[bug] why-selected ignores account policies: a paused account is reported as the winning candidate

[choa712]

Summary

codex-multi-auth why-selected recomputes candidate scores without applying the account-policy layer that the runtime rotation proxy actually uses (blockedAccountIndexes / scoreBoostByAccount from lib/policy/runtime-policy.ts). As a result it can report a paused account as the winning candidate — while the runtime proxy correctly refuses to route to it. During incident response this diagnostic actively misleads: we spent time believing the router was selecting a revoked account when it was not.

Reproduction

1. Have 3 accounts registered. Pause one of them: codex-multi-auth account pause 2
2. Confirm the policy is stored: codex-multi-auth account policy list --json (shows "paused": true)
3. Run codex-multi-auth why-selected --last
4. Send a real request through the runtime rotation proxy and check ~/.codex/multi-auth/app-bind/runtime-rotation-app-bind-status.json → lastAccountLabel

Expected Behavior

why-selected mirrors the runtime selection inputs: paused/drained accounts are excluded (or at least rendered as blocked (policy: paused)), and per-account policy weight boosts ((weight - 1) * 2) are included in the displayed scores, so the diagnostic agrees with what the proxy does.

Actual Behavior

The paused account is counted in "Available: 2 of 3 account(s)" and printed as Selected: with health=100, top score. Policy weight boosts are absent from the scores (an account with policy weight 10 shows no boost). Meanwhile the actual runtime proxy correctly serves requests from a different, non-paused account (lastAccountLabel confirms).

Sanitized output on 2.8.1, with account B paused (its refresh token is also revoked upstream):

why-selected: Last selection (live recomputation; no persistent tracker)
Available: 2 of 3 account(s)

Selected: account 2 <account-B>
  score: 463.98
  health: 100.0
  ...
Candidates (sorted by score desc):
  * 2 <account-B>: score=463.98 health=100 ...   <- paused via account policy
  x 3 <account-C>: ... (unavailable (rate-limited, cooling down, or circuit open))
    1 <account-A>: score=450.20 ...              <- policy weight 10, no boost shown

Code pointers: lib/codex-manager/commands/why-selected.ts builds candidates from the rotation scorer only; it never calls loadRuntimePolicyState / evaluateRuntimePolicy, whereas the proxy path (lib/runtime/rotation-account-selection.ts) checks policy.blockedAccountIndexes and feeds scoreBoostByAccount into the hybrid selector.

Environment

• codex --version: codex-cli 0.146.0
• codex-multi-auth status output summary: 3 accounts, storage healthy, runtime rotation app-bind in use
• codex-multi-auth rotation status output summary: proxy enabled, app bind running
• npm ls -g codex-multi-auth: 2.8.1 (installed from a checkout of the 2.8.1 release commit; not fork-modified)
• OS: macOS (Apple Silicon)
• Node.js: 26.4.0

Diagnostic Outputs

Covered above (sanitized why-selected + account policy list). Happy to attach more on request.

Compliance Confirmation

• I am using this project for personal development workflows.
• This report does not request policy bypasses or prohibited usage.
• I removed any secrets/tokens from this report.

[coderabbitai]

🔗 Related PRs

#480 - Fix runtime rotation pool diagnostics [merged]
#510 - feat(rotation): sequential drain-first account scheduling mode (#509) [merged]
#607 - fix(rotation): break stale-recovery deadlock on transient account state [closed]
#608 - fix(rotation): persist cooldown when account has no resolvable accountId [closed]
#624 - feat(codex): force a specific account per invocation with --account (#623) [merged]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!