ingest: unknown --kind is silently stored as a memory

[Vaibhavtripathi7]

CLI Version

hydradb-cli 0.2.0 (installed from source, main at d7f263d)

Python Version

Python 3.14.4

Operating System

EndeavourOS (Arch Linux), kernel 6.18

Bug Description

hydradb ingest does not validate --kind. The only value it checks for is knowledge; anything else falls through to the memory path. So a misspelled or wrongly-cased value (--kind knowlegde, --kind Knowledge) stores the text as a memory, prints ✓ Memory added, and exits 0.

Expected: the same rejection every other command gives. query, list, delete, relations and subgraph all fail with --kind must be one of: knowledge, memory for the same input. The README also documents --kind for ingest as exactly memory (default) or knowledge.

The cause is the branch in src/hydradb_cli/commands/canonical.py, ingest():

if kind == "knowledge":
    _impl.do_ingest_knowledge_text(...)
    return
_impl.do_ingest_memory(...)

Nothing before it checks kind against _impl.VALID_KINDS, and do_ingest_memory hard-codes kind="memory", so the bad value is never seen again.

Steps to Reproduce

1. Have an API key and a default database configured (hydradb login --database <db>).
2. Run hydradb ingest --kind knowlegde --text "quarterly notes".
3. Observe ✓ Memory added, exit code 0. The request sent to /context/ingest carries type=memory and a memories field, so the text is filed as a memory, not knowledge.
4. Compare: hydradb query "x" --kind knowlegde exits 1 with --kind must be one of: knowledge, memory. Got 'knowlegde'.

CLI Output

$ hydradb ingest --kind knowledge --text "quarterly notes"
  ✓ Knowledge source uploaded to database db
  "quarterly notes"
  Source ID: src_1
$ echo $?
0

$ hydradb ingest --kind knowlegde --text "quarterly notes"
  ✓ Memory added (1 success, 0 failed)
  "quarterly notes"
  Source ID: src_1 (ok)
$ echo $?
0

$ hydradb query "x" --kind knowlegde
  ✗ Error: --kind must be one of: knowledge, memory. Got 'knowlegde'.
$ echo $?
1

(Captured with HYDRADB_BASE_URL pointed at a local stub of the ingest endpoint so the request body could be inspected; the src_1 id is the stub's. The CLI's branching is identical against the real API.)

Additional Context

The deprecated aliases (memories add, knowledge upload-text) are not affected; they pass a fixed kind into _impl directly.

Since this is a silent success rather than an error, a script that builds the --kind value and gets it slightly wrong will keep filing knowledge into the memory corpus with no signal until a later query --kind knowledge comes back empty.

I have a fix ready (one guard in ingest() reusing _impl.VALID_KINDS, plus tests) and will open a PR referencing this issue.