Home › CLI › Utilities › Flags

Common flags

A handful of flags appear on multiple commands. Listing them once here so you don't have to memorise them per command. Each flag below works on /run, /apply, /run-apply, /sync, and /profile unless otherwise noted.

Prerequisites

• AMX installed.
• An active DB and LLM profile (run /setup if not).

Flag reference

Flag	Type	Default	What it does
--db-profile <name>	string	active profile	Override the active DB profile for this command only
--llm-profile <name>	string	active profile	Override the active LLM profile for this command only
--scope <kind>	database/schema/table/column	inferred	Force the scope picker level instead of inferring from positional args
--apply	flag	off	Auto-/apply after /run finishes, using the review strategy you pick
--csv <path>	string	—	Write the run's review queue to a CSV file in addition to (or instead of) the in-memory store
--csv-only	flag	off	Skip the in-memory store; writes only to the --csv path
--debug	flag	off	Verbose logging — every SQL statement, every LLM call, every retry
--skip-network	flag	off	Skip the connectivity check at command start (useful when /doctor already ran)
--dry-run	flag	off	On /apply: print the SQL that would run; don't execute
--limit <n>	int	—	On /run: limit to the first N tables in scope. Useful for incremental runs
--filter <expr>	string	—	On /run: restrict to tables whose name matches the regex
--llm-batch-size <n>	int	profile default (10)	Override column_batch_size for one run
--n-alternatives <n>	int (1-5)	profile default	On /run: number of drafts to surface per asset
--verbosity <mode>	terse/balanced/verbose	balanced	On /run: target output length
--temperature <f>	float [0.0, 2.0]	profile default	Sampling temperature for this command only
--prompt-detail <mode>	low/auto/high	auto	How much profile/RAG context to feed the LLM

REPL-only commands

A few slash commands are not flags but they shape the same dimensions:

Command	Effect
/max-tokens	Show the active LLM profile's per-call output budget
/max-tokens <n>	Set the budget. Reasoning models (Claude extended-thinking, GPT-5 / o-series, DeepSeek-reasoner, Kimi K2.x) automatically get an additional 32 k reasoning floor on top
/temperature <f>	Show or set the active LLM profile's default temperature
/n-alternatives <n>	Show or set the default number of drafts
/profiling [full\|sampled\|metadata] [max_rows] [sample_size]	Configure the profiling mode and guardrails per profile (there is no --profiling-mode flag)

The review strategy (one-by-one, accept-all-high, accept-all, reject-all) is chosen interactively at the review step, not via a flag.

Common combinations

Sweep a whole warehouse cheaply

> /profiling metadata          # set the profile's mode first — no per-run flag
> /run \
  --llm-profile openai-mini \
  --csv ~/amx-output/sweep-2026-05-03.csv

metadata mode skips row scans entirely (no warehouse cost) and --llm-profile openai-mini uses the cheap model; the CSV gives you a portable record outside the AMX history store. At the review step, pick the accept-all-high strategy to skip ahead through high-confidence rows.

Re-run only the tables that failed last time

> /run \
  --filter '^(stg_|fct_|dim_)' \
  --limit 50 \
  --debug

The regex matches your dbt-style staging / fact / dimension tables; --limit 50 keeps the run bounded; --debug surfaces the exact failure if it happens again.

Apply high-confidence drafts in one shot

> /run-apply --apply

/run-apply is /run immediately followed by /apply. At the review step pick the accept-all-high strategy: high-confidence drafts are accepted and everything else is queued for later interactive review.

Be careful with --apply on a freshly-tuned LLM profile

--apply writes back whatever the chosen review strategy accepts. Until you've confirmed the LLM profile produces good descriptions for your schema, review one-by-one. Re-add --apply only after one or two successful manual sweeps.

Test profiling cost without LLM calls

> /profiling full              # set the profile's mode first — no per-run flag
> /profile sales.customer \
  --skip-network

/profile runs only the Profile agent — it samples the table and prints stats but doesn't call the LLM. Use it to estimate how long full-mode profiling would take on your warehouse before unleashing /run.

Sample config — defaults that act like flags

Most flags have a per-profile default that lives in ~/.amx/config.yml:

db_profiles:
  prod-pg:
    backend: postgresql
    # ...
    profiling_mode: sampled            # set via /profiling
    profiling_sample_size: 5000        # set via /profiling

llm_profiles:
  openai-prod:
    provider: openai
    model: gpt-4o
    column_batch_size: 10              # → --llm-batch-size default
    n_alternatives: 3
    temperature: 0.2
    logprob_high: 0.85
    logprob_medium: 0.50

Set sensible defaults per profile so you only need flags for one-off overrides.