Skip to content

Latest commit

 

History

History
501 lines (451 loc) · 40.4 KB

File metadata and controls

501 lines (451 loc) · 40.4 KB

CLI Reference

This page lists the broader Vault Agent Memory CLI surface. Most people should not memorize it. Start with the small human surface, then let agents and scheduled jobs use the wider toolbox.

For the compact map, run:

vault guide
vault guide --intent faq
vault guide --mode agent
vault guide --mode maintenance

See also: docs/agent_first_usage.md.

Minimal Human Surface

Intent Command
Install or connect an agent vault quickstart
Read beginner FAQ vault guide --intent faq
See the small command map vault guide
Browse locally vault gui
Give agents one safe HTTP entrypoint vault gateway serve
Self-host a central memory server vault remote-server serve
Search memory vault search "query"
Propose memory safely vault remember "Title" --content "..." --reason "..."
Continue a task vault task start/update/handoff
Check health vault doctor / vault security doctor

Central Memory Station Surface

Use this surface when multiple agents, hosts, or devices share memory through a trusted central store. The older detailed commands remain available for scripts and debugging.

Intent Command
Start or orient vault start
Run the daily memory loop vault daily-loop run --write-report --json
Rebuild the latest daily-loop report without ingestion vault daily-loop report --refresh --write-report --json
Check central sync state vault memory-sync status --json
Submit candidate memory to the Supabase central inbox vault memory-sync push --central-store --title "..." --content "..." --reason "..."
Submit candidate memory to a self-host inbox vault memory-sync push --central-backend self-host --title "..." --content "..."
Pull central candidates into the local review queue vault memory-sync pull --central-backend supabase --apply --json
Pull self-host central candidates into local review vault memory-sync pull --central-backend self-host --apply --json
Dry-run one scheduled-sync pass vault memory-sync run-once --push-read-copy --push-central-store --pull-candidates --dry-run --json
Run one trusted scheduled-sync pass vault memory-sync run-once --push-read-copy --push-central-store --pull-candidates --apply --json
Preview candidate inbox migration between backends vault memory-sync migrate-candidates --direction self-host-to-supabase --json
Export reviewed memory snapshots to a bundle vault memory-sync export-snapshots --bundle ./vault-snapshots.json --json
Verify a snapshot bundle before DR import vault memory-sync verify-snapshots --bundle ./vault-snapshots.json --require-content --json
Import a snapshot bundle as review candidates vault memory-sync import-snapshots --bundle ./vault-snapshots.json --apply --json
Review candidates and conflicts vault memory-review inbox --json
Run conservative review automation vault memory-review run --mode conservative --json
Resolve a conflict explicitly vault memory-review resolve <conflict_id> --action keep_active --reason "..."
Inspect lifecycle pressure vault memory-lifecycle status --json
Run Dream as a report vault memory-lifecycle dream --mode report --json
Preview archive / cold-store work vault memory-lifecycle archive --json / vault memory-lifecycle cold-store --json
Run operations checks vault ops status --json / vault ops doctor / vault ops security --json

Agent And Operator Workflow

These commands are still useful, but they are mostly for agents, scripts, scheduled jobs, or explicit maintenance sessions.

