opencode-codex-meter handles OAuth credentials and usage data. This document describes what sensitive data the plugin touches and how it protects it.

• auth.json — the OpenCode credential file at ~/.local/share/opencode/auth.json (or $XDG_DATA_HOME/opencode/auth.json). The plugin reads only the OpenAI OAuth entry's access, expires, and accountId fields. It never reads, stores, logs, or returns the refresh token.

• Session messages — the plugin reads assistant message token counts (input, output, reasoning, cache.read, cache.write) via the OpenCode SDK's session.messages() API. It does not read message text, tool inputs/outputs, or file contents.

• Quota response — the plugin fetches usage data from the unsupported https://chatgpt.com/backend-api/wham/usage endpoint. The response contains usage percentages and reset times, not credentials.

• Never writes to auth.json — the plugin has no auth-write capability. The AuthReader only reads.
• Never refreshes OAuth tokens — the plugin does not send refresh requests. OpenCode owns the credential lifecycle.
• Never logs access tokens, refresh tokens, JWTs, account IDs, or Authorization headers — all log and error paths are sanitized by the centralized redact.ts module.
• Never sends telemetry — no analytics, no usage reporting, no phone-home.
• Never makes unexpected network requests — the only network destination is https://chatgpt.com/backend-api/wham/usage, and only when credentials are available.
• Never executes install-time code — the package has no postinstall, preinstall, or other lifecycle scripts.

The src/redact.ts module provides:

• redact(input: string): string — replaces JWT-like strings, Bearer tokens, refresh tokens (rt_...), account IDs (acct_...), API keys (sk-...), and generic long token-like strings.
• redactDeep(value: unknown): unknown — recursively redacts objects and arrays, and skips known secret field names (access, refresh, authorization, accountId, key, token).
• sanitizeError(err: unknown) — extracts and redacts error messages and codes.

The https://chatgpt.com/backend-api/wham/usage endpoint is undocumented and unsupported by OpenAI. It may change shape, move, or disappear without notice. The plugin:

• Validates the response at runtime with a tolerant Zod schema.
• Identifies windows by duration (not response position).
• Preserves unknown windows rather than discarding them.
• Treats any failure (401/403/429/5xx/timeout/malformed) as non-fatal — session token reporting continues independently.
• Does not cache unauthenticated for the full TTL (uses a shorter 30-second negative cache).

The test suite includes an automated secret-leak scan that checks:

• All test output and snapshots for known secret patterns.
• All built files in dist/ for embedded credentials.
• The packed tarball contents for leaked secrets.

The scan uses synthetic secret fixtures (ey_fake_access, rt_fake_refresh, acct_fake) and verifies they never appear in output, logs, or packaged artifacts.

If you discover a security vulnerability, please report it privately by opening a GitHub security advisory. Do not file a public issue.