Runtime RPC command reference
This reference covers all 82 top-level commands, 32 subcommands and 18 MCP tools in the published 1.0.17-dev.14 release. See Runtime RPC for transport fields and runnable clients.
Arguments, defaults, accepted values and aliases below are the release binary's per-command help, checked against its published SHA-256 build manifest. The original help is retained with only trailing whitespace removed so it can be compared with an installed binary. <VALUE> denotes a placeholder you must replace; templates are not a script to execute in bulk.
This is the complete catalog for the inspected version, not a claim that every command exists in the older stable baseline. DEV diary-fragment, snapshot, feedback, witness and host-management features require checking the running bin; v: 1 alone is insufficient.
Calling boundaries
| Label | Meaning |
|---|---|
RPC | Send an argument array through cli.exec; account rights, service modes, validation and operation authorization still apply. |
ADMIN | Includes configuration, identity, external-file or plugin operations. Use the read subset for integration and review each mutation. |
HOST | A third-party client should delegate this to the host. Explicitly rejected RPC branches are identified below; this label itself is not a server ACL. |
RPC argument arrays omit the executable name and contain --json, command, subcommand and arguments. All paths resolve on the runtime host: import, export, backup, snapshot, retitle-many, share-import, extract and injection do not upload or download client files. Do not send stdin, TTY, shell pipes or interactive pickers. Closing an HTTP connection cannot cancel a command that has no dedicated cancellation API.
Read each purpose and boundary, then expand the original help for all arguments. --yes, --go, --confirm and --force have different meanings; do not add them automatically just to obtain success. HTTP 200, exit 0 or absent error prose cannot replace checking the business outcome.
Global options and command index (release help)
跨设备跨软件统一 AI 记忆系统（local-first）
Usage: rsrs [OPTIONS] [COMMAND]
Commands:
runtime Host-owned foreground runtime and opt-in current-user sign-in startup
restart Restart the resident runtime; kill the verified process if graceful shutdown times out
prompt Read the full memory workflow before storing or maintaining memories
v Print the embedded CLI version (same as `--version` / `-v`) [alias: version]
grant Manage independent read-only subtree grants for local tools
session List or revoke independent cloud login sessions; does not change local decrypt keys
keygen Generate local key material (offline; writes local session/keyfile - no server, no auth)
remember Store a memory (local authority store; encrypts + embeds automatically). Judge-then-store (dedup first, then choose): without --force, an internal recall runs first - high similarity -> suggest merge (`--merge-ids "id1,id2"` deletes old, writes a combined entry); medium -> suggest attach (`--parent` as cause); no candidate or --force writes directly. `--parent` = causal attach; parent/child is cause/effect
import Import a JSON-array dump: keep fields, rebuild parent links, sync once after the batch
recall Semantic recall through the private Core
query-log Query log: each recall's candidates + adoption + model self-grade (local DPO/SFT raw material)
bench Retrieval quality benchmark: run an evalset for hit-rate/MRR, optional baseline compare; mine drafts an evalset from query-log
run-log Record a procedure run verdict (ok|fail|confounded|no_effect) for a skill/knowledge entry; local-only, feeds how-to ordering and audit
classify Batch classify via JEV (TypeSafe): ask which top-level category each entry belongs to, compare with the current parent (read-only)
list List recent entries
forget Delete (tombstone; propagates with sync)
purge Clear current-version ciphertext; pending operations and unresolved conflicts are kept
restore Restore a deleted memory by full ID (keeps body and parent; propagates with sync)
promote Causal tree: promote - this entry becomes a more-upstream cause (reattach to grandparent/root; effects follow)
demote Causal tree: demote - attach this entry under --parent as its effect (extends the chain by one; cycle-safe)
show Show one entry in full (cause/effect complete, not truncated)
chain Causal-chain deep dive: fetch the whole chain at once - ancestor bodies (root first) + this entry + descendants to --depth. After an AI sees a cause-chain name, it can dig without layer-by-layer show
attach Attach a child under a parent (explicit attach is intent; no judge-then-store; auto-syncs)
tree Causal tree: render a subtree (--from start, default root forest); --outline prints a full-depth id+title directory (AI tidy reads this first)
resort Memory tidy: batch reparent (spec={"ops":[{"id":"child","parent":"parent"}]}; id / first 8 chars / root title all work); dry-run by default, --go writes, auto-syncs, and resets the counter
agent-config AI behavior config (agent.json): no args = print the whole file (AI reads it to decide); --set key=value writes a key
retitle Retitle: decrypt -> change title -> reseal (ciphertext payload updates) -> dirty -> auto-sync
retitle-many Batch retitle (JSON array [{"id":...,"title":...}]; one process, one model load)
sync Two-way sync: pull cloud delta -> land locally -> push local changes (LWW)
sync-conflicts Show conflicts that still need a decision; --all includes resolved and old-generation history
sync-resolve Resolve one retained revision; mark it resolved only after the remote confirms
sync-history Show retained sync revisions; --remote reads server history without local caching
sync-restore Restore a kept revision as a new local edit
sync-reset Rebuild after a server backup restore; keep current recovery snapshots and pending edits
status Status
memory-revision Read the initialized library's opaque snapshot token without loading a model
capabilities Report supported CLI capabilities without a profile, runtime or model download
reembed Recompute all embeddings (for embedder upgrades / dim changes; re-embed by semantics)
defrag Stock analysis: cluster BGE vectors, flag lone leaves, suggest a tree (read-only)
split Split a mixed node (AI decides): no flags prints material JSON for the AI to plan; --go --spec '<json>' applies
tree-deepen Deepen the tree: cluster flat children of a large root into sub-roots (--root to inspect/apply; --auto deepens every flat large root)
inject Distribute the inject source: --targets lists 14 hosts; --id <target> injects; --remove --id <target> uninstalls; no args in a TTY opens the picker
tree-cure Tree hygiene: root-size bill + lone-leaf attach suggestions (read-only; --id --parent actually attaches)
tree-float Heat float: an entry with more hits than its parent becomes the grandparent's child (effects follow; dry-run report, --go applies)
doctor Self-check: model/store/lock/remote/inject/scope in one pass (first stop for install issues)
audit Whole-store health audit: orphans / illegal importance / duplicate titles / truncated titles / deep chains; structured report for an AI audit
witness Verify the local ciphertext-digest witness chain (not an external trust anchor)
web Open the hosted user dashboard without starting the local runtime
classify-config Inspect or configure classification without running a model
model Model management: install, verify or uninstall BGE-M3
config View/set local config (client.json: data dir / server address / auto-sync / auto-cure)
candidates Judge-then-store candidate bill (read-only): high similarity suggests merge, medium suggests attach, similar candidates suggest merging them
passport Memory passport: ASCII-art card (user / memory count / connected agents / devices) for sharing
history Entry change history (audit stream: create/update/delete/restore/purge; SQLite triggers record it)
plugin Plugin hooks (plugins.json): list shows config; test <EVENT> fires an event
mcp MCP server over stdio JSON-RPC. The internal runtime serves HTTP/SSE at `/mcp` and `/sse`
update-check Check npm for a newer CLI (24h throttle; --force skips it; --clear drops the cache)
repack Backfill: reseal ciphertext so payload.parent_id matches the plaintext column (fixes dual-source drift; idempotent)
diary Diary = whole-store time chain (the trivial zone is merged in; not filtered by importance): recent (--limit), a day (--date, today/yesterday), a range (--from/--to inclusive), keyword (--contains)
diary-search Explicit cross-date diary fragment retrieval; never changes normal recall
diary-fragment Read one immutable diary fragment locator and optional immediate neighbors
taxonomy Built-in top-level catalog (20 human-domain + 3 AI-domain = 23, embedded at compile): --list prints it; --ensure <root,...> creates roots; no args shows current state
root-create Create a root outside the catalog (needs the user to say yes): AI proposes -> user allows -> run with --yes; the AI must not create one on its own
