fix(async-context-compression): v1.7.3 — fresh-PostgreSQL summary persist + reasoning-model inlet reuse (#98)

[Fu-Jie]

Summary

Bumps async-context-compression to v1.7.3. This release fixes two community-reported regressions and adds an end-to-end verification suite that replays real OpenWebUI request-construction code against the filter.

Fixes

1. Summary persistence on fresh PostgreSQL databases

Symptom: ⚠️ Summary generated but was not persisted on freshly deployed PostgreSQL; backend log showed psycopg2.errors.DuplicateTable: relation "ix_chat_summary_chat_id" already exists.

Root cause: The shared SQLAlchemy owui_Base.metadata retained a stale unique index definition with the same name as the new non-unique index. extend_existing=True merges column definitions but does not clean up stale index objects, so CREATE TABLE emitted two colliding CREATE INDEX statements.

Fix (ef434e8 + a84f7bb):

• New _dedup_chat_summary_metadata_indexes() scans ChatSummary.__table__.indexes and drops stale unique-name duplicates before CREATE TABLE.
• Idempotent DROP INDEX IF EXISTS for ix_chat_summary_chat_id, ix_chat_summary_branch_tip_id, ix_chat_summary_updated_at as a belt-and-suspenders cleanup.

2. Outlet summary reuse for idless plain-chat branches (496cabf)

When the outlet request had no stable message refs but carried a chat_id, the filter could not attach refs to the generated summary. It now reads the active DB branch and aligns the body via _compatible_db_branch_for_body_ref_fallback, so plain-chat summaries persist and are reused on subsequent turns.

3. Reasoning-model inlet reuse — issue #98 (9f15b1c)

Symptom: For reasoning models (DeepSeek-R1, o1-like, Qwen3-thinking, etc.), the cached summary was rejected every turn and the inlet re-sent the full uncompressed history.

Root cause: Reasoning models store assistant content with folded <details type="reasoning"> blocks in the DB, but process_messages_with_output → convert_output_to_messages(output, reasoning_format=...) reconstructs the request body with reasoning stripped (OpenAI-compatible), re-tagged as <think> (Ollama), or routed to a reasoning_content field (llama.cpp). Both existing alignment paths (folded content match / unfolded content match) therefore failed on content inequality.

Fix — Path 3 (position-based fallback): When Path 1 and Path 2 both fail, accept the snapshot iff:

• body and DB branches have equal length,
• all body messages are idless (frontend sends plain chat-completion messages),
• roles / tool_calls / tool_call_id match position-by-position,
• DB messages that carry an output array are exempted from content comparison (their content was reconstructed),
• DB messages without output still require exact content equality (fail-closed against edits).

Rejected cases (fail-closed preserved):

• Edited body content when DB has no output → rejected
• Tampered tool_calls → rejected
• Mismatched tool_call_id / role drift → rejected
• Tool-call expansion (body count > DB count) → rejected (position map no longer aligns)

4. DB branch anchor + failed-assistant alignment — issue #98 (9a3689a + c3aefb2)

Symptom: Path 3 still rejected with role mismatch at index=N even after the anchor fix, on long chats (~800 messages) where the live body and the filter's DB walk were equal length but diverged mid-sequence.

Root cause (two layers):

• Anchor mismatch (9a3689a): the DB walk used history.currentId as its walk anchor, but OpenWebUI's load_messages_from_db walks from metadata['user_message_id']. After a regeneration/edit these point at different branches, so the reconstructed DB sequence did not match the live body. Fixed by passing user_message_id as the walk anchor.
• Failed-assistant filtering (c3aefb2): _load_full_chat_messages applied _filter_model_visible_history_messages, which dropped assistant messages carrying an error field, on the assumption that OpenWebUI's middleware filters them too. It does not — load_messages_from_db only strips fields to (role, content, output, files) and keeps every node on the parentId chain. Whenever the active chain contained a failed/regenerated assistant turn, the DB walk silently dropped it while the live body kept it, shifting every subsequent index. Fixed by removing the filter call; the DB walk now reproduces load_messages_from_db verbatim.

Inlet diagnostics (__event_call__ → browser DevTools console):

[Inlet] 📍 DB walk anchor: user_message_id=<id>
[Inlet] 🧭 DB walk used <source>: anchor=<id> currentId=<id> (same branch as currentId | ⚠️ TRUE DIVERGENCE)
[Inlet] 📦 Summary snapshot: FOUND (will inject) | NONE (full context sent)
[Inlet] ⚠️ Path 3 rejected at index=<N>: <reason>     # only when rejected

The 🧭 line walks parentId up from currentId to decide whether user_message_id is on the same branch (normal) or a diverged fork (needs a different fix).