Command Purpose
vault init Initialize a project vault
vault setup-agent / vault install-agent Run the agent installer wizard and optional Obsidian sync template generator
vault remember "Title" --content "..." --reason "..." Propose candidate memory for review
vault capture discover --pretty List likely session transcript exports without reading transcript contents
vault capture session codex-session.jsonl --pretty Preview candidate memories extracted from an agent session transcript
vault capture session codex-session.jsonl --write-candidates Write extracted session lessons into the candidate queue, not active knowledge
vault memory pipeline --search-dir sessions --write-candidates --write-report Run the automatic conversation-memory pipeline: discover transcripts, extract lessons, gate them, write candidates, and leave a compact ingestion receipt
vault task start "Ship feature" Create a Task Ledger working set for a long-running task without writing active knowledge
vault task start "Ship feature" --priority P1 --due-at 2026-07-01 Create a task with explicit urgency and a due/reminder timestamp
vault task update <task_id> --decision "..." --next-action "..." Update task-runtime state for handoff/resume; reusable lessons still go through candidates
vault task update <task_id> --priority P0 --due-at 2026-06-30 Reprioritize the current working set without promoting it to long-term memory
vault task handoff <task_id> Print a compact continuation note for the next agent/session
vault memory temporal status / vault memory temporal list --state past Inspect current, past, future, and timeless fact windows
vault candidates List pending candidate memories without dumping full raw content
vault promote <candidate_id> --confirm Promote reviewed candidate memory
vault compile Compile raw/ into SQLite and generated artifacts
vault compile --no-embed Compile without optional embedding generation; recommended for the most portable first run
vault search "query" Search the vault
vault search "query" --exclude-expired Search currently valid temporal facts while keeping past facts available through temporal list
vault map read <id> --lines 10-30 Read a bounded source range for citation
vault gui Start the local read-only Vault Console
vault gateway serve Start the token-protected HTTP Gateway for agent search, bounded read, and candidate-first memory submission
vault gateway health --json Return machine-readable Gateway readiness without starting the HTTP server
vault gateway openapi --json Return the stable Gateway HTTP contract for Coze, n8n, OpenClaw, and future remote adapters
vault gateway serve --auth-token "$VAULT_GATEWAY_TOKEN" Use a stable token instead of printing a generated token at startup
vault gateway serve --shutdown-timeout-seconds 30 Drain on SIGINT/SIGTERM: stop new requests and wait for active requests before exit
vault gateway serve --allow-shared-candidates Allow agents to submit shared-scope candidates; they still do not write active knowledge
vault remote-server health --json Check self-hosted remote-server readiness without starting the server
vault remote-server openapi --json Return the same Gateway contract with self-hosted remote-server metadata
vault remote-server serve --host 0.0.0.0 Start a self-hosted central memory entrypoint; requires VAULT_GATEWAY_TOKEN or --auth-token
vault security doctor Check local GUI/MCP security posture plus Supabase service-role and self-host Remote Server boundary warnings
vault remote status --json Offline check for local source-of-truth, Supabase read-copy/candidate-inbox setup, sync freshness hints, and Agent sharing policy files
vault remote smoke --agent-id remote-agent --query "deployment SOP" --json Verify Supabase remote reader credentials and the vault_search_readable RPC
vault remote doctor --agent-id remote-agent --query "deployment SOP" --json Diagnose the full Supabase remote reader path: search, readable-entry RPCs, Document Map, claims, content, map, and bounded read
vault remote search "query" --agent-id remote-agent --json Search the Supabase read-only memory view through vault_search_readable
vault remote map <id> --compact --json Inspect remote synced Document Map rows
vault remote read <id> --node-uid <node> --json Read remote bounded evidence from synced content/claims
vault remote submit-candidate --from-agent remote-agent --title "..." --content "..." --trust 0.8 --json Submit a remote candidate request through Supabase; does not write active knowledge
vault remote pull-candidates --apply --json Trusted sync host pulls submitted remote requests into local memory_candidates
vault remote pull-candidates --apply --auto-promote-low-risk --json Trusted sync host pulls remote requests and promotes only this pull's candidates that match automation_policy.yaml low-risk rules
vault remote hmac-keys --json Show active remote candidate HMAC key ids for rotation without printing secrets
vault sync revisions --json Inspect the local revision graph created by remote candidate imports and promotions
vault sync conflicts --json List open local sync conflicts, such as same-title remote candidates that differ from active reviewed knowledge
vault sync audit --json Inspect local sync audit events without reading private memory content
vault sync resolve-conflict <conflict_id> --resolution keep_local --agent-id work-agent Keep local knowledge and reject the conflicting remote candidate
vault sync resolve-conflict <conflict_id> --resolution accept_remote --apply-memory-change --agent-id work-agent Promote the remote candidate and archive the conflicting local row; old content stays auditable
vault remove <id> --confirm Remove a knowledge entry after reviewing its ID

