mirror of
https://github.com/openai/codex-security.git
synced 2026-11-23 01:54:36 +08:00
* fix(publish): explain when Cloud import is unavailable `publish findings --to cloud` starts with the repository catalog. For a ChatGPT account or deployment without Codex Security Cloud import, that request returns 404 and the CLI printed only "Cloud import failed (HTTP 404). Not found", which does not say whether the repository, the file, or the account was missing. Keep the server's text and add the same explanation `publish scan` gives for this case: Cloud import is not available for this account or deployment. Other statuses and endpoints are unchanged. * fix(publish): handle an empty catalog error response * docs(imports): correct AWS Continuum and mktemp guidance - Name the first AWS CLI v2 release with `aws continuum` (2.38.5); earlier v2 releases reject the command. - Say that validation results come with findings Continuum detects, while findings submitted to Continuum can omit them. - macOS `mktemp -d` ignores TMPDIR, so the advice to set TMPDIR did not move report directories there. Suggest a template instead. * docs(imports): honor temporary report paths and clarify validation evidence * fix(imports): read Continuum weakness IDs from subCategory Findings that AWS Continuum detects itself omit cweIds and name their weakness only in subCategory, for example "CWE-89: Improper Neutralization ...". The reader mapped weakness_ids from cweIds alone, so every native Continuum code finding imported without a CWE. Fall back to the leading CWE ID in subCategory when cweIds is absent. Explicit cweIds still win, and the previous-release replay projection keeps reading cweIds only. * fix(imports): preserve Continuum replay and evidence admission * fix(cli): point native sign-in hints at codex-security login After starting browser sign-in, native Codex prints "On a remote or headless machine? Use `codex login --device-auth` instead." Running that command signs in to the regular Codex home, not the Codex Security credential home, so the suggested recovery does not help. For the sign-in commands that pass native output straight through (login, login status and logout), forward native stderr through a small streaming rewrite of `codex login` to `codex-security login`. Stdin and stdout stay attached to the terminal, other output is unchanged, and only a possible partial match is held back between chunks. * fix(cli): forward login hints and reap interrupted sign-ins * fix(auth): sign in with the credential storage set in Codex config On macOS, a fresh `codex-security login` could not be used to publish to Cloud. Scans write `cli_auth_credentials_store = "auto"` into the Codex Security credential home, so native login saved the token in the system keyring and removed auth.json, while Cloud publication reads file-backed credentials only. The documented fix, setting `cli_auth_credentials_store = "file"` in Codex config.toml and signing in again, had no effect: `login` used whatever the last scan wrote, and SDK scans ignored the user's Codex config for sign-in settings (only skill commands mirrored it). - Layer the user's Codex config sign-in settings (cli_auth_credentials_store, forced_login_method, forced_chatgpt_workspace_id) between the wrapper defaults and explicit overrides in mergedCodexConfig, so scans, their runtime config and Deep Scan workers use them. A Codex config that cannot be read or parsed sets no preference here; the commands that use it still report why. - Before `login` and `login status`, write the keys that config sets into the credential home. Native argv is unchanged, so managed policy still wins. Unset keys and `logout` leave the stored location alone. - Make the Cloud sign-in error and CLI docs name the setting and command. * fix(cli): exit interrupted login when stderr stops draining * fix(auth): keep sign-in, credential cleanup, and Cloud publication consistent * fix(sdk): cancel queued logins without aborting settled operations * test(sdk): guard the budget cancellation wait after usage is recorded --------- Co-authored-by: mldangelo <michael.l.dangelo@gmail.com>