Integrating Ruby with temper-rb

For integrators. How to call Temper from a Rails application: which credential a Puma request uses, which one a
Sidekiq job uses, why the two write under different names, and what happens between minting a
machine credential and your first successful write.

Audience: a Ruby developer wiring temper-rb into a Rails app — a web tier serving signed-in
users and a background tier doing unattended work.

Scope: the SDK's own behaviour — credentials, the token lifecycle, attribution, the error
taxonomy — plus the onboarding cliff you hit the first time a machine principal calls the API. The
operator's side of a machine credential (who may mint one, reach containment, revocation) is
standing-up-a-machine-credential; this guide points at it rather than restating it.

Every pattern below is pinned by a spec in clients/temper-rb/spec/, and CI runs that suite (rubocop, rspec, codegen-drift checks) on every change to the gem. Where a claim rests on
server behaviour rather than the gem's, this guide says so.

Install and configure

Ruby >= 3.1. There is no native extension — one source gem, no compiler on the install box.

The connection is process-global and memoized: one ApiClient, one Faraday connection, one
net-http-persistent pool. The token is per call, resolved from fiber-local storage at request
time. That split is the whole design — a fresh client per request would mean a fresh TLS handshake
per request, and a shared client with a shared token would leak one user's token into another
user's call.

A token never escapes its call. Temper.with_token restores the previous value on the way
out — including when the block raises — and a thread spawned inside a token scope sees no
token at all. Do not hand a Temper::Client to a thread and expect its credential to travel.

The two callers

Which credential you construct is determined by who is acting, and it is your explicit choice —
the gem never sniffs the environment to decide.

	Puma request	Sidekiq / ActiveJob worker
Who is acting	a signed-in human	the application itself
Class	`Temper::Credentials::BearerToken`	`Temper::Credentials::ClientCredentials`
Where the token comes from	the user's session — you already hold it	the SDK mints it, `client_credentials` grant
I/O at construction	none	none; the first mint is lazy, on first use
Can it refresh?	no — `refresh!` raises `Unauthorized`	yes — mints, caches, re-mints
Writes land as	the user's profile	the machine's agent profile

A Puma request — a token the caller already holds

A BearerToken does no I/O and cannot refresh — refresh! raises Unauthorized rather than
attempting a mint. That is deliberate, not an omission: a user's token is not the application's to
re-mint. On a 401 the exception propagates on the first request, with no retry.

A Sidekiq worker — a machine principal

Every field is required to be a non-empty String, and a missing one raises ArgumentError at
construction. The credential throws rather than defaulting — a machine that boots with a blank
secret and discovers it on its first write is strictly worse than one that refuses to boot.

audience is the one exception: it is optional, because it is Auth0's, not the protocol's. Absent
is a supported configuration; present-and-empty is still a caller bug and raises.

The two mint paths

Temper can be fronted by two different issuers, and ClientCredentials works against both. The
only thing that changes is your config.

	Auth0-issued (`temper admin machine provision`)	Temper-issued (`temper admin machine issue`)
Who holds the secret	your Auth0 tenant	Temper — it is the authorization server
`client_id`	the Auth0 M2M application's client id	minted by Temper, prefixed `tmpr_`
`client_secret`	from Auth0	printed once at `issue`; stored only as a hash
`token_url`	your Auth0 tenant's `/oauth/token`	your own instance's `/oauth/token`
`audience`	required — must equal the API's `AUTH_AUDIENCE`	omit it
Rotate the secret with	Auth0 (no Temper action needed)	`temper admin machine rotate-secret`

The request itself is identical on both paths, and that is the point: the SDK sends a
form-encoded body (application/x-www-form-urlencoded), which RFC 6749 §4 mandates at the token
endpoint. Auth0 also tolerates JSON; Temper's own authorization server reads the body with
req.formData() and does not. Form-encoding is the shape both accept.

Omit audience for a temper-issued credential — do not pass an empty string. Temper's AS mints
with its server-side AS_AUDIENCE and ignores a request-supplied audience entirely, so the SDK
leaves the parameter off the wire rather than sending a lie. Passing audience: '' raises at
construction.

The wire shape is pinned as a cross-language contract at tests/contracts/m2m-token-request.json,
asserted from both ends — the gem's spec proves it emits that shape, and temper-cloud's integration
test proves the AS accepts it. Neither test alone would catch a mismatch, which is exactly how the
gem once shipped a JSON mint against a form parser with both suites green.

The token lifecycle you get for free

ClientCredentials is not a thin wrapper over a POST. Three behaviours come with it, each of them a
bug that a hand-rolled minter tends to ship.