Knowledge Ingestion

Command Purpose
vault add "Title" --content "..." Add one active knowledge entry directly
vault add "Title" --file note.md Add an entry from a Markdown file
vault import long-doc.md Import and chunk a long document
vault import memory --source ./chatbox-export.json --format auto --dry-run --json --pretty Preview an external memory export from Chatbox, ChatGPT, Claude, Markdown, JSON, CSV, transcript, or OKF
vault import memory --source ./chatbox-export.json --write-candidates --only summaries,decisions,preferences Write external memories into memory_candidates for review; never writes active knowledge directly
vault import okf --bundle ./okf-bundle --dry-run --json --pretty Preview candidate-first import from an OKF-style Markdown bundle
vault import okf --bundle ./okf-bundle --scope shared --owner-agent work-agent Write OKF concepts into memory_candidates for review, not active knowledge
vault import obsidian --vault /path/to/ObsidianVault --dry-run Preview importing existing Obsidian notes into raw/obsidian/
vault import obsidian --vault /path/to/ObsidianVault --compile Import changed Obsidian notes and compile them into vault.db
vault export obsidian --vault /path/to/ObsidianVault --dry-run --json Preview read-only Markdown notes for Obsidian browsing with machine-readable output
vault export obsidian --vault /path/to/ObsidianVault --include-graph-overview --include-review-inbox Write Vault Home, Graph Overview, and review inbox notes so Obsidian can act as the human GUI
vault okf validate ./okf-bundle --json --pretty Validate an OKF-style Markdown knowledge bundle before import, export, or publication
vault export okf --bundle ./okf-bundle --dry-run --json --pretty Preview a read-only OKF-style export; private and restricted memories are excluded by default
vault export okf --bundle ./okf-bundle --include-private --include-restricted Explicitly include protected memories when creating a private/internal bundle
vault export markdown --bundle ./memory-md --dry-run --json --pretty Preview a batch Markdown export with one active memory per file plus manifest.json
vault export json --bundle ./memory-json --json --pretty Write knowledge.json and manifest.json for machine-readable backup or migration work

Prefer vault remember over vault add for autonomous agents or unreviewed memory.

Prefer vault gateway serve when many local agents or scripts need the same memory access pattern. Gateway keeps the public surface small: /search, /read-range, /submit-candidate, /health, and /openapi.json. It requires agent_id, hides private memory by default, writes only candidates, and records compact audit rows in reports/gateway/audit.jsonl. Use vault gateway health --json in installers and CI when you only need to confirm the project vault is ready for Gateway clients. Use vault gateway openapi --json when an adapter needs the request/response shape without starting a server. Use vault remote-server serve when the user wants a self-hosted central memory host instead of Supabase. It reuses the Gateway contract, requires a stable token, and still writes remote contributions as review candidates rather than active memory.

Use vault import memory when moving memory from another product, an exported chat transcript, a JSON/CSV archive, or a folder of Markdown notes. The import is candidate-first: preview is the default, --write-candidates writes review candidates, and promotion into active knowledge still requires the normal review gates.

vault import obsidian skips .obsidian/, .trash/, .git/, and 00-Vault-Knowledge/ by default. The generated raw notes include obsidian_source_path and obsidian_source_hash, so repeated imports update changed notes without duplicating unchanged ones.

