@sanqianx/project-knowledge 5.0.18 → 5.0.20-rc.1

package/README.md

@@ -1,349 +1,349 @@
-<p align="center">
-  <img src="docs/assets/logo.svg" alt="project-knowledge" width="540">
-</p>
-
-<p align="center">
-  <strong>Local, Git-triggered knowledge for software projects.</strong><br>
-  Markdown is the source of truth; the local vector index is always derived and rebuildable.
-</p>
-
-## Install
-
-### npm
-
-Install the web product and start the local service:
-
-```bash
-npm install -g @sanqianx/project-knowledge@latest
-project-knowledge-integrations install
-project-knowledge
-```
-
-1. `npm install -g @sanqianx/project-knowledge@latest` installs the shell,
-   three pinned module dependencies and the bundled knowledge-engine.
-2. `project-knowledge-integrations install` is the one-time Integration Setup.
-   For every detected client (Claude Code / Codex / OpenCode) it installs two
-   independent capabilities, reported separately:
-   - **Knowledge Integration** — plugin / MCP / Skill so agents can query the
-     knowledge base;
-   - **Development Capture** — managed Bridge hooks / notify / plugin so the
-     agent conversations in real Git workspaces are captured into each
-     imported project's Development Conversation.
-   It also registers the host-level `project-knowledge` Bridge consumer.
-   ZCode is supported for Development Capture only; an unverified knowledge
-   plugin is never reported as installed. Use `install --ide zcode --capture-only`
-   or System Status → 外部 Agent 采集 in the web UI. Per-client repair/removal
-   preserves unrelated hooks and existing records. Failures include an operation
-   ID and persist to the local system diagnostic log.
-   No client UI needs to be opened; third-party hooks and configs are
-   preserved. Codex setup repairs this product's stale notify paths, keeps live
-   desktop notification wrappers and existing compatible Radar notifications,
-   and refuses to overwrite unknown third-party commands. Changed notify
-   settings are backed up beside `config.toml` without copying other settings.
-3. `project-knowledge` starts a fresh v5 local backend. Import projects again;
-   older registries, knowledge, sessions and unfinished tasks are not migrated.
-
-Notes:
-
-- Restart already-running Claude Code sessions once after installing
-  Development Capture — hooks load at session start. Only conversations after
-  setup are captured; history is never backfilled.
-- To narrow the setup: `--ide claude`, `--capture-only`, `--knowledge-only`;
-  to remove capture per client: `uninstall --capture-only`; to disable
-  capture globally (including the Bridge consumer): `disable-capture`.
-- To repair Codex capture after an npm or Codex desktop upgrade, use System
-  Status → 外部 Agent 采集 → 接入/修复, or
-  `project-knowledge-integrations install --ide codex --capture-only`.
-  Repeating setup does not add notification wrappers. Removing capture restores
-  the preserved notification chain and leaves captured records on disk.
-- Capture repair replaces the known old standalone Bridge entry with the
-  current product Bridge and keeps one capture hook per event. Claude Code
-  and ZCode mixed-group third-party hooks and knowledge-context hooks remain
-  intact. Duplicate or stale entries are reported as needing repair rather
-  than healthy; existing journals are not moved or deleted.
-
-The local service listens on `127.0.0.1:5757` by default. The CLI can select a
-nearby free port and records the active loopback endpoint so managed Git Hooks
-do not depend on a hard-coded port.
-
-```bash
-project-knowledge              # start in the background and open the UI
-project-knowledge --fg         # run in the foreground
-project-knowledge status       # show the active local backend
-project-knowledge stop         # stop the background backend
-project-knowledge status --all # list verified local instances and data directories
-project-knowledge stop --port 5804 # stop a specific verified instance
-project-knowledge --data "D:/data/knowledge" # select an independent data directory
-project-knowledge --port 9000
-project-knowledge --no-open
-```
-
-Requires Node.js 22 or newer and Git on `PATH`.
-
-Commands such as `stop`, `STOP`, and `Stop` are equivalent. Unknown commands and
-invalid arguments fail without starting a service. Default start/status/stop are
-scoped to the selected data directory; stopping another directory requires an
-explicit port or `--data`. The CLI uses a local control channel, not a shared PID
-file or blind process termination. Stop succeeds only after the backend and its
-owned modules finish shutdown; independently started modules are preserved.
