Greptile Summary

This PR addresses two separate concerns: (1) the primary fix prevents isolated cron jobs whose sessionKey matches the parent session key from advancing updatedAt past the daily reset boundary, and (2) a secondary change adds Bailian API (OpenAI-compatible nested usage object) support in the usage normalization layer.

Key changes:

• src/cron/isolated-agent/run.ts: persistSessionEntry now uses mergeSessionEntryPreserveActivity for writes to agentSessionKey in both in-memory and on-disk paths, correctly preserving the parent session's updatedAt without disturbing session content fields.
• src/agents/usage.ts: normalizeUsage and hasNonzeroUsage are extended to unwrap a nested usage sub-object (raw.usage ?? raw) for Bailian API compatibility. All existing fields retain their priority via ?? chaining.
• src/agents/usage.normalization.test.ts: Two well-written test cases cover the new nested usage path.

One minor gap noted: hasNonzeroUsage checks usageObj.prompt_tokens, usageObj.completion_tokens, and usageObj.total_tokens from the nested object, but omits usageObj.cached_tokens — inconsistent with normalizeUsage, which does extract that field as cacheRead. In practice this edge case (non-zero cache reads with zero other tokens) is unlikely, but worth aligning for consistency.

Confidence Score: 4/5

• Safe to merge; the cron fix is logically sound and the usage normalization change is additive with no breaking impact on existing paths.
• The primary run.ts fix is correct: mergeSessionEntryPreserveActivity properly preserves parent session updatedAt while still applying all other session fields. Edge cases (no existing entry, runSessionKey !== agentSessionKey) are handled correctly. The usage changes are purely additive. Score is 4 rather than 5 due to the minor usageObj.cached_tokens gap in hasNonzeroUsage and the lack of automated test coverage for the cron fix path.
• No files require special attention, though src/agents/usage.ts has a minor inconsistency in hasNonzeroUsage worth addressing.

This is a comment left during a code review.
Path: src/agents/usage.ts
Line: 93-102

Comment:
**Missing `cached_tokens` in nested usage check**

`hasNonzeroUsage` checks `usageObj.prompt_tokens`, `usageObj.completion_tokens`, and `usageObj.total_tokens` from the nested object, but omits `usageObj.cached_tokens`. This creates an inconsistency with `normalizeUsage`, which correctly maps `usageObj.cached_tokens` → `cacheRead`.

In the unlikely but possible case where a Bailian API response returns non-zero `cached_tokens` with zero prompt/completion/total tokens, `hasNonzeroUsage` would return `false` while `normalizeUsage` would return a non-undefined result with a non-zero `cacheRead`. Callers that guard behind `hasNonzeroUsage` (e.g. `pi-embedded-subscribe.ts:274`) would silently discard that usage.

```suggestion
  return [
    usage.input,
    usage.output,
    usage.cacheRead,
    usage.cacheWrite,
    usage.total,
    usageObj.prompt_tokens,
    usageObj.completion_tokens,
    usageObj.total_tokens,
    usageObj.cached_tokens,
  ].some(
```

How can I resolve this? If you propose a fix, please make it concise.

_{Last reviewed commit: "fix(cron): preserve ..."}

Thanks for documenting the isolated-cron updatedAt vector. I’m closing this as superseded by #71845.

What changed in the replacement:

• daily/idle reset freshness no longer depends on updatedAt, so cron writes cannot mask a stale session by refreshing that bookkeeping field
• cron session handling now preserves/sets lifecycle timestamps explicitly
• lastInteractionAt is the idle-reset authority, while sessionStartedAt is the daily-reset authority
• the fix covers the broader duplicate family: heartbeat, cron, exec, gateway-agent, and ordinary channel-session metadata writes
• regression tests cover cron session freshness plus the shared session-store write behavior

This PR’s root diagnosis was correct, but the replacement removes updatedAt from the reset decision entirely instead of preserving it one write site at a time. Closing in favor of #71845.