Agent Setup

Command Purpose
vault quickstart Small first-run wizard for agent-assisted users; asks only language, independent/shared memory, optional Obsidian/Supabase connections, and daily report time
vault quickstart --non-interactive --project ~/Vaults/my-project --json Script the consumer defaults without exposing advanced setup-agent flags
vault setup-agent Ask for scope, setup language, optional features, Obsidian import, sync templates, and generate agent-install/local-smoke.sh plus common Agent adapter startup templates
vault setup-agent --non-interactive --agent codex --scope shared --agent-project-dir ~/Vaults/my-project --features core,mcp,obsidian_import Agent-friendly scripted install
vault setup-agent --non-interactive --agent codex --scope shared --agent-project-dir ~/Vaults/my-project --memory-layout hybrid --features core,mcp --json Create a shared project vault plus private Agent vault and write hybrid-vault-layout.json
vault setup-agent --non-interactive --agent codex --scope shared --agent-project-dir ~/Vaults/my-project --features core,mcp,semantic,supabase,headroom --language en --install-optional-deps --install-embedding-model mix --supabase-setup simple --json Install selected optional dependencies and configure a local semantic model
vault setup-agent --non-interactive --agent maintenance-agent --scope shared --agent-project-dir ~/Vaults/my-project --features core,mcp,memory_agents --language zh-Hant --json Generate generic memory-maintenance guidance with report-only and candidate-only defaults
vault setup-agent --non-interactive --agent profile-agent --scope shared --agent-project-dir ~/Vaults/my-project --features core,mcp,supabase --language zh-Hant --install-optional-deps --supabase-setup simple --supabase-sync cron --remote-reader all --json Generate guided Supabase setup, daily sync templates, and shell/n8n/Coze remote reader templates
vault setup-agent --non-interactive --agent profile-agent --scope shared --agent-project-dir ~/Vaults/my-project --features core,mcp,supabase,memory_agents --agent-roster profile-agent:profile,work-agent:work,remote-agent:remote,n8n:automation --validation-pack all --json Generate a multi-agent access matrix plus live Supabase/n8n/Coze validation checklists
vault setup-agent --non-interactive --agent codex --scope shared --agent-project-dir ~/Vaults/my-project --features core,mcp,supabase --write-stable-venv-script --json Generate agent-install/setup-stable-venv.sh for reboot-safe scheduled jobs and MCP commands
vault setup-agent --non-interactive --agent automation-agent --scope shared --agent-project-dir ~/Vaults/my-project --features core,mcp,memory_agents --automation-schedule all --automation-mode balanced --automation-write-workspace --automation-include-transcripts --json Generate cron, LaunchAgent, and n8n templates for the candidate-first memory closed loop: pipeline, reflection, automation cycle, handoff, review-summary cards, and learning-health output
vault setup-agent --non-interactive --agent automation-agent --scope shared --agent-project-dir ~/Vaults/my-project --features core,mcp,memory_agents --automation-schedule cron --automation-apply --automation-auto-promote-low-risk --json Write automation_policy.yaml for low-risk session-capture auto-promotion and generate a scheduled cycle that can apply it
vault setup-agent --obsidian-vault ~/Documents/ObsidianVault --import-obsidian --obsidian-sync all Run first Obsidian import and write cron, LaunchAgent, and n8n templates
vault update-status Show installed Vault version, local Agent registry, project vaults, per-Agent update notices, and startup handoff commands
vault upgrade --check Check PyPI for a newer Vault release and print the correct pip, pipx, uv tool, or editable-install guidance without changing the environment
vault upgrade --check --json Return the same check-only upgrade plan as machine-readable JSON
vault upgrade --latest-version 0.10.2 Compare against a manually supplied version when PyPI is unavailable
vault update-status --check-pypi --write-status --json Check PyPI for the latest release and write the machine-level update notice to ~/.vault-for-llm/update-status.json
vault update-status --read-status --json Read the existing machine-level update notice without recomputing or contacting PyPI
vault update-status --read-status --agent codex --json Read the shared notice and add a focused startup checklist for one Agent/runtime
vault update-status --doctor --json Check whether the shared update notice exists, is fresh, includes every registered Agent, and shows runtime attention
vault agent register --agent codex --project ~/Vaults/my-project --scope shared Manually register an Agent/runtime in the local multi-agent registry
vault agent register --agent codex --project ~/Vaults/my-project --skills review-helper@1.0.0,task-helper Register which local skills this Agent expects, so update status can make skill usage visible
vault agent list List Agents registered on this machine
vault agent status --latest-version 0.7.21 Show the same registry/update status without contacting the network
vault agent doctor --json Run the same shared update-distribution health check through the Agent registry namespace
vault agent startup-doctor --template-dir ./agent-install --json Check generated startup contracts plus minimal Codex/Claude/Hermes/OpenClaw/Coze/n8n configs
vault agent install-runtime-template --runtime codex --target ./AGENTS.md Preview applying the generated Codex startup template into a runtime instruction file
vault agent install-runtime-template --runtime codex --target ./AGENTS.md --apply Apply the marked startup block and create a backup before changing an existing file

vault quickstart is the preferred first-run command for new users. It wraps the consumer setup path and deliberately hides the advanced feature matrix.

