v0.2.32 — CRITICAL HOTFIX for 0.2.31 (unterminated docstring)

Critical hotfix — 0.2.31 was completely unusable

Fixed

• memnet/serve_client.py unterminated module docstring. The 0.2.31 client-API documentation update (#2) added closing paragraph text to the module docstring but omitted the closing """, which wrapped from __future__ import annotations and the rest of the module inside the string literal. Every memnet invocation — including memnet --help — raised:

  SyntaxError: unterminated triple-quoted string literal
  

  Closes #9.

Added

• tests/test_packaging_sanity.py — ast.parse's every .py file under src/, imports the top-level memnet/memnet_mcp modules, and walks + imports every submodule of memnet via pkgutil.walk_packages. This class of bug (valid-looking edit, broken at actual import time, never exercised by unit tests that only test individual functions) can no longer ship silently.
• Pre-publish verification: this release was smoke-tested by installing the built wheel into a clean virtualenv and running import memnet + memnet --help before uploading, not just twine check (which does not execute code).

Action required

If you installed memnet-llm==0.2.31, upgrade immediately:

pip install -U memnet-llm==0.2.32

PyPI: https://pypi.org/project/memnet-llm/0.2.32/

v0.2.31 — 8 issues resolved (registry, client API, snapshot diagnostics, ingest rollback, exit_code, read_list, sliding TTL)

8 issues resolved — MemNet core reliability pass

All changes preserve public API compatibility. Core test suites pass.

Fixed

• #1 registry naming — get/remove → get_entry/remove_entry (avoids builtin shadowing).
• #2 client surface — serve_client.dispatch is now the documented canonical entry point; low-level TCP helpers in serve are explicitly internal.
• #3 remote snapshot diagnostics — session_load / load_snapshot now emit snapshot_not_found|<path>|serve_cwd=<...> on FileNotFoundError, making cross-host path mismatches obvious.
• #4 ingest rollback cost — replaced full-store to_jsonl_rows snapshot with O(batch) incremental journal (added ids + replaced Records). Rollback is now proportional to batch size, not store size (estimated 50–500× win on large stores).
• #5 exit_code on per-line errors — both inline (_run_inline) and TCP serve (_handle_request) paths now capture the app(...) return value for typer.Exit, ensuring exit_code is non-zero when any line is rejected (FIELD_COUNT, unknown_tag, etc.).
• #7 enumeration path — added read_list MCP tool (and documented read list --tag) as the supported way to list rows by tag without prior IDs. query_walk docstrings redirect to it.
• #8 TTL expiry — get_session implements sliding TTL: every successful access extends expires_at by the original ttl_minutes. Long-lived sessions (novel-rpg world state) no longer expire silently.

Changed (by design)

• #6 TagMap immutability — confirmed: schema is fixed at session open. Runtime mutation is unsupported and dangerous (desync risk with parser/emitter, snapshot ambiguity). Closed with rationale; the hand-edit workaround remains unsupported.

Install

pip install -U memnet-llm==0.2.31

PyPI: https://pypi.org/project/memnet-llm/0.2.31/

v0.2.29

Summary

• 10-beat live smoke passed on session mn_1d2a37a6 (script+prose full pipeline).
• Beat errors propagate to UI/CLI instead of silent None.
• HUD built after successful beat_turn_finish; VIT01 honours update_lines with auto satiety downgrade.
• beat_turn_finish warm uses NOVEL_WARM_DEPTH so slot-6 library options validate correctly.
• Opening beat: mobile sends { start: true }; continue reserved for prose-stage resume.

Test plan

• pytest novel suite (59 tests)
• python scripts/run_beats_smoke.py 10 shenjia_caifa — 0 failures

v0.2.28 — beat progression & seed-driven prompts

Summary

• Fix beat chapter落盤：world 路徑同步、缺 chapter 硬失敗、continuation anchor 優先世界目錄
• Seed-driven HUD（body_state）與飽食/正文衝突檢查
• beat_prompt 不再 hardcode 魂穿/圖書館/文風；改由 presentation.contracts 注入
• 種子補 @chp tag map（修復 bootstrap 整批 seed 載入失敗）
• 武學 catalog 全名驗證與邊界列更名

See CHANGELOG.md for full details.

v0.2.27 — EDG acquaintance & multi-world mobile

Added

• entity_knowledge — @edg SSOT for holder acquaintance; depth-gated POV name masking
• party_sheet — roster panel with EDG-gated NPC names
• novel-mobile multi-world — world registry/slot, optional auth, E2E script

Changed

• knowledge_graph — EDG-based view (@knh deprecated)
• presentation / beat_prompt — acquaintance masking; prose POV (旁白「你」、內心才「我」)
• novel_mobile — per-world sessions, beat jobs, mobile UX fixes

See CHANGELOG.md for full notes.

v0.2.25

Added

• NOVEL_WARM_MAX_ROWS (150) — warm reads no longer truncate USR21 prose advisory at 55 rows.