update Update a memory (any of title/body/tags/kind/importance); decrypt, reseal, auto-sync
export Export the whole store as plaintext JSON (local backup)
backup Back up the local store file (whole SQLite copy, ciphertext and index)
snapshot Create or inspect verified local snapshots; never rolls back cloud state
secret Show account keys: Account Secret / five-keys (manual cross-device join; leak = loss of the store)
keys-export Export decrypt keys (super password + Secret Key + vault material) - how a new machine logs in; leak = loss of the store
super-reset Reset the super password: unlock URK with the current code, issue a new one, upload. Data is unchanged
fivekeys Five-keys join (another machine): verify and write a local session; a non-empty --addr also logs in for a token
register Register an account (first device): login password for auth + super password wrapping keys. Address defaults to https://api.rsrs.rs; use --addr to override it. Missing user/password drops into an interactive prompt (TTY only)
account Multiple accounts on one machine: each account has its own data dir (local store/keys/session switch as a unit). Logging into a new account creates a profile; `account list` lists them; `account <name>` switches (main = primary); `remove <name> --yes` deletes
migrate Explicitly copy a selected old library into a new account; no startup migration
space Spaces (virtual accounts): an owner can create several virtual accounts; each is a space with its own super key and data dir. Join = a member builds that space's profile on their machine from an invite; they can switch freely; leaving is `kick`
logout Log out: clear addr/token and drop the cloud (identity and keys stay; login brings them back); --full forgets this machine's identity
login Log in through browser authorization, or use --interactive for password and TOTP prompts. Address defaults to the last server this machine used (kept in session.json by register/logout). Missing user/password drops into an interactive prompt (TTY only)
book-material Book material: a root subtree -> volume/chapter full text (for an AI to draft)
portrait-material Portrait material: whole-store portrait aggregate (for an AI to draft a portrait)
share Share a subtree: emit a prompt another AI can paste (payload is plaintext; the other AI imports from the prompt)
share-import Import a shared subtree: without --go, print attach-candidate bill (for the AI to judge); --go writes the store
extract Extract local coding-agent sessions (Codex, OpenCode, Claude Code, …) into rsrs memories
help Print this message or the help of the given subcommand(s)
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print help
-V, --version Print version
Sandbox: use --client-only or RSRS_CLIENT_ONLY=1 to connect to the host HTTP runtime without managing its lifecycle.Diagnostics and status
status
Status.
Calling boundary: RPC.
RPC args template: ["--json","status"]
All arguments, defaults and help
Status
Usage: rsrs status [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpmemory-revision
Read the initialized library's opaque snapshot token without loading a model.
Calling boundary: RPC.
RPC args template: ["--json","memory-revision"]
All arguments, defaults and help
Read the initialized library's opaque snapshot token without loading a model
Usage: rsrs memory-revision [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpcapabilities
Report supported CLI capabilities without a profile, runtime or model download.
Calling boundary: RPC.
RPC args template: ["--json","capabilities"]
All arguments, defaults and help
Report supported CLI capabilities without a profile, runtime or model download
Usage: rsrs capabilities [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpv
Print the embedded CLI version (same as --version / -v).
Calling boundary: RPC.
RPC args template: ["--json","v"]
All arguments, defaults and help
Print the embedded CLI version (same as `--version` / `-v`)
Usage: rsrs v [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpprompt
Read the full memory workflow before storing or maintaining memories.
Calling boundary: RPC.
RPC args template: ["--json","prompt"]
All arguments, defaults and help
Read the full memory workflow before storing or maintaining memories
Usage: rsrs prompt [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpdoctor
Self-check: model/store/lock/remote/inject/scope in one pass (first stop for install issues).
Calling boundary: RPC.
RPC args template: ["--json","doctor"]
All arguments, defaults and help
Self-check: model/store/lock/remote/inject/scope in one pass (first stop for install issues)
Usage: rsrs doctor [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--remote Also probe remote server reachability (GET /health, ~1–3s)
--check-update Also check npm for a newer CLI (~1–5s; default reads the 24h cache, no network)
--fix Install a missing or unusable embedder; default only reports diagnostics
-h, --help Print helpaudit
Whole-store health audit: orphans / illegal importance / duplicate titles / truncated titles / deep chains; structured report for an AI audit.
Calling boundary: RPC.
RPC args template: ["--json","audit"]
All arguments, defaults and help
Whole-store health audit: orphans / illegal importance / duplicate titles / truncated titles / deep chains; structured report for an AI audit
Usage: rsrs audit [OPTIONS]
Options:
--json JSON output (client/AI consumption)
-h, --help Print helpwitness
Verify the local ciphertext-digest witness chain (not an external trust anchor).
Calling boundary: RPC.
RPC args template: ["--json","witness"]
All arguments, defaults and help
Verify the local ciphertext-digest witness chain (not an external trust anchor)
Usage: rsrs witness [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--verify
-h, --help Print helppassport
Memory passport: ASCII-art card (user / memory count / connected agents / devices) for sharing.
Calling boundary: RPC.
RPC args template: ["--json","passport"]
All arguments, defaults and help
Memory passport: ASCII-art card (user / memory count / connected agents / devices) for sharing
Usage: rsrs passport [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helphistory
Entry change history (audit stream: create/update/delete/restore/purge; SQLite triggers record it).
Calling boundary: RPC.
RPC args template: ["--json","history"]
All arguments, defaults and help
Entry change history (audit stream: create/update/delete/restore/purge; SQLite triggers record it)
Usage: rsrs history [OPTIONS] [ID]
Arguments:
[ID] Entry id (8-char prefix ok); omit = recent changes for the whole store
Options:
--json Machine-readable output: JSON only, without human progress
--limit <LIMIT> [default: 20]
-h, --help Print helpupdate-check
Check npm for a newer CLI (24h throttle; --force skips it; --clear drops the cache).
Calling boundary: RPC.
RPC args template: ["--json","update-check"]
All arguments, defaults and help
Check npm for a newer CLI (24h throttle; --force skips it; --clear drops the cache)
Usage: rsrs update-check [OPTIONS]
Options:
--force Skip the throttle and query the network
--json Machine-readable output: JSON only, without human progress
--clear Clear the cache (the next check hits the network)
-h, --help Print helpRead and retrieve
recall
Semantic recall through the private Core.
Calling boundary: RPC.
RPC args template: ["--json","recall","<QUERY>"]
All arguments, defaults and help
Semantic recall through the private Core
Usage: rsrs recall [OPTIONS] <QUERY>
Arguments:
<QUERY>
Options:
--json Machine-readable output: JSON only, without human progress
--mode <MODE> Override the recall mode for this request without changing saved settings [possible values: fast, quality]
--limit <LIMIT> [default: 20]
--type <TYPE>
--project <PROJECT> Restrict to this project (use when other projects interfere)
--trace Recall trace: stderr dumps the full routing path (candidates -> scope -> two zones -> per-axis Top10)
--titles Return compact IDs and titles in JSON; use show to read full memory content
--no-related Disable associations for this request (original recall results only)
--intent <INTENT> Recall intent: auto (keyword), howto (extra skill pass, keeps one procedure visible), fact (single pass) [default: auto] [possible values: auto, howto, fact]
-h, --help Print helplist
List recent entries.
Calling boundary: RPC.
RPC args template: ["--json","list"]
All arguments, defaults and help
List recent entries
Usage: rsrs list [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--limit <LIMIT> [default: 20]
--since <SINCE> Only entries created after this time (RFC3339 or YYYY-MM-DD)
--since-resort Only entries created since the last tidy (maintenance.json resort_at) - the §3.8 tidy intake: tidy this batch, do not pull the whole store
-h, --help Print helpshow
Show one entry in full (cause/effect complete, not truncated).