vault install-agent is an alias for vault setup-agent. Interactive setup asks before installing optional dependencies. Non-interactive agents must pass --install-optional-deps; semantic model download is opt-in with --install-embedding-model zh|en|mix. vault setup-agent also registers the Agent in ~/.vault-for-llm/agent-registry.json, so other local runtimes can discover the same project vault and startup handoff path through vault update-status. The status payload includes agent_update_notices; if one Agent updates Vault and writes status, other local runtimes can see whether their registered environment is behind the current or latest known version. Use vault agent doctor after upgrades to verify the shared notice is fresh and all registered runtimes appear in the update notice. vault upgrade is intentionally check-only in this release. It never invokes a package manager or migrates a database; review and run the recommended command, then use vault --version and vault doctor to verify the upgraded runtime. MCP-capable runtimes can call vault_update_status and vault_automation_handoff from the core profile for the same startup path. They can also call vault_update_status with doctor=true and their agent_id for the same update-distribution health check without adding another MCP tool. vault agent install-runtime-template is dry-run by default. It writes a marked Vault block only with --apply, replaces that block on later runs, and creates a .bak.<timestamp> backup for existing target files unless --no-backup is used. The default memory layout is hybrid: setup creates the shared project vault plus a private Agent vault under ~/Vaults/agents/<agent>/private-memory, then writes agent-install/hybrid-vault-layout.json. Use --write-stable-venv-script for the default long-lived venv path, or --stable-venv PATH when the user chooses a custom stable virtualenv. Memory-maintenance guidance is opt-in with the compatibility feature ID --features memory_agents; it writes README-memory-agents.md and does not install a model, auto-promote memory, or construct identity/personality/ relationship/life-phase/human models. Memory automation templates can opt into transcript ingestion with --automation-capture-transcripts; use it together with --automation-apply only when scheduled jobs are allowed to read discovered transcript files and write review candidates. Supabase sync templates are opt-in with --supabase-sync cron|launchagent|n8n|realtime|all and use python -m scripts.sync_to_supabase --db <project>/vault.db. The consolidated Central Memory Station worker is vault memory-sync run-once or python -m scripts.central_memory_sync; it writes reports/central-memory-sync-latest.json, can push the reviewed read copy with --push-read-copy, can write Central Memory Station tables with --push-central-store, can write reviewed safe-summary embeddings into the central pgvector table with --push-central-vectors, and can pull central candidates into local review with --pull-candidates --apply. Use --central-backend self-host when the candidate inbox is local vault-central.db behind Gateway / Remote Server instead of Supabase. For one-off Supabase candidate submissions, vault memory-sync push --central-store ... writes the new vault_memory_candidates_central inbox; omitting --central-store keeps the older guarded RPC path for compatibility. vault memory-sync migrate-candidates --direction supabase-to-self-host|self-host-to-supabase copies only pending Central Memory Station candidate inbox rows between backend adapters. It previews by default and requires --apply before writing the target inbox. It does not migrate active memory, local review candidates, or promotion decisions. vault memory-sync export-snapshots --bundle ./vault-snapshots.json writes a reviewed active-memory snapshot bundle. By default the bundle excludes raw memory content; --include-content should be used only on a trusted, encrypted transfer path. The bundle includes a manifest with snapshot count, snapshot digest, content policy, and metadata-only local revision/audit counts. vault memory-sync verify-snapshots --bundle ./vault-snapshots.json validates the bundle manifest and content hashes without writing memory; add --require-content before a disaster-recovery candidate import. vault memory-sync import-snapshots --bundle ./vault-snapshots.json previews by default and, with --apply, writes local