Recipes: from an actual task to one CLI call

简体中文

Choose a recipe, write the contract, and replace only the collector or input path. The agent should see the resulting packet—not the full input before filtering.

1. Custom command output

Create analysis.json:

{
  "mode": "filter",
  "context": {
    "scope": "Current request attempts only",
    "success": "Keep establishment failures that have not recovered",
    "exclusions": "Past failures followed by a successful current attempt"
  },
  "required_context": ["scope", "success", "exclusions"],
  "requirements": [
    {"id":"network","statement":"The current request fails during DNS, TCP or TLS establishment.","expected":true},
    {"id":"response","statement":"The current request received HTTP response headers.","expected":false}
  ],
  "fields": ["source_id", "text", "answers", "source"]
}

Your collector should emit records like this JSON array:

[
  {"id":"req-17","text":"TLS verification failed before any HTTP response.","request_id":"req-17"},
  {"id":"req-18","text":"An earlier timeout recovered. Current request returned HTTP 200.","request_id":"req-18"}
]

Then run:

jev-filter exec --analysis analysis.json \
  --task 'Find current network establishment failures' \
  -- your-collector --json

your-collector is the command you already use; it is not shipped by this package. For a runnable fixture, save the array as records.json and replace the command with cat records.json. Add --plan before -- to collect once without model inference. --plan still executes the collector. For arbitrary text output use self-contained paragraphs; split lines only when each line has enough context to stand alone.

2. Select with context

Reuse choose.json with candidates.json:

jev-filter query --input examples/candidates.json \
  --analysis examples/choose.json \
  --task 'Choose the JSON export belonging to project Beta'

The context distinguishes otherwise similar candidates. required_record_fields prevents selection when a competitor's project/format is missing. Supply only observed metadata; a field existing is not proof that its value is correct.

3. Search whole code symbols

jev-filter code-search 'retry|backoff' --root ./src \
  --task 'Find the implementation that retries transient network failures'

The lexical search narrows candidates; the collector expands hits to full functions or methods for semantic judgment. Python works out of the box; native npm bundles also include JS/TS/Go parsers. A lexical shortlist is not a complete semantic index. Use native rg directly if you already know the exact symbol.

4. Find a browser control

Use an existing local Camofox server and observed session/tab IDs:

jev-filter locate --session YOUR_SESSION --tab YOUR_TAB \
  --origin https://example.com \
  --task 'Select the enabled Continue control in the checkout form'

The locator only observes and verifies a candidate. It never navigates, clicks or submits. Keep the returned selector with its freshness information and reverify before action. Duplicate or changed targets require review. Domain-specific session restrictions remain the responsibility of your installed workflow.

5. Triage correlated events

jev-filter triage --input examples/events.jsonl \
  --analysis examples/triage.json \
  --task 'Find current network establishment failures'

Grouping preserves event order within a request. A historical timeout followed by a current success should not become a new unresolved incident. Malformed rows and incomplete evidence remain visible. A label does not establish live recovery.

Output customization

Use questions for typed Choice/Noul/Score decisions and output for a nested projection. With an intent Choice question, this returns your own field names:

{"output":{"record":"source_id","classification":"answers.intent.choice"}}

This is a fragment to combine with a complete question contract, not a standalone analysis file. See the context contract for required fields, filter/review rules and the JSON schema.