-An older or unverifiable instance is reported explicitly, never treated as safely
-stopped. Stop the previous running version before upgrading.
-
-### Windows desktop
-
-Version 5.0.0 delivers the web product through npm. A new Windows installer
-and desktop auto-update are not part of this release.
-
-## Runtime model
-
-The shell connects and projects; knowledge-engine owns the analysis pipeline:
-
-```text
-Git Hook → event-bridge journal → durable engine claim → isolated Agent
-         → staging validation → Markdown promotion → vector-hub → retrieval
-```
-
-Import establishes a Git tracking baseline, installs and verifies the managed
-Hook, and creates project metadata. It does not run AI, infer requirements from
-the repository, or generate placeholder knowledge. For an empty repository,
-the first later Commit is eligible for analysis.
-
-The Hook writes commit evidence to the bridge journal and only notifies the
-local service. Notifications may be lost; journal consumption and Git startup
-reconciliation recover eligible commits. A Hook failure never blocks a commit.
-Import does not trigger a costly historical analysis.
-
-Within a project, Commits run oldest-first and stop at the first failure.
-Different projects may analyze concurrently (maximum two). Duplicate events
-do not re-analyze completed commits. Closing the browser does not stop work.
-
-## Knowledge safety
-
-AI receives a frozen claim containing the Commit SHA, real patch evidence,
-bound requirement IDs, prompt hash, and fixed knowledge path. It can write only
-to an internal per-run staging directory. A manifest must validate before any
-file reaches the final project knowledge directory.
-
-Promotion uses verified hashes, backups, and a durable journal. Only after
-Markdown promotion succeeds does the analyzed-Commit pointer advance and mark
-the index dirty. Index failure never rolls back Markdown or reruns AI; it stays
-visible as dirty state and is retried.
-
-vector-hub owns the derived index; the shell no longer runs a second index
-writer. Failed indexing leaves promoted Markdown intact and repairable.
-
-Managed vectors follow the configured knowledge root at `<knowledge>/.indexes`.
-The system drawer displays the actual directory and pending wiring separately.
-Root changes leave project Markdown and old indexes untouched, drain existing
-index writes, switch subsequent writes, and rebuild registered knowledge with
-the configured embedding model. Credentials and small runtime metadata remain
-in the product data directory; no expensive historical Agent analysis is started.
-
-The application does not create, edit, refresh, or delete `CLAUDE.md`.
-
-## Requirement capture
-
-### Automatic development context
-
-Claude Code's bundled plugin has a synchronous `UserPromptSubmit` hook. It verifies the current local product instance and asks `/api/engine/context` for project/domain facts plus applicable confirmed project, domain and global memories before the model processes a normal request. No reminder to use MCP is required. The complete context is bounded to 8,000 characters; memories have an independent 3,000-character budget. MCP and the Skill remain available for deeper reads. Workbench uses the same engine endpoint and suppresses duplicate native injection.
-
-Knowledge Integration, Automatic Context and Development Capture are separate capabilities in System Status. An old or disabled plugin is not automatic-context readiness: npm upgrades do not silently upgrade third-party clients, and the existing integration repair action updates stale product plugins. Reopen external sessions after updating. The native-client candidate adds Codex plugin `UserPromptSubmit` and ZCode user-configuration process hooks. Codex requires native `/hooks` review and trust of the current definition; installation never writes trust or uses a bypass flag. Hook trust is read from the native app-server and checked for the selected repository. ZCode context and conversation capture are separate entries, preserving model and third-party configuration. OpenCode keeps its existing tool integration; automatic injection is not claimed for it.
-
-Context reading has a ten-second total client deadline and fails open with a visible warning. A vector failure is reported as file fallback, not vector success; an unknown project is ignored rather than borrowing a different project's context. `knowledge.context.prepared` and `agent.context.output` logs contain only metadata and citations, not the request or knowledge body. An output log does not itself prove model delivery; native-client acceptance must verify that separately. Internal analysis and Workbench-owned context suppress repeated client injection.