Behaviour	What it does	Why
Absolute-expiry cache	caches until 60s before the token's absolute expiry	a duration cannot survive being cached; only an absolute `expires_at` can
Mutex-guarded mint	many threads racing a cold cache mint once	under Puma every in-flight thread hits expiry together
Re-mint on 401	re-mints once and retries the call, then gives up	a job holding a token across a long unit of work outlives it

Refresh-ahead-of-expiry alone is not sufficient, and this is the subtle one. A worker resolves its
token at the top of a long unit of work; the token expires mid-job; the next call takes a 401 that no
amount of expiry-checking recovers. So the client repairs a 401 by re-minting once and replaying
the request — for reads and writes alike, because re-authenticating is not re-submitting. A second
401 raises.

A BearerToken takes no part in any of this: refresh! raises, so the repair path terminates
immediately and the Unauthorized propagates out of refresh! itself. One request, one exception.

Attribution across the enqueue boundary

A Puma request writes as the signed-in user. The Sidekiq job it enqueues writes as the machine
principal — a different profile, with different reach. So a single logical operation that spans the
enqueue boundary lands in the ledger under two different authors:

Where the write happens	Credential	Author
the Puma request	`BearerToken` (the user's token)	`dana@sdk`
the Sidekiq job it enqueued	`ClientCredentials` (the machine)	`acme-app@sdk`

That is honest, not a bug. The machine really did make the second write; the user was not in the
room. Attributing it to the user would be a lie the ledger cannot later distinguish from the truth.

What stitches the two together is a caller-supplied correlation id — a bare UUID that outlives any
credential. Mint it in the request, serialize it into the job arguments, stamp it again in the worker.

Temper::Act renames correlation: to the wire key correlation_id (and invocation: to
invocation_id) for you. The act's keys flatten into the write body — and onto the query string
for a delete, which takes its act context as query parameters rather than a body.

Correlation is provenance, never authorization. Nothing gates on it, and an act that supplies none
self-roots to its own event id — so you may always omit it. That is why Act accepts correlation:
with no confidence:, while it refuses reasoning:, rationale:, persona:, or model: without
one: those are authorship, the server's AgentAuthorship.confidence is non-optional, and the call
would earn a 400. The gem rejects it locally instead of paying the round trip.

The gem also stamps X-Temper-Surface: sdk on every request for you — once, on the client, not per
call. That is where the @sdk suffix on both author names comes from.

Errors, and what retries

Every exception is a Temper::Error carrying #status, #code, #message, and #details.

Temper::Error
├─ Temper::TransientError          # let it escape → Sidekiq retries
│  ├─ RateLimited      (429, #retry_after)
│  ├─ ServerError      (5xx)
│  └─ ConnectionError  (timeout / refused)
└─ Temper::PermanentError          # rescue it → dead-letter
   ├─ Unauthorized     (401)
   ├─ Forbidden        (403)
   │  └─ SystemAccessRequired      (403 with code SYSTEM_ACCESS_REQUIRED)
   ├─ NotFound         (404)
   ├─ Conflict         (409)
   └─ BadRequest       (400 / 422)

The split is load-bearing, not decorative. Sidekiq retries a job whose exception escapes, so a 409
classified transient would spin forever and a 503 classified permanent would be silently dropped.

Call	On a `TransientError`	On a `PermanentError`
Idempotent read (`show`, `list`, `search`, `whoami`)	retried, 3 attempts, backoff 200ms then 400ms	raised on the first attempt
Write (`create`, `update`, `delete`, `assert_relationship`, `set_facet`)	never auto-retried — raised	raised

Writes are never auto-retried, even on a 503. The SDK classifies; it does not decide to
re-submit. Retrying a write it cannot prove failed is how you get two postmortems. A 401 is the one
exception, and it is not a re-submission — re-minting a token and replaying the call is repairing
authentication, not retrying the operation.

Two classifications are not derivable from the HTTP status alone:

SystemAccessRequired discriminates off error.code == "SYSTEM_ACCESS_REQUIRED", not off the
403. A plain 403 maps to Forbidden. It means the caller authenticated but has no reach — the
machine profile exists and is registered, and is missing a grant.

ConnectionError is inferred from an ApiError with a nil status. The generated client
rescues Faraday timeouts and connection failures and re-raises them code-less; without that branch
every timeout would fall through to a bare Temper::Error and Sidekiq would dead-letter it.

A response that is not the server's {"error":{code,message,details}} envelope — an HTML 502 from a
proxy, say — degrades to a raw body on #details rather than raising inside the error path.

Going live

Authentication is not authorization. This is the wall every new machine caller hits, and it hits
it in two stages.