Calling boundary: RPC.
RPC args template: ["--json","show","<ID>"]
All arguments, defaults and help
Show one entry in full (cause/effect complete, not truncated)
Usage: rsrs show [OPTIONS] <ID>
Arguments:
<ID>
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpchain
Causal-chain deep dive: fetch the whole chain at once - ancestor bodies (root first) + this entry + descendants to --depth. After an AI sees a cause-chain name, it can dig without layer-by-layer show.
Calling boundary: RPC.
RPC args template: ["--json","chain","<ID>"]
All arguments, defaults and help
Causal-chain deep dive: fetch the whole chain at once - ancestor bodies (root first) + this entry + descendants to --depth. After an AI sees a cause-chain name, it can dig without layer-by-layer show
Usage: rsrs chain [OPTIONS] <ID>
Arguments:
<ID>
Options:
--depth <DEPTH> Descendant depth (default 3; 0 = ancestors + this entry only) [default: 3]
--json Machine-readable output: JSON only, without human progress
-h, --help Print helptree
Causal tree: render a subtree (--from start, default root forest); --outline prints a full-depth id+title directory (AI tidy reads this first).
Calling boundary: RPC.
RPC args template: ["--json","tree"]
All arguments, defaults and help
Causal tree: render a subtree (--from start, default root forest); --outline prints a full-depth id+title directory (AI tidy reads this first)
Usage: rsrs tree [OPTIONS]
Options:
--from <FROM> [default: ""]
--json Machine-readable output: JSON only, without human progress
--depth <DEPTH> [default: 3]
--outline
--material <MATERIAL> Export this node's subtree as Markdown material (same as the client install dialog)
-h, --help Print helpdiary
Diary = whole-store time chain (the trivial zone is merged in; not filtered by importance): recent (--limit), a day (--date, today/yesterday), a range (--from/--to inclusive), keyword (--contains).
Calling boundary: RPC.
RPC args template: ["--json","diary"]
All arguments, defaults and help
Diary = whole-store time chain (the trivial zone is merged in; not filtered by importance): recent (--limit), a day (--date, today/yesterday), a range (--from/--to inclusive), keyword (--contains)
Usage: rsrs diary [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--limit <LIMIT> [default: 10]
--date <DATE> Flip a day (YYYY-MM-DD or today/yesterday, local timezone)
--from <FROM> Range start (YYYY-MM-DD, inclusive; mutually exclusive with --date)
--to <TO> Range end (YYYY-MM-DD, inclusive; mutually exclusive with --date; --to without --from means "up to that day")
--contains <CONTAINS> Full-text keyword filter (title and decrypted full body); returns a bounded matching excerpt
-h, --help Print helpdiary-search
Explicit cross-date diary fragment retrieval; never changes normal recall.
Calling boundary: RPC.
Query length is at most 500 Unicode scalars; limit is 1–20 and max-chars is 512–12000. Budget includes metadata and JSON escaping. truncated can mean budget exhaustion; semantic_ready=false is incomplete semantic coverage, not an indexing wait instruction.
RPC args template: ["--json","diary-search","<QUERY>"]
All arguments, defaults and help
Explicit cross-date diary fragment retrieval; never changes normal recall
Usage: rsrs diary-search [OPTIONS] <QUERY>
Arguments:
<QUERY>
Options:
--json Machine-readable output: JSON only, without human progress
--limit <LIMIT> [default: 5]
--max-chars <MAX_CHARS> Budget for fragment result JSON, in Unicode characters (not tokens) [default: 4000]
--mode <MODE> [default: hybrid] [possible values: lexical, hybrid]
-h, --help Print helpdiary-fragment
Read one immutable diary fragment locator and optional immediate neighbors.
Calling boundary: RPC.
neighbors is 0 or 1. Preserve source_id, fragment_id and source_version. Offsets are Unicode scalar positions, not JavaScript UTF-16 indexes. If the source changed, search once again for a current locator; do not silently read the full diary instead.
RPC args template: ["--json","diary-fragment","<SOURCE_ID>","<FRAGMENT_ID>"]
All arguments, defaults and help
Read one immutable diary fragment locator and optional immediate neighbors
Usage: rsrs diary-fragment [OPTIONS] <SOURCE_ID> <FRAGMENT_ID>
Arguments:
<SOURCE_ID>
<FRAGMENT_ID>
Options:
--json Machine-readable output: JSON only, without human progress
--neighbors <NEIGHBORS> [default: 0]
--max-chars <MAX_CHARS> [default: 4000]
-h, --help Print helpbook-material
Book material: a root subtree -> volume/chapter full text (for an AI to draft).
Calling boundary: RPC.
The result envelope uses command tree-material and details.material; do not derive the envelope command mechanically from the CLI spelling.
RPC args template: ["--json","book-material","<ROOT>"]
All arguments, defaults and help
Book material: a root subtree -> volume/chapter full text (for an AI to draft)
Usage: rsrs book-material [OPTIONS] <ROOT>
Arguments:
<ROOT>
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpportrait-material
Portrait material: whole-store portrait aggregate (for an AI to draft a portrait).
Calling boundary: RPC.
RPC args template: ["--json","portrait-material"]
All arguments, defaults and help
Portrait material: whole-store portrait aggregate (for an AI to draft a portrait)
Usage: rsrs portrait-material [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--limit <LIMIT> [default: 40]
-h, --help Print helpWrite, import and preserve
remember
Store a memory in the local authority store, encrypting and embedding it automatically. The original help below simplifies judge-then-store: in this version, an internal recall runs only when --force, --merge-ids, --supersedes and --parent are absent, --see-also is empty, and importance is not trivial. Merge or parent candidates return pending without writing; otherwise the command continues to the write path. Explicit merge, parent and association options skip this candidate check.
Calling boundary: RPC.
Inspect the returned action/ID and any pending candidate decision. Merge, supersede and see-also have different effects; merge deletes its sources. A pre-remember hook can return pending without a write.
RPC args template: ["--json","remember","<CONTENT>"]
All arguments, defaults and help
Store a memory (local authority store; encrypts + embeds automatically). Judge-then-store (dedup first, then choose): without --force, an internal recall runs first - high similarity -> suggest merge (`--merge-ids "id1,id2"` deletes old, writes a combined entry); medium -> suggest attach (`--parent` as cause); no candidate or --force writes directly. `--parent` = causal attach; parent/child is cause/effect
Usage: rsrs remember [OPTIONS] <CONTENT>
Arguments:
<CONTENT>
Options:
--json Machine-readable output: JSON only, without human progress
--type <TYPE> [default: context]
--tags <TAGS> [default: ""]
--title <TITLE> [default: ""]
--project <PROJECT> [default: ""]
--computer <COMPUTER> [default: ""]
--emotion <EMOTION> [default: -1]
--importance <IMPORTANCE> Importance (AI-judged): important (reusable, recall primary zone) | trivial (diary; default). The normal tier is retired [default: trivial]
--parent <PARENT> [default: ""]
--force Skip judge-then-store and write immediately
--merge-ids <MERGE_IDS> Merge: delete the listed old memories (comma-separated ids) and store this content as the combined entry (inherits the first cause chain)
--supersedes <SUPERSEDES> This new conclusion replaces a live important memory
--see-also <SEE_ALSO> Bidirectional links to live important memories
-h, --help Print helpupdate
Update a memory (any of title/body/tags/kind/importance); decrypt, reseal, auto-sync.
Calling boundary: RPC.