-
-Coding-agent integrations can append the user's real request before code is
-changed. Requirement capture writes bounded metadata to the selected project's
-`requirements.jsonl`; it does not start analysis or write knowledge.
-
-The MCP server exposes:
-
-- read-only `resolve`, `search`, `ask`, `get`, and `history` tools;
-- one write-only `record_requirement` metadata tool.
-
-```bash
-npx @sanqianx/project-knowledge@latest install
-npx @sanqianx/project-knowledge@latest install --ide claude
-npx @sanqianx/project-knowledge@latest install --ide codex
-npx @sanqianx/project-knowledge@latest install --ide opencode
-```
-
-Read-only CLI examples:
-
-```bash
-project-knowledge-kb search --project <projectId> --query "refresh token policy" --json
-project-knowledge-kb ask --project <projectId> --query "what was decided about login?"
-project-knowledge-kb get --project <projectId> --entry modules/auth.md --json
-project-knowledge-kb history --project <projectId> --json
-```
-
-Queries never create or rewrite configuration. If the internal index is
-unavailable, search falls back to current Markdown. Project, domain and brain
-scopes are nested; unconfirmed memory candidates are not task preferences.
-
-For engine-wired projects, the CLI and MCP delegate `search`, `ask`, `get` and
-`history` to the same engine used by the web UI, including commit provenance.
-They verify the running instance against the selected data directory; an
-unavailable or mismatched instance is an explicit failure, not a second local
-index with different facts. Only projects without engine wiring retain the
-local legacy read contract. Use `--scope domain` or `--scope brain` to broaden
-CLI searches; MCP accepts the equivalent `scope`. Shared-file reads accept
-`--space <spaceId>` (MCP: `space` or `spaceId`).
-
-## Storage
-
-Version 5 is a fresh start, not a data upgrade. Stop the previous process,
-remove any old `KB_DATA_DIR` override, configure models and import projects into
-new empty knowledge directories. Old files remain untouched but are not read.
-See [fresh v5 setup](docs/fresh-v5-start.md). Queued tasks can be cancelled by
-unbinding; only an active knowledge/index writer blocks removal.
-
-Default internal data directory: `~/.project-knowledge-v5/`
-
-Override it with `KB_DATA_DIR`:
-
-```powershell
-$env:KB_DATA_DIR = 'D:\data\project-knowledge'
-project-knowledge
-```
-
-The user selects a global knowledge root before importing projects. That root
-chooses directories for future imports, new shared scopes, and derived vector
-indexes (`<knowledge root>/.indexes`). Each imported project stores an
-absolute, fixed `knowledgePath`; changing the global root does not move existing
-projects.
-
-Existing project and shared-knowledge directories do not move when the root
-changes. Brain/domain directory bindings are retained in the shell settings
-after the engine acknowledges them. Existing engine bindings are reused when
-first recording those fields. A failed registration is shown as pending, not
-healthy. Old index directories are retained; indexes rebuild at the new root.
-Settings, registration metadata, caches, runtime state, logs, and recovery
-assets remain in the internal data directory; vector payloads do not:
-
-```text
-~/.project-knowledge-v5/
-├── settings.json
-├── projects.json                         # minimal project ID/order index
-├── projects/<projectId>/
-│   ├── config.json                       # fixed repoPath/knowledgePath
-│   ├── state.json                        # tracking, claim, index, Hook state
-│   └── requirements.jsonl                # created only when used
-├── knowledge-engine/                    # durable claims, runs and cursors
-├── cache/                                # models, exports, context packs
-├── runtime/                              # claims, staging, journals, locks
-├── logs/
-│   ├── app/
-│   ├── projects/<projectId>/
-│   └── hooks/
-└── recovery/                             # migrations, promotion/index backups
-
-<selected knowledge root>/
-├── .indexes/                            # rebuildable vector payloads
-├── _shared/<brain or domain ID>/         # fixed shared-knowledge bindings
-└── <project storage name>/
-    ├── README.md                         # when produced by verified knowledge
-    ├── GOAL.md
-    ├── ARCHITECTURE.md
-    ├── modules/*.md
-    └── changes/*.md
-```
-
-Team knowledge is an explicit exception: an import may bind an existing
-checked-out team-store subdirectory. Its final absolute path is still fixed in
-the project config, must remain inside the selected store, and is treated as
-externally owned when the project registration is deleted.