memory_candidates only. It does not write active knowledge, does not promote candidates, and does not turn the target backend into a multi-master memory database. When setup-agent generates Supabase sync templates, it also writes central-memory-sync.cron, com.zycaskevin.vault-for-llm.central-memory-sync.plist, n8n-central-memory-sync.workflow.json, and README-central-memory-sync.md for the 30-60 minute Central Memory Station loop. Use realtime only when the local machine is trusted to hold a service-role key; it runs python -m scripts.watch_supabase_sync and pushes local changes after a short debounce. Supabase remote reader templates are opt-in with --remote-reader shell|n8n|coze|all; validate credentials with vault remote smoke --agent-id <agent> --json. Before live validation, run vault remote status --json from the project vault. It does not contact Supabase. It tells agents that the local SQLite vault is the source of truth, Supabase is a reviewed read copy plus candidate request inbox, near-realtime active-memory sync is still one-way local-to-Supabase push, and remote freshness is unknown unless a local sync report exists. Remote candidate sync is intentionally candidate-first: hosted agents use vault remote submit-candidate with anon/publishable credentials, while a trusted local host uses vault remote pull-candidates --apply with the service role key to write local memory_candidates, not active knowledge. --auto-promote-low-risk is available only with --apply; it still requires automation_policy.yaml, passing gates, allowed source/type/scope/sensitivity, trust above threshold, and a source reference. vault sync revisions/conflicts/audit is the local safety surface for the next multi-host phase. It records what arrived, what was promoted, what conflicted, and who marked a conflict resolved. accept_remote requires --apply-memory-change because it changes active memory by promoting the remote candidate and archiving the local row. This is still candidate-first sync, not multi-master active-knowledge sync. Multi-agent roster templates are opt-in with --agent-roster; each entry uses agent_id:role[:scope[:max_sensitivity]], for example profile-agent:profile,work-agent:work,remote-agent:remote,n8n:automation. Allowed roles remain work, profile, care, dream, remote, automation, and observer for compatibility. The profile/care/dream names are access labels, not Vault-owned modeling domains. Live external validation files are opt-in with --validation-pack remote|n8n|coze|all. Memory automation schedule files are opt-in with --automation-schedule cron|launchagent|n8n|all; generated jobs default to vault daily-loop run and stay report-first unless --automation-apply is explicitly set. Use --automation-command run for maintenance-only schedules. Add --automation-write-workspace when those scheduled cycle jobs should write reports/automation/cycle-latest.json and reports/automation/cycle-latest.md as the next-agent workbench. Generated memory automation jobs also write reports/automation/pipeline-latest.json plus .md, reports/automation/review-summary-latest.json plus .md, and reports/automation/learning-health-latest.json and .md after each pass. Add --automation-include-transcripts when those scheduled jobs should write metadata-only uncaptured transcript hints into the inbox handoff. Add --automation-auto-promote-low-risk when setup should write the narrow low-risk auto-promote policy; pair it with --automation-apply for scheduled promotion, otherwise scheduled runs preview only. Manual interactive setup asks for en, zh-Hant, or zh-CN; non-interactive agent installs can pass --language. Supabase setup guide generation is opt-in with --supabase-setup none|simple|advanced; keep simple as the default path unless the user asks for RLS or multi-agent permissions. For MCP remote readers, use vault-mcp --tool-profile remote and the sequence vault_remote_search -> vault_remote_map_show -> vault_remote_read_range after applying docs/supabase_read_policy.sql in Supabase. Remote IDs can be local integers or Supabase UUIDs; pass the ID returned by search directly into map/read. Use vault remote doctor --agent-id <agent> --json when search works but map/read fails, or after applying remote-reader SQL. It returns safe status checks, counts, and next actions without dumping raw synced content. For per-tool MCP examples, see docs/mcp_tool_reference.md.