RPC args template: ["--json","update","<ID>"]
All arguments, defaults and help
Update a memory (any of title/body/tags/kind/importance); decrypt, reseal, auto-sync
Usage: rsrs update [OPTIONS] <ID>
Arguments:
<ID>
Options:
--json Machine-readable output: JSON only, without human progress
--title <TITLE>
--content <CONTENT>
--tags <TAGS>
--kind <KIND>
--importance <IMPORTANCE> Importance revision: important | normal | trivial (reseals ciphertext payload and the index column)
-h, --help Print helpretitle
Retitle: decrypt -> change title -> reseal (ciphertext payload updates) -> dirty -> auto-sync.
Calling boundary: RPC.
RPC args template: ["--json","retitle","<ID>","--title","<TITLE>"]
All arguments, defaults and help
Retitle: decrypt -> change title -> reseal (ciphertext payload updates) -> dirty -> auto-sync
Usage: rsrs retitle [OPTIONS] <ID>
Arguments:
<ID>
Options:
--json Machine-readable output: JSON only, without human progress
--title <TITLE> New title (required - omitting it would blank the title; fixed in the 2026-09-20 audit)
-h, --help Print helpretitle-many
Batch retitle (JSON array [{"id":...,"title":...}]; one process, one model load).
Calling boundary: RPC.
RPC args template: ["--json","retitle-many","<FILE>"]
All arguments, defaults and help
Batch retitle (JSON array [{"id":...,"title":...}]; one process, one model load)
Usage: rsrs retitle-many [OPTIONS] <FILE>
Arguments:
<FILE>
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpforget
Delete (tombstone; propagates with sync).
Calling boundary: RPC.
RPC args template: ["--json","forget","<ID>"]
All arguments, defaults and help
Delete (tombstone; propagates with sync)
Usage: rsrs forget [OPTIONS] <ID>
Arguments:
<ID>
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helppurge
Clear current-version ciphertext; pending operations and unresolved conflicts are kept.
Calling boundary: RPC.
RPC args template: ["--json","purge","<ID>"]
All arguments, defaults and help
Clear current-version ciphertext; pending operations and unresolved conflicts are kept
Usage: rsrs purge [OPTIONS] <ID>
Arguments:
<ID>
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helprestore
Restore a deleted memory by full ID (keeps body and parent; propagates with sync).
Calling boundary: RPC.
RPC args template: ["--json","restore","<ID>"]
All arguments, defaults and help
Restore a deleted memory by full ID (keeps body and parent; propagates with sync)
Usage: rsrs restore [OPTIONS] <ID>
Arguments:
<ID>
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpimport
Import a JSON-array dump: keep fields, rebuild parent links, sync once after the batch.
Calling boundary: RPC.
RPC args template: ["--json","import","<FILE>"]
All arguments, defaults and help
Import a JSON-array dump: keep fields, rebuild parent links, sync once after the batch
Usage: rsrs import [OPTIONS] <FILE>
Arguments:
<FILE>
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpexport
Export the whole store as plaintext JSON (local backup).
Calling boundary: RPC.
RPC args template: ["--json","export","<FILE>"]
All arguments, defaults and help
Export the whole store as plaintext JSON (local backup)
Usage: rsrs export [OPTIONS] <FILE>
Arguments:
<FILE>
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpbackup
Back up the local store file (whole SQLite copy, ciphertext and index).
Calling boundary: RPC.
RPC args template: ["--json","backup","<FILE>"]
All arguments, defaults and help
Back up the local store file (whole SQLite copy, ciphertext and index)
Usage: rsrs backup [OPTIONS] <FILE>
Arguments:
<FILE>
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpsnapshot
Create or inspect verified local snapshots; never rolls back cloud state.
Calling boundary: RPC.
Create/list are runtime operations. Restore/recover are explicitly rejected inside a runtime worker and require the host CLI. Restore quarantines synchronization and never rolls back cloud state. Resume-sync requires the exact reviewed local revision.
All arguments, defaults and help
Create or inspect verified local snapshots; never rolls back cloud state
Usage: rsrs snapshot [OPTIONS] <COMMAND>
Commands:
create Create a consistent private database snapshot, excluding access grants
list List snapshots after validating file digests and SQLite integrity
restore Restore one local snapshot after a mandatory pre-backup; cloud state is unchanged
recover Recover an interrupted local snapshot transaction (host only)
resume-sync Resume sync only after reviewing local/remote differences at this exact local revision
help Print this message or the help of the given subcommand(s)
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpsnapshot create
Calling boundary: RPC.
RPC args template: ["--json","snapshot","create"]
All arguments, defaults and help
Create a consistent private database snapshot, excluding access grants
Usage: rsrs snapshot create [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--note <NOTE> [default: ""]
-h, --help Print helpsnapshot list
Calling boundary: RPC.
RPC args template: ["--json","snapshot","list"]
All arguments, defaults and help
List snapshots after validating file digests and SQLite integrity
Usage: rsrs snapshot list [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpsnapshot restore
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","snapshot","restore","<ID>"]
All arguments, defaults and help
Restore one local snapshot after a mandatory pre-backup; cloud state is unchanged
Usage: rsrs snapshot restore [OPTIONS] <ID>
Arguments:
<ID>
Options:
--confirm
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpsnapshot recover
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","snapshot","recover"]
All arguments, defaults and help
Recover an interrupted local snapshot transaction (host only)
Usage: rsrs snapshot recover [OPTIONS]
Options:
--confirm
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpsnapshot resume-sync
Calling boundary: RPC.
RPC args template: ["--json","snapshot","resume-sync","--reviewed-revision","<REVIEWED_REVISION>"]
All arguments, defaults and help
Resume sync only after reviewing local/remote differences at this exact local revision
Usage: rsrs snapshot resume-sync [OPTIONS] --reviewed-revision <REVIEWED_REVISION>
Options:
--json
Machine-readable output: JSON only, without human progress
--reviewed-revision <REVIEWED_REVISION>
-h, --help
Print helpextract
Extract local coding-agent sessions (Codex, OpenCode, Claude Code, …) into rsrs memories.
Calling boundary: RPC.
--dry-run and --targets are inspections; --force reimports previously seen sessions, and --include-subagents expands the scan. Imported source paths and session cursors belong to the host. --tui requires a terminal.
RPC args template: ["--json","extract","--targets"]
All arguments, defaults and help
Extract local coding-agent sessions (Codex, OpenCode, Claude Code, …) into rsrs memories
Usage: rsrs extract [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--targets List agent sources and whether a local session store exists
--tui Force the interactive TUI picker (needs a TTY)
--all Extract from every detected source (default when --id is omitted)
--id <ID> Extract from one source (codex/opencode/claude/…); see --targets
--dry-run Scan and summarize without writing any memory
--since <SINCE> Only sessions updated since this bound (today|yesterday|7d|YYYY-MM-DD)
--limit <LIMIT> Maximum sessions to import per source (newest first)
--cwd <CWD> Filter by working-directory substring
--model <MODEL> Filter by model/provider substring
--session <SESSION> Filter by session id substring
--kind <KIND> Memory kind for imported sessions [default: context]
--importance <IMPORTANCE> Importance: important (default) or trivial (diary) [default: important]
--parent <PARENT> Optional pre-attach parent (id / tree / anchor title); omit to ingest as roots for a later tidy [default: ""]
--include-subagents Include subagent/child sessions (hidden by default)
--force Re-import sessions already recorded in the cursor
-h, --help Print helpOrganize the library
candidates
Judge-then-store candidate bill (read-only): high similarity suggests merge, medium suggests attach, similar candidates suggest merging them.
Calling boundary: RPC.
RPC args template: ["--json","candidates","<CONTENT>"]
All arguments, defaults and help
Judge-then-store candidate bill (read-only): high similarity suggests merge, medium suggests attach, similar candidates suggest merging them
Usage: rsrs candidates [OPTIONS] <CONTENT>
Arguments:
<CONTENT>
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpattach
Attach a child under a parent (explicit attach is intent; no judge-then-store; auto-syncs).