-
-## Migration
-
-The versioned `layout-v2` migration discovers legacy assets without modifying
-them, creates a journal and centralized recovery backup, stages settings and
-per-project metadata, validates preserved paths/secrets/pointers and any legacy
-index, then activates files with the minimal registry last. The completion
-marker is written only after the new stores reopen successfully.
-
-Interrupted or failed migration keeps the legacy reader path available and
-leaves source knowledge, old logs, configuration files, and backups intact for
-retry or diagnosis.
-
-## Product UI
-
-The six primary pages are Brain, Start work, Knowledge search, Long-term
-memory, Domains/projects and Activity. Start work embeds the real Workbench,
-not a second chat implementation. Import projects from Domains/projects;
-confirm both code and knowledge directories and retry only failed wiring.
-
-Domain knowledge and confirmed brain memories live alongside existing project
-knowledge. Adding a brain does not move project Markdown. Model credentials
-are owned by Workbench; maintenance has a persisted default independent of
-temporary chat selection. Runtime credentials never enter engine state.
-
-System status groups four-module health, configuration and diagnostics.
-Activity retains the structured-log capabilities:
-
-- trace/debug/info/warn/error/fatal filtering;
-- local-date, project, component, event, operation, Commit, and text filters;
-- newest-first cursor pagination, pause/auto-refresh, and filtered export;
-- operation-flow and structured error/stack details;
-- logger health/degraded state and Hook/index/project read-only status;
-- matching light/dark themes across the shell and native module frames.
-
-Logs use `log/v2` JSONL, daily files and bounded segments. Retention defaults to
-365 days; `0` disables time-based deletion. Capacity cleanup is deterministic.
-Secrets are recursively redacted during write, query, error serialization, and
-export. The log root is fixed under the internal data directory.
-
-## HTTP security
-
-The default loopback server accepts same-origin/loopback browser requests. It
-does not use wildcard CORS. Binding to a non-loopback address requires
-`KB_SITE_AUTH_TOKEN`, and requests must also pass Origin validation. Public AI
-profile responses expose only masked secret metadata; generic errors return an
-operation ID without a stack. There is no raw-file route.
-
-## Repository layout
-
-```text
-_site/server.js                       # thin executable entry
-_site/lib/server-app.js               # HTTP/runtime composition
-_site/lib/storage-layout.js           # canonical internal paths
-_site/lib/project-*-store.js          # v2 registry/config/state stores
-_site/lib/project-lifecycle-service.js
-_site/lib/commit-reconciler.js
-_site/lib/knowledge-promotion.js
-_site/lib/index-service.js
-_site/lib/structured-logger.js
-_site/lib/knowledge-tool-runtime.js
-_site/scripts/hook-trigger.js
-ui/index.html                         # unified production web UI
-ui/product.js, ui/product.css          # shell navigation and state projection
-_modules/knowledge-engine/            # verified dependency-free engine
-bin/project-knowledge*.js             # CLI and MCP entry points
-desktop/                              # Windows Electron package
-```
-
-## Testing
-
-```bash
-npm ci
-npm test -- --no-report
-npm test --prefix desktop
-npm pack --dry-run --json
-```
-
-The automated suite includes real Git Hooks, paths with spaces/non-ASCII text,
-online and offline Commit reconciliation, crash-lock recovery, atomic index
-replacement, migration fault injection, packaged-startup contracts, API
-security and log redaction/cursors. Product browser acceptance uses one
-1920×1080 desktop viewport and controlled model/embedding doubles; real-model
-and real-index smoke tests are reported separately. No mobile matrix or new
-Windows installer is claimed. Full crash-stage combinations and the final
-switch of a user's existing data require separate acceptance.
-
-See the [Leader / GLM Workbench playbook](docs/leader-glm-workbench-playbook.md)
-for task dispatch and independent review, and the
-[5.0.16 acceptance record](docs/release-5.0.16-acceptance.md) for this round's
-actual results and remaining gates. These documents are not a release approval.
