kiwime-store 0.1.0__tar.gz

kiwime_store-0.1.0/.gitignore

@@ -0,0 +1,78 @@
+# OS
+.DS_Store
+Thumbs.db
+
+# Editors
+.vscode/*
+!.vscode/settings.json
+.idea/
+*.swp
+*.swo
+
+# Rust
+target/
+Cargo.lock.bak
+
+# Node
+node_modules/
+dist/
+.pnpm-store/
+*.log
+
+# Python
+__pycache__/
+*.py[cod]
+*.egg-info/
+.venv/
+.pytest_cache/
+.mypy_cache/
+.ruff_cache/
+.coverage
+htmlcov/
+
+# Tauri
+apps/desktop/src-tauri/target/
+
+# Frozen Python sidecars (build artifacts, see ADR 0006)
+apps/desktop/src-tauri/binaries/
+target/pyinstaller/
+
+# uv
+.python-version
+
+# Frontend build copied into the host package by the publish workflow
+apps/host/src/kiwime_host/_ui/
+
+# Local data
+*.sqlite
+*.sqlite-journal
+*.db
+data/
+blobs/
+
+# Secrets
+.env
+.env.local
+*.pem
+
+# Per-user Claude Code settings (personal permission allowlists, local paths)
+.claude/settings.local.json
+
+# TypeScript incremental build state
+*.tsbuildinfo
+apps/desktop/dist-types-node/
+
+# Internal planning notes (kept out of the public tree; see docs/ for the
+# published roadmap, ADRs, and release docs)
+.internal-notes/
+docs/release/C_NOTES.md
+docs/release/D_NOTES.md
+docs/release/oss-readiness-review.md
+docs/release/implementation-plan-*.md
+docs/release/webification-plan.md
+
+# Tool caches / scratch
+.mypy_cache/
+.pytest_cache/
+.ruff_cache/
+.a1/

kiwime_store-0.1.0/PKG-INFO

@@ -0,0 +1,9 @@
+Metadata-Version: 2.4
+Name: kiwime-store
+Version: 0.1.0
+Summary: kiwiMe store sidecar — owns SQLite, the blob store, and the vector store.
+Project-URL: Homepage, https://github.com/kiwiberry-ai/kiwime
+License-Expression: Apache-2.0
+Requires-Python: >=3.11
+Requires-Dist: kiwime-sdk==0.1.0
+Requires-Dist: sqlite-vec>=0.1.9

kiwime_store-0.1.0/README.md

@@ -0,0 +1,56 @@
+# kiwime-store
+
+Python sidecar that owns SQLite, the on-disk blob store, and (later) the
+vector store. Other sidecars never touch these files directly — they go
+through this worker over JSON-RPC.
+
+See `docs/architecture/ipc-protocol.md` §4 for the wire methods, and
+`docs/schema/core.md` for the record shapes.
+
+## Run standalone
+
+```sh
+# Default DB path: ~/Library/Application Support/kiwiMe/store.sqlite (macOS)
+# or $XDG_DATA_HOME/kiwiMe/store.sqlite (Linux).
+python -m kiwime_store --db /tmp/store.sqlite
+```
+
+It speaks JSON-RPC 2.0, one frame per line, on stdin/stdout. Logs go to
+stderr only.
+
+A trivial smoke test from a shell:
+
+```sh
+printf '%s\n' '{"jsonrpc":"2.0","id":"1","method":"hello"}' \
+  | python -m kiwime_store --db /tmp/store.sqlite
+```
+
+## Inspect the DB
+
+```sh
+sqlite3 /tmp/store.sqlite '.tables'
+sqlite3 /tmp/store.sqlite 'SELECT version, name, applied_at FROM schema_migration;'
+sqlite3 /tmp/store.sqlite 'SELECT id, source_id, version FROM document;'
+```
+
+Blob layout:
+
+```
+<db_dir>/blobs/<sha256[:2]>/<sha256>
+```
+
+## Methods
+
+Implements every `workers/store` method from ipc-protocol.md §4:
+`store.put_record`, `store.get_document`, `store.list_pending_for_compile`,
+`store.put_memory`, `store.put_capability`, `store.put_graph`,
+`store.query_sources`, `store.query_memories`, `store.query_capabilities`,
+`store.query_graph`, `store.migrate`. Plus `store.create_profile` and
+`store.put_source` for bootstrap. All sidecars also expose `hello`,
+`shutdown`, `health` from `kiwime_sdk.Server`.
+
+## Tests
+
+```sh
+cd workers/store && python -m pytest
+```