Calling boundary: RPC.
RPC args template: ["--json","attach","--parent","<PARENT>","<ID>"]
All arguments, defaults and help
Attach a child under a parent (explicit attach is intent; no judge-then-store; auto-syncs)
Usage: rsrs attach [OPTIONS] --parent <PARENT> <ID>
Arguments:
<ID>
Options:
--json Machine-readable output: JSON only, without human progress
--parent <PARENT>
-h, --help Print helppromote
Causal tree: promote - this entry becomes a more-upstream cause (reattach to grandparent/root; effects follow).
Calling boundary: RPC.
RPC args template: ["--json","promote","<ID>"]
All arguments, defaults and help
Causal tree: promote - this entry becomes a more-upstream cause (reattach to grandparent/root; effects follow)
Usage: rsrs promote [OPTIONS] <ID>
Arguments:
<ID>
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpdemote
Causal tree: demote - attach this entry under --parent as its effect (extends the chain by one; cycle-safe).
Calling boundary: RPC.
RPC args template: ["--json","demote","--parent","<PARENT>","<ID>"]
All arguments, defaults and help
Causal tree: demote - attach this entry under --parent as its effect (extends the chain by one; cycle-safe)
Usage: rsrs demote [OPTIONS] --parent <PARENT> <ID>
Arguments:
<ID>
Options:
--json Machine-readable output: JSON only, without human progress
--parent <PARENT>
-h, --help Print helpresort
Memory tidy: batch reparent (spec={"ops":[{"id":"child","parent":"parent"}]}; id / first 8 chars / root title all work); dry-run by default, --go writes, auto-syncs, and resets the counter.
Calling boundary: RPC.
RPC args template: ["--json","resort"]
All arguments, defaults and help
Memory tidy: batch reparent (spec={"ops":[{"id":"child","parent":"parent"}]}; id / first 8 chars / root title all work); dry-run by default, --go writes, auto-syncs, and resets the counter
Usage: rsrs resort [OPTIONS]
Options:
--go
--json Machine-readable output: JSON only, without human progress
--spec <SPEC>
--status Show tidy counter (adds since last tidy / threshold)
--reset Manually reset the counter (e.g. after tidying another way)
--threshold <THRESHOLD> Set the alert threshold (new-entry count)
-h, --help Print helpdefrag
Stock analysis: cluster BGE vectors, flag lone leaves, suggest a tree (read-only).
Calling boundary: RPC.
RPC args template: ["--json","defrag"]
All arguments, defaults and help
Stock analysis: cluster BGE vectors, flag lone leaves, suggest a tree (read-only)
Usage: rsrs defrag [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--min <MIN> User-selected similarity floor for duplicate suggestions [default: 0.6]
--top <TOP> [default: 20]
-h, --help Print helpsplit
Split a mixed node (AI decides): no flags prints material JSON for the AI to plan; --go --spec '<json>' applies.
Calling boundary: RPC.
RPC args template: ["--json","split","<ID>"]
All arguments, defaults and help
Split a mixed node (AI decides): no flags prints material JSON for the AI to plan; --go --spec '<json>' applies
Usage: rsrs split [OPTIONS] <ID>
Arguments:
<ID>
Options:
--go Apply the split plan
--json Machine-readable output: JSON only, without human progress
--spec <SPEC> AI-authored split-plan JSON (run with no flags first to get material and field docs)
-h, --help Print helptree-deepen
Deepen the tree: cluster flat children of a large root into sub-roots (--root to inspect/apply; --auto deepens every flat large root).
Calling boundary: RPC.
RPC args template: ["--json","tree-deepen"]
All arguments, defaults and help
Deepen the tree: cluster flat children of a large root into sub-roots (--root to inspect/apply; --auto deepens every flat large root)
Usage: rsrs tree-deepen [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--root <ROOT> Target root id (8-char prefix ok); default = every flat large root
--go Apply the plan (pair with --titles)
--titles <TITLES> Sub-root title array JSON (draft order; default = longest member title in the cluster)
--auto One-shot: auto-deepen every flat large root (sub-root title = longest member title; retitle later)
--min <MIN> User-selected similarity floor for Core grouping [default: 0.55]
-h, --help Print helptree-cure
Tree hygiene: root-size bill + lone-leaf attach suggestions (read-only; --id --parent actually attaches).
Calling boundary: RPC.
RPC args template: ["--json","tree-cure"]
All arguments, defaults and help
Tree hygiene: root-size bill + lone-leaf attach suggestions (read-only; --id --parent actually attaches)
Usage: rsrs tree-cure [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--top <TOP> [default: 15]
--id <ID> Apply: attach --id under --parent (skip suggestions, just do it)
--parent <PARENT>
--auto Auto-attach: apply every suggestion, skip failures, then summarize
--min <MIN> User-selected similarity floor for Core attachment suggestions [default: 0.5]
-h, --help Print helptree-float
Heat float: an entry with more hits than its parent becomes the grandparent's child (effects follow; dry-run report, --go applies).
Calling boundary: RPC.
RPC args template: ["--json","tree-float"]
All arguments, defaults and help
Heat float: an entry with more hits than its parent becomes the grandparent's child (effects follow; dry-run report, --go applies)
Usage: rsrs tree-float [OPTIONS]
Options:
--go Apply: float each reported entry (dirty; propagates with sync)
--json Machine-readable output: JSON only, without human progress
--min <MIN> Hit-count floor to float (default 3; relative heat - must be strictly above the parent's hits) [default: 3]
-h, --help Print helptaxonomy
Built-in top-level catalog (20 human-domain + 3 AI-domain = 23, embedded at compile): --list prints it; --ensure <root,...> creates roots; no args shows current state.
Calling boundary: RPC.
RPC args template: ["--json","taxonomy","--list"]
All arguments, defaults and help
Built-in top-level catalog (20 human-domain + 3 AI-domain = 23, embedded at compile): --list prints it; --ensure <root,...> creates roots; no args shows current state
Usage: rsrs taxonomy [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--list List every built-in category (domain / title / surface-word count / gist)
--ensure <ENSURE> Create/fill roots (idempotent): titles must match the catalog, comma-separated; default = fill every missing root
-h, --help Print helproot-create
Create a root outside the catalog (needs the user to say yes): AI proposes -> user allows -> run with --yes; the AI must not create one on its own.
Calling boundary: RPC.
RPC args template: ["--json","root-create","<TITLE>"]
All arguments, defaults and help
Create a root outside the catalog (needs the user to say yes): AI proposes -> user allows -> run with --yes; the AI must not create one on its own
Usage: rsrs root-create [OPTIONS] <TITLE>
Arguments:
<TITLE>
Options:
--content <CONTENT> One-line gist (default = the title)
--json Machine-readable output: JSON only, without human progress
--yes User already allowed this (turn their "yes" into this flag; without it, refuse to write)
-h, --help Print helpclassify
Batch classify via JEV (TypeSafe): ask which top-level category each entry belongs to, compare with the current parent (read-only).
Calling boundary: RPC.
The short help's read-only wording does not apply to --auto: --auto applies selected changes. A configured remote classifier can receive memory text. Prefer --plan for a local preview and review the explicit scope before authorizing external processing.