-
-## License
-
-[Apache-2.0](LICENSE). See [NOTICE](NOTICE).
+<p align="center">
+  <img src="docs/assets/logo.svg" alt="project-knowledge" width="540">
+</p>
+
+<p align="center">
+  <strong>Local, Git-triggered knowledge for software projects.</strong><br>
+  Markdown is the source of truth; the local vector index is always derived and rebuildable.
+</p>
+
+## Install
+
+### npm
+
+Install the web product and start the local service:
+
+```bash
+npm install -g @sanqianx/project-knowledge@latest
+project-knowledge-integrations install
+project-knowledge
+```
+
+1. `npm install -g @sanqianx/project-knowledge@latest` installs the shell,
+   three pinned module dependencies and the bundled knowledge-engine.
+2. `project-knowledge-integrations install` is the one-time Integration Setup.
+   For every detected client (Claude Code / Codex / OpenCode) it installs two
+   independent capabilities, reported separately:
+   - **Knowledge Integration** — plugin / MCP / Skill so agents can query the
+     knowledge base;
+   - **Development Capture** — managed Bridge hooks / notify / plugin so the
+     agent conversations in real Git workspaces are captured into each
+     imported project's Development Conversation.
+   It also registers the host-level `project-knowledge` Bridge consumer.
+   ZCode is supported for Development Capture only; an unverified knowledge
+   plugin is never reported as installed. Use `install --ide zcode --capture-only`
+   or System Status → 外部 Agent 采集 in the web UI. Per-client repair/removal
+   preserves unrelated hooks and existing records. Failures include an operation
+   ID and persist to the local system diagnostic log.
+   No client UI needs to be opened; third-party hooks and configs are
+   preserved. Codex setup repairs this product's stale notify paths, keeps live
+   desktop notification wrappers and existing compatible Radar notifications,
+   and refuses to overwrite unknown third-party commands. Changed notify
+   settings are backed up beside `config.toml` without copying other settings.
+3. `project-knowledge` starts a fresh v5 local backend. Import projects again;
+   older registries, knowledge, sessions and unfinished tasks are not migrated.
+
+Notes:
+
+- Restart already-running Claude Code sessions once after installing
+  Development Capture — hooks load at session start. Only conversations after
+  setup are captured; history is never backfilled.
+- To narrow the setup: `--ide claude`, `--capture-only`, `--knowledge-only`;
+  to remove capture per client: `uninstall --capture-only`; to disable
+  capture globally (including the Bridge consumer): `disable-capture`.
+- To repair Codex capture after an npm or Codex desktop upgrade, use System
+  Status → 外部 Agent 采集 → 接入/修复, or
+  `project-knowledge-integrations install --ide codex --capture-only`.
+  Repeating setup does not add notification wrappers. Removing capture restores
+  the preserved notification chain and leaves captured records on disk.
+- Capture repair replaces the known old standalone Bridge entry with the
+  current product Bridge and keeps one capture hook per event. Claude Code
+  and ZCode mixed-group third-party hooks and knowledge-context hooks remain
+  intact. Duplicate or stale entries are reported as needing repair rather
+  than healthy; existing journals are not moved or deleted.
+
+The local service listens on `127.0.0.1:5757` by default. The CLI can select a
+nearby free port and records the active loopback endpoint so managed Git Hooks
+do not depend on a hard-coded port.
+
+```bash
+project-knowledge              # start in the background and open the UI
+project-knowledge --fg         # run in the foreground
+project-knowledge status       # show the active local backend
+project-knowledge stop         # stop the background backend
+project-knowledge status --all # list verified local instances and data directories
+project-knowledge stop --port 5804 # stop a specific verified instance
+project-knowledge --data "D:/data/knowledge" # select an independent data directory
+project-knowledge --port 9000
+project-knowledge --no-open
+```
+
+Requires Node.js 22 or newer and Git on `PATH`.
+
+Commands such as `stop`, `STOP`, and `Stop` are equivalent. Unknown commands and
+invalid arguments fail without starting a service. Default start/status/stop are
+scoped to the selected data directory; stopping another directory requires an
+explicit port or `--data`. The CLI uses a local control channel, not a shared PID
+file or blind process termination. Stop succeeds only after the backend and its
+owned modules finish shutdown; independently started modules are preserved.
