testing.md

Testing

← Documentation

pytest app/tests -v

1,192 tests, no network access and no credentials required. PyGithub is replaced with fakes, and Jira, Confluence and Slack with an httpx.MockTransport, all of which record which API calls were made — which is how the suite proves that ignored files are never downloaded, that the Jira issue cap shrinks the request rather than trimming the answer, that every Confluence page request carries the resolved space ID, that every Slack request names the one channel and that conversations.replies is never called at all, and that a token never reaches a response, a log or an error message.

Module	Covers
test_file_filter.py	every include/exclude rule, plus segment- and suffix-matching regressions
test_typescript_parser.py	each symbol kind, parent links, exact source spans, line ranges, TSX, fallbacks, syntax errors, the line-index regression
test_github_connector.py	branch resolution, filter-before-fetch, binary/UTF-8 skips, error mapping, token containment
test_ingestion_service.py	the real GitHub pipeline end to end with only the network faked
test_api.py	request validation, response projection, HTTP status mapping, token never echoed
test_jira_adf.py	every ADF node kind, nesting, unknown nodes, malformed shapes, recursion depth — all exact-string assertions
test_jira_parser.py	each field and each fallback, null tolerance, the one fatal case (no issue key)
test_jira_chunker.py	the chunk template, omitted lines, one-chunk-per-issue, an Epic never carrying a child's description
test_jira_connector.py	cloud-ID resolution and the gateway, the JQL, the six fields, pagination and its three loop guards, the cap, error mapping, credential containment
test_jira_ingestion_service.py	the real Jira pipeline end to end; linking, orphans, and the no-N+1 guarantee
test_jira_api.py	request validation, sampling vs truncation, HTTP status mapping, token never echoed
test_confluence_storage.py	every storage element kind, macros, nesting, noise removal, malformed markup, recursion depth — all exact-string assertions
test_confluence_parser.py	each field and each fallback, null tolerance, the one fatal case (no page id)
test_confluence_chunker.py	the chunk template, omitted lines, one-chunk-per-page, ids kept out of the text
test_confluence_connector.py	cloud-ID resolution and the gateway, the space lookup, the space-ID confinement, cursor pagination and its guards, the cap, error mapping, credential containment
test_confluence_ingestion_service.py	the real Confluence pipeline end to end with only the network faked
test_confluence_api.py	request validation, sampling vs truncation, HTTP status mapping, token never echoed
test_slack_connector.py	the channel confinement and the one-method guarantee, bearer auth, cursor pagination and its guards, the cap, ok: false mapping, the no-sleep rate limit, chronological ordering, token containment
test_slack_parser.py	every filter rule and its near-miss twin, the three escapes, markup left verbatim, reactions/files/attachments/blocks ignored, a filtered message recording no error
test_slack_chunker.py	the chunk is the message text alone, ids kept out of it, one-message-one-chunk
test_slack_ingestion_service.py	the real Slack pipeline end to end with only the network faked; the funnel gap, and the connector closed before parsing
test_slack_api.py	request validation, sampling vs truncation vs filtering, HTTP status mapping, token never echoed
test_ingestion_controller.py	the common endpoint: source resolution, the per-source config checks, what the ExternalDataSource is built from, and — running the pipeline directly — that permissions reach every item and chunk and that the token reaches neither the response nor the run file
test_source_retrievers.py	the four retrievers over a faked search service: each searches its own source, every SourceType has one, query/top_k/access pass through unchanged
test_knowledge_search_service.py	the query embedded once and that vector reaching the repository, a session per search closed on every path including failure, never committing, and the three ways of having nothing to search costing neither a call nor a connection
test_answer_generator.py	the last stage over a faked model: that nothing is reranked — every retrieved chunk shown, in plan order, and the returned sources being the ones its numbering points at — plus the three context-window trims (a duplicate chunk, an over-long chunk cut for the model but returned whole, the cap), the history window, and a model failure never carrying the vendor's words into the response
test_chunk_search_query.py	the vector query's two permission filters, compiled to real SQL with literal_binds: team source ownership, the access-scope disjunction read from resources and not the chunk's stale copy, another team's ids appearing nowhere, unembedded chunks skipped, cosine ranking and top_k — plus the row mapping, and that no vector, token or connector config is ever selected

The Jira, Confluence and Slack connector tests drive a real httpx.Client through a mock transport, so base-URL joining, the auth header and query-string encoding are all genuinely exercised rather than patched out.

Verifying against real sources

The automated suite never touches the network. Each connector page ends with a by-hand checklist for running it against a real account:

• GitHub — against a real repository
• Jira — against a real project
• Confluence — against a real space
• Slack — against a real channel