fieldflow: Skill for Codex

FieldFlow CLI

Use this skill to keep large JSON CLI output out of model context.

Qualify The Command

Use fieldflow-cli only when all of these are true:

• The command is read-only.
• The command is external or service-facing, not a tiny local shell command.
• The command can emit JSON on stdout.
• The expected output is likely large enough that raw output would pollute context.

Do not use this skill for commands like pwd, date, ls, git status, rg, or any mutating command such as deploy, apply, delete, or create.

Inspect First

Run fieldflow-cli inspect before choosing selectors unless you already have a manifest for the exact same wrapped command.

fieldflow-cli inspect --sample-items 100 -- <wrapped command>

The inspect step writes a compact field catalog under .fieldflow/inspect/ and prints the manifest to stdout. Treat that manifest as the source of truth for valid selectors.

The manifest is deterministic and intentionally small:

• path
• types

It does not store raw command output.

Pick Minimal Fields

Choose the smallest field set that answers the user’s question.

Prefer fields like:

• timestamps
• severity or status
• identifiers or names
• URLs
• concise message fields
• latency, count, or state fields

Avoid broad selectors such as [] or whole nested objects unless the task truly needs them.

Run The Reduced Command

After choosing selectors, rerun the command through fieldflow-cli.

fieldflow-cli \
  --field "[].timestamp" \
  --field "[].severity" \
  --field "[].jsonPayload.message" \
  -- \
  <wrapped command>

If the result is too narrow, broaden the selectors and rerun the reduced call. Do not fall back to raw output unless the user explicitly asks for it.

JSON Output Rules

Prefer the CLI’s native JSON mode:

• gcloud: --format=json
• kubectl: -o json
• gh: --json ...
• aws: JSON is already standard, or use --output json when needed

If the command cannot emit JSON, do not use this skill.