Factory docs, home
Page navigation

See the input → responsibility → output diagram.

Independent adoption draft: Discovery and Service adoption exposes the existing Access, discovery and catalogue contracts as one Service-access family. Caller-owned trees, names-only captures and optional catalogue injection work without Portfolio or a Git checkout. Count/CoverageStatus now come from shared reporting types; runtime, authority exclusions and recorded protocols are unchanged. Current acceptance and review evidence: build review.

What it does for you: It keeps a map of the outside services each Product uses (such as Vercel, Clerk, Sentry, Cloudflare, App Store Connect or Google Play) and of where every credential for them lives: which GitHub secret or variable, which workflow job uses it, which template names it, and which provider it is passed on to. It is built to record names, places, dates and whether something is present, never a secret itself; its tests prove that for the kinds of file and value they plant, not for every file anyone could write. From that map it answers one question for other parts of the Factory: can this need (a service, an environment, a purpose such as "publish a build") be met now, and by which route? The answer is ready, missing, unknown or refused. It says missing only when it has seen a complete list, so it never asks you to register something that may already exist, and it flags things worth fixing, such as a public value stored as a secret or a secret passed on a command line, as proposals.

A slate-blue bob carrying two distinct keys.

Meet the custodian. Know where access belongs. Maps services and credential references by name, never by secret value. Meet the whole Plumb family.

How it links: It stores everything in the Command journal and shares the same counting contract as Portfolio, where only an exact count shows zero. Each repository or 1Password vault belongs to exactly one Product, so one Product's services and credentials are never visible from another's. Nothing uses it yet; later steps will read your repositories' secret names, prepare registration steps for you to carry out yourself, and check afterwards that the credential arrived.

Honest limits:

  • Nothing reads a live source yet. Everything is proven on made-up test files shaped like Tally; the real Tally map comes in the next step and may differ.
  • It can now check which items exist in a Product's 1Password vault (step 4, independently reviewed and integrated, not yet accepted). It runs one listing command only, and every command that could show a secret is refused before anything runs. It keeps each item's id, version, last-edit date and title, the title only when it follows the naming of the Factory's item templates. A listing never shows what fields an item holds, so an item can count as there but is never treated as ready to use. An item counts as gone only when a full, recent listing of its vault lacks it; a listing older than a day leaves the answer unknown. When you are signed out, the app is locked, or the output is blank or cannot be read (an impossible date included), the answer stays unknown. It has been tested only against a stand-in for 1Password's command, never the real one, and nothing runs it yet.
  • It sees names, never values, so it cannot tell whether a credential works or when it expires. Other branches, older history, values held inside Vercel or Convex, your Keychain and your local files stay unknown.
  • A file path, secret name or other name that looks like a key is never recorded, and the scan is marked incomplete instead. This is a check on shapes, not a secret scanner: a secret that looks like an ordinary name cannot be told apart, and an unusually long name with digits in it is skipped too.
  • Where a workflow or settings template is written in a way that could be read more than one way (a value spread over several lines, an unusually long line, YAML shortcuts such as anchors), it stops reading or marks the scan incomplete rather than guess, so text inside a value is never mistaken for a name.
  • A name absent from repository and environment listings stays unknown for an organisation-owned repository until captured organisation listings can prove absence. Missing owner type also leaves absence unproven; a known user-owned repository does not need organisation evidence.
  • A name is only called missing when every list it could be in was read in full and every place that uses it shows which list that is. A job whose environment it cannot read, or a workflow that other workflows call (whose credentials come from the caller), leaves the answer unknown instead, and such a called workflow never counts as a ready route on its own.
  • A Product with no secrets yet shows "unknown" rather than "missing", because an empty list alone never proves that nothing exists; a later step will let you confirm it.
  • Each Product can have one repository for now. A second is refused rather than overwriting the first repository's map.
  • It spots services from a reviewed list of thirteen kinds. Cloudflare's steps so far are independently reviewed but not yet accepted; Datadog is the newest, and Sentry and PostHog have new revisions, still awaiting independent review. For Cloudflare it reads only names. It opens Wrangler's settings files (wrangler.toml, wrangler.json and wrangler.jsonc) with a careful reader of its own, which keeps the names of environments, Workers, web addresses and connections and never a value, comment or account ID; a file it cannot read whole is marked as partly read. It never opens Cloudflare's cloudflare.config.ts, Wrangler's local secrets (.dev.vars) or Wrangler's working folder (.wrangler/). It flags a secret's name among Wrangler's plain variables, and the account-wide Global API Key wherever it is found, declared or confirmed, since a narrower token should replace it. Anything else shows as unclassified until a new kind is reviewed, and a name the list gets wrong is corrected in the list, not per Product. Its findings are suggestions, not verdicts.
  • For reading Sentry, PostHog and Datadog health signals later, it now knows the read-only credentials each needs and where each one lives: only in the Product's own 1Password vault, never in GitHub. It also lists the steps you take to create them, in the order the Factory prefers its tools: the vendor's own agentic tool, then its API, then its command-line tool, and its website last. For these credentials every step that makes one is on the website, since the other tools would print the new key where an agent could see it. Nothing reads with them yet, and none exists. Register signal read credentials lists the steps.
  • Registering for a service, accepting terms, creating keys and anything that costs money stay yours: the Factory will only prepare the steps and check the result. When the answer is that only you can do something (at the service's own website), it says so, so no part of the Factory tries to act on it.
  • Its records are attribution: anyone who can write the journal could forge them, and nothing here proves who you are.

Status: Accepted local implementation: independent reviews and combined checks passed on fixtures; required CI remains pending. Acceptance and evidence.

Contract and limits

Shared value implementation: common contracts. Owning errors, limits, stored identities and authority remain as specified above.

Source: docs/guide/service-access.md