RPC args template: ["--json","classify","--plan"]
All arguments, defaults and help
Batch classify via JEV (TypeSafe): ask which top-level category each entry belongs to, compare with the current parent (read-only)
Usage: rsrs classify [OPTIONS]
Options:
--json
Machine-readable output: JSON only, without human progress
--limit <LIMIT>
Process the latest N entries (default 50, so a full-store run is not accidental; --all overrides) [default: 50]
--all
Whole store (overrides --limit)
--root <ROOT>
Restrict to a subtree (id or 8-char prefix; includes the node itself and descendants)
--max-chars <MAX_CHARS>
Truncate each entry body to this many chars (API payload cap) [default: 600]
--min-confidence <MIN_CONFIDENCE>
Below this confidence, list as "low confidence, needs a human" (default 0.5) [default: 0.5]
--save <SAVE>
Save suggestion JSON (audit / later apply)
--dry-run
Dry run: preview final actions without network access or local writes
--ds [<DS>]
Use the DS (OpenAI-compatible) backend to emulate JEV: `--ds` uses a stored key; `--ds <key>` takes the key and stores it
--backend <BACKEND>
Classify backend: jev (TypeSafe official Jev, default) | ds (DeepSeek-compatible emulate) - same as --ds; pick one
--tree
Tree mode: pick a suggested parent among **existing tree nodes** (not a linear 23-root assignment) - pair with --batch for speed
--causal
Causal-tree segment rewrite: parent=cause, child=effect; walk by level and let the AI judge order (prints a resort --spec compatible reparent plan; does not write the store)
--out <OUT>
Causal mode: write the reparent plan to a file (resort --spec can consume it)
--auto
Apply the final organization actions selected by Core
--plan
Preview local structural repair actions; text only, no store writes, no network
--api-base <API_BASE>
API base (default TypeSafe; later may switch to a rsrs gateway)
--model <MODEL>
Model name [default: jev-latest]
-h, --help
Print helprepack
Backfill: reseal ciphertext so payload.parent_id matches the plaintext column (fixes dual-source drift; idempotent).
Calling boundary: RPC.
RPC args template: ["--json","repack"]
All arguments, defaults and help
Backfill: reseal ciphertext so payload.parent_id matches the plaintext column (fixes dual-source drift; idempotent)
Usage: rsrs repack [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpSynchronize and resolve conflicts
sync
Two-way sync: pull cloud delta -> land locally -> push local changes (LWW).
Calling boundary: RPC.
RPC args template: ["--json","sync"]
All arguments, defaults and help
Two-way sync: pull cloud delta -> land locally -> push local changes (LWW)
Usage: rsrs sync [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpsync-conflicts
Show conflicts that still need a decision; --all includes resolved and old-generation history.
Calling boundary: RPC.
RPC args template: ["--json","sync-conflicts"]
All arguments, defaults and help
Show conflicts that still need a decision; --all includes resolved and old-generation history
Usage: rsrs sync-conflicts [OPTIONS]
Options:
--id <ID>
--json Machine-readable output: JSON only, without human progress
--all
--refresh Sync latest content and cross-device resolution state first
-h, --help Print helpsync-resolve
Resolve one retained revision; mark it resolved only after the remote confirms.
Calling boundary: RPC.
Choose keep-current, take-incoming or merge only after reading both revisions. head-rev binds the decision to the observed head; merge needs explicit content. A transport failure is not confirmation of remote resolution.
RPC args template: ["--json","sync-resolve","--epoch","<EPOCH>","--rev","<REV>","--head-rev","<HEAD_REV>","--action","<ACTION>"]
All arguments, defaults and help
Resolve one retained revision; mark it resolved only after the remote confirms
Usage: rsrs sync-resolve [OPTIONS] --epoch <EPOCH> --rev <REV> --head-rev <HEAD_REV> --action <ACTION>
Options:
--epoch <EPOCH>
--json Machine-readable output: JSON only, without human progress
--rev <REV>
--head-rev <HEAD_REV> current.rev shown by sync-conflicts, so the head cannot change under you after you compared
--action <ACTION> [possible values: keep-current, take-incoming, merge]
--content <CONTENT> For merge, the explicit merged body; metadata stays on the current version
-h, --help Print helpsync-history
Show retained sync revisions; --remote reads server history without local caching.
Calling boundary: RPC.
RPC args template: ["--json","sync-history"]
All arguments, defaults and help
Show retained sync revisions; --remote reads server history without local caching
Usage: rsrs sync-history [OPTIONS]
Options:
--id <ID>
--json Machine-readable output: JSON only, without human progress
--remote
-h, --help Print helpsync-restore
Restore a kept revision as a new local edit.
Calling boundary: RPC.
Use op-id, or rev together with epoch; do not mix the two selectors. A pending result represents a local edit awaiting synchronization.
RPC args template: ["--json","sync-restore","--op-id","<OP_ID>"]
All arguments, defaults and help
Restore a kept revision as a new local edit
Usage: rsrs sync-restore [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--op-id <OP_ID>
--rev <REV>
--epoch <EPOCH>
-h, --help Print helpsync-reset
Rebuild after a server backup restore; keep current recovery snapshots and pending edits.
Calling boundary: RPC.
RPC args template: ["--json","sync-reset"]
All arguments, defaults and help
Rebuild after a server backup restore; keep current recovery snapshots and pending edits
Usage: rsrs sync-reset [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpFeedback and evaluation
query-log
Query log: each recall's candidates + adoption + model self-grade (local DPO/SFT raw material).
Calling boundary: RPC.
Feedback must describe a real verdict, not a successful tool call. With explicit feedback enabled, mark requires query-log-id from recall.summary.query_log_id and an appropriate source; legacy grading has narrower compatibility rules.
All arguments, defaults and help
Query log: each recall's candidates + adoption + model self-grade (local DPO/SFT raw material)
Usage: rsrs query-log [OPTIONS] [COMMAND]
Commands:
mark Report a self-grade: true hit --good (chosen) / misleading --bad (rejected). ids are comma-separated; 8-char prefixes work; matches recall candidates from the last 10 minutes
help Print this message or the help of the given subcommand(s)
Options:
--limit <LIMIT> [default: 20]
--stats Stats only: query count, adoption rate, empty-candidate rate, good/bad self-grade, top entries
--json JSON output
-h, --help Print helpquery-log mark
Calling boundary: RPC.
RPC args template: ["--json","query-log","mark","<IDS>","--good","--query-log-id","<QUERY_LOG_ID>","--source","model"]
All arguments, defaults and help
Report a self-grade: true hit --good (chosen) / misleading --bad (rejected). ids are comma-separated; 8-char prefixes work; matches recall candidates from the last 10 minutes
Usage: rsrs query-log mark [OPTIONS] <IDS>
Arguments:
<IDS> Candidate entry ids (comma-separated; prefixes ok)
Options:
--good Mark as a true hit (post-training chosen)
--bad Mark as misleading (post-training rejected)
--query-log-id <QUERY_LOG_ID> Explicit request identity from recall.summary.query_log_id (required with feedback enabled)
--source <SOURCE> Origin of this explicit verdict; never inferred from a successful command [default: model] [possible values: human, model, task_outcome]
-h, --help Print helprun-log
Record a procedure run verdict (ok|fail|confounded|no_effect) for a skill/knowledge entry; local-only, feeds how-to ordering and audit.
Calling boundary: RPC.
RPC args template: ["--json","run-log","<ID>"]
All arguments, defaults and help
Record a procedure run verdict (ok|fail|confounded|no_effect) for a skill/knowledge entry; local-only, feeds how-to ordering and audit
Usage: rsrs run-log [OPTIONS] <ID>
Arguments:
<ID> Skill/knowledge entry id (full id or unique prefix)
Options:
--json Machine-readable output: JSON only, without human progress
--status <STATUS> Outcome: ok | fail (plain success/failure) | confounded (other changes interfered, cannot attribute) | no_effect (worked, but changed nothing observable) [default: ok] [possible values: ok, fail, confounded, no_effect]
--note <NOTE> Short note: what deviated or which command failed (local-only, never synced) [default: ""]
-h, --help Print helpbench
Retrieval quality benchmark: run an evalset for hit-rate/MRR, optional baseline compare; mine drafts an evalset from query-log.
