feat(init): AskUserQuestion interactive flow + remote-init mode + dashboard 0.0.0.0

[lightzt99]

Summary

• Replace shell-level interactive prompts in init.sh with a 3-round AskUserQuestion flow (Server → Storage+LLM → Embedding) driven from the init SKILL, so decisions are made in the Claude Code UI instead of the terminal.
• Split the Server question (Round 1) into three options: Local — loopback (127.0.0.1), Local — all interfaces (0.0.0.0), and Remote. Loopback remains the script default; all-interface binding is an explicit user choice, mapped to POWERMEM_SERVER_HOST.
• Add a remote-init short-circuit: when the user supplies a remote PowerMem server URL, init.sh skips .env / uvx / PID management and writes only runtime.env and/or .mcp.json per POWERMEM_INIT_CONNECTION_MODE (hook | mcp | both).
• Make status.sh / status SKILL mode-aware (hook | mcp | both | none) by reading both runtime.env (hook URL) and ${PLUGIN_ROOT}/.mcp.json (MCP URL).
• Extract shared model-cache / CN-mirror download helpers (is_model_cached, detect_country, download_via_modelscope, bridge_modelscope_to_hf_cache) from pyseekdb_default.py into a new _model_cache.py module so both pyseekdb_default and huggingface share the same logic.

Changes

Plugin scripts (apps/claude-code-plugin/scripts/)

• common.sh — shared helpers: is_remote_url, is_healthy, write_runtime_remote.
• init.sh — remote-init short-circuit; dashboard host ${POWERMEM_SERVER_HOST:-127.0.0.1} (loopback default, 0.0.0.0 opt-in via env or Question 0); connection-mode switch.
• status.sh — mode-aware reporting from runtime.env + .mcp.json.

SKILL files (apps/claude-code-plugin/skills/)

• init/SKILL.md — 3-round AskUserQuestion flow; Round 1 splits Local into loopback / all-interfaces options; skip Storage question when server healthy; CLAUDE_PLUGIN_ROOT resolved per-invocation.
• status/SKILL.md — documents the new hook|mcp|both|none mode reporting.

Embedding refactor (src/powermem/integrations/embeddings/)

• _model_cache.py (new) — shared cache/mirror helpers extracted from pyseekdb_default.py.
• pyseekdb_default.py — imports from _model_cache.
• huggingface.py — adopts the shared cache/mirror helpers.
• tests/unit/test_pyseekdb_default_download.py — patches _model_cache instead of pyseekdb_default internals.

Tests

• tests/unit/test_claude_plugin_uv_install.py — updated uvx-launcher assertion to the parameterized form --host "${POWERMEM_SERVER_HOST:-127.0.0.1}" to match the new default + override contract.

Test plan

• pytest tests/unit/test_claude_plugin_uv_install.py — uvx-launcher assertion matches the new parameterized host form.
• pytest tests/unit/test_pyseekdb_default_download.py — download / CN-mirror tests pass against the extracted module.
• CI test (3.11), test (3.12), claude-hook-regression all green on 049627a.
• Manual: install plugin from this branch, run /init, confirm Server → Storage+LLM → Embedding questions appear in the Claude Code UI (not the terminal).
• Manual: with POWERMEM_INIT_BASE_URL=https://remote:8848, confirm init.sh writes runtime.env and/or .mcp.json without starting a local server.
• Manual: confirm dashboard binds 127.0.0.1:8848 by default; choosing "Local — all interfaces" in Round 1 (or setting POWERMEM_SERVER_HOST=0.0.0.0) binds 0.0.0.0 and is reachable from another host.
• Manual: /status reports mode=hook, mode=mcp, mode=both, or mode=none correctly.

[wayyoungboy]

Thanks for putting the interactive / remote init flow together. I cannot mark this ready yet for two reasons:

• CI is red: both test (3.11) and test (3.12) fail in tests/unit/test_claude_plugin_uv_install.py::test_init_uses_uvx_launcher_instead_of_plugin_venv_install. The test still asserts local uvx launch contains powermem-server --host 127.0.0.1 --port "$port", while this PR now launches with ${POWERMEM_SERVER_HOST:-0.0.0.0}. Please either update the test if this behavior is intentional, or keep loopback as the local default and make all-interface binding an explicit opt-in.
• The new init skill text still describes Local mode as starting a server on 127.0.0.1:8848, but the script and generated config now default local launch to 0.0.0.0. Please make the docs, generated config, and launch behavior agree, especially because binding all interfaces changes the local exposure model.

Once the host contract is settled and the Python test jobs are green, I can re-review the updated head.