+An older or unverifiable instance is reported explicitly, never treated as safely
+stopped. Stop the previous running version before upgrading.
+
+### Windows desktop
+
+Version 5.0.0 delivers the web product through npm. A new Windows installer
+and desktop auto-update are not part of this release.
+
+## Runtime model
+
+The shell connects and projects; knowledge-engine owns the analysis pipeline:
+
+```text
+Git Hook → event-bridge journal → durable engine claim → isolated Agent
+         → staging validation → Markdown promotion → vector-hub → retrieval
+```
+
+Import establishes a Git tracking baseline, installs and verifies the managed
+Hook, and creates project metadata. It does not run AI, infer requirements from
+the repository, or generate placeholder knowledge. For an empty repository,
+the first later Commit is eligible for analysis.
+
+The Hook writes commit evidence to the bridge journal and only notifies the
+local service. Notifications may be lost; journal consumption and Git startup
+reconciliation recover eligible commits. A Hook failure never blocks a commit.
+Import does not trigger a costly historical analysis.
+
+Within a project, Commits run oldest-first and stop at the first failure.
+Different projects may analyze concurrently (maximum two). Duplicate events
+do not re-analyze completed commits. Closing the browser does not stop work.
+
+## Knowledge safety
+
+AI receives a frozen claim containing the Commit SHA, real patch evidence,
+bound requirement IDs, prompt hash, and fixed knowledge path. It can write only
+to an internal per-run staging directory. A manifest must validate before any
+file reaches the final project knowledge directory.
+
+Promotion uses verified hashes, backups, and a durable journal. Only after
+Markdown promotion succeeds does the analyzed-Commit pointer advance and mark
+the index dirty. Index failure never rolls back Markdown or reruns AI; it stays
+visible as dirty state and is retried.
+
+vector-hub owns the derived index; the shell no longer runs a second index
+writer. Failed indexing leaves promoted Markdown intact and repairable.
+
+Managed vectors follow the configured knowledge root at `<knowledge>/.indexes`.
+The system drawer displays the actual directory and pending wiring separately.
+Root changes leave project Markdown and old indexes untouched, drain existing
+index writes, switch subsequent writes, and rebuild registered knowledge with
+the configured embedding model. Credentials and small runtime metadata remain
+in the product data directory; no expensive historical Agent analysis is started.
+
+The application does not create, edit, refresh, or delete `CLAUDE.md`.
+
+## Requirement capture
+
+### Automatic development context
+
+Claude Code's bundled plugin has a synchronous `UserPromptSubmit` hook. It verifies the current local product instance and asks `/api/engine/context` for project/domain facts plus applicable confirmed project, domain and global memories before the model processes a normal request. No reminder to use MCP is required. The complete context is bounded to 8,000 characters; memories have an independent 3,000-character budget. MCP and the Skill remain available for deeper reads. Workbench uses the same engine endpoint and suppresses duplicate native injection.
+
+Knowledge Integration, Automatic Context and Development Capture are separate capabilities in System Status. An old or disabled plugin is not automatic-context readiness: npm upgrades do not silently upgrade third-party clients, and the existing integration repair action updates stale product plugins. Reopen external sessions after updating. The native-client candidate adds Codex plugin `UserPromptSubmit` and ZCode user-configuration process hooks. Codex requires native `/hooks` review and trust of the current definition; installation never writes trust or uses a bypass flag. Hook trust is read from the native app-server and checked for the selected repository. ZCode context and conversation capture are separate entries, preserving model and third-party configuration. OpenCode keeps its existing tool integration; automatic injection is not claimed for it.
+
+Context reading has a ten-second total client deadline and fails open with a visible warning. A vector failure is reported as file fallback, not vector success; an unknown project is ignored rather than borrowing a different project's context. `knowledge.context.prepared` and `agent.context.output` logs contain only metadata and citations, not the request or knowledge body. An output log does not itself prove model delivery; native-client acceptance must verify that separately. Internal analysis and Workbench-owned context suppress repeated client injection.
