Read-back recipes

Attest reads back with the credentials the agent already holds — in-process, never through the cloud.

at = Attest(readers={"gmail": gmail_service_or_token, "slack": web_client_or_token, "hubspot": client_or_token},
            http_get=lambda url, params=None: session.get(url, params=params).json())

system	write	read-back	compares
gmail	send / reply	messages.get	SENT label, every intended recipient in To/Cc, subject, reply thread
gmail	create draft	drafts.get	exists, recipients, subject
gmail	update labels	messages.get	added ⊂ labels, removed ∩ labels = ∅
slack	send	conversations.history / .replies	ts, text, thread_ts
slack	create channel	conversations.info	exists, name, is_private
hubspot	create / update any object	GET crm/v3/objects/{type}/{id}	id, every intended property
calendar	create / update event	events.get	confirmed, summary, start / end, attendees
drive	file / share	files.get / permissions.list	name, parents, not trashed / email present, role
docs	create / append	documents.get	title, appended text present
sheets	create / write values	spreadsheets.get / values.get	title, every written row present
notion	page / database	pages.retrieve / databases.retrieve	properties (title, status, select, text…), parent
linear	issue / project / comment	GraphQL	title, priority, state, assignee, team, body
outlook	send / event	sent-items search / me/events/{id}	recipients, subject / start, attendees
teams	send	channel / chat message	text
any REST API	create / update	convention GET <url>/<id>	id, every intended field the record carries
any REST API with a spec	create / update	OpenApiDriver(spec, http_get)	spec-derived read path; never guesses
any MCP server	create / update	tool pair create_X ⇒ get_X	id, overlapping fields

Readers accept the vendor SDK object (duck-typed, no vendor import), a bearer token (stdlib HTTP), or any fetch(path, params) callable. Every reader also accepts a fetch(path, params) callable, so any HTTP client (or a test double) works.

Live verification against real accounts

tests/live proves the ladder against real Gmail, Slack, HubSpot, Notion and Linear — including the case that matters most: an action the API accepted whose record contradicts the intent must come back unverified.

export ATTEST_LIVE_GMAIL_TOKEN=…  ATTEST_LIVE_SLACK_TOKEN=…  ATTEST_LIVE_SLACK_CHANNEL=C…
pytest tests/live -q --live-report        # only what you have credentials for; prints the ledger it produced

Each system has a positive test (record matches ⇒ verified) and a contradiction test (claim ≠ record ⇒ unverified). Tokens stay in your shell: the suite reads back with the same token it writes with.