fix(hindsight): flush buffered turns on session end

[kumaxs]

Problem

HindsightMemoryProvider accumulates conversation turns in _session_turns and only flushes them to the Hindsight API every _retain_every_n_turns turns (default 5). If a session ends (idle TTL eviction, /reset, cap enforcement) before reaching the threshold, the buffered turns are silently discarded — the message content is lost forever.

The base MemoryProvider.on_session_end() is a no-op, and Hindsight does not override it, so session boundaries never trigger a flush.

Fix

Override on_session_end(messages) to flush any remaining _session_turns into the Hindsight API via the same aretain_batch + writer queue path used by the periodic flush and on_session_switch flush.

The implementation:

• Snapshots current _session_turns, then clears the buffer (re-entrancy safe)
• Resolves the document target (append-mode for v0.5.0+, fallback to per-process document for legacy)
• Enqueues a _flush job to the existing serial writer queue, so it's ordered behind any in-flight retains
• Logs success/failure at INFO/WARNING level

Related

Depends on the gateway-side change in PR #55045 which calls on_session_end before evicting cached agents. Without that PR, this hook would never be reached on idle TTL eviction (only on explicit /reset and the existing shutdown path).

[alt-glitch]

This was generated by AI during triage.

Duplicate of #36219 — same fix (HindsightMemoryProvider.on_session_end() flushing buffered _session_turns through the writer queue, both Fixes #36216) with the identical title. #36219 is the earlier open canonical. Companion to your #55045 (the gateway-side hook this depends on). Flagging the cluster so a maintainer can pick one of the on_session_end flush PRs.

[teknium1]

Thanks for targeting a real data-loss path. Current main buffers sub-threshold turns in plugins/memory/hindsight/__init__.py:1621-1629, and shutdown only drains work already enqueued (plugins/memory/hindsight/__init__.py:1905-1925), so an end-of-session flush is needed.

Problems

• The proposed flush sends all _session_turns after resolving update_mode. On append-capable APIs, current periodic retention deliberately sends only _session_turns[_last_retained_turn_count:] and advances that watermark (plugins/memory/hindsight/__init__.py:1631-1696). A completed batch followed by a partial tail would therefore append the completed batch a second time.
• tests/plugins/memory/test_hindsight_provider.py has switch-flush coverage (:1196-1437) but no Hindsight on_session_end() coverage for this new lifecycle path.

Suggested changes

• Send only the unretained watermark delta in append mode; retain the full-buffer behavior for legacy overwrite mode.
• Add session-end tests for sub-threshold flushing, append-mode tail-only flushing after a prior batch, no-op guards, and queue-before-shutdown ordering.
• Consider sharing the corrected flush helper with on_session_switch() (plugins/memory/hindsight/__init__.py:1822-1867), which currently also flushes its full buffer.

This is an automated hermes-sweeper review.

[kumaxs]

@teknium1 已修复 append-mode 重复 flush 问题：

修复内容 (5fc7ff2)：

• on_session_end 现在先 resolve update_mode，append 模式下使用 watermark delta _session_turns[_last_retained_turn_count:]，只发送未被 periodic retain 覆盖的新 turns
• 新增 early return：如果所有 turns 已被 periodic flush 覆盖，跳过 flush
• overwrite 模式保持全量发送（legacy 兼容）

还需要补充测试覆盖（sub-threshold flushing、append-mode tail-only、no-op guard）——我接下来处理。

[alt-glitch]

This was generated by AI during triage.

Related to #36219 and #55936 rather than a duplicate: the current head adds the append-mode watermark delta that the earlier duplicate review found missing. The remaining review gap is dedicated on_session_end regression coverage.