Changed

• novel-mcp / beat pipeline defaults use 150; operator docs require cursor_beat.py for play (MCP beat tools for SDK/tests only).

See CHANGELOG.md.

v0.2.24

[0.2.24] ??2026-06-27

Added

• novel-seed-spec.md ??normative SEED planning principles (禮2 layering, three EDG planes, opening scene contract, god-realm vs play).
• Player setup pipeline ??player_profile.py, player_setup.py, setup_graph.py, setup_constants.py, opening_loadout.py; MCP read_player_setup, commit_player_profile, commit_opening_pick.
• Generic catalog schema ??catalog_schema.py, martial_catalog_expand.py, applications/novel_cursor/catalog_specs/; instance-driven genre validation (no Jin Yong hardcoding in novel_mcp core).
• Warm enrich ??warm_supplement.py stage-aware merge; presentation opening_scene / biz / scn_code; warm_index BIZ/SCN rows.
• cursor_beat ??--setup, --name, --gender, --arts; beat_orchestrator.py, catalog_expand.py, chat_thread.py, llm_client.py, wire_parse.py.
• Shenjia martial catalog ??application-notes/novel-shenjia-martial-catalog.md; seed USR69/USR70 opening contract.
• Tests ??opening loadout, player profile, setup guidance, warm supplement, seed LAW budget, martial catalog expand, wire parse, chat thread, edg_time.

Changed

• scripts/novel_bootstrap.py ??generic --app bootstrap; optional catalog expand; removed bootstrap_shenjia.py / shenjia_bootstrap.py.
• novel-shenjia-initial-state.md ??Engine/World split maintenance; opening scene domain wiring (USR70, LAW-OLN01 opening_scn); setup USRs 60??7; LAW budget ??0.
• Tag map ??@ART, @WUX, @MWU for martial loadout.
• beat_pipeline ??setup gate integration; prose-stage SCR enrich from read list.

Fixed

• Setup wire ??USR/PLR 4-field shape; PLR without extra recycle field; EDG add vs update on bootstrap.
• catalog_expand ??complete_messages API for LLM expand.
• test_beat_pipeline ??patch play_context.run_memnet for enrich isolation.

v0.2.23 — dual-loop SDK beat agents

Added

• Dual-loop Cursor SDK beat agents — separate persistent script (編劇) and prose (作者) agents; ids under novel-output/<slug>/agents/
• agent_session.py — create+primer or Agent.resume, stale-id recovery
• play_context.py — script_beat_prepare / prose_beat_prepare (USR23 handoff gate)
• MCP tools — script_beat_prepare, prose_beat_prepare
• chapter_io — last_committed_paragraph for continuation_anchor
• tests — test_play_context, test_agent_session_paths, test_beat_prompts

Changed

• cursor_beat.py — dual-phase orchestrator; --script-only, --prose-only, --reset-agents, --continue
• beat_prompt.py — split script/prose primers and turns
• llm-novel-cursor-sdk.md — dual-loop operator SSOT

Full notes: CHANGELOG.md

v0.2.22

Added

• applications/novel_cursor/ — generic Cursor SDK beat runner; \NOVEL_BEAT_RESULT\ contract; paths from seed USR14/USR15
• application-notes/llm-novel-cursor-sdk.md — operator guide
• tests/test_novel_cursor_config.py

Fixed

• novel_mcp.beat_pipeline — USR23 beat_stage persist (_usr_beat_stage_row)

Changed

• Generic novel-writer chat rule; shenjia_caifa thin shim

Full notes: CHANGELOG.md

v0.2.21 — fix mem_store test fixtures

Fixed

• tests/test_mem_store.py — remove @LAW from user tag-map fixtures; LAW is a fixed tag. Restores linked-law scope, context walk, and all-law scope tests.

Full pytest suite (149 tests) green.