Search And Navigation

Command Purpose
vault search "query" Search the vault; use --min-score to tune weak-match suppression
vault search "query" --json Return machine-readable search results for agents, scripts, and automation smoke tests
vault search "query" --graph-expand 2 Search with graph expansion
vault map build Build/backfill Document Map rows
vault map show <id> Show a knowledge entry's section map
vault map read <id> --lines 10-30 Read a bounded source range
vault map build/show/read/query ... --json Return machine-readable Document Map status, nodes, bounded lines, or claim matches
vault list List knowledge entries
vault remove <id> --confirm / vault delete <id> --confirm Delete a knowledge entry by ID
vault stats Show vault statistics
vault usage stats / vault usage archive-expired --apply Inspect retrieval usage and archive expired unused memories without deleting them
vault usage cold-store-expired --apply Summarize and archive expired-but-used memories; skips private/high/restricted/L0/L1 rows and retains original content

Task Ledger / Working Set

Task Ledger is the task-runtime workbench. It is not L2, not L4, and not active knowledge. Use it to keep long-running tasks resumable across Codex, Hermes, Claude Code, OpenClaw, or other agents without dumping whole chat history into the context window.

Command Purpose
vault task start "Repair benchmark" --plan "inspect failing tests" --next-action "run focused pytest" Create a task working set
vault task start "Repair benchmark" --priority P1 --due-at 2026-07-01 Create a task with urgency and a due/reminder timestamp
vault task update <task_id> --done "schema added" --decision "Task Ledger is not L2" Append completed work and hard decisions
vault task update <task_id> --priority P0 --due-at 2026-06-30 Change task urgency without touching active memory
vault task update <task_id> --blocker "waiting for CI" --question "sync task fields to Supabase?" Track blockers and open questions
vault task status <task_id> --json Read the current working set for an agent
vault task status --status active List active task ledgers
vault task resume <task_id> Alias for reading one task with resume semantics
vault task handoff <task_id> Print compact Markdown for the next agent/session
vault task send-handoff <task_id> --from-agent codex --to-agent hermes --message "continue from tests" Create a directed handoff packet in the shared task inbox
vault task inbox --agent-id hermes List pending handoff packets addressed to an agent
vault task claim-handoff <handoff_id> --agent-id hermes Mark a handoff as claimed by the receiving agent
vault task complete <task_id> --summary "..." Mark the task completed without promoting task events

At task phase end, extract only reusable lessons with the normal candidate-first workflow. Failed attempts and temporary task events should remain task history, not knowledge rows.

Task lists sort by priority first (P0 to P3), then by due date, then by recent updates. Use P0 for active incidents, P1 for urgent release/blocking work, P2 for normal planned work, and P3 for backlog or optional follow-up.

Directed handoff packets are the local multi-Agent handoff protocol. They are task-runtime inbox items, not private memory sharing. A sender can address a handoff to another agent, the receiver can claim it, and the task event log keeps the audit trail. The packet contains the task snapshot and sender note; it should reference shared evidence instead of copying another agent's private memory.

Skill Registry

The local Skill registry stores the latest skill row plus revision history. It is for agent capability handoff and upgrade planning, not for silently changing an agent's runtime.

Command Purpose
vault skill push --file SKILL.md --name review-helper --version 1.0.0 Register one Skill and store its first revision
vault skill push --file SKILL.md --name review-helper --version 1.1.0 Publish a newer local revision and make it the latest version
vault skill versions review-helper List known versions without raw Skill content
vault skill diff review-helper --from-version 1.0.0 --to-version 1.1.0 Compare compact metadata/content hashes between two versions
vault skill upgrade-plan --installed '{"review-helper":"1.0.0"}' --json Compare caller-installed skill versions with the registry latest versions
vault skill upgrade-plan --installed-file installed-skills.json --outdated-only Compare an installed Skill manifest and show only upgrade/drift/local-newer items
vault skill pull review-helper Write the latest Skill content into the local skills cache

MCP exposes Skill tools outside core; this keeps daily agent startup small. review can inspect versions, upgrade plans, and sync manifests. maintenance and full can also use permission-gated registry writes:

  • vault_skill_push writes or revises a Skill in the local registry only.
  • vault_skill_sync_manifest gives trusted sync workers hashes and metadata, and exports bounded content only when explicitly allowed.
  • vault_skill_mark_synced records that an external sync worker succeeded.

These MCP tools do not install, overwrite, or delete runtime Skill files. Agent runtime updates still require an explicit user/operator-approved installer or adapter step.

Installed manifests may use either compact versions:

{"review-helper": "1.0.0"}

or version plus content hash:

{"review-helper": {"version": "1.0.0", "content_hash": "abc123"}}

The richer form lets Vault distinguish normal upgrades from same-version content drift.