Calling boundary: RPC.
All arguments, defaults and help
Retrieval quality benchmark: run an evalset for hit-rate/MRR, optional baseline compare; mine drafts an evalset from query-log
Usage: rsrs bench [OPTIONS] <COMMAND>
Commands:
run Run an evalset: JSONL one case per line {"query","expect":[8-char prefix or full id...],"project?","note?"}; empty expect = negative
mine Mine an evalset draft from query-log (self-grade good is the main signal, adopted is weak)
help Print this message or the help of the given subcommand(s)
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpbench run
Calling boundary: RPC.
RPC args template: ["--json","bench","run","<FILE>"]
All arguments, defaults and help
Run an evalset: JSONL one case per line {"query","expect":[8-char prefix or full id...],"project?","note?"}; empty expect = negative
Usage: rsrs bench run [OPTIONS] <FILE>
Arguments:
<FILE> Evalset file (JSONL)
Options:
--json Machine-readable output: JSON only, without human progress
--topk <TOPK> Eval depth: whether the expected hit is in the top K (default 5) [default: 5]
--save <SAVE> Archive this run as JSON (for a later --baseline compare)
--baseline <BASELINE> Compare against a baseline (JSON from a previous --save): per-case up/down/same
--verbose Per-case detail (default lists only misses)
-h, --help Print helpbench mine
Calling boundary: RPC.
RPC args template: ["--json","bench","mine","--out","<OUT>"]
All arguments, defaults and help
Mine an evalset draft from query-log (self-grade good is the main signal, adopted is weak)
Usage: rsrs bench mine [OPTIONS] --out <OUT>
Options:
--json Machine-readable output: JSON only, without human progress
--out <OUT> Output JSONL file
--limit <LIMIT> Max cases to export [default: 100]
--strict Strict: use self-grade good only, ignore adopted as a weak signal
-h, --help Print helpIdentity, access and sharing
grant
Manage independent read-only subtree grants for local tools.
Calling boundary: ADMIN.
All arguments, defaults and help
Manage independent read-only subtree grants for local tools
Usage: rsrs grant [OPTIONS] <COMMAND>
Commands:
create Create a grant; the token is shown only this once - keep it
list
revoke
help Print this message or the help of the given subcommand(s)
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpgrant create
Calling boundary: ADMIN.
RPC args template: ["--json","grant","create","--root","<ROOT>","--label","<LABEL>"]
All arguments, defaults and help
Create a grant; the token is shown only this once - keep it
Usage: rsrs grant create [OPTIONS] --root <ROOT> --label <LABEL>
Options:
--json Machine-readable output: JSON only, without human progress
--root <ROOT>
--label <LABEL>
-h, --help Print helpgrant list
Calling boundary: ADMIN.
RPC args template: ["--json","grant","list"]
All arguments, defaults and help
Usage: rsrs grant list [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpgrant revoke
Calling boundary: ADMIN.
RPC args template: ["--json","grant","revoke","<ID>"]
All arguments, defaults and help
Usage: rsrs grant revoke [OPTIONS] <ID>
Arguments:
<ID>
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpsession
List or revoke independent cloud login sessions; does not change local decrypt keys.
Calling boundary: ADMIN.
All arguments, defaults and help
List or revoke independent cloud login sessions; does not change local decrypt keys
Usage: rsrs session [OPTIONS] <COMMAND>
Commands:
list List independent sessions for this account (the current one is marked current)
revoke Revoke cloud access for a session; the full UUID comes from session list
help Print this message or the help of the given subcommand(s)
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpsession list
Calling boundary: ADMIN.
RPC args template: ["--json","session","list"]
All arguments, defaults and help
List independent sessions for this account (the current one is marked current)
Usage: rsrs session list [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpsession revoke
Calling boundary: ADMIN.
RPC args template: ["--json","session","revoke","<ID>"]
All arguments, defaults and help
Revoke cloud access for a session; the full UUID comes from session list
Usage: rsrs session revoke [OPTIONS] <ID>
Arguments:
<ID>
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpkeygen
Generate local key material (offline; writes local session/keyfile - no server, no auth).
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","keygen"]
All arguments, defaults and help
Generate local key material (offline; writes local session/keyfile - no server, no auth)
Usage: rsrs keygen [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--pass <PASS> Compatibility flag: unused for encryption since v4 (the super password is system-generated)
--force Force overwrite with offline keys even if this machine already has cloud key material (drops the cloud session)
-h, --help Print helpsecret
Show account keys: Account Secret / five-keys (manual cross-device join; leak = loss of the store).
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","secret"]
All arguments, defaults and help
Show account keys: Account Secret / five-keys (manual cross-device join; leak = loss of the store)
Usage: rsrs secret [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--reveal Show the full Secret and five-keys (default is masked)
-h, --help Print helpkeys-export
Export decrypt keys (super password + Secret Key + vault material) - how a new machine logs in; leak = loss of the store.
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","keys-export"]
All arguments, defaults and help
Export decrypt keys (super password + Secret Key + vault material) - how a new machine logs in; leak = loss of the store
Usage: rsrs keys-export [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--out <OUT> Write a file (mode 600) instead of printing; the file holds every decrypt secret - keep it safe
-h, --help Print helpsuper-reset
Reset the super password: unlock URK with the current code, issue a new one, upload. Data is unchanged.
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","super-reset"]
All arguments, defaults and help
Reset the super password: unlock URK with the current code, issue a new one, upload. Data is unchanged
Usage: rsrs super-reset [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--super <SUPER_PASS> Current super password (omit if already in the local keyring)
-h, --help Print helpfivekeys
Five-keys join (another machine): verify and write a local session; a non-empty --addr also logs in for a token.
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","fivekeys","--pass","<PASS>","--kdf-salt","<KDF_SALT>","--wrapped-urk","<WRAPPED_URK>","--urk-nonce","<URK_NONCE>"]
All arguments, defaults and help
Five-keys join (another machine): verify and write a local session; a non-empty --addr also logs in for a token
Usage: rsrs fivekeys [OPTIONS] --pass <PASS> --kdf-salt <KDF_SALT> --wrapped-urk <WRAPPED_URK> --urk-nonce <URK_NONCE>
Options:
--addr <ADDR> [default: ""]
--json Machine-readable output: JSON only, without human progress
--user <USER> [default: ""]
--pass <PASS>
--secret <SECRET> [default: ""]
--kdf-salt <KDF_SALT>
--wrapped-urk <WRAPPED_URK>
--urk-nonce <URK_NONCE>
--super <SUPER_PASS> Super password (new wrap). If set, login password + Account Secret are not used to unwrap [default: ""]
-h, --help Print helpregister
Register an account (first device): login password for auth + super password wrapping keys. Address defaults to https://api.rsrs.rs; use --addr to override it. Missing user/password drops into an interactive prompt (TTY only).
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","register"]
All arguments, defaults and help
Register an account (first device): login password for auth + super password wrapping keys. Address defaults to https://api.rsrs.rs; use --addr to override it. Missing user/password drops into an interactive prompt (TTY only)
Usage: rsrs register [OPTIONS]
Options:
--addr <ADDR>
--json Machine-readable output: JSON only, without human progress
--user <USER>
--pass <PASS>
--super <SUPER_PASS> Super password (for decrypt; separate from the login password; forgotten = memories unrecoverable)
-h, --help Print helpaccount
Multiple accounts on one machine: each account has its own data dir (local store/keys/session switch as a unit). Logging into a new account creates a profile; account list lists them; account <name> switches (main = primary); remove <name> --yes deletes.
Calling boundary: ADMIN.
Actions: list; use <name>; <name> as shorthand for use; remove <name> --yes. Read listing through RPC; use the host for switching or deleting profiles. A name resolves on the runtime host, not in the caller's filesystem.
