logicsrc/packages/openerrand/README.md
Anthony Ettinger dde596b276
OpenErrand reference runner: @logicsrc/openerrand and logicsrc errand (#228)
* OpenErrand 0.1: an errand on a website with no API, with the human steps kept human

docs/openerrand.md mints OpenErrand: one JSON file per errand (register an
account, download a transcript) naming the site, the inputs with a
sensitivity class and ordered sources (document, vault, prompt, generate,
derive, candidate, literal), field rules matched by id then label, page and
wait steps, five human gates a runner never performs (declare,
identity-proofing, code, mail, captcha), outcomes, the never-retried shared
secret, vault and download outputs, hand-off cards that may name only public
inputs, the publisher index at /.well-known/openerrand.json, and thirteen
runner rules. The worked example is the MyFTB business registration that
cli-tools `ftb` performs (profullstack/cli-tools#125), with no personal data.

- @logicsrc/schemas: openerrand + openerrand-index schemas and fixtures
- @logicsrc/validators: semantic checks (references, templates, no personal
  or secret input on a card) and tests that validate the spec's own examples
- logicsrc-web: registry entry (process family), /openerrand landing page,
  the example and the index served as static files, contract tests

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* OpenErrand: hand-off cards stay on the surface that owns the data

Anthony's ruling: tax and finance data never touches a social or promotion
tool, and nothing is sent to a CPA or preparer.

- Hand-off cards are delivered only on the surface that owns the errand's
  data (for a tax or finance errand, the principal's finance app through its
  CLI, PWA, MCP server or API, such as CoinPay, or the runner's terminal),
  never a social, promotion or third-party posting service, and never to
  anyone but the principal. A card for an errand with personal or secret
  inputs does not leave that surface. Runner rule 9 says the same.
- The run record and the sample run name the card by an opaque id
  (pin-letter/7f3k2q) instead of a mynaposter.com URL; the myna mention is gone.
- `principal: represented` no longer cites a preparer with a power of attorney.
- The FTB card's last step no longer suggests sending the PIN to someone else.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* OpenErrand: user-agent rule, captcha solver policy, reference runner note

Anthony's answers on #227 ("go with your recommendations"):

- Rule 11: a runner may run headless with a normal desktop browser user agent
  (dropping HeadlessChrome) and nothing more: no fingerprint spoofing beyond
  the UA string, no stealth plugins, no solving or evading a bot challenge.
  A challenge the browser completes itself is a wait step; any other is a
  captcha gate.
- Captcha solvers: new site.sector and captcha step `solver`
  (forbidden by default | allowed). Never allowed on government, tax,
  financial, healthcare or identity-provider sites, nor on any errand with a
  declare or identity-proofing step or a secret input; elsewhere only when the
  file says so, with every use logged. The validator rejects `allowed` in the
  forbidden set or without a stated sector; six new tests. The FTB example
  states sector "tax".
- Reference runner: @logicsrc/openerrand / `logicsrc errand run`, marked in
  progress; ftb stays the runner the example was taken from.
- Name stays OpenErrand; family stays Agents and process.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* OpenErrand reference runner: @logicsrc/openerrand and logicsrc errand

Ship the runner docs/openerrand.md promised. `logicsrc errand run <file>`
reads an OpenErrand 0.1 file, validates it with @logicsrc/validators, and
drives headless Chrome through it under the spec's thirteen rules;
`errand validate` shows what a file will ask of you and `errand status`
shows the last run of each errand, its card and any lockout.

The engine is generalised from cli-tools `ftb` (PR #125) with no dependency
on cli-tools: the CDP client and Chrome finder from wcag.ts, the page reader,
native-setter fill and forward-button picker from ftb-run.ts, and the rule
matcher, throttle and outcome logic from ftb.ts, all now driven by the file.

- Inputs: document (an extractor hook; the one shipped runs a local command
  that reads JSON requests and prints records), vault (teams or OpenCreds),
  prompt (no echo for secrets), generate, derive, candidate, literal.
  `--input name=value` wins. Shared-secret candidates are ranked as the spec
  says, one is submitted, and a rejection lists the others for --candidate.
- Rules: id before label, step rules first, choices before text, an id match
  final, an unmatched required field stops the run naming it.
- Gates: declare only with --declare after the values are shown; identity
  proofing never touched (URL only) and handed over or stopped on; code from
  the terminal or a code file, used once, a wrong code waits for the next;
  mail ends the run waiting with the card; captcha is the person's, and a
  CaptchaSolver interface is called only where the spec permits (no solver is
  bundled); wait steps are polled, never solved.
- Throttle: 2 runs per errand and account in 30 minutes, 4 a day, 2 minutes
  between runs on a site, lockouts from metadata.lockout or a default, held
  per site and account, never lifted by --force.
- Outputs: credentials written before success to a teams vault by pull,
  merge, push (metadata.vault or --vault), else a 0600 file said aloud;
  downloads type-checked and never overwritten with different bytes; cards
  only in the local run record.
- The user agent is Chrome's own with HeadlessChrome replaced, given at
  launch: a CDP override did not reach a navigation the page's own script
  started, which is exactly the proof-of-work interstitial case.

Tests: 85 in the package (rule engine, inputs, gates, throttle, outcomes,
captcha gating, vaults, outputs, commands) including an integration test
that runs the published FTB example unchanged in real headless Chrome
against a local HTTPS fake site (Chrome maps webapp.ftb.ca.gov to it and
every other host to NOTFOUND; all data fictional), and 2 in the CLI.