Quality And Curation

Command Purpose
vault lint Run quality checks
vault dream Produce report-first memory curation summaries
vault candidates --include-gates Review candidate-memory queue and gate details before promotion
vault candidate-review <id> --outcome rejected --reason "..." Record rejected/blocked candidate feedback without promoting memory
vault capture session <transcript> --write-candidates Capture decisions, pitfalls, workflows, and source-of-truth lines from JSONL/Markdown/text transcripts as gated candidates
vault memory reflection --write-candidates Run Dream plus lifecycle automation as a report-first reflection pass
vault daily-loop run --write-report Run the safe daily memory loop: sync dry-run/freshness, automation cycle handoff, inbox, review-summary, learning-health, and the human daily report
vault daily-loop report --refresh --write-report Rebuild reports/daily-loop/daily-loop-latest.* from read-only status surfaces without capture, reflection, sync writes, or new candidate writes
vault automation plan --write-policy Create a policy-based maintenance plan and starter automation_policy.yaml
vault automation run / vault automation run --apply Run report-first memory automation; reports include a dry-run diff and action ledger, and --apply only performs policy-allowed reversible actions or explicit low-risk auto-promote matches that are not downgraded by learned review feedback
vault automation cycle --apply Run one safe feedback-to-curation loop: evaluate reviewed candidate outcomes, write learning_policy.json, then run policy-based automation so Dream can consume the latest hints
vault automation cycle --write-workspace --include-transcripts Write reports/automation/cycle-latest.json and .md: compact candidate review, metadata-only transcript paths, priority brief, suggested next tasks, and learning-policy summary for the next agent
vault automation activity --pretty Show a compact read-only activity feed for recent auto-promote previews, promotions, and skipped reasons
vault automation brief --pretty Show one compact intelligence brief: learning hints, explainable memory importance, forgetting pressure, shared agent health, and the 5% human-review queue
vault automation brief --write-brief Write reports/automation/brief-latest.json and .md for dashboards or the next agent
vault automation review-summary --write-summary Write the shortest human approval cards to reports/automation/review-summary-latest.json and .md
vault automation review-feedback --kind memory_importance --card-id 12 --decision accept --reason "Correct card" --write-learning-policy Record feedback for one review-summary card, refresh bounded learning hints, and rewrite the next review-summary plus learning-health reports
vault automation learning-health --write-health Write a compact dashboard-safe health panel for feedback learning to reports/automation/learning-health-latest.json and .md
vault automation fleet-health --write-health Write a shared multi-Agent automation health panel to reports/automation/fleet-health-latest.json and .md
vault automation handoff Print the latest compact automation handoff for the next agent; surfaces fleet-health-latest first when present, then prefers cycle-latest.md
vault automation inbox --limit 5 --write-handoff Show the shortest read-only review queue and write reports/automation/inbox-latest.json
vault automation inbox --include-transcripts --write-handoff Add metadata-only uncaptured transcript hints to the inbox handoff
vault automation report / vault automation eval --write-learning-policy / vault automation doctor Review automation reports, evaluate candidate-outcome feedback into bounded curation hints, and check scheduled-job readiness
vault freshness Experimental freshness/review scheduling
vault usage archive-expired Preview TTL-based archive actions; add --apply to mark eligible memories archived
vault usage cold-store-expired Preview summarize-then-cold-store actions for expired-but-used memories; add --apply to write summaries and archive eligible rows
vault dedup Detect or merge duplicate entries
vault converge Experimental convergence/self-questioning check
vault cross-validate Experimental cross-model validation
vault search-qa run / vault search-qa compare Run Search QA snapshots, hard-negative checks, and before/after comparisons

Storage And Maintenance

Command Purpose
vault doctor Check local environment and optional dependencies
vault doctor --json Return machine-readable environment checks with top-level ok, status, and next_action
vault db status / vault db migrate Inspect or update local SQLite schema
vault db backup / vault db verify-backup / vault db restore Create, verify, and restore local SQLite backups
vault graph build / vault graph show Build or inspect the inferred knowledge graph
vault graph build/show --json Return machine-readable graph build results or graph summary

Optional Semantic Workflow

Command Purpose
vault install-embedding Install a local embedding model
vault config set embedding.provider ollama Configure an embedding provider; also supports openai, cohere, and voyage when the matching API key environment variable is set
vault semantic rebuild Rebuild semantic vector rows after configuring a real embedding provider
vault semantic rebuild --changed-only --persist-cache Refresh only missing or stale semantic vector rows
vault semantic warm Precompute QA query embeddings without writing vector rows
vault semantic smoke Rebuild, warm, and run a Search QA smoke snapshot
vault semantic cache-stats / vault semantic cache-prune Inspect or prune the durable embedding cache
vault semantic startup / vault semantic daemon Run startup or bounded daemon lifecycle hooks

Experimental Local Skills

Command Purpose
vault skill search "query" Search local experimental skill registry entries

Run vault <command> --help for command-specific options.