RPC args template: ["--json","account","list"]
All arguments, defaults and help
Multiple accounts on one machine: each account has its own data dir (local store/keys/session switch as a unit). Logging into a new account creates a profile; `account list` lists them; `account <name>` switches (main = primary); `remove <name> --yes` deletes
Usage: rsrs account [OPTIONS] <ACTION> [NAME]
Arguments:
<ACTION> Action: list | remove | or a profile name (same as use; main is the primary profile)
[NAME] Profile name (required for remove)
Options:
--json Machine-readable output: JSON only, without human progress
--yes Confirm a real delete on remove (store and keys are unrecoverable)
-h, --help Print helpmigrate
Explicitly copy a selected old library into a new account; no startup migration.
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","migrate"]
All arguments, defaults and help
Explicitly copy a selected old library into a new account; no startup migration
Usage: rsrs migrate [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--source <SOURCE>
--account <ACCOUNT>
--vault Upgrade only the explicitly selected legacy account vault, preserving its data key
--addr <ADDR>
--user <USER>
--pass <PASS>
--super <SUPER_PASS>
--secret-key <SECRET_KEY>
--new-super <NEW_SUPER> Explicit v4 recovery code for the unchanged data key (required on a headless host without keyring)
-h, --help Print helpspace
Spaces (virtual accounts): an owner can create several virtual accounts; each is a space with its own super key and data dir. Join = a member builds that space's profile on their machine from an invite; they can switch freely; leaving is kick.
Calling boundary: ADMIN.
Actions: list; create <name>; use <name>; invite; join <code> or join --code <code>; members [name]; kick --session <id> or kick --all; remove <name> --yes. Invite supports note and readonly. A readonly invite is server-enforced; joining does not imply switching to that profile.
RPC args template: ["--json","space","list"]
All arguments, defaults and help
Spaces (virtual accounts): an owner can create several virtual accounts; each is a space with its own super key and data dir. Join = a member builds that space's profile on their machine from an invite; they can switch freely; leaving is `kick`
Usage: rsrs space [OPTIONS] <ACTION> [NAME]
Arguments:
<ACTION> Action: list | create <name> | use <name> | invite | join <code> | members | kick | remove
[NAME] Space name (needed for create/use/remove)
Options:
--json Machine-readable output: JSON only, without human progress
--note <NOTE> Note on invite (written to the member session's device_name, so you can tell who is who later)
--readonly invite --readonly: sign a read-only invite (that member can recall, not write; enforced server-side)
--code <CODE> Invite code for join (the name positional also accepts it)
--session <SESSION> Member session id to kick
--all kick --all: revoke every member session in this space
--yes Confirm a real delete on remove (store and keys are unrecoverable)
-h, --help Print helplogout
Log out: clear addr/token and drop the cloud (identity and keys stay; login brings them back); --full forgets this machine's identity.
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","logout"]
All arguments, defaults and help
Log out: clear addr/token and drop the cloud (identity and keys stay; login brings them back); --full forgets this machine's identity
Usage: rsrs logout [OPTIONS]
Options:
--full Delete the whole session.json (including key material - rejoin needs register or five-keys)
--json Machine-readable output: JSON only, without human progress
-h, --help Print helplogin
Log in through browser authorization, or use --interactive for password and TOTP prompts. Address defaults to the last server this machine used (kept in session.json by register/logout). Missing user/password drops into an interactive prompt (TTY only).
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","login"]
All arguments, defaults and help
Log in through browser authorization, or use --interactive for password and TOTP prompts. Address defaults to the last server this machine used (kept in session.json by register/logout). Missing user/password drops into an interactive prompt (TTY only)
Usage: rsrs login [OPTIONS]
Options:
--addr <ADDR>
--json
Machine-readable output: JSON only, without human progress
--user <USER>
--pass <PASS>
--super <SUPER_PASS>
--interactive
Choose OAuth or password/TOTP in this terminal (a supplied --pass selects password login)
--oauth
Explicitly select browser authorization (also the default)
--dashboard <DASHBOARD>
Dashboard origin for a custom server
--no-open
Display the authorization URL without opening a browser
--migrate-local
Copy selected local memories into this server account without plaintext export files
-h, --help
Print help (see a summary with '-h')
Vault recovery: cloud authorization does not decrypt memories. Use the existing recovery code or a trusted device's keys-export, then retry login. Missing or incorrect recovery material must not replace the selected account. super-reset rotates an already unlocked vault; it cannot recover a lost data key. keygen --force switches to unrelated offline keys and is not cloud recovery. If all recovery material is lost, existing encrypted data cannot be recovered by resetting a password. See docs/vault-recovery.md.share
Share a subtree: emit a prompt another AI can paste (payload is plaintext; the other AI imports from the prompt).
Calling boundary: RPC.
The result includes plaintext subtree material. --raw is a terminal presentation mode; RPC still returns its captured envelope, not a raw stdout stream. The returned envelope command can be share-export.
RPC args template: ["--json","share"]
All arguments, defaults and help
Share a subtree: emit a prompt another AI can paste (payload is plaintext; the other AI imports from the prompt)
Usage: rsrs share [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--root <ROOT> Subtree root (8-char prefix ok); default is this machine's scope root, else error
--raw Print only the prompt body (pipes/clipboard), no hint lines
--out <OUT> Write a file instead of printing (prefer <file>.txt with the full prompt)
-h, --help Print helpshare-import
Import a shared subtree: without --go, print attach-candidate bill (for the AI to judge); --go writes the store.
Calling boundary: RPC.
RPC args template: ["--json","share-import","<FILE>"]
All arguments, defaults and help
Import a shared subtree: without --go, print attach-candidate bill (for the AI to judge); --go writes the store
Usage: rsrs share-import [OPTIONS] <FILE>
Arguments:
<FILE> Share text file (full prompt or bare payload)
Options:
--json Machine-readable output: JSON only, without human progress
--parent <PARENT> Attach point: hang this subtree under that existing memory (8-char prefix ok); default is a new root [default: ""]
--go Write switch: omit to only print the candidate bill
--title <TITLE> Override the root title (default = the sharer's original title)
--force Bypass Core-reported conflicts and explicitly create a sibling entry
-h, --help Print helpConfiguration and integrations
agent-config
AI behavior config (agent.json): no args = print the whole file (AI reads it to decide); --set key=value writes a key.
Calling boundary: ADMIN.
RPC args template: ["--json","agent-config"]
All arguments, defaults and help
AI behavior config (agent.json): no args = print the whole file (AI reads it to decide); --set key=value writes a key
Usage: rsrs agent-config [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--set <SET> Set a key, format key=value (e.g. diary_mode=verbose)
-h, --help Print helpconfig
View/set local config (client.json: data dir / server address / auto-sync / auto-cure).
Calling boundary: ADMIN.
Read configuration through RPC. Set data-dir and account/server identity changes through the host coordinator. rpc-parallelism applies after host restart; changing your client's environment does not change running worker capacity.
RPC args template: ["--json","config"]
All arguments, defaults and help
View/set local config (client.json: data dir / server address / auto-sync / auto-cure)
Usage: rsrs config [OPTIONS]
Options:
--data-dir <DATA_DIR>
Set the data directory (absolute path; empty string clears back to default ~/.rsrs)
--json
Machine-readable output: JSON only, without human progress
--addr <ADDR>
Set the server address
--autosync <AUTOSYNC>
Set post-write auto-sync [possible values: true, false]
--cure-auto <CURE_AUTO>
Set periodic auto-cure [possible values: true, false]
--rpc-parallelism <RPC_PARALLELISM>
Max concurrent ordinary runtime jobs. 0 follows available CPU parallelism; explicit values cannot exceed it. Applies after runtime restart
-h, --help
Print helpclassify-config
Inspect or configure classification