET and DELETE are refused.
Source schema detail: get_source_schema details at most 20 tables per call. Call it again for more, or use cassis schema pull for a full local snapshot.
Error convention: Handle three levels: returned error fields, MCP tool errors (including parameter validation and service failures), and HTTP authentication failures such as 401.
Roles: Reads need an active account in the project’s organization. update_issue_status needs the editor or admin role. Chats are author-only over MCP: an org admin can read a colleague’s chat in the web app but not through the server, and only the chat’s author can rate it with submit_feedback.
Credential expiry: Expired or revoked credentials return HTTP 401. OAuth clients re-run the flow automatically; an API key has to be replaced.
Unreachable project: Querying a project the authenticated user cannot access returns an error rather than an empty result.

Code	Meaning
`0`	Success. For `check`, advisory warnings never change this. For `test`, every probe completed, whatever the outcome. For `dbt check`, with or without findings
`1`	The thing failed on its own terms: validation findings, failed eval cases (with `--repeat`, a case that failed every attempt), a duplicate eval case, a case id that does not exist, a failed or stale schema plan, a rejected dbt upload, or `verify` stopping at a failing step
`2`	Usage error: missing API key or project, no context directory, an unreadable file, a tree over the size limits, `context upload` or `schema push` run outside a git checkout or with uncommitted context files, or a dbt command without a `manifest.json` or with an artifact over the size limits
`3`	Transport or API error: unreachable API, invalid key, inaccessible project, another run already active, out of credits, a timeout, or Cassis failing while it processes a dbt upload (push again)
`130`	Ctrl-C during `eval run`. The run is cancelled server-side

Tree size: Commands that send the local tree accept up to 20,000 context files (YAML plus domains/**/README.md) and 100 MB total, sized for roughly 10,000 modeled tables. Beyond that the CLI fails fast with exit 2 before uploading anything. If you hit it, check --base-path. On cassis-cli older than 1.5.1 the client stops at the previous ceiling of 2,000 files and 5 MB, whatever the server accepts.
DDL size: cassis schema plan, apply and push send the file inline, and the API rejects a DDL over 10 MB. A whole-warehouse export above that is planned schema by schema, which the commands handle: a file speaks only for the schemas it contains. See Review schema updates.
Eval run timeout: 30 minutes by default, --timeout to change it. The run continues server-side if the CLI stops waiting.
Schema plan timeout: 30 minutes for schema plan, apply and push, --timeout to change it. The plan keeps computing server-side, and cassis schema apply --plan <id> resumes the wait.
Probe duration: context test is one full agent run per question, roughly 30 to 90 seconds each.

Failure is safe: On malformed YAML or a missing required field, Cassis keeps its previous context. Nothing is half-imported.
Idempotence: Re-uploading identical content with the CLI creates no new version. A GitHub App merge that touches the context path always creates one, recording the merged commit.
Complete replacement: Every import and publish replaces the complete context. This is why a project should have one editing path: see Choose how your context is managed.

Failure-by-failure fixes for the git loop are in Troubleshoot git and publishing.