kiwime_store-0.1.0/migrations/0001_init.sql

@@ -0,0 +1,203 @@
+-- kiwiMe initial schema, v0.
+-- See docs/schema/core.md for the normative contract this implements.
+
+PRAGMA foreign_keys = ON;
+PRAGMA journal_mode = WAL;
+
+-- ============================================================================
+-- Profiles
+-- ============================================================================
+CREATE TABLE profile (
+    id           TEXT PRIMARY KEY,
+    display_name TEXT NOT NULL,
+    created_at   TEXT NOT NULL
+);
+
+-- ============================================================================
+-- Ingestion layer
+-- ============================================================================
+CREATE TABLE source (
+    id            TEXT PRIMARY KEY,
+    profile_id    TEXT NOT NULL REFERENCES profile(id) ON DELETE CASCADE,
+    connector     TEXT NOT NULL,
+    display_name  TEXT NOT NULL,
+    kind          TEXT NOT NULL CHECK (kind IN ('knowledge','activity')),
+    scope         TEXT NOT NULL,           -- JSON
+    state         TEXT NOT NULL CHECK (state IN ('active','paused','error')) DEFAULT 'active',
+    cursor        TEXT,
+    installed_at  TEXT NOT NULL,
+    last_sync_at  TEXT
+);
+CREATE INDEX idx_source_profile ON source(profile_id);
+
+CREATE TABLE document (
+    id            TEXT PRIMARY KEY,
+    source_id     TEXT NOT NULL REFERENCES source(id) ON DELETE CASCADE,
+    external_id   TEXT NOT NULL,
+    title         TEXT,
+    mime_type     TEXT NOT NULL,
+    text          TEXT,
+    blob_ref      TEXT,                    -- sha256:<hex>
+    metadata      TEXT NOT NULL,           -- JSON
+    created_at    TEXT,                    -- provider-side
+    updated_at    TEXT NOT NULL,
+    ingested_at   TEXT NOT NULL,
+    version       INTEGER NOT NULL DEFAULT 1,
+    CHECK (text IS NOT NULL OR blob_ref IS NOT NULL),
+    UNIQUE(source_id, external_id)
+);
+CREATE INDEX idx_document_source ON document(source_id);
+CREATE INDEX idx_document_updated ON document(updated_at);
+
+CREATE TABLE activity_event (
+    id            TEXT PRIMARY KEY,
+    source_id     TEXT NOT NULL REFERENCES source(id) ON DELETE CASCADE,
+    kind          TEXT NOT NULL,
+    actor         TEXT,
+    subject_ref   TEXT,                    -- optional document id
+    payload       TEXT NOT NULL,           -- JSON
+    occurred_at   TEXT NOT NULL,
+    ingested_at   TEXT NOT NULL
+);
+CREATE INDEX idx_event_source ON activity_event(source_id);
+CREATE INDEX idx_event_occurred ON activity_event(occurred_at);
+
+CREATE TABLE asset (
+    id            TEXT PRIMARY KEY,
+    source_id     TEXT NOT NULL REFERENCES source(id) ON DELETE CASCADE,
+    kind          TEXT NOT NULL,
+    name          TEXT NOT NULL,
+    description   TEXT,
+    metadata      TEXT NOT NULL            -- JSON
+);
+
+CREATE TABLE relation_candidate (
+    id            TEXT PRIMARY KEY,
+    from_ref      TEXT NOT NULL,
+    to_ref        TEXT NOT NULL,
+    kind          TEXT NOT NULL,
+    confidence    REAL NOT NULL CHECK (confidence BETWEEN 0 AND 1),
+    metadata      TEXT
+);
+
+-- ============================================================================
+-- Compilation layer
+-- ============================================================================
+CREATE TABLE memory (
+    id             TEXT PRIMARY KEY,
+    profile_id     TEXT NOT NULL REFERENCES profile(id) ON DELETE CASCADE,
+    kind           TEXT NOT NULL CHECK (kind IN ('identity','semantic','episodic')),
+    text           TEXT NOT NULL,
+    summary        TEXT,
+    embedding_ref  TEXT NOT NULL,          -- pointer into vector store
+    confidence     REAL NOT NULL CHECK (confidence BETWEEN 0 AND 1),
+    created_at     TEXT NOT NULL,
+    superseded_by  TEXT REFERENCES memory(id)
+);
+CREATE INDEX idx_memory_profile ON memory(profile_id);
+CREATE INDEX idx_memory_active ON memory(profile_id, superseded_by) WHERE superseded_by IS NULL;
+
+CREATE TABLE capability (
+    id            TEXT PRIMARY KEY,
+    profile_id    TEXT NOT NULL REFERENCES profile(id) ON DELETE CASCADE,
+    name          TEXT NOT NULL,
+    description   TEXT NOT NULL,
+    confidence    REAL NOT NULL CHECK (confidence BETWEEN 0 AND 1),
+    recency       TEXT NOT NULL CHECK (recency IN ('recent','stable','dormant'))
+);
+
+CREATE TABLE topic (
