Repository navigation
Description
User Story
I use OpenShell and build automation against its SDKs. While reviewing sdk/ I hit this behavior and reproduced it with the minimal test below against main.
Problem Statement
The Go SDK persists refresh_token but contains no refresh_token grant. oidc.readToken returns nil for an expired access token even when a valid refresh token is stored, so Login starts a full browser flow, and gateway.NewClient OIDC auth cannot renew at all. The CLI refreshes in this case (ensure_valid_oidc_token_bundle).
Impact / Why This Matters
Go programs and services must re-run an interactive browser login every time the access token expires (often hourly), which is unusable for unattended automation.
Acceptance Criteria
- With an expired access token and a valid refresh token on disk,
Loginreturns a refreshed token without user interaction and persists it. - OIDC auth created by
gateway.NewClientrenews via the refresh token (or the stored file) before expiry.
Reproduction Steps
Write {"access_token":"a","refresh_token":"still-good","expiry":"2020-01-01T00:00:00Z"} to oidc_token.json; readToken(dir) returns nil, nil, and Login proceeds to the interactive flow.
Suggested UX (if applicable)
No response
Environment
- OpenShell:
main@ 82977e2 (2026-10-13) - SDK: Go (
sdk/go/openshell/v1/oidc,gateway) - Toolchain: Go 1.25.4 linux/amd64, Node 22
Logs
BUG: refresh token "still-good" dropped; caller is forced into interactive login
Activity
rhuss commentedon Oct 14, 2026
Thanks for the report.
Confirmed on
main. The Go SDK implements theauthorization_code,device_codeandclient_credentialsgrants but norefresh_tokengrant.oidc.readTokentreats an expired access token as missing regardless of the stored refresh token, and the OIDC auth thatgateway.NewClientbuilds (RefreshableTokenover the disk token source) only re-readsoidc_token.json. It renews only if the CLI happens to refresh the file in the meantime.Adding the grant is straightforward. Deciding where the refreshed token goes is the harder part, because the refresh token in
oidc_token.jsonis one credential shared with the CLI (see the discussion on #4508). If the SDK refreshes and keeps the result elsewhere, providers that rotate refresh tokens on use leave the CLI holding a revoked token. That leaves two workable designs:oidc_token.jsonin the CLI's format, under a file lock, re-reading the file before refreshing (another process may already have done it) and replacing it atomically via a temp file and rename. This also fixes an existing race in the CLI itself:store_oidc_tokenwrites the file in place without locking, so two CLI processes refreshing at once can clobber each other.For a standalone
oidc.Loginwithout a gateway (WithInMemory), an in-memory refresh is fine either way. This depends on #4508 for theissuerandclient_idfields that the refresh request needs, and on #4513 for the scopes. Next step: maintainer acceptance decision, including which of the two designs to follow.rhuss commentedon Oct 14, 2026
One more data point that narrows the design question: the Python SDK already implements the shared-file approach (
python/openshell/sandbox.py, the OIDC refresher). Before refreshing, it re-readsoidc_token.jsonand adopts the disk bundle if itsexpires_atis later, so it never refreshes with a refresh token another process has already rotated. It writes the result back atomically. If the refresh still fails withinvalid_grant, it re-reads once and retries with a peer's newer refresh token. The Go SDK should mirror that behavior rather than introduce a third variant, and the CLI'sstore_oidc_tokenshould switch to an atomic write so concurrent readers never see a partial file.Related:
refreshableAuth.isTokenValid()treats a zeroExpiryas valid forever. Combined with #4508 (CLI-writtenexpires_atis ignored, so every loaded token has a zero expiry), the Go client caches the first access token for the life of the process and never re-reads the file. The fix for #4508 should include a test that an expired cached token actually triggers a re-read.shvma commentedon Nov 12, 2026
@rhuss I'd like to take this. Since the move to ai-openshell/sdk-go (#5461), the bug still reproduces there at
2900b06, currentmain.refresh_tokengrant, onlyauthorization_code(authcode.go#L183),device_code(device.go#L224) andclient_credentials(credentials.go#L127).Loginwith an expired access token and a stored refresh token goes straight to the interactive flow (oidc.go#L62-L72).gateway.NewClientwraps the disk token source inRefreshableToken(gateway.go#L173-L174), and that source returns an error for an expired token instead of renewing it (token.go#L127-L130).Two tests I wrote (below) fail on
2900b06. The rest of both packages passes. Output, trimmed:The second test calls the disk token source directly, so it assumes the refresh lives there. I'll move it if you'd rather put the refresh elsewhere.
TestLogin_ExpiredTokenTriggersFlowstoresRefreshToken: "old-refresh"and expects the interactive flow, but its mock token endpoint answers any grant with the same token. After the fix it would still pass while no longer covering the flow, so I'd drop the refresh token from that fixture.The reproduction steps in the issue body are out of date:
readTokenno longer returnsnilfor an expired token. Since #4830 the bundle also carries theissuerandclient_idthe refresh request needs.My plan:
expires_atis later; write the result with the same temp-file-and-rename write asoidc.writeToken; oninvalid_grant, re-read once and retry if a peer wrote a different refresh token. The refresh will send the gateway scopes, as the CLI'soidc_refresh_tokendoes. Python doesn't (bug(python): OIDC refresh drops the gateway's oidc_scopes #4919). The CLI bundle gainedclient_credentialsin fix(cli): renew client credentials OIDC tokens #4389 (oidc_token.rs#L36-L39). The GoBundledoesn't model it yet; I'll add it so a write-back keeps it, and skip the refresh when there's no refresh token (client-credentials bundles normally have none; the CLI renews those by repeating the grant).Part of NVIDIA/OpenShell#4514. fix(sdk/go): re-read edge token on every request ai-openshell/sdk-go#11 also touchesgateway/token.go, so I'll rebase on it if it lands first.store_oidc_token(still an in-placestd::fs::write), as a separate PR here withCloses #4514. I'll open it after the SDK PR merges, so the issue doesn't close with half the fix in.I read your second comment as replacing the file lock from the first design: the CLI and Python take no lock, so a Go-only lock wouldn't exclude them. I'll do the Python-style
invalid_grantrecovery without a lock unless you'd rather have one.Could someone with triage access assign this issue to me, so the CLI PR counts as accepted work?
Reproduction tests
openshell/v1/oidc/zz_repro_4514_test.goopenshell/v1/gateway/zz_repro_4514_test.gorhuss commentedon Nov 12, 2026
@shvma Thanks for rechecking this against the extracted Go SDK and providing the reproductions. The issue remains valid, and your proposed split makes sense: the Go SDK PR belongs in
ai-openshell/sdk-go, and the CLI atomic-write PR belongs here.Let's keep this issue as the tracking point for both and close it once both changes have merged. There's no need to recreate the issue in the SDK repository; please link the SDK PR back here as you proposed.
rhuss commentedon Nov 12, 2026
(unfortunately I don't have the karma to assign you the issue, but maybe @johntmyers @SDAChess @elezar could do ?)
test(oidc): reword refresh comments and assert no token on persist fa…
fix(oidc): honor context at the refresh lock, retry replace on Windows