CLI 0.6.0 -> 0.7.0; @logicsrc/schemas and @logicsrc/validators 0.3.0 ->
0.4.0 (the OpenErrand schemas, and the vocabularies now exported for
runners); PRD 0009; the spec's Reference runner section and the landing
page say it ships.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-04 09:17:06 -07:00

6.6 KiB

@logicsrc/openerrand

The reference runner for OpenErrand: one JSON file that describes an errand on a website with no API, run in headless Chrome, stopping at every step that belongs to a person.

It reads an errand file, validates it with @logicsrc/validators, resolves every input before the browser opens, then for each page finds its step, tests the outcomes, fills the form from the rules (id first, label second, choices before text), hands every gate to a person, and presses the forward button, until the site answers or the runner has to stop.

Install

curl -fsSL https://logicsrc.com/install.sh | sh
logicsrc errand --help

The commands live in this package and the umbrella CLI mounts them as logicsrc errand. To embed the runner:

npm install @logicsrc/openerrand

Commands

logicsrc errand validate ftb-register-business.json
logicsrc errand run ftb-register-business.json --extractor "python3 extract.py ~/taxes" --dry-run
logicsrc errand run ftb-register-business.json --extractor "python3 extract.py ~/taxes" --declare
logicsrc errand status
Flag What
--input name=value Give an input's value (repeatable). Wins over every source.
--dry-run Fill every page up to the first that holds a declaration or a shared secret, show it with secrets masked and the declaration word for word, and stop before its forward button. Not counted by the throttle.
--declare Your consent, for this run, to tick the declarations the errand names. Never read from a file, the environment or a saved default.
--headful A visible window, so you can take identity proofing or a captcha yourself.
--chrome PATH The Chrome or Chromium binary (default: CHROME_PATH, then PATH, then the Puppeteer and Playwright caches).
--vault TARGET teams:<team>/<project>/<env>, opencreds[:<field>] or file:<path>. Default: the file's metadata.vault, else a 0600 file.
--extractor CMD The local command that reads your documents (see below).
--candidate N Which shared-secret candidate to submit; a rejection lists the others.
--account NAME Which account at the site, for the throttle.
--force Lift the per-window, per-day and spacing caps. Never a lockout.
--yes Run a file whose SHA-256 changed since its last run without asking.
--json Print the run record on stdout.

Exit codes: 0 success, waiting or dry run; 1 rejected; 2 invalid file or usage; 3 stopped (a gate, an unmatched field, a loop, off-site, a timeout); 4 refused by the throttle.

Inputs

Source Served by
document --extractor: a command on this machine. It gets { "errand", "documents": [{ "input", "form", "field", "match" }] } on stdin and prints [{ "form", "field", "value", "year", "label", "file", "page" }]. The runner applies match, years, pick and transform, and keeps the values in memory only. No extractor is bundled.
vault The vault target, read-only here.
prompt The terminal; a secret does not echo. With no terminal the next source is tried.
generate crypto.randomInt, at least one character of each class, a letter first when letters are allowed.
derive, candidate, literal The file and the other inputs.

A shared-secret input's document sources become a ranked list of candidates (sources in order, newest year first within each). A run submits one; a rejection ends the run and lists the others for you to choose with --candidate. Nothing is ever retried on its own.

Gates

Gate What the runner does
declare Stops on the page and prints the statement word for word, unless you passed --declare; then it shows the values on the page (secrets masked) and ticks the box.
identity-proofing Never touches the provider's pages (it reads only the URL). Headless, it stops and tells you; with --headful at a terminal it waits until the page is back on the site.
code The terminal prompt, or the code file ~/.local/share/logicsrc/errand/codes/<name>.code written by whoever holds the phone. Used once, never logged. A wrong code waits for the next one.
mail Ends the run waiting, with the deadline and the hand-off card in the run record.
captcha Yours: a visible window with --headful, or the run stops. A program may pass a CaptchaSolver, and it is called only when the step says solver: allowed and the errand is outside the forbidden set (a stated sector that is not government, tax, financial, healthcare or identity-provider; no declaration or identity proofing; no secret input). Each use is logged with the time, URL and service. No solver ships with this package.
wait Polled until its match no longer fits, then the run goes on. Never solved or bypassed.

What it keeps

Under $LOGICSRC_ERRAND_HOME, else ~/.local/share/logicsrc/errand (files 0600, directories 0700):

Path What
throttle.json Attempts and lockouts. 2 runs of an errand per account in 30 minutes, 4 a day; 2 minutes between runs on one site; nothing during a lockout. A lockout is the file's metadata.lockout.text pattern (lasting metadata.lockout.duration) or a default pattern (35 minutes), and it holds every errand on that site and account.
approvals.json The SHA-256 of each errand file last run. A changed file is shown and not run without --yes or a yes at the terminal.
runs/ One run record per run: outcome, page, candidate (no value), card, deadline. Never an input value.
pages/<name>.jsonl The page log: each page's fields (selector, type, label, required, options) and a result page's text. Never a value.
profiles/<host>/ One Chrome profile per site, kept so a passed bot check stays passed.
codes/<name>.code Where a code may be written.
credentials/<name>.env Credentials when no writable vault is named.

Hand-off cards stay in the run record and errand status. They are never posted anywhere, and a card that names a personal or secret input is refused.

Browser

Chrome over the DevTools protocol, no dependency. The user agent is Chrome's own with HeadlessChrome replaced by Chrome, given at launch so it reaches every request; nothing else about the browser is changed.

Library

import { loadErrand, resolveInputs, runErrand, openCdpDriver, Store, fileVault } from "@logicsrc/openerrand";

runErrand takes a Driver factory, so an embedder can bring its own browser; the tests drive it with a scripted fake and one integration test runs the FTB example in real headless Chrome against a local fake site.

License

MIT