Sort PR review comments

Pick-one per review comment, four ways.

Input

reviews.txt:

nit: rename x to count
This loop never terminates when the list is empty
LGTM, nice work
Could we reuse the retry helper here instead of a new one?
Looks good to me
This builds the SQL with string concatenation from user input: SQL injection

Call the local server

curl -s localhost:21118/v1/systemone -d '{"state": "nit: rename x to count", "questions": {"kind": {"type": "choice", "instructions": "What kind of review comment is this?", "criteria": {"must": "a real bug or risk that must be fixed", "should": "a reasonable change request", "nit": "style or naming only", "none": "praise or approval"}}}}'

Real answer for the first input (openjevx v0.4.0 8-bit model, CPU):

{"answers":{"kind":{"action":{"act_probability":1},"answer_confidence":0.5296,"choice":"nit","confidence":0.5296,"probabilities":{"must":0.0552,"nit":0.5296,"none":0.1049,"should":0.3103},"type":"choice"}},"model":"openjevx","usage":{"input_tokens":53,"output_tokens":0}}

Same thing with jevx

jevx ask --profile openjevx --no-context --lines reviews.txt --choice kind="What kind of review comment is this?|must=a real bug or risk that must be fixed;should=a reasonable change request;nit=style or naming only;none=praise or approval"

Real output:

VERDICT  P     INPUT
nit      0.62  nit: rename x to count
must     0.80  This loop never terminates when the list is empty
none     0.77  LGTM, nice work
unsure   0.40  Could we reuse the retry helper here instead of a new one?
unsure   0.59  Looks good to me
must     0.86  This builds the SQL with string concatenation from user input: SQL injection

The table output needs jevx v0.11.0 or newer.

What to do with the answer

Handle must first, reply to should, batch the nits.

How the local model did

Four of six are right: the nit (0.62), the endless loop (must, 0.80), LGTM (none, 0.77) and the SQL injection (must, 0.86). The retry-helper request (0.40) and "Looks good to me" (0.59) come out unsure.