+
+Coding-agent integrations can append the user's real request before code is
+changed. Requirement capture writes bounded metadata to the selected project's
+`requirements.jsonl`; it does not start analysis or write knowledge.
+
+The MCP server exposes:
+
+- read-only `resolve`, `search`, `ask`, `get`, and `history` tools;
+- one write-only `record_requirement` metadata tool.
+
+```bash
+npx @sanqianx/project-knowledge@latest install
+npx @sanqianx/project-knowledge@latest install --ide claude
+npx @sanqianx/project-knowledge@latest install --ide codex
+npx @sanqianx/project-knowledge@latest install --ide opencode
+```
+
+Read-only CLI examples:
+
+```bash
+project-knowledge-kb search --project <projectId> --query "refresh token policy" --json
+project-knowledge-kb ask --project <projectId> --query "what was decided about login?"
+project-knowledge-kb get --project <projectId> --entry modules/auth.md --json
+project-knowledge-kb history --project <projectId> --json
+```
+
+Queries never create or rewrite configuration. If the internal index is
+unavailable, search falls back to current Markdown. Project, domain and brain
+scopes are nested; unconfirmed memory candidates are not task preferences.
+
+For engine-wired projects, the CLI and MCP delegate `search`, `ask`, `get` and
+`history` to the same engine used by the web UI, including commit provenance.
+They verify the running instance against the selected data directory; an
+unavailable or mismatched instance is an explicit failure, not a second local
+index with different facts. Only projects without engine wiring retain the
+local legacy read contract. Use `--scope domain` or `--scope brain` to broaden
+CLI searches; MCP accepts the equivalent `scope`. Shared-file reads accept
+`--space <spaceId>` (MCP: `space` or `spaceId`).
+
+## Storage
+
+Version 5 is a fresh start, not a data upgrade. Stop the previous process,
+remove any old `KB_DATA_DIR` override, configure models and import projects into
+new empty knowledge directories. Old files remain untouched but are not read.
+See [fresh v5 setup](docs/fresh-v5-start.md). Queued tasks can be cancelled by
+unbinding; only an active knowledge/index writer blocks removal.
+
+Default internal data directory: `~/.project-knowledge-v5/`
+
+Override it with `KB_DATA_DIR`:
+
+```powershell
+$env:KB_DATA_DIR = 'D:\data\project-knowledge'
+project-knowledge
+```
+
+The user selects a global knowledge root before importing projects. That root
+chooses directories for future imports, new shared scopes, and derived vector
+indexes (`<knowledge root>/.indexes`). Each imported project stores an
+absolute, fixed `knowledgePath`; changing the global root does not move existing
+projects.
+
+Existing project and shared-knowledge directories do not move when the root
+changes. Brain/domain directory bindings are retained in the shell settings
+after the engine acknowledges them. Existing engine bindings are reused when
+first recording those fields. A failed registration is shown as pending, not
+healthy. Old index directories are retained; indexes rebuild at the new root.
+Settings, registration metadata, caches, runtime state, logs, and recovery
+assets remain in the internal data directory; vector payloads do not:
+
+```text
+~/.project-knowledge-v5/
+├── settings.json
+├── projects.json                         # minimal project ID/order index
+├── projects/<projectId>/
+│   ├── config.json                       # fixed repoPath/knowledgePath
+│   ├── state.json                        # tracking, claim, index, Hook state
+│   └── requirements.jsonl                # created only when used
+├── knowledge-engine/                    # durable claims, runs and cursors
+├── cache/                                # models, exports, context packs
+├── runtime/                              # claims, staging, journals, locks
+├── logs/
+│   ├── app/
+│   ├── projects/<projectId>/
+│   └── hooks/
+└── recovery/                             # migrations, promotion/index backups
+
+<selected knowledge root>/
+├── .indexes/                            # rebuildable vector payloads
+├── _shared/<brain or domain ID>/         # fixed shared-knowledge bindings
+└── <project storage name>/
+    ├── README.md                         # when produced by verified knowledge
+    ├── GOAL.md
+    ├── ARCHITECTURE.md
+    ├── modules/*.md
+    └── changes/*.md
+```
+
+Team knowledge is an explicit exception: an import may bind an existing
+checked-out team-store subdirectory. Its final absolute path is still fixed in
+the project config, must remain inside the selected store, and is treated as
+externally owned when the project registration is deleted.