+    id            TEXT PRIMARY KEY,
+    profile_id    TEXT NOT NULL REFERENCES profile(id) ON DELETE CASCADE,
+    label         TEXT NOT NULL,
+    description   TEXT
+);
+
+CREATE TABLE topic_member (
+    topic_id      TEXT NOT NULL REFERENCES topic(id) ON DELETE CASCADE,
+    member_ref    TEXT NOT NULL,
+    PRIMARY KEY (topic_id, member_ref)
+);
+
+CREATE TABLE capability_topic (
+    capability_id TEXT NOT NULL REFERENCES capability(id) ON DELETE CASCADE,
+    topic_id      TEXT NOT NULL REFERENCES topic(id) ON DELETE CASCADE,
+    PRIMARY KEY (capability_id, topic_id)
+);
+
+CREATE TABLE case_record (
+    id            TEXT PRIMARY KEY,
+    profile_id    TEXT NOT NULL REFERENCES profile(id) ON DELETE CASCADE,
+    title         TEXT NOT NULL,
+    summary       TEXT NOT NULL,
+    outcome       TEXT,
+    confidence    REAL NOT NULL CHECK (confidence BETWEEN 0 AND 1),
+    status        TEXT NOT NULL CHECK (status IN ('candidate','confirmed','dismissed')) DEFAULT 'candidate'
+);
+
+CREATE TABLE case_capability (
+    case_id        TEXT NOT NULL REFERENCES case_record(id) ON DELETE CASCADE,
+    capability_id  TEXT NOT NULL REFERENCES capability(id) ON DELETE CASCADE,
+    PRIMARY KEY (case_id, capability_id)
+);
+
+-- ============================================================================
+-- Provenance — links any compilation-layer object to its evidence
+-- ============================================================================
+CREATE TABLE provenance (
+    id            TEXT PRIMARY KEY,
+    target_id     TEXT NOT NULL,            -- memory/capability/case/edge id
+    target_kind   TEXT NOT NULL CHECK (target_kind IN ('memory','capability','case','edge')),
+    ref           TEXT NOT NULL,            -- document or event id
+    span_start    INTEGER,
+    span_end      INTEGER,
+    weight        REAL NOT NULL CHECK (weight BETWEEN 0 AND 1)
+);
+CREATE INDEX idx_provenance_target ON provenance(target_id, target_kind);
+CREATE INDEX idx_provenance_ref ON provenance(ref);
+
+-- ============================================================================
+-- Knowledge graph (projected from capabilities/topics/cases/assets/sources)
+-- ============================================================================
+CREATE TABLE graph_node (
+    id            TEXT PRIMARY KEY,
+    profile_id    TEXT NOT NULL REFERENCES profile(id) ON DELETE CASCADE,
+    kind          TEXT NOT NULL CHECK (kind IN ('capability','topic','case','asset','source')),
+    ref           TEXT NOT NULL,            -- id of the underlying object
+    label         TEXT NOT NULL,
+    weight        REAL NOT NULL DEFAULT 1.0
+);
+CREATE INDEX idx_node_profile ON graph_node(profile_id);
+
+CREATE TABLE graph_edge (
+    id            TEXT PRIMARY KEY,
+    profile_id    TEXT NOT NULL REFERENCES profile(id) ON DELETE CASCADE,
+    from_node     TEXT NOT NULL REFERENCES graph_node(id) ON DELETE CASCADE,
+    to_node       TEXT NOT NULL REFERENCES graph_node(id) ON DELETE CASCADE,
+    kind          TEXT NOT NULL,
+    weight        REAL NOT NULL DEFAULT 1.0
+);
+CREATE INDEX idx_edge_profile ON graph_edge(profile_id);
+CREATE INDEX idx_edge_from ON graph_edge(from_node);
+CREATE INDEX idx_edge_to ON graph_edge(to_node);
+
+-- ============================================================================
+-- Provider config (per-profile model assignments — see ADR 0004)
+-- ============================================================================
+CREATE TABLE provider_config (
+    profile_id    TEXT NOT NULL REFERENCES profile(id) ON DELETE CASCADE,
+    role          TEXT NOT NULL,            -- generator | embedder | reranker
+    provider      TEXT NOT NULL,            -- bedrock | ollama | ...
+    config        TEXT NOT NULL,            -- JSON
+    PRIMARY KEY (profile_id, role)
+);
+
+-- ============================================================================
+-- Schema version (Yoyo-style: just a record of applied migrations)
+-- ============================================================================
+CREATE TABLE schema_migration (
+    version       INTEGER PRIMARY KEY,
+    name          TEXT NOT NULL,
+    applied_at    TEXT NOT NULL
+);
+INSERT INTO schema_migration (version, name, applied_at) VALUES (1, '0001_init', strftime('%Y-%m-%dT%H:%M:%fZ','now'));

