Metadata-Version: 2.4
Name: hermes-plugin-qdrant-memory
Version: 0.7.1
Summary: Durable Qdrant memory for Hermes, standalone CLI and local MCP
License-Expression: MIT
Project-URL: Homepage, https://github.com/cnkang/hermes-plugin-qdrant-memory
Project-URL: Repository, https://github.com/cnkang/hermes-plugin-qdrant-memory
Project-URL: Issues, https://github.com/cnkang/hermes-plugin-qdrant-memory/issues
Project-URL: Changelog, https://github.com/cnkang/hermes-plugin-qdrant-memory/blob/main/CHANGELOG.md
Project-URL: MCP, https://github.com/cnkang/qdrant-memory-mcp
Project-URL: Skill, https://github.com/cnkang/skills/tree/main/qdrant-memory
Requires-Python: >=3.11
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: qdrant-client<2,>=1.15
Requires-Dist: httpx<1,>=0.28
Requires-Dist: portalocker<4,>=3.2
Requires-Dist: jsonschema<5,>=4.23
Requires-Dist: pywin32<313,>=311; sys_platform == "win32"
Provides-Extra: mcp
Requires-Dist: mcp<3,>=1.30; extra == "mcp"
Provides-Extra: remote
Requires-Dist: mcp<3,>=1.30; extra == "remote"
Requires-Dist: uvicorn<1,>=0.35; extra == "remote"
Provides-Extra: local
Requires-Dist: fastembed==0.9.1; extra == "local"
Dynamic: license-file

# Runtime 0.7.1: keep the recovery point and stop on unresolved receipts

Runtime 0.7.1 fixes two upgrade-safety defects reported against the published
0.7.0 wheel. It does not change dense retrieval, point identity, stored
payloads, the embedding fingerprint, authorization, confirmed writes or the
four scoped MCP tools.

## Interrupted upgrades keep their recovery point

In 0.7.0, a normal `upgrade plan` after an interrupted apply replaced the saved
state with a fresh `planned` record. The backup directory still existed, but
`status` lost its path and the step history. The same `plan_id` could return as
if nothing had run.

Runtime 0.7.1 reads the saved record inside the existing writer lease. A plan
that has executed a step and has not reached `ready` keeps its phase, failure,
step records and recovery-point reference. `upgrade plan` reports the existing
plan identifier and tells the operator to resume it.

If the plan is stale, `upgrade plan --replace` archives the old plan and its
backup reference before it saves a new plan. An operator never has to delete a
control-plane file by hand.

`apply` and `resume` re-read the plan after they take the lease. If another pass
replaced the plan while the lock was acquired, the older plan is refused and the
newer record is not overwritten. A repeated resume reports `already_applied`
without a second backup or policy generation.

When an embedded backup exists without its step record, `resume` reopens the
copy and checks the planned point count. It then adopts the existing path. An
interrupted activation is reported complete only after the committed policy
record matches the target and the live destination still satisfies the verify
step.

## Shared Owner receipts block planning and apply

In 0.7.0, an admitted Shared Owner lifecycle receipt was visible to the normal
write interlock but not to the upgrade planner. The receipt had no terminal event. `plan` could omit the blocker and `apply` could report `ready`.

Runtime 0.7.1 reuses the read-only condition from
`qdrant_memory/shared_state.py` for both call sites. It covers
`owner_events_v1` receipts without a terminal event and incomplete
`shared_requests_v1` rows. A receipt whose event is already terminal remains a
crash leftover and does not block.

`plan` reports the blocker. `apply` and `resume` recheck after taking the lease,
so a receipt that appears after planning is still caught. Existing reset,
scoped-delete, failed-event and quarantined-event blockers are unchanged.
Maintenance commands never clear receipts, re-extract a turn or replay a pending
event. Complete the existing Owner recovery flow first, then resume the plan
when it is still valid, or use `upgrade plan --replace` when it is stale.

## Compatibility

Runtime 0.7.1 keeps the 0.7.0 storage and policy schema. No data migration,
re-embedding, index rebuild or policy default change is included. Qdrant Server
1.20.1 remains the recommended self-hosted baseline, and the published
`qdrant-client` range is unchanged.

The separate `qdrant-memory-mcp` 0.3.4 distribution pins Runtime 0.7.0 and does
not receive this fix automatically. A later MCP patch can select 0.7.1. This
Runtime release does not change the Hermes Catalog pin, and it does not request
an upstream Catalog update.

## Installation

Check [PyPI](https://pypi.org/project/hermes-plugin-qdrant-memory/0.7.1/) for
availability before installing in a private environment:

```sh
python -m pip install 'hermes-plugin-qdrant-memory[mcp]==0.7.1'
```

For an existing destination, keep the same profile, collection and embedding
pipeline. Run `upgrade plan`, review the blockers and index status, then apply
inside a short maintenance window. A failed or interrupted pass is resumed with
its original plan identifier, not replanned by default.

See the [0.7.0 release](0.7.0.md) for the initial policy upgrade design.
