fix(anthropic): recover Claude Code OAuth via CLI refresh fallback

[woriwka-ai]

Summary

Adds a bounded last-resort recovery path for Claude Code OAuth credentials in resolve_anthropic_token():

• keep the existing fast path: env token, CLAUDE_CODE_OAUTH_TOKEN, Claude Code credential store + pure refresh;
• if Hermes still cannot resolve a token, run the official Claude Code CLI once with a tiny non-secret prompt;
• re-read the Claude Code credential store and continue to credential-pool/API-key fallback if CLI recovery did not produce fresh creds;
• throttle CLI refresh attempts with a 5-minute in-process cooldown;
• allow operators to disable the fallback via HERMES_ANTHROPIC_DISABLE_CLAUDE_CLI_REFRESH=1.

This addresses the degraded state where both persisted Claude Code sources look expired to Hermes, Hermes' direct/pure refresh loses or is blocked, but the official claude CLI can still refresh its own Keychain/session state. In that state Hermes/Fable can fall through to No Anthropic credentials found even though claude -p ... immediately restores a usable token.

Related: #6347, #48534, #55051.

Testing

python -m py_compile agent/anthropic_adapter.py
python -m pytest tests/agent/test_anthropic_claude_code_refresh_fallback.py -q

Result:

7 passed

Live local smoke before upstreaming, with tokens redacted/not printed:

Claude Code credential source present
credentials valid
Anthropic token resolved
fable_guard=ok · smoke=ok
claude-fable-5 clean-env smoke -> OK
pytest tests/agent/test_anthropic_claude_code_refresh_fallback.py -q -> 7 passed

Safety notes

• The CLI output is ignored and never logged, so tokens are not printed.
• The fallback only runs after the existing direct Claude Code credential/pure-refresh path fails.
• Nonzero/timeout/missing CLI falls through to the existing pool/API-key behavior.
• The env kill switch prevents the CLI subprocess entirely.

[kshitijk4poor]

Thanks @woriwka-ai — appreciate the thorough recovery path. Closing in favor of a root-cause fix merged via #58178.

The underlying failure was that Hermes' own OAuth token exchange/refresh was blocked at the token endpoint: Anthropic 429s any /v1/oauth/token request whose User-Agent starts with claude-code/. Once that UA is corrected to axios/ (matching the real Claude Code CLI's exchange client), Hermes' direct refresh works again and the CLI-subprocess fallback isn't needed. Verified live: claude-code/ → 429, axios/ → 400 invalid_grant (reached validation).

Two additional reasons for the root-cause fix over the fallback here:

• Shelling out to the claude CLI as a refresh mechanism couples Hermes' auth to an external binary and adds ~250 LOC of subprocess plumbing for a problem that's a one-line UA change at the source.
• The HERMES_ANTHROPIC_DISABLE_CLAUDE_CLI_REFRESH env var is a behavioral (non-secret) config flag. Per our contribution rubric, .env / HERMES_* env vars are for secrets/credentials only — behavioral settings belong in config.yaml. (Not a knock on the PR — just why we don't take the flag.)

Fixed in #58178: #58178