kiwime_store-0.1.0/migrations/0002_tombstone_and_id_prefixes.sql

@@ -0,0 +1,73 @@
+-- 0002 — Tombstones for soft-delete + id-prefix CHECK constraints.
+--
+-- Why: connectors can re-emit a record set after the source has dropped some
+-- items (file deleted on disk, GitHub issue closed and removed from scope,
+-- Notion page archived). We need to mark them gone without losing provenance
+-- that already cites them. Sweep with ``deleted_at IS NOT NULL`` later.
+--
+-- Why prefix CHECKs: docs/schema/core.md §IDs makes a hard contract that every
+-- id is URL-safe and prefixed. We have been trusting connectors to do the
+-- right thing; lock it at the database boundary so a buggy connector fails
+-- fast at insert time instead of producing a wrong-shaped ``ref`` that tools
+-- elsewhere have to defensively parse. Asset/relation prefixes (ast_/rel_)
+-- are added here because they were missing from the original docs table.
+
+PRAGMA foreign_keys = ON;
+
+-- ---------------------------------------------------------------------------
+-- Tombstone columns. Existing rows get NULL (i.e. live).
+-- ---------------------------------------------------------------------------
+ALTER TABLE document          ADD COLUMN deleted_at TEXT;
+ALTER TABLE activity_event    ADD COLUMN deleted_at TEXT;
+ALTER TABLE asset             ADD COLUMN deleted_at TEXT;
+ALTER TABLE relation_candidate ADD COLUMN deleted_at TEXT;
+
+-- Indexes that match how the compiler scans live rows. Without these, every
+-- compile run would do a full scan and filter ``deleted_at IS NULL`` in app
+-- code; with them, SQLite uses a partial index and skips the tombstoned rows.
+CREATE INDEX idx_document_live ON document(source_id) WHERE deleted_at IS NULL;
+CREATE INDEX idx_event_live ON activity_event(source_id, occurred_at) WHERE deleted_at IS NULL;
+CREATE INDEX idx_asset_live ON asset(source_id) WHERE deleted_at IS NULL;
+CREATE INDEX idx_relation_live ON relation_candidate(from_ref) WHERE deleted_at IS NULL;
+
+-- ---------------------------------------------------------------------------
+-- ID-prefix CHECK constraints.
+--
+-- SQLite can't ALTER TABLE ADD CONSTRAINT, so for tables where we want to
+-- enforce a prefix we rebuild the table. We do this for the *new* tables
+-- (asset, relation_candidate) where prefix conventions weren't enforced
+-- before; the older tables (document, activity_event, source) are left as-is
+-- to avoid a costly rebuild — connectors already prefix their ids correctly,
+-- and the contract is documented. New tables built from here on must include
+-- the CHECK at creation time.
+-- ---------------------------------------------------------------------------
+
+CREATE TABLE asset_new (
+    id            TEXT PRIMARY KEY CHECK (id LIKE 'ast_%'),
+    source_id     TEXT NOT NULL REFERENCES source(id) ON DELETE CASCADE,
+    kind          TEXT NOT NULL,
+    name          TEXT NOT NULL,
+    description   TEXT,
+    metadata      TEXT NOT NULL,
+    deleted_at    TEXT
+);
+INSERT INTO asset_new (id, source_id, kind, name, description, metadata, deleted_at)
+SELECT id, source_id, kind, name, description, metadata, deleted_at FROM asset;
+DROP TABLE asset;
+ALTER TABLE asset_new RENAME TO asset;
+CREATE INDEX idx_asset_live ON asset(source_id) WHERE deleted_at IS NULL;
+
+CREATE TABLE relation_candidate_new (
+    id            TEXT PRIMARY KEY CHECK (id LIKE 'rel_%'),
+    from_ref      TEXT NOT NULL,
+    to_ref        TEXT NOT NULL,
+    kind          TEXT NOT NULL,
+    confidence    REAL NOT NULL CHECK (confidence BETWEEN 0 AND 1),
+    metadata      TEXT,
+    deleted_at    TEXT
+);
+INSERT INTO relation_candidate_new (id, from_ref, to_ref, kind, confidence, metadata, deleted_at)
+SELECT id, from_ref, to_ref, kind, confidence, metadata, deleted_at FROM relation_candidate;
+DROP TABLE relation_candidate;
+ALTER TABLE relation_candidate_new RENAME TO relation_candidate;
+CREATE INDEX idx_relation_live ON relation_candidate(from_ref) WHERE deleted_at IS NULL;

kiwime_store-0.1.0/migrations/0003_skill_profile.sql

@@ -0,0 +1,67 @@
+-- 0003 — Skill Profile + Skill Manager (B1).
+--
+-- Why: A2 (MCP server) + A3 (MCP client) gave kiwiMe a dual MCP role,
+-- but the *configuration* of those — which external MCP servers to ingest
+-- from, what model to compile with, what scope to apply — has no schema
+-- yet. This migration adds skill profiles so users can save and switch
+-- between named bundles of (model preference + MCP servers + source scope).
+--
+-- A "skill profile" is NOT the same as ``profile`` (which represents the
+-- person whose context is being compiled). One person has one ``profile``
+-- and many ``skill_profile`` rows ("Backend FDE skill", "Demo prep skill",
+-- "Personal-research skill"). At compile / export time the app picks
+-- exactly one skill_profile to drive provider selection and MCP-source
+-- inclusion.
+
+PRAGMA foreign_keys = ON;
+
+-- ---------------------------------------------------------------------------
+-- Skill profile: a named bundle the user can switch between.
+-- ---------------------------------------------------------------------------
+CREATE TABLE IF NOT EXISTS skill_profile (
+    id              TEXT PRIMARY KEY CHECK (id LIKE 'skp_%'),
+    profile_id      TEXT NOT NULL REFERENCES profile(id) ON DELETE CASCADE,
+    name            TEXT NOT NULL,
+    description     TEXT,
+    -- ``model_pref`` is JSON: ``{"generator": {"provider": "ollama", ...},
+    -- "embedder": {...}}``. Mirrors provider_config but per-skill, so a user
+    -- can have one skill that uses Bedrock and another Ollama without
+    -- editing global state.
+    model_pref      TEXT NOT NULL DEFAULT '{}',
+    -- ``scope`` is JSON: ``{"include_sources": [source_id, ...],
+    -- "exclude_topics": [topic_id, ...]}``. Empty/missing keys mean
+    -- "no restriction".
+    scope           TEXT NOT NULL DEFAULT '{}',
+    is_active       INTEGER NOT NULL DEFAULT 0,
+    created_at      TEXT NOT NULL,
+    updated_at      TEXT NOT NULL,
+    UNIQUE (profile_id, name)
+);
+CREATE INDEX IF NOT EXISTS idx_skill_profile_profile ON skill_profile(profile_id);
+-- Partial unique index: at most one active skill per profile. This is
+-- the one place the application *requires* exclusivity, so enforce it
+-- at the DB instead of trusting app code to dedupe.
+CREATE UNIQUE INDEX IF NOT EXISTS idx_skill_profile_active
+    ON skill_profile(profile_id) WHERE is_active = 1;
+
+-- ---------------------------------------------------------------------------
+-- MCP servers attached to a skill: third-party MCP endpoints to ingest from.
+-- These map 1:1 to MCPClientConnector install configs (A3).
+-- ---------------------------------------------------------------------------
+CREATE TABLE IF NOT EXISTS skill_mcp_server (
+    id              TEXT PRIMARY KEY CHECK (id LIKE 'mcp_%'),
+    skill_id        TEXT NOT NULL REFERENCES skill_profile(id) ON DELETE CASCADE,
+    display_name    TEXT NOT NULL,
+    command         TEXT NOT NULL,
+    args            TEXT NOT NULL DEFAULT '[]',           -- JSON array
+    -- ``env_secret_refs``: JSON ``{ENV_NAME: {namespace, key}}``. Resolved
+    -- against the secret worker at spawn time so cleartext tokens never
+    -- live in this table. Cleartext env (non-secret values like a custom
+    -- endpoint URL) goes in ``env_plain``.
+    env_plain       TEXT NOT NULL DEFAULT '{}',           -- JSON {str: str}
+    env_secret_refs TEXT NOT NULL DEFAULT '{}',           -- JSON
+    enabled         INTEGER NOT NULL DEFAULT 1,
+    created_at      TEXT NOT NULL,
+    UNIQUE (skill_id, display_name)
+);
+CREATE INDEX IF NOT EXISTS idx_skill_mcp_skill ON skill_mcp_server(skill_id);

kiwime_store-0.1.0/migrations/0004_skill_definition.sql

@@ -0,0 +1,72 @@
+-- 0004 — Skill definitions (importable / AI-generated SKILL.md content).
+--
+-- Why this exists separately from ``skill_profile``:
+--   ``skill_profile`` (0003) is a *configuration bundle* — model pref +
+--   external MCP servers + scope. ``skill_definition`` is the actual
+--   transferable artifact: a SKILL.md document (frontmatter + body) plus
+--   any referenced resource files. Profiles can optionally point at one
+--   definition to "load" it, but a definition is meaningful on its own —
+--   you can have 30 imported skills and only one active profile.
+--
+-- Source kinds:
+--   - ``file_import`` — user picked a single SKILL.md / *.skill.md file.
+--   - ``path_scan``   — recursively scanned a directory and imported each
+--                       SKILL.md found there.
+--   - ``ai_generated``— produced by an in-app generator chat using the
+--                       active profile's model preference.
+--   - ``builtin``     — shipped with the app; reserved, not used yet.
+--
+-- Format follows the emerging "Anthropic Claude Skills" / Cursor rules
+-- convention: a YAML frontmatter block (name, description, optional
+-- allowed-tools / model / etc) followed by markdown body.
+--
+-- Dedup is by (profile_id, content_sha256). Re-importing the same file
+-- updates ``imported_at`` rather than creating a duplicate row.
+
+PRAGMA foreign_keys = ON;
+
+CREATE TABLE IF NOT EXISTS skill_definition (
+    id              TEXT PRIMARY KEY CHECK (id LIKE 'skd_%'),
+    profile_id      TEXT NOT NULL REFERENCES profile(id) ON DELETE CASCADE,
+    name            TEXT NOT NULL,
+    description     TEXT,
+    source_kind     TEXT NOT NULL CHECK (source_kind IN
+                        ('file_import', 'path_scan', 'ai_generated', 'builtin')),
+    -- Original location (file path, scan root, or chat session id). Kept
+    -- for traceability and so a "re-import from disk" UX can find the
+    -- file again without nagging the user for the path.
+    source_uri      TEXT,
+    -- Markdown body (without frontmatter). Frontmatter goes in its own
+    -- column as JSON so we can query it (e.g. "skills allowing Read tool").
+    content_md      TEXT NOT NULL,
+    frontmatter     TEXT NOT NULL DEFAULT '{}',  -- JSON
+    -- Resource references (relative paths next to the SKILL.md, copied
+    -- into the blob store for portability):
+    -- ``[{"path": "scripts/foo.py", "sha256": "...", "bytes": 123}]``
+    resources       TEXT NOT NULL DEFAULT '[]',  -- JSON array
+    -- sha256 of the canonical (frontmatter + body) text — the dedup key.
+    content_sha256  TEXT NOT NULL,
+    imported_at     TEXT NOT NULL,
+    updated_at      TEXT NOT NULL,
+    UNIQUE (profile_id, content_sha256)
+);
+
+CREATE INDEX IF NOT EXISTS idx_skill_definition_profile
+    ON skill_definition(profile_id);
+CREATE INDEX IF NOT EXISTS idx_skill_definition_source
+    ON skill_definition(source_kind);
+
+-- ``skill_profile_definition`` is a 0..1 association: a profile can load
+-- one definition. Modeled as a separate table (instead of ALTER TABLE
+-- ADD COLUMN) because SQLite has no ``ALTER TABLE ... IF NOT EXISTS``,
+-- so this keeps the migration safely idempotent. ``UNIQUE(profile_id)``
+-- enforces the at-most-one rule at the DB.
+CREATE TABLE IF NOT EXISTS skill_profile_definition (
+    profile_id      TEXT NOT NULL REFERENCES skill_profile(id) ON DELETE CASCADE,
+    definition_id   TEXT NOT NULL REFERENCES skill_definition(id) ON DELETE CASCADE,
+    attached_at     TEXT NOT NULL,
+    PRIMARY KEY (profile_id),
+    UNIQUE (profile_id)
+);
+CREATE INDEX IF NOT EXISTS idx_skill_profile_definition_def
+    ON skill_profile_definition(definition_id);

