Skip to content

docs: retarget OIDC/OAuth after DID and local-first - #1326

Open
joepio wants to merge 13 commits into
developfrom
claude/pr-1310-review-oxr9xm
Open

docs: retarget OIDC/OAuth after DID and local-first#1326
joepio wants to merge 13 commits into
developfrom
claude/pr-1310-review-oxr9xm

Conversation

@joepio

@joepio joepio commented Sep 1, 2026

Copy link
Copy Markdown
Member

Identity is now did:ad:agent:{pubkey}; writes are client-signed.
Record that OIDC belongs on the control-plane session (a better
magic-link), not on atomic-server AUTH or commits. Connector OAuth
stays a separate importer problem.

Co-authored-by: joepmeindertsma joepmeindertsma@gmail.com

cursoragent and others added 13 commits September 1, 2026 20:37
Identity is now did:ad:agent:{pubkey}; writes are client-signed.
Record that OIDC belongs on the control-plane session (a better
magic-link), not on atomic-server AUTH or commits. Connector OAuth
stays a separate importer problem.

Co-authored-by: joepmeindertsma <joepmeindertsma@gmail.com>
SSO against the operator's IdP, connector refresh tokens, and a later
OAuth AS all belong on always-on atomic-server. That is local config,
not a control-plane client. Commits and resource AUTH stay Ed25519;
OIDC only gates the envelope index. HostMode enrollment is unchanged
until an explicit grant says otherwise.

Co-authored-by: joepmeindertsma <joepmeindertsma@gmail.com>
Drop closed-product names, billing, and control-plane strategy from the
OIDC/OAuth write-up. The feature is optional IdP config on this node.

Co-authored-by: joepmeindertsma <joepmeindertsma@gmail.com>
Authorization-code flow with the node as confidential RP. Two cookies:
OIDC session fetches the envelope; Ed25519 atomic_session is still
resource AUTH. First visit mints locally; a returning device unwraps
with passkey or recovery.

Co-authored-by: joepmeindertsma <joepmeindertsma@gmail.com>
Sign in with the IdP must not also demand a passkey or recovery code.
After a valid ID token the node unwraps with a node-held KEK and
releases the Agent secret. Commits stay client-signed; the operator of
this node is trusted with that identity.

Co-authored-by: joepmeindertsma <joepmeindertsma@gmail.com>
Record the threat table. Prefer a short-lived issued/session agent
behind the same button; GET /oidc/secret of the root Agent is the
implementable fallback and a key-exfiltration API.

Co-authored-by: joepmeindertsma <joepmeindertsma@gmail.com>
JWT-in-signature and node-as-signing-oracle fight forever-commits,
peer verify, and offline. The Atomic-shaped version is a short-lived
issued agent: same one-click OIDC UX, Ed25519 still on the wire.

Co-authored-by: joepmeindertsma <joepmeindertsma@gmail.com>
OIDC-only login: node holds the root, browser generates a short-lived
keypair, root signs a SessionCert. Commits stay Ed25519. write lists
still name the root. Iroh AUTH must treat cert-chain as the person.

Co-authored-by: joepmeindertsma <joepmeindertsma@gmail.com>
Authorization-code + PKCE against the operator IdP is unchanged.
A valid id_token mints a SessionCert instead of becoming a Bearer
header on /commit.

Co-authored-by: joepmeindertsma <joepmeindertsma@gmail.com>
Leaving Google is a Settings action with step-up OIDC, not the login
path. Unlink deletes the node's custodial wrap and must not mint a
second identity if they click the button again.

Co-authored-by: joepmeindertsma <joepmeindertsma@gmail.com>
SessionCert in validate_signature and AUTH is the hard part; the OIDC
dance is ordinary. Absent config, nothing changes. Record Google and
generic issuer setup for the operator, not the end user.

Co-authored-by: joepmeindertsma <joepmeindertsma@gmail.com>
Spell out the verifier path: session Ed25519, then cert, then rights
for the root via one effective_agent helper. Maps to validate_signature,
AUTH, check_write, admit, and genesis write-list insert.

Co-authored-by: joepmeindertsma <joepmeindertsma@gmail.com>
… develop

Rewrite planning/oidc-oauth.md at about half its length and fix the
issues found in review:

- The cert window is checked against createdAt, which the signer picks,
  and no path rejects an old timestamp (utils::check_timestamp_in_past
  only rejects future stamps; replica ingest skips validate_timestamp).
  A stolen session key was therefore unbounded on the commit path. Add a
  live-ingest `now <= notAfter + skew` bound, state that replay and
  catch-up stay exposed, and point at auth-checkpoint receipts as the
  real fix. Drop the "bounded credential" claim where it did not hold.
- Drop the claim that Iroh AUTH is byte-equal same-agent; sync::peer and
  sync::engine already gate on per-subject rights and the commit
  signature. Only own_agent_update_frame needs the root remap.
- GenesisCert::private_drive_subject signs with the root private key, so
  the node must mint the personal Drive on first OIDC visit and return
  its subject; the browser cannot derive it.
- State plainly that node disk + KEK theft is permanent identity theft
  for every OIDC user who never exported, since roots cannot rotate.
- Name a TTL (24h, ATOMIC_OIDC_SESSION_TTL) and explain why sign-at-drain
  makes it harmless for offline editing.
- Add "After login: what an OIDC user can and cannot do": sharing is by
  invite link in v1, deprovision lags to notAfter, identity is per node,
  Flutter later, no API bearer tokens, one IdP per node.
- Always inline the cert on commits; no hash-only variant.

User-facing docs (authentication.md, solid.md) now describe OIDC login as
planned, not shipped. Collapse the duplicated changelog entries into one
that says planning only.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Vjnz9exHsA8dKVi3aA9Jfj
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants