memorix 1.3.4 → 1.4.1

package/plugins/codex/memorix/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "memorix",
-  "version": "1.3.4",
+  "version": "1.4.1",
   "description": "Shared workspace memory for Codex and other AI coding agents.",
   "author": {
     "name": "AVIDS2",

package/plugins/copilot/memorix/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "memorix",
-  "version": "1.3.4",
+  "version": "1.4.1",
   "description": "Shared local workspace memory for GitHub Copilot CLI and other AI coding agents.",
   "author": {
     "name": "AVIDS2",

package/plugins/gemini/memorix/gemini-extension.json

@@ -1,6 +1,6 @@
 {
   "name": "memorix",
-  "version": "1.3.4",
+  "version": "1.4.1",
   "mcpServers": {
     "memorix": {
       "command": "memorix",

package/plugins/omp/memorix/package.json

@@ -1,6 +1,6 @@
 {
   "name": "memorix-omp-package",
-  "version": "1.3.4",
+  "version": "1.4.1",
   "description": "Memorix shared workspace memory package for Oh-my-Pi.",
   "license": "Apache-2.0",
   "keywords": [

package/plugins/openclaw/memorix/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "memorix",
-  "version": "1.3.4",
+  "version": "1.4.1",
   "description": "Shared workspace memory for OpenClaw and other AI coding agents.",
   "author": {
     "name": "AVIDS2",

package/plugins/pi/memorix/package.json

@@ -1,6 +1,6 @@
 {
   "name": "memorix-pi-package",
-  "version": "1.3.4",
+  "version": "1.4.1",
   "description": "Memorix shared workspace memory package for Pi coding agent.",
   "license": "Apache-2.0",
   "keywords": [

package/server.json

@@ -7,7 +7,7 @@
     "url": "https://github.com/AVIDS2/memorix",
     "source": "github"
   },
-  "version": "1.3.4",
+  "version": "1.4.1",
   "websiteUrl": "https://github.com/AVIDS2/memorix#readme",
   "icons": [
     {
@@ -22,7 +22,7 @@
     {
       "registryType": "npm",
       "identifier": "memorix",
-      "version": "1.3.4",
+      "version": "1.4.1",
       "runtimeHint": "npx",
       "packageArguments": [
         {

package/src/codegraph/current-facts.ts

@@ -129,6 +129,7 @@ function detectStaleProgressNotes(input: {
   now: Date;
 }): ProjectStaleNote[] {
   const candidates = [
+    'ACTIVE_WORK.md',
     'progress.txt',
     'docs/dev-log/progress.txt',
     'docs/memcode/dev-log/progress.txt',

package/docs/dev-log/progress.txt

@@ -1,388 +0,0 @@
-# Memorix Development Progress
-
-> Current handoff note for new agent sessions. Code and release facts outrank
-> older notes when they conflict.
-
-## Current State
-- Phase: 1.2.5 retrieval performance and reliability release candidate
-- Branch: `codex/1.2.5-retrieval-performance`
-- Last updated: 2026-07-27
-- Released baseline: `memorix@1.2.4`, tag `v1.2.4`, `origin/main` at
-  `24f3b9b`.
-- Goal: make ordinary retrieval deterministic and bounded without changing the
-  local SQLite + in-process Orama architecture.
-
-## 1.2.5 Retrieval Performance and Reliability
-- Retrieval now has three explicit profiles: `fast` is fully local, `balanced`
-  is the default vector-capable path without LLM query work, and `thorough`
-  explicitly opts into configured LLM query rewrite/reranking.
-- `OPENROUTER_API_KEY` remains an embedding fallback only. It activates the
-  memory LLM lane only when that lane explicitly names an OpenRouter provider
-  or base URL, preventing embedding-only configurations from adding hidden
-  network work to normal search.
-- Embedding, rerank, MCP search, and maintenance share one timeout helper that
-  clears watchdog timers after settlement. This removes the previously measured
-  avoidable CLI exit delay from a completed rerank path while preserving real
-  timeouts for stalled work.
-- Read-only memory CLI commands skip session-store and maintenance-target setup
-  while retaining the existing project/identity visibility boundary.
-- Local verification completed before release: TypeScript check; focused
-  timeout/config/search/CLI/SDK tests; a full serial suite that reached a
-  single stale Codex plugin-manifest version assertion, followed by its
-  dedicated passing recheck after the template was synced; production build;
-  package SDK benchmark; real CLI smoke; and an in-memory MCP client smoke
-  confirming `memorix_search.quality` and `fast` mode.
-- Current benchmark evidence on Windows/Node 22: a 1,000-record hot in-process
-  SDK lexical fixture measured specific lookup p50 44.603 ms / p95 96.085 ms
-  and broad-query p50 160.903 ms / p95 171.878 ms. This excludes CLI startup,
-  MCP transport, remote embeddings, and LLM work by design; the two query
-  shapes are reported separately rather than averaged into a marketing number.
-
-## 1.2.4 Persistent Memory Delivery
-- Continuation is now a bounded delivery projection over existing sessions and
-  durable observations, not a second handoff database. `TaskWorkset` keeps its
-  task lens (`bugfix`, `feature`, and so on) while independently recognizing a
-  continuation request.
-- A continuation Workset includes at most the latest meaningful completed
-  session summary plus three readable durable memory anchors. Each item is
-  visible in the existing `ContextReceipt`; ordinary tasks do not receive it.
-- `memorix resume "<task>"` is the CLI-equivalent continuation entry. The
-  existing MCP `memorix_project_context` receives the same behavior from the
-  user's task text, so the micro MCP profile stays at seven tools.
-- Setup-generated rules now use a single fallback call when MCP is genuinely
-  unavailable: `memorix resume` for prior work or `memorix context` for new
-  work. They explicitly forbid help probing, command enumeration, and broad
-  search loops before the first brief.
-- The visibility reader is carried into continuation retrieval, preserving
-  project aliases while failing closed for other agents' personal memory.
-- Continuation text preserves whole technical identifiers instead of emitting a
-  misleading partial token. MCP Project Context also establishes a short-lived
-  delivery boundary: a follow-up search/detail request does not re-send an
-  already represented record unless the caller names a missing fact or the user
-  explicitly asks for the underlying detail.
-- A read-only or "do not modify files" Autopilot task blocks `memorix_store`
-  by default. The only override is an explicit user request to persist a
-  record, represented by `overrideReadOnly: true`.
-- A real Codex user-path smoke, in a fresh Git fixture seeded only through the
-  public CLI, made one natural-language continuation request. Codex received
-  the bounded brief, used one specific Context Pack expansion, was stopped
-  from re-reading already-covered memory, made no `memorix_store` write, and
-  returned the correct read-only next step.
-- Completed local gates: TypeScript check, production build, focused
-  continuation/hook/CLI regressions, and the full suite (`237` files / `2677`
-  tests). A fresh packed install with optional native dependencies omitted used
-  Node SQLite to persist the same `memorix.db`; CLI resume retained
-  `AUTH_REFRESH_V2`; the micro MCP package surface exposed exactly seven tools
-  and enforced duplicate-retrieval plus read-only-write boundaries; and an
-  isolated Codex global setup wrote/activated the versioned Personal-marketplace
-  plugin with the matching skill guidance. Pending release gates: CI and
-  publish.
-
-## 1.2.3 Session Context Visibility Hotfix
-- A real installed-package CLI smoke found that a second coordination identity
-  could receive the first identity's personal observation in automatic session
-  context. The root cause was a reader-less core session-context call.
-- `getSessionContext()` now accepts an `ObservationReader`; CLI, MCP, and TUI
-  session entry points pass their resolved caller identity. Alias-backed
-  project-visible records remain readable, while personal/team records fail
-  closed.
-- Graph-neighbor routing is omitted whenever a reader is present because graph
-  relations are not yet visibility-scoped. Core, CLI, MCP, and alias regression
-  tests cover the boundary.
-
-## 1.1.13 Closeout
-- Codex setup now stamps plugin versions and doctor distinguishes bundle,
-  marketplace, installed/enabled state, declared hooks, and hook trust.
-- Root package publishing is isolated from internal `@memorix/*` workspaces.
-- Full tests, build, pack inspection, isolated MCP smoke, isolated Codex smoke,
-  and Ubuntu/Windows CI all passed before release.
-
-## 1.2 Accepted Direction
-- The [1.2 Product Story](../1.2.0-PRODUCT-STORY.md) is the canonical
-  user-facing narrative; the Development Charter is its implementation
-  contract.
-- The [Knowledge and Workflow Research](../1.2.0-KNOWLEDGE-WORKFLOW-RESEARCH.md)
-  establishes Knowledge Workspace and Workflow Inheritance as 1.2 core product
-  layers, with an explicit boundary against claiming the existing memory map is
-  already a semantic graph or maintained LLM Wiki.
-- The [1.2 Implementation Plan](../1.2.0-IMPLEMENTATION-PLAN.md) is the
-  execution contract for the single complete 1.2 goal. All phases are internal
-  gates in one continuous delivery path, not separate user approval points.
-- Use the [1.2 Development Charter](../1.2.0-DEVELOPMENT-CHARTER.md) as the
-  working product contract.
-- Keep the current SQLite CodeGraph tables and incremental refresh as the base;
-  do not discard working 1.1 Code Memory.
-- Add versioned Code State and evidence-quality semantics before new parsers,
-  external providers, or user-visible tool surfaces.
-- Treat natural-language `memorix_project_context` as the product surface;
-  internal graph/provider complexity must stay behind it.
-
-## Phase 0 Evaluation Foundation
-- Existing UI/API projections now identify themselves honestly as a read-only
-  Memory Overview and deterministic Memory Map. They are not presented as a
-  maintained wiki or semantic graph.
-- The Workset evaluation harness has fixture repositories for TypeScript,
-  Python, Go, docs-only, dirty worktree, deleted-symbol, and incomplete-scan
-  cases. Its seed manifest covers observations, Git facts, test evidence,
-  mini-skills, and explicit graph relations.
-- The harness compares memory-only, current-context, and candidate Worksets
-  against source/evidence/caution requirements and variant token ceilings.
-- The current Auto Context path is smoke-tested against real local TypeScript
-  and dirty-worktree fixtures with embeddings disabled. The candidate budget is
-  180 tokens; the current baseline budget is 320 tokens.
-- The former memory map remains separate from the formal Claim Ledger. Phase 2
-  owns claim provenance, conflict, and supersession semantics; Phase 0 only
-  prevents the old views from being mislabeled as those capabilities.
-
-## Phase 1 Code State
-- Lite refresh now appends a Code State Snapshot with bounded Git revision and
-  worktree metadata, a source epoch, a prior-snapshot link, and explicit scan
-  completeness.
-- Current files, symbols, edges, and current observation-code refs are marked
-  with the latest completed snapshot. Existing CodeGraph rows survive the
-  additive schema migration.
-- Project Context, the CodeGraph MCP status payload, and the codegraph CLI
-  status output expose the latest snapshot. A non-Git directory reports
-  unavailable instead of clean.
-
-## Phase 2 Claim Ledger
-- SQLite now has additive, transactional tables for claims, evidence
-  references, and lifecycle events. Every stored claim requires at least one
-  source reference and credential values are sanitized before durable writes.
-- Exact normalized assertions reuse their record and attach more evidence;
-  competing approved assertions remain disputed rather than being overwritten.
-  Evidence-aware supersession preserves an audit trail.
-- Explicit memories and Git-ingested facts can derive conservative claims.
-  Hook/model output does not become approved knowledge automatically.
-- Observation code references inherit the latest Code State Snapshot when they
-  are created. CodeGraph refresh requalifies bound claims: changed or deleted
-  symbols become unknown, and incomplete scans force review instead of false
-  certainty.
-- Unit and integration tests cover evidence validation, redaction, conflict,
-  supersession, model-review policy, snapshot binding, changed-symbol
-  requalification, and compact task selection.
-
-## Phase 3 Knowledge Workspace
-- Memorix now initializes a private local Markdown workspace by default. A
-  versioned workspace requires an explicit absolute path inside the Git project
-  and refuses ignored or unsafe target directories.
-- Approved active claims compile into Markdown proposals with frontmatter,
-  source claim ids, evidence references, snapshot context, a catalog, and an
-  append-only change log. Proposed pages are not silently published.
-- The knowledge init, status, compile, apply, and lint commands provide the
-  operator path. Applying a proposal protects a manual page edit by default;
-  an explicit force flag is the only overwrite path and is logged.
-- Workspace lint detects malformed frontmatter, broken links, orphan pages,
-  missing claims/evidence, superseded claims, unresolved conflicts, stale code
-  snapshots, and protected manual edits.
-- The compiled CLI was smoke-tested in an isolated Git project through init,
-  compile, apply, and lint with a source-backed claim.
-
-## 1.2 Delivery Complete, Released As 1.2.1
-- Phase 4 added canonical Markdown workflows, safe adapter preview/apply, import,
-  selection, and verification receipts. Agent instruction files are render
-  targets, not the workflow source of truth.
-- Phase 5 routes Project Context and Context Pack through one bounded Workset
-  assembler. It ranks current facts, claims, knowledge pages, workflows, and
-  cautions instead of dumping unrelated narrative memory.
-- Phase 6 makes refresh, claim derivation/requalification, knowledge
-  compile/lint, workflow indexing, retention, and consolidation durable
-  maintenance work. Interactive MCP requests do not scan an entire corpus.
-- Phase 7 adds an optional local semantic CodeGraph boundary. Lite remains the
-  durable local fallback; external output is used only from a healthy existing
-  `.codegraph` index, is validated, bounded, and never persisted as raw code.
-- The advanced `memorix_knowledge` MCP action tool now exposes explicit
-  workspace/proposal/workflow management only in `team` and `full` profiles.
-- Completed local release gates: lint, full test suite, production build,
-  package-content audit, isolated Windows npm install, stdio MCP smoke, and
-  HTTP dashboard/lifecycle smoke.
-- Release audit also incorporated #130 before publication: the optional SQLite
-  runtime now uses the Node-26-compatible `better-sqlite3` 12.x line, with a
-  dedicated Node 26 CI smoke that opens an in-memory database.
-- The release was completed as `memorix@1.2.1`. Its next development line is
-  [1.2.2 Memory Control Plane](../1.2.2-MEMORY-CONTROL-PLANE.md), which
-  consolidates these capabilities instead of adding another disconnected
-  memory feature.
-
-## 1.2.2 Control Plane
-- `TaskWorkset` already has bounded rendering and source-aware claims, but
-  SessionStart handoff and legacy context surfaces need one shared admission,
-  budget, and receipt contract.
-- First implementation slice complete: the shared `ContextReceipt` records
-  delivery target, selected source ids/kinds, trust/freshness, exact budget
-  omissions, queued refreshes, and Workset selection time without entering the
-  agent prompt. Project Context, Context Pack, and full SessionStart injection
-  now carry the same delivery-target semantics; `memorix explain` exposes the
-  receipt in text and JSON.
-- Budget selection now keeps a knowledge page dependent on a Claim out of the
-  prompt when its qualifying Claim could not fit. The receipt records both as
-  withheld instead of presenting an orphaned page path as usable knowledge.
-- Phase 3 admission slice complete: automatic observations now persist an
-  `ephemeral`, `candidate`, or `qualified` state and a reason. Hooks and the
-  internal orchestrator capture concrete evidence as candidates, keep routine
-  success/telemetry ephemeral, and drop low-signal activity without writing it.
-  Automatic capture stays quiet in the host agent.
-- A deduplicated `observation-qualify` maintenance job runs after CodeGraph
-  refresh. It requires a scan no older than the capture and at least one
-  current code reference before upgrading a candidate. Qualification does not
-  create a Claim; explicit saves and Git remain the Claim boundary.
-- Auto Project Context, automatic session L1 delivery, Wiki/Knowledge Graph
-  compilation, and graph-context packets exclude candidates and ephemeral
-  traces. Ordinary search prefers qualified material, while explicit lookup
-  can still inspect pending evidence.
-- Pending automatic evidence is also excluded from automatic consolidation so
-  a later qualification can still inspect its original source grain. Retention
-  explanations no longer describe an unqualified `core` candidate as immune.
-- Actor and visibility enforcement is now one policy shared by MCP retrieval,
-  auto context, CLI/TUI, dashboards, knowledge/graph projections, and scoped
-  maintenance. Legacy observations remain project-visible; new records may be
-  `project`, `team`, or `personal` with an explicit creator and recipients.
-  Missing identity never grants private or team visibility.
-- Automatic hooks and orchestrator traces begin personal. A verified hook
-  candidate may become project evidence only when qualification proves a
-  current code reference. Targeted handoffs are readable by sender and
-  recipient but writable only by sender; broadcasts require an active project
-  team member.
-- Automatic orchestration lesson lookup is explicitly project-scoped, so its
-  prompt injection cannot read personal or team-scoped records.
-- Topic-key writes now check mutation scope before storage, and the
-  project-attribution guard receives only memories visible to its current
-  reader. This closes both guessed-key overwrite and cross-project existence
-  hints without hiding historical project knowledge after upgrade.
-- The embedded SDK remains project-visible when unbound. CLI and the local TUI
-  now share the same explicit actor reader: an unbound terminal sees project
-  memory only, while `memorix identity join|use` (or one-shot `--as`) grants
-  only that active project member's personal/team visibility and coordination
-  ergonomics. `identity clear` returns to project scope; private evidence
-  cannot be promoted into shared skills or generated project skills.
-- CLI control-plane ergonomics now include `--cwd`, kebab-case flag aliases,
-  action-oriented namespace help, explicit `memorix workbench`, and
-  file/stdin transfer import/export. Root memory aliases use the same
-  canonical command paths as `memorix memory`; CLI and MCP exports are filtered
-  by the same active reader so an unbound backup cannot contain private/team
-  observations.
-- Direct CLI and foreground Code Memory refreshes now enqueue qualification as
-  well as claim requalification, and they register the local project target
-  needed by the isolated control-plane worker. The host hook path persists its
-  candidate before scheduling a refresh, avoiding a fast-worker race where a
-  scan could predate the capture.
-- Compiled-artifact smoke: an isolated Codex `PostToolUse` hook returned only
-  `continue: true`, stored a `candidate`, and the isolated background control
-  plane upgraded it to `qualified` after a real Code Memory scan with one
-  current reference. Embeddings were deliberately disabled for this
-  deterministic local smoke. The full test suite, TypeScript check, and
-  production build also passed after the admission policy was reconciled with
-  OpenCode's concise technical response events.
-- #135 remains a contributor-owned session-handoff candidate for the bounded
-  `HandoffPacket` phase. It is not copied or silently replaced.
-- #133/#134 remain open pending an official provider-capability layer and a
-  safe agent-facing media path. The existing OpenRouter text embedding path is
-  the supported baseline.
-- See [1.2.2 Memory Control Plane](../1.2.2-MEMORY-CONTROL-PLANE.md) for
-  phases, acceptance gates, non-goals, and release criteria.
-
-## 1.2.6 Pi Overlay Integrity
-- Replaced the stale Memcode TUI command list with one registry derived from
-  the executable Pi-compatible built-ins plus Memorix-native memory actions.
-  `/commands` and editor completion no longer advertise non-dispatched `/git`,
-  `/help`, `/config`, or synthetic session commands; nested `/memory `
-  completion is available in the live input path.
-- Made `/memory delete <id>` execute the intended project-memory resolution,
-  including fresh project-store initialization. The no-argument form remains a
-  safe list-and-select prompt.
-- Adopted Pi's current model-initiated bash session contract without changing
-  general harness behavior: Memcode supplies current `PI_*` session/model
-  metadata by default, removes stale inherited values, and lets an extension
-  opt out explicitly.
-- Added a pinned, read-only Pi upstream audit manifest and maintainer command.
-  Large upstream comparisons fall back from GitHub's capped file list to a
-  recursive official Git-tree diff before any manual adoption decision.
-- Completed the Node built-in SQLite compatibility path instead of treating it
-  as a degraded happy-path fallback: prepared statements now accept complete
-  row objects like `better-sqlite3`, and synchronous transactions implement
-  deferred, immediate, and exclusive entry points. This unblocks task claims,
-  coordination, and Knowledge Workspace review when the optional native module
-  is absent.
-- Release polish: bundled `memcode --version` resolves the verified public
-  Memorix package version; `memorix context --task "..."` no longer requires a
-  redundant positional argument; and every shipped versioned plugin manifest
-  is aligned with the public release version.
-- Final verification after the last Memcode change: `npm run lint`, the root
-  production build, the root test suite, and the complete Memcode suite all
-  passed. The Memcode suite reported 158 passing files / 1474 passing tests
-  (with its 5 opt-in skipped files / 43 skipped tests). Pi upstream audit
-  through `v0.82.1` also passed.
-- A fresh isolated installation of the final tarball verified `memorix` and
-  `memcode` version parity, all seven versioned plugin manifests, both Context
-  CLI forms, and a real stdio MCP handshake: 43 tools were discoverable and
-  `memorix_project_context` returned an Autopilot Brief.
-
-## 1.2.7 Native Context Continuity
-- Added a project-scoped SQLite checkpoint store that separates host-native
-  compaction evidence from Observations, claims, Knowledge Workspace pages,
-  and transcript storage. It captures preflight markers, native summaries
-  where the host exposes them, lifecycle-only completion where it does not,
-  idempotent retry handling, one-time delivery state, and archival.
-- Pi and Oh-my-Pi extensions now pass through their real `compactionEntry`
-  fields. Codex and Claude Code use their documented compact lifecycle paths
-  without fabricating a host summary or changing the host compactor.
-- Codex consumes one bounded checkpoint at compact `SessionStart`; Claude Code
-  consumes one on the next official `UserPromptSubmit` delivery point. Explicit
-  CLI/MCP continuation can carry at most one recent source-labelled checkpoint
-  across configured agents, while ordinary new tasks remain free of it.
-- Added `memorix checkpoint list|show|context|archive` and the full-profile
-  `memorix_compaction_checkpoint` MCP action. The default MCP profile remains
-  unchanged to protect agent context budgets.
-- Final verification passed: focused compaction, hook, Workset, CLI, and MCP
-  tests; the full root test suite; TypeScript check; production build; npm
-  package dry-run; an isolated installed-artifact Codex/CLI/MCP smoke; and an
-  isolated Claude Code handoff/continuation smoke. `npm publish --dry-run
-  --ignore-scripts` accepted the public 1.2.7 tarball; the full prepublish
-  build/test path is covered by the root release gates above.
-
-## 1.2.8 Vector Recovery and Codex Installation Clarity
-- Reworked vector backfill failure handling so transient provider or local
-  index failures leave one durable, diagnosable retry with bounded backoff
-  instead of consuming the maintenance retry budget in each new MCP process.
-  Healthy work clears stale diagnostics, revives an old failed vector job when
-  needed, and resolves superseded historical failure rows once recovery is
-  complete. Provider calls are batched while process-local Orama writes remain
-  serialized.
-- Corrected dashboard and HTTP control-plane vector reporting. A process that
-  does not own a hydrated MCP search index now states that vector status belongs
-  to the active MCP session rather than showing a misleading green `0 / 0`.
-- Corrected Codex install diagnostics: an enabled official Codex plugin owns
-  its `.mcp.json` `memorix serve` endpoint, while first-use hook approval is an
-  explicit host consent step rather than a broken installation. Plugin setup
-  removes only the legacy source-path `memorix` entry after successful plugin
-  installation; custom user-managed MCP entries are preserved.
-- Release verification: focused regression tests (96 assertions), `npm run
-  lint`, production build, and the full root test suite (all passed, 280s).
-  A live OpenRouter `qwen/qwen3-embedding-8b` batch returned two 4096d vectors.
-  An isolated tarball install reported CLI version `1.2.8`; its stdio MCP
-  handshake exposed 44 tools and completed `memorix_project_context`, and its
-  CLI `context --json` path returned a structured Workset. `npm pack --dry-run`
-  and `npm publish --dry-run --ignore-scripts` both accepted the final tarball
-  without manifest normalization warnings.
-
-## 1.3 Curated Long-Term Memory
-- Added an additive SQLite long-term-memory layer with explicit episodic,
-  semantic, and procedural kinds; independent scope and portability; source
-  evidence; and an auditable `candidate -> qualified -> approved ->
-  archived/superseded` lifecycle. Existing Observation, Claim, Wiki, Workflow,
-  Session, and Code State stores remain their own canonical sources.
-- Added the complete CLI lifecycle surface and extended the existing MCP store/
-  detail routes without increasing the default seven-tool micro profile. Task
-  Worksets expose at most three relevant approved or qualified records as short
-  `durable:<id>` references for progressive expansion.
-- Explicit `user + portable` manual/user-confirmed records can be reused across
-  local projects for the same owner. Code, Git, tests, sessions, workflows,
-  observations, claims, and team/project evidence remain project-bound.
-- Semantic durable retrieval is an optional cross-language/paraphrase fallback
-  after lexical selection. Provider initialization and the one batch request use
-  a 1.8-second no-retry budget; API request cancellation is propagated to the
-  remote fetch so an unavailable embedding service cannot retain a CLI process.
-- Verification in progress: focused lifecycle/CLI/MCP/embedding tests, a real
-  packed-package cross-project Chinese-to-English retrieval smoke, and an
-  isolated Claude Code plugin user test that returned the reviewed release
-  procedure without inspecting project files.