Commands

Run

leakdown <url>

A plain run, with menus for the brain, model and who visits.

leakdown

No arguments: a guided flow for everything.

--ladder

The full run: cheap models visit every persona, sonnet verifies the sessions that agree, opus re-walks the hardest one and writes the report. About an hour.

--wide "<spec>"

With --ladder, splits the sweep, e.g. "haiku:5,opencode/<model>:5".

--stop <stage>

End after one stage: site, map, personas, visit, report or fix.

--flow "<intent>"

Drafts checkpoints for a journey and scores each session against them (FLOW.md).

--flow-file <id|path>

Runs against a flow file you wrote.

--validate-flow [file]

Checks flow files without a browser and lists what is wrong. Exits 1 on errors.

--variant <slug>

Labels a run for an A/B test, e.g. control or new-pricing.

--compare <site>

The newest run of each variant side by side, in COMPARE.md.

--plan

Re-reads the site and rebuilds its personas.

--no-map

Skips the re-crawl that checks whether the site changed.

--force

Regenerates outputs that are already up to date.

Who visits

--persona <list>

An explicit queue, e.g. cold,warm,hot. Max 10.

--random <n>

n prospects chosen at random. Max 10.

--count <n>

How many prospects the personas stage builds, 2 to 10. Default 10.

Goal tests

--goal "<text>"

Every persona gets this goal. Exit 0 only if every session completes it.

--steps <n>

Step cap per session, 1 to 50. Default: each persona's own patience.

--expect "label=value"

Repeatable. The page must show the value before a completion counts.

How it runs

--brain <claude|opencode|codex>

Which AI CLI plays the prospect.

--model <name>

Pin a model. Lists are read live from the CLI.

--effort <level>

Reasoning effort. claude: low to max. codex: low, medium, high.

--time <minutes>

Wall-clock ceiling per session. Default 20. Waiting on mail and pauses does not count.

--headless

No visible browser window.

--serial / --parallel

One persona at a time (default) or all at once. Parallel needs a browser and AI CLI per persona.

--mobile

Phone viewport, 390 × 844 with touch.

--yes

Never prompt; take the default for every question.

--version

Print the version. Put it in bug reports.

-h, --help

The flag reference.

On its own

--history [site]

Every run, one line each: date, the one number, who sat in which seat.

--report [what…]

Rebuild the reports for those sessions.

--fix [what…]

Run the expert panel over those sessions (FIXES.md).

--replication [what…]

Element refs cited across sessions: replicated versus single-source.

--pdf [sites…]

One shareable PDF per site: the funnel and every fix.

--doctor

Checks Node, Chromium, your AI CLI, mail and ffmpeg.

--list-personas

Every persona: built-in, yours and per site.

--new-persona "Name"

Build one persona by answering a few questions.

--mailtest

Tests mailbox create, receive, extract and destroy.

Website orders

--orders [--all]

Run requests left on the website: new ones, or every status.

--order <id>

Run one order here and email the PDF.

--order <id> --reject "why"

Decline an order.