End-to-end verification (test_issue98_e2e.py)

Not a simple unit test. This module inlines verbatim from the OpenWebUI main branch:

• convert_output_to_messages (misc.py)
• process_messages_with_output (middleware.py)
• reconcile_tool_pairs

It then reconstructs the request body for 4 reasoning-model scenarios by running the real OpenWebUI pipeline (process_messages_with_output + id stripping to simulate the frontend), and asserts behavior at both the alignment-method level and the full inlet() entry point:

Scenario	reasoning_format	Body content shape
OpenAI-compatible	None	reasoning stripped
Ollama	think_tags	<think> tags inline
llama.cpp	reasoning_content	separate reasoning_content field
Tool calls	None + function_call	body count > DB count (expansion)

14 e2e tests cover acceptance (Path 3 admits reasoning content mismatch) and rejection (edited content with no DB output, tampered tool_calls, tool-call count expansion), plus full inlet() summary-injection assertions for OpenAI-compatible and Ollama.

Test results

• test_issue98_e2e.py alone: 14/14 pass
• test_async_context_compression.py: 129/129 pass (the previously-failing sqlalchemy-dependent test now passes once sqlalchemy is installed in the environment; unrelated to logic changes)

Files changed

• async_context_compression.py — version bump 1.7.2 → 1.7.3 (logic changes were in prior commits on this branch)
• v1.7.3.md / v1.7.3_CN.md — release notes (new)
• README.md / README_CN.md — what's-new + changelog pointer updates
• _diag.py — removed (temporary diagnostic script)

Upgrade notes

No database migration required. On first launch after upgrade, _init_database will clean any colliding legacy indexes automatically.

Related issue

Closes #98 (reasoning-model inlet reuse). Also resolves the Summary generated but was not persisted regression on fresh PostgreSQL.

Note: This PR is opened as a draft to invite the issue reporter and other affected users to test the branch before merge. Please leave feedback in #98 or here.

[github-actions]

✅ Plugin Version Check / 插件版本检查

版本更新检测通过！PR 包含版本变化和更新说明。

Version check passed! PR contains version changes and update description.

---

版本变化 / Version Changes

New Plugins

• Metadata ID Debug v0.2.0
  • Minimal inlet/outlet filter that prints all ID-related fields from metadata to the BROWSER console (DevTools) via event_emitter execute events. Use to verify which IDs OpenWebUI actually sends to filters (chat_id, message_id, user_message_id, assistant_message_id, etc.).
  • 📖 README

Plugin Updates

• Async Context Compression: v1.7.2 → v1.7.3 | 📖 README

---

This comment was generated automatically. / 此评论由自动生成。

[gemini-code-assist]

Code Review

此拉取请求将插件更新至 v1.7.3 版本，主要解决了在新部署的 PostgreSQL 数据库上 summary 持久化失败的问题，并引入了基于位置的兜底路径（Path 3）以支持推理模型（Reasoning Models）的 inlet 缓存复用（解决 issue #98）。此外，还增加了端到端验证测试。审查中发现了两个改进点：一是在进行基于位置的比较时，应添加防御性类型检查以确保消息对象为字典，防止潜在的 AttributeError 崩溃；二是在 PostgreSQL 的原始 SQL 中对 owui_schema 变量进行双引号处理，以确保与混大小写 Schema 部署的兼容性。

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

To prevent potential AttributeError crashes during position-based comparison, add defensive type checks to ensure both body_message and db_message are dictionaries before calling .get() on them.

Suggested change

	if self._normalize_role(body_message.get("role")) != self._normalize_role(
	db_message.get("role")
	):
	if not isinstance(body_message, dict) or not isinstance(db_message, dict):
	return False
	if self._normalize_role(body_message.get("role")) != self._normalize_role(
	db_message.get("role")
	):

[gemini-code-assist]

In PostgreSQL, mixed-case schema names must be double-quoted to prevent case-folding issues. Double-quote the owui_schema variable in the raw SQL string to ensure compatibility with mixed-case schema deployments.

Suggested change

	qualified = (
	f'{owui_schema}."{name}"' if owui_schema else f'"{name}"'
	)
	qualified = (
	f'"{owui_schema}"."{name}"' if owui_schema else f'"{name}"'
	)