Ask

jevx ask is the full command. The shortcuts (is, pick, filter, rank) are ask with one question already shaped.

$ jevx ask -h
Usage of ask:
  -choice value
      NAME="QUESTION|key=desc;key2=desc" (repeat)
  -context value
      background for this call: TEXT or @file (repeat); added after the ## Jev sections
  -in string
      one input: text, a JSON object, @file or - (default: stdin)
  -json
      one input: print JSON instead of lines
  -lines string
      batch: a text file (or -), one input per line
  -min float
      override: choice / score confidence below is unsure (default -1)
  -no float
      override: noul P at or below is no (default -1)
  -no-context
      skip the ## Jev sections of the global and folder agent files
  -noul value
      NAME="QUESTION" (repeat); or NAME="QUESTION|true means|false means"
  -parallel int
      override: concurrent requests in a batch
  -profile string
      endpoint profile
  -questions string
      a question-set file: {NAME: {type, instructions, criteria}}
  -raw
      one input: print the server's full response
  -score value
      NAME="QUESTION|low;mid;high" (repeat, lowest first)
  -states string
      batch: a JSONL file (or -), one JSON object or string per line
  -yes float
      override: noul P at or above is yes (default -1)

Questions

• Saved names: jevx ask urgent,team < ticket.txt (see Saved questions).
• Inline yes/no: --noul NAME="QUESTION". A second form spells out the poles: NAME="QUESTION|true means|false means".
• Inline pick-one: --choice NAME="QUESTION|key=desc;key2=desc". The verdict is one key.
• Inline rating: --score NAME="QUESTION|low;mid;high", lowest first. The verdict is one level.
• A file: --questions set.json, a JSON object {NAME: {type, instructions, criteria}}.

Any number of questions go in one call and one request (up to chunk, default 32, per request).

$ jevx ask --noul urgent="Is this urgent for the person receiving it?" \
    --choice team="Which team should handle this?|web=frontend or UI;api=backend;billing=money" \
    --in "Checkout is down, customers are being charged twice"
team             billing    0.72
urgent           yes        0.95

Output lines are sorted by question name, not by the order you gave them.

Input

• stdin (the default), or --in "text", --in @file, --in '{"json": "object"}', --in -.
• --lines FILE (or -): a batch, one input per line of text.
• --states FILE.jsonl (or -): a batch, one JSON object or string per line. A JSON input keeps its shape; context is merged into its "context" field.

Output for one input

One line per question: NAME VERDICT P. A yes/no verdict is yes, no or unsure (P at or above yes 0.8, at or below no 0.2, or between). A choice or score is the key or level, or unsure when its confidence is below min_confidence 0.6.

--json prints the same as one JSON object; --raw prints the server’s reply untouched.

Terminal

jevx ask --json --noul spam="Is this message spam?" --in "Congratulations, you have won a prize, click here"

output

{"answers":{"spam":{"answer":{"type":"noul","noul":0.95},"verdict":"yes"}},"model":"jev-1.13.0","profile":"jev"}

Output for a batch

JSONL in input order, one object per input with line, input and answers. Each answer carries verdict and, for yes/no, p.

Terminal

jevx ask --lines app.log --noul err="Is this line an error or failure?"

output

{"line":1,"input":"INFO  09:12:01 request served in 12ms","answers":{"err":{"answer":{"type":"noul","noul":0.02},"p":0.02,"verdict":"no"}}}
{"line":2,"input":"ERROR 09:12:04 payment gateway timeout after 30s","answers":{"err":{"answer":{"type":"noul","noul":0.98},"p":0.98,"verdict":"yes"}}}
{"line":3,"input":"INFO  09:12:05 cache warmed","answers":{"err":{"answer":{"type":"noul","noul":0.02},"p":0.02,"verdict":"no"}}}

Filter it with jq:

Terminal

jevx ask --lines app.log --noul err="Is this line an error or failure?" \
  | jq -c 'select(.answers.err.verdict=="yes") | .input'

Batches run parallel requests at a time (default 8; --parallel N for one call). A batch exits 0, or 4 if any input failed; that line carries an "error" field instead of answers.

Overrides for one call

--yes, --no, --min, --parallel, --profile, --context TEXT|@file (repeatable), --no-context. Every command also takes --cwd DIR to run as if from that directory (it picks up that folder’s .jevx/ and ## Jev section).

Precedence: built-in < config < profile < the question’s own thresholds < flags. See Settings.

judge: before handing a turn back

judge asks four fixed questions about a proposed final message: would the user accept it, did the agent stop short, the likely reaction, and a 0-4 satisfaction. The questions come from a question pack (JSON); the built-in one is neutral, and a profile can point at its own with --questions pack.json.

$ jevx judge -h
Usage of judge:
  -action value
      one tool call as a short line (repeat)
  -context value
      background for this call: TEXT or @file (repeat)
  -json
      raw answers
  -no-context
      skip the ## Jev sections of the global and folder agent files
  -profile string
      endpoint profile
  -proposal string
      the agent's final message
  -request string
      the user's request

$ jevx judge --request "Fix the failing test in parser_test.go" \
    --proposal "I found the off-by-one in parseRange and fixed it; go test ./... passes." \
    --action "Edit: core/parser.go" --action "Bash: go test ./..."
accept        0.60
wanted more   0.73
reaction      accept (0.93, confidence 0.91)
satisfaction  2.9 / 4  (jev)

The skill tells the agent: below accept_min 0.35 the user would likely push back, so verify and show evidence; above more_max 0.65 on “wanted more”, finish the missing part. The call above scored 0.73 on “wanted more” for a one-line proposal with no diff shown, which is the intended nudge.

Caution

judge is a heuristic from one model. There is no published accuracy number for it.

Writing a question that works

• Say whose decision it is and what counts as yes: “Would a senior on-call engineer page someone for this line?” beats “bad?”.
• The model sees only the question, the input and the context; nothing from your conversation.
• Put the facts that decide it in context, not in the question.
• Look at a few verdicts before you loop over 5,000 lines. Tune thresholds per question if the default band is wrong for it.