Provider credential and account files

sac accounts list treats the provider as part of account identity. For example, claude-code:person-example-com and openai:person-example-com are distinct accounts even when their email-derived slugs match.

OpenAI Codex

Run sac accounts sync-openai to collect the active Codex login into ~/.scitex/agent-container/accounts/openai/<account-slug>/auth.json. The provider directory is part of storage identity, so openai:<account-slug> cannot collide with the legacy Claude store at accounts/<account-slug>/. The source login is read from $CODEX_HOME/auth.json, falling back to ~/.codex/auth.json. For ChatGPT login mode, SAC decodes display-only claims from the local ID token: email, display name, ChatGPT account ID, plan, organization, subscription dates, and last refresh time. This decoding is for status display only and is never used as an authorization decision.

The extractor returns an explicit allowlist. It never returns the ID, access, or refresh token, and never returns OPENAI_API_KEY. API-key login mode shows only that the mode is configured because no user identity claims are available.

OpenAI contributes every collected Codex identity to the combined account view. SCITEX_GENAI_CODEX_HOMES remains an explicit override. Rotation of OpenAI accounts is performed by the gateway; Claude Code OAuth rotation remains SAC-managed. The gateway always invokes its rotation selector, including when the candidate list contains exactly one account.

Claude Code harness with Codex subscriptions

SAC uses the nested spec.claude.provider backend axis for this mode. Do not set the top-level spec.provider: openai: that selects the OpenAI Agents SDK and replaces the Claude Code harness.

Start the scitex-genai gateway with its account homes and a local gateway key:

pip install 'scitex-agent-container[codex]'
sac accounts sync-openai
export SCITEX_GENAI_GATEWAY_API_KEY="$(openssl rand -hex 32)"
scitex-genai-gateway --host 127.0.0.1 --port 18765

Then declare the backend in an agent spec:

spec:
  # Omit top-level provider, or keep its default `anthropic` value.
  claude:
    provider: codex
    model: gpt-5.6-sol
    flags:
      - --dangerously-skip-permissions
      - --effort=medium

The gateway discovers the collected Codex homes, keeps sessions sticky, ranks accounts by available usage-window headroom, spreads concurrent sessions, and rotates away from temporary rate limits. Therefore Codex account files are configured on the gateway, not repeated in each agent’s claude.credentials_files. The gateway key authenticates only the local SAC to gateway hop; Codex OAuth tokens remain in each auth.json.

This document describes the on-disk files Claude Code manages for a user and which fields scitex-agent-container is allowed to read and surface.

Claude Code files

`~/.claude.json`

The main per-user settings JSON managed by Claude Code itself. Top-level keys relevant to agent orchestration:

oauthAccount (subdict): accountUuid, emailAddress, organizationUuid, organizationName, billingType, accountCreatedAt, subscriptionCreatedAt, hasExtraUsageEnabled, displayName, organizationRole.
hasAvailableSubscription (bool)
cachedExtraUsageDisabledReason (str, e.g. "out_of_credits")
overageCreditGrantCache (obj)
numStartups (int)
installMethod (str)
claudeCodeFirstTokenDate (str)
firstStartTime (str)
hasCompletedOnboarding (bool)
passesEligibilityCache (obj)
changelogLastFetched (str)
lastReleaseNotesSeen (str)
skillUsage (obj)

Any other keys (model caches, feature flags, editor state, MCP server definitions, per-project history) are considered opaque and MUST NOT be surfaced by our tooling.

`~/.claude/.credentials.json`

OAuth tokens for Claude.ai. Contains (inside a claudeAiOauth subdict): accessToken, refreshToken, expiresAt, scopes, subscriptionType, rateLimitTier.

RULE: this file MUST NEVER be read or emitted by scitex-agent-container tooling except for the non-secret strings subscriptionType and rateLimitTier. The extractor must not load, log, cache, or transmit any other field from this file. Tokens are the highest-sensitivity material on the host.

`~/.claude/settings.json`

Per-user Claude Code settings. Common keys: permissions, statusLine (command used to render the bottom bar, often claude-hud), enabledPlugins. Contains no secrets but may reveal which plugins / skills are enabled.

Fleet hosts

Each fleet host runs exactly one Claude Code OAuth identity shared by all tmux-managed agents on that host:

Host	Domain role	Credential home
MBA	scitex-orochi.com hub	`~/.claude/`
NAS	scitex.ai	`~/.claude/`
spartan	GPU worker	`~/.claude/`
ywata-note-win	Windows/WSL	`~/.claude/`

All tmux panes on a host inherit the same ~/.claude.json + ~/.credentials.json, so any head-agent view of “Claude account” is per-host, not per-agent.

What NOT to emit

The extraction layer MUST strip any field whose key or stringified value contains any of these substrings (case-insensitive):

accessToken
refreshToken
sk-ant-
Bearer (with trailing space)
secret
apiKey
claudeAiOauth

A post-extraction guard asserts the returned dict contains none of the above in either keys or values, and raises if violated.

Safe metadata fields (whitelist)

read_credentials_metadata() returns a flat dict with exactly these keys. Fields unavailable on disk are returned as None.

From ~/.claude.json oauthAccount:

account_uuid
email_address
organization_uuid
organization_name
billing_type
account_created_at
subscription_created_at
has_extra_usage_enabled
display_name
organization_role

From ~/.claude.json top level:

has_available_subscription
cached_extra_usage_disabled_reason
num_startups
install_method
claude_code_first_token_date
first_start_time
has_completed_onboarding

From ~/.claude/.credentials.json claudeAiOauth (only these two):

subscription_type
rate_limit_tier

From ~/.claude/settings.json:

status_line_command
enabled_plugins

Any addition to this list requires updating both this doc and the whitelist in src/scitex_agent_container/credentials.py.