[lightzt99]

Thanks for the review @wayyoungboy. Both points are addressed in 049627a:

1. CI red — test assertion mismatch

Loopback is restored as the local default. init.sh now uses ${POWERMEM_SERVER_HOST:-127.0.0.1} for both .env generation (line 377) and the powermem-server launch lines (786, 792). The uvx-launcher test assertion in tests/unit/test_claude_plugin_uv_install.py:552 is updated to the parameterized form:

assert "powermem-server --host \"${POWERMEM_SERVER_HOST:-127.0.0.1}\" --port \"$port\"" in script

test (3.11), test (3.12), and claude-hook-regression are all green on 049627a.

2. Docs / generated config / launch behavior consistency

All three now agree on 127.0.0.1 as the local default. All-interface binding is an explicit opt-in via Round 1 of the AskUserQuestion flow: the "Local" option is split into two choices — "Local — loopback" (127.0.0.1) and "Local — all interfaces" (0.0.0.0) — and the answer is mapped to POWERMEM_SERVER_HOST before invoking init.sh. The SKILL text, generated .env, and launch command all reflect the chosen host. The exposure model is now an explicit user decision rather than a silent default change.

PR description updated to match. Ready for re-review when you have a moment.

[wayyoungboy]

Thanks for the updates. The CI jobs I checked are green now, and the loopback/all-interface default issue from the earlier review looks addressed. I still cannot approve the latest head because the new remote connection modes have a few correctness issues:

• runtime.env writes remote URL/API key values as raw shell assignments, but the hook/status scripts later source that file. URLs or keys containing shell metacharacters or spaces can break the config, and command substitution is evaluated while sourcing.
• init.sh now registers MCP through claude mcp add --scope user, but status.sh still reads ${PLUGIN_ROOT}/.mcp.json, so remote mcp mode reports none and both reports only hook.
• Remote mcp mode says hooks are disabled, but it neither removes/stamps runtime.env nor changes the hook launcher/native hook behavior. Existing hook config can remain active, and with no runtime file the native hook still falls back to http://localhost:8848.
• git diff --check upstream/main...HEAD currently fails on trailing whitespace at EOF in apps/claude-code-plugin/scripts/common.sh.

Local checks run: shell syntax check passed; pytest tests/unit/test_claude_plugin_uv_install.py tests/unit/test_pyseekdb_default_download.py passed with 43 tests.

[lightzt99]

@wayyoungboy Thanks for the review. All four issues are fixed in 54a25f8:

1. runtime.env quoting — write_runtime_remote now single-quotes both POWERMEM_BASE_URL and POWERMEM_API_KEY values, with embedded ' escaped via the standard '\'' trick. URLs/keys with $, ;, spaces, backticks, and quotes survive being sourced by run-hook.sh and status.sh verbatim. Regression test: test_write_runtime_remote_quotes_metacharacters.

2. status.sh reads the wrong MCP file — switched from $PLUGIN_ROOT/.mcp.json (volatile plugin-cache, always empty since the user-scope move) to ~/.claude.json top-level mcpServers.powermem.url, which is where claude mcp add --scope user actually writes. mcp mode now reports mcp, both reports both.

3. MCP-only mode didn't disable the hook — init.sh's mcp branch now calls write_runtime_hook_disabled(), which overwrites runtime.env with POWERMEM_HOOK_DISABLED=1 (replacing any stale POWERMEM_BASE_URL). run-hook.sh checks that marker after load_env_file and exits 0 before exec'ing the native binary, so the hook can no longer fall back to http://localhost:8848. Regression tests: test_write_runtime_hook_disabled_writes_marker, test_run_hook_exits_early_when_hook_disabled, test_init_mcp_branch_writes_hook_disabled_marker.

4. git diff --check trailing whitespace — stripped the trailing blank lines at EOF in common.sh; git diff --check upstream/main...HEAD is now clean.

Local: pytest tests/unit/test_claude_plugin_uv_install.py — 23 passed (was 19). sh -n clean on all four touched scripts. The SKILL.md Round 1 wording was also tightened (dedup'd URL placeholder options, clarified Hook mode applies when auth is disabled or a key was supplied; Both retained). Ready for re-review.

[lightzt99]

AskUserQuestion Round 1

AskUserQuestion Round 2 (Remote)

1. provide server URL

2. provide API key if required

3. choose connection mode (hook/mcp/both)

AskUserQuestion Round 2 (Local)

1. choose storage backend (oceanbase/sqlite)

2. enable LLM or not

AskUserQuestion Round 3 (storage backend = sqlite)