kiwime_store-0.1.0/migrations/0005_custom_mcp_source.sql

@@ -0,0 +1,61 @@
+-- 0005 — Custom MCP source (data source backed by an external MCP server).
+--
+-- A "custom MCP source" lets the user point kiwiMe at any process that
+-- speaks Model Context Protocol over stdio (e.g. ``npx mcp-server-foo``)
+-- and treat its ``resources/list`` + ``resources/read`` output as another
+-- corpus alongside GitHub / Notion / Drive / local files.
+--
+-- This is the *inbound* (mcp-client) half of kiwiMe's MCP dual-role —
+-- the *outbound* (mcp-server) half exposes our compiled assets the other
+-- way around and lives in a different module.
+--
+-- Scheduling
+--   ``sync_interval_minutes`` is the cadence; NULL means "manual only".
+--   ``next_sync_at`` is precomputed on insert/update and on each completed
+--   sync, so the supervisor cron can do a cheap range scan over the
+--   partial index instead of re-deriving deadlines for every row.
+--
+-- Secrets
+--   ``env_secret_refs`` stores opaque pointers (``{"VAR":{"namespace":...,
+--   "key":...}}``) into the secret worker. The plain ``env_plain`` map is
+--   for non-sensitive values (URLs, flags). Both get merged at spawn time
+--   by the connector runtime, so tokens never sit in this DB.
+--
+-- Why a separate table (vs. reusing ``source``)
+--   ``source`` is the unified surface the UI / compile pipeline sees, but
+--   it doesn't know how to spawn child processes. We mirror each row into
+--   ``source`` (with kind='custom_mcp') from the worker layer; this table
+--   owns the spawn config + cron state.
+
+PRAGMA foreign_keys = ON;
+
+CREATE TABLE IF NOT EXISTS custom_mcp_source (
+    id                       TEXT PRIMARY KEY CHECK (id LIKE 'cms_%'),
+    profile_id               TEXT NOT NULL REFERENCES profile(id) ON DELETE CASCADE,
+    display_name             TEXT NOT NULL,
+    command                  TEXT NOT NULL,
+    args                     TEXT NOT NULL DEFAULT '[]',   -- JSON array of strings
+    env_plain                TEXT NOT NULL DEFAULT '{}',   -- JSON {VAR: value}
+    env_secret_refs          TEXT NOT NULL DEFAULT '{}',   -- JSON {VAR: {namespace, key}}
+    -- NULL = manual sync only. >0 = run every N minutes via supervisor cron.
+    sync_interval_minutes    INTEGER,
+    last_sync_at             TEXT,
+    next_sync_at             TEXT,
+    state                    TEXT NOT NULL DEFAULT 'active'
+                                CHECK (state IN ('active', 'paused', 'error')),
+    last_error               TEXT,
+    created_at               TEXT NOT NULL,
+    updated_at               TEXT NOT NULL,
+    UNIQUE (profile_id, display_name)
+);
+
+-- Partial index: the cron loop only cares about active rows that actually
+-- have a deadline. Skipping paused/error/manual rows here keeps the scan
+-- proportional to "how many timed sources are configured" rather than to
+-- the full table.
+CREATE INDEX IF NOT EXISTS idx_custom_mcp_next_sync
+    ON custom_mcp_source(next_sync_at)
+    WHERE state = 'active' AND next_sync_at IS NOT NULL;
+
+CREATE INDEX IF NOT EXISTS idx_custom_mcp_profile
+    ON custom_mcp_source(profile_id);

