feat(providers): add shared ChatGPT accounts for Codex

[aredddd]

Summary

Codex Agents can use a shared ChatGPT account registered on their Provider. The Provider page supports multiple accounts; the Agent editor can select an existing account or connect another before saving its binding. API-key and server-local CLI sessions keep their existing behavior.

Sign-in runs through the user's local a2wave CLI and a loopback callback, so it works when the platform is hosted on a different computer. The server owns PKCE and token exchange, verifies identity, encrypts credentials, and refreshes them with a lease and revision check. Authenticated teammates may bind accounts; only the creator or an administrator may manage them. Removing an account requires clearing its Agent bindings and revoking its upstream session.

Subscription execution uses Codex app-server with the selected account. Sessions are scoped to that account. A structured authentication failure permits one refresh and continuation of the saved thread within the original deadline; failed recovery cannot replay the original task through generic fallback. Cancellation also covers credential preparation before a CLI process exists.

Includes SQLite and PostgreSQL migrations, API/CLI/OpenAPI support, account-specific model discovery, import/export/clone handling, audit labels, and English/Chinese manuals.

Type of change

• New feature (non-breaking change that adds functionality)

Testing checklist

• Lint: 0 errors, 739 warnings (baseline: 740); explicit repository paths supplied because the worktree's parent directory is excluded by Biome's root glob.
• pnpm typecheck passes for every workspace, including tests; rerun by the pre-push hook on the final commit.
• pnpm test: 10,728 passing tests including repository scripts, with 2 existing skipped API tests. Workspaces ran sequentially with VITEST_MAX_WORKERS=4 after an unconstrained concurrent run hit three import-hook timeouts; no test timeouts or assertions were relaxed.
• E2E updated for Provider login and Agent selection/save; both new cases pass.
• User-facing docs, both locales, and the in-app manuals updated.

Additional validation:

• Real ChatGPT consent completed through the built CLI. Model discovery returned five account-authorized models. Two actual Agent turns using the locked Codex 0.154.0 binary and gpt-5.6-luna executed a terminal tool and retained conversation context. After the final session-isolation fix, a second two-turn acceptance run confirmed the account-scoped session, completed tool records, and persisted usage.
• Isolated HTTP + SQLite binding checks passed 14/14: top-level and chain bindings, updates, unbinding, cloning/export/import, and referenced-account deletion.
• Full Playwright run: 189 passed, 16 skipped, 8 failed in pre-existing Runs/session-drawer tests. All eight reproduced at the same assertions in a separate archive of unchanged main (27bf128), with a new database and fake CLIs. They expect the previous Runs endpoint/drawer navigation. The failures are reported rather than waived or hidden.
• PostgreSQL schema regeneration produced identical output; secret scanning, architecture, file-size, and Docker-context gates pass.
• Coverage suites pass: API 87.99% lines (7,622 passing tests), CLI 88.44%, Web 59.54%. The final API coverage run used four workers.
• pnpm build passes; restart-recovery was rerun on the final commit and passes all 4 scenarios.
• The final commit's GitHub CI passes lint, typecheck, all four API shards, CLI/Web tests, PostgreSQL schema generation, SCM storage integration, secret scanning, and license inventory.

AI assistance disclosure

This implementation was substantially AI-generated and reviewed with parallel agents. The checks above distinguish real account execution from fixtures and protocol tests; the maintainer should review the change before merging.

Additional notes

The aggregate CI result remains red for two reasons: the migrations require the repository's ci-reviewed maintainer label, and dependency auditing reports five high-severity findings in existing locked undici (7.29.0/6.28.0) and brace-expansion (5.0.9). This PR changes no dependency manifests or lockfile; those findings are also reproducible locally.

The account-specific PostgreSQL concurrency paths and real upstream refresh/revocation were not exercised against live services; protocol, locking, rotation, cancellation, and recovery behavior are covered by automated tests. Runtime authentication recovery is limited to one continuation attempt per turn. No production deployment is included.