A valid M2M token — one your issuer happily minted, correctly signed, correctly audienced — gets you a
clean 401 until someone registers your client_id with the instance:

machine client 'k3Fq9xZ' is not registered with this instance.
An administrator must run: temper admin machine provision --client-id k3Fq9xZ --label <label>

Temper keeps a registration allowlist and the gate is lookup-or-reject — there is no first-call
auto-provisioning. The SDK classifies that 401 as a PermanentError, so a Sidekiq worker
dead-letters it rather than retrying forever against a wall.

Once registered, you clear authentication and meet the second wall: a 403
(SystemAccessRequired) on anything you have no reach for. Registration creates the agent profile;
it does not, by itself, grant it anything.

Register the client, and grant its reach, in one command

Both mint paths register in the same step that mints. Reach is plural and explicit — repeat
--team and --cogmap — and is never inferred from --owner-team, which records who owns the
machine and is never consulted for authorization.

Each --team <ref>[:role] gives the machine team membership (role defaults to member) for read
reach. Each --cogmap <ref>[:ro] applies a cogmap write grant on that map (:ro for read-only).

This is not necessarily an operator ticket. Minting is authorized by is_system_admin or
ownership of the team that will own the machine — so a team owner can register their own team's
machine with no admin in the loop, bounded to reach they could already confer on a human. A
teamless machine (no --owner-team) is admin-only, because the empty owning team fails closed.
The full rule, and what a team owner may and may not confer, is
standing-up-a-machine-credential.

Assert it at boot, not on the first write

An unregistered client surfaces as Unauthorized naming the client id. A
registered-but-under-granted one surfaces as Forbidden / SystemAccessRequired naming the
missing grant. Both come from the server's error.details — you get the diagnosis, not a bare 4xx.

Rotation — two different operations

These are not the same thing and conflating them will cost you your authorship history.

	`rotate-secret`	`rebind`
What changes	the secret — same `client_id`	a new `client_id`, same agent profile
For	a temper-issued (`tmpr_`) credential	rotating the external IdP application behind an Auth0 machine
Who may run it	admin, or owner of the machine's team	system admin only
Downtime	none — a grace window (default 24h) keeps the previous secret valid	none if you pass `--no-revoke-old` for an overlap
What you redeploy with	the new `TEMPER_M2M_CLIENT_SECRET`	a new `client_id` and secret

The grace window is what makes this a non-event for a running fleet: mint the new secret, redeploy the
app with it in the environment, and the old secret expires under you with nothing dropped. At most two
secrets are ever live at once.

Rotating an Auth0 secret needs no Temper action at all — the client_id is unchanged, so
authorship history is continuous and Temper never saw the secret. Only rotating the Auth0
application (a genuinely new client_id) needs rebind, which transplants the existing agent
profile's identity — and its full accumulated reach — onto the new id. That inherited reach can exceed
a team owner's own authority, which is why rebind alone keeps the system-admin bar.

Fork safety

The connection holds sockets. Clear the memo in every forked worker:

Strictly speaking the sockets are already safe: connection_pool (>= 2.4, which
net-http-persistent pools through) drops pooled connections from a Process._fork hook, and the
gem's suite measures it — a forked child that skips the reset still causes a fresh TCP accept on a
keep-alive server. The hooks above clear this gem's memoized client rather than relying on a
transitive dependency's default. Keep them.

Addressing, and what the SDK will not do

Resources are addressed by ref: a bare UUID, or the decorated sluggify(title)-<uuid> form.
Resolution is trailing-UUID-only, so a stale slug half is harmless and there is no by-slug lookup.

Temper.parse_ref is available directly and never touches the network. Unparseable input raises
before a request is made — no fuzzy matching, never a guess.

Bulk cogmap reconcile is deliberately absent. PUT /api/cognitive-maps/{id} takes a pre-embedded
desired-state manifest whose chunks_packed is a required, client-computed 768-dimension BGE
embedding carried verbatim, with no server-side fallback. Ruby has no embedder, so that path is
physically out of reach — it is a CLI operator's job, not a Rails request's. Incremental authoring
(cognitive_maps.author, assert_relationship, set_facet) is fully supported, and on every one of
those paths the server chunks and embeds for you.

Integrating Ruby with temper-rb

Install and configure#

The two callers#

A Puma request — a token the caller already holds#

A Sidekiq worker — a machine principal#

The two mint paths#

The token lifecycle you get for free#

Attribution across the enqueue boundary#

Errors, and what retries#

Going live#

Register the client, and grant its reach, in one command#

Assert it at boot, not on the first write#

Rotation — two different operations#

Fork safety#

Addressing, and what the SDK will not do#

See also#