+
+## Migration
+
+The versioned `layout-v2` migration discovers legacy assets without modifying
+them, creates a journal and centralized recovery backup, stages settings and
+per-project metadata, validates preserved paths/secrets/pointers and any legacy
+index, then activates files with the minimal registry last. The completion
+marker is written only after the new stores reopen successfully.
+
+Interrupted or failed migration keeps the legacy reader path available and
+leaves source knowledge, old logs, configuration files, and backups intact for
+retry or diagnosis.
+
+## Product UI
+
+The six primary pages are Brain, Start work, Knowledge search, Long-term
+memory, Domains/projects and Activity. Start work embeds the real Workbench,
+not a second chat implementation. Import projects from Domains/projects;
+confirm both code and knowledge directories and retry only failed wiring.
+
+Domain knowledge and confirmed brain memories live alongside existing project
+knowledge. Adding a brain does not move project Markdown. Model credentials
+are owned by Workbench; maintenance has a persisted default independent of
+temporary chat selection. Runtime credentials never enter engine state.
+
+System status groups four-module health, configuration and diagnostics.
+Activity retains the structured-log capabilities:
+
+- trace/debug/info/warn/error/fatal filtering;
+- local-date, project, component, event, operation, Commit, and text filters;
+- newest-first cursor pagination, pause/auto-refresh, and filtered export;
+- operation-flow and structured error/stack details;
+- logger health/degraded state and Hook/index/project read-only status;
+- matching light/dark themes across the shell and native module frames.
+
+Logs use `log/v2` JSONL, daily files and bounded segments. Retention defaults to
+365 days; `0` disables time-based deletion. Capacity cleanup is deterministic.
+Secrets are recursively redacted during write, query, error serialization, and
+export. The log root is fixed under the internal data directory.
+
+## HTTP security
+
+The default loopback server accepts same-origin/loopback browser requests. It
+does not use wildcard CORS. Binding to a non-loopback address requires
+`KB_SITE_AUTH_TOKEN`, and requests must also pass Origin validation. Public AI
+profile responses expose only masked secret metadata; generic errors return an
+operation ID without a stack. There is no raw-file route.
+
+## Repository layout
+
+```text
+_site/server.js                       # thin executable entry
+_site/lib/server-app.js               # HTTP/runtime composition
+_site/lib/storage-layout.js           # canonical internal paths
+_site/lib/project-*-store.js          # v2 registry/config/state stores
+_site/lib/project-lifecycle-service.js
+_site/lib/commit-reconciler.js
+_site/lib/knowledge-promotion.js
+_site/lib/index-service.js
+_site/lib/structured-logger.js
+_site/lib/knowledge-tool-runtime.js
+_site/scripts/hook-trigger.js
+ui/index.html                         # unified production web UI
+ui/product.js, ui/product.css          # shell navigation and state projection
+_modules/knowledge-engine/            # verified dependency-free engine
+bin/project-knowledge*.js             # CLI and MCP entry points
+desktop/                              # Windows Electron package
+```
+
+## Testing
+
+```bash
+npm ci
+npm test -- --no-report
+npm test --prefix desktop
+npm pack --dry-run --json
+```
+
+The automated suite includes real Git Hooks, paths with spaces/non-ASCII text,
+online and offline Commit reconciliation, crash-lock recovery, atomic index
+replacement, migration fault injection, packaged-startup contracts, API
+security and log redaction/cursors. Product browser acceptance uses one
+1920×1080 desktop viewport and controlled model/embedding doubles; real-model
+and real-index smoke tests are reported separately. No mobile matrix or new
+Windows installer is claimed. Full crash-stage combinations and the final
+switch of a user's existing data require separate acceptance.
+
+See the [Leader / GLM Workbench playbook](docs/leader-glm-workbench-playbook.md)
+for task dispatch and independent review, and the
+[5.0.16 acceptance record](docs/release-5.0.16-acceptance.md) for this round's
+actual results and remaining gates. These documents are not a release approval.
+
+## License
+
+[Apache-2.0](LICENSE). See [NOTICE](NOTICE).