kiwime_store-0.1.0/migrations/0006_vector_store.sql

@@ -0,0 +1,55 @@
+-- 0006_vector_store.sql
+--
+-- Real semantic-retrieval storage, replacing the placeholder `emb:{id}:{dim}`
+-- string the compiler used to write into `memory.embedding_ref`. See ADR 0007.
+--
+-- Two tables, by design:
+--
+--   vec_memory   -- sqlite-vec virtual table holding the actual float vectors.
+--                   vec0 fixes the dimension at table-creation time, so this
+--                   table is the canonical 768-dim store (nomic-embed-text,
+--                   the default embedder). Only vectors of the canonical
+--                   dimension live here and are KNN-searchable.
+--
+--   embedding_meta -- ordinary table recording, per target, which
+--                     provider/model/dimension produced the vector, plus a
+--                     `needs_reindex` flag. This is the source of truth for
+--                     "what produced this embedding" and survives even when a
+--                     vector cannot enter vec_memory because its dimension
+--                     differs from the canonical store (model switch). It is
+--                     also what a future reindex tool reads (Iteration 5).
+--
+-- The vec0 extension must be loaded on the connection before this migration
+-- runs; the store's Database loads it (see db.py). If the extension is
+-- unavailable the CREATE VIRTUAL TABLE fails loudly rather than silently
+-- degrading to keyword-only search.
+
+-- Canonical 768-dim vector store. memory_id is the TEXT primary key so we can
+-- upsert/delete by memory id and join back to the `memory` table.
+CREATE VIRTUAL TABLE vec_memory USING vec0(
+    memory_id TEXT PRIMARY KEY,
+    embedding float[768]
+);
+
+-- Per-embedding provenance + lifecycle. One row per (target_id, target_kind).
+-- target_kind is 'memory' today; the column leaves room for embedding other
+-- governed objects later without another migration.
+CREATE TABLE embedding_meta (
+    target_id      TEXT NOT NULL,
+    target_kind    TEXT NOT NULL DEFAULT 'memory'
+                       CHECK (target_kind IN ('memory')),
+    provider       TEXT NOT NULL,           -- e.g. 'ollama', 'openai_compat'
+    model          TEXT NOT NULL,           -- e.g. 'nomic-embed-text', 'bge-m3'
+    dimension      INTEGER NOT NULL,        -- vector length the model produced
+    in_vec_store   INTEGER NOT NULL DEFAULT 0
+                       CHECK (in_vec_store IN (0, 1)),  -- 1 iff present in vec_memory
+    needs_reindex  INTEGER NOT NULL DEFAULT 0
+                       CHECK (needs_reindex IN (0, 1)),  -- 1 iff dim != canonical
+    created_at     TEXT NOT NULL,
+    PRIMARY KEY (target_id, target_kind)
+);
+
+CREATE INDEX idx_embedding_meta_reindex
+    ON embedding_meta(needs_reindex) WHERE needs_reindex = 1;
+CREATE INDEX idx_embedding_meta_model
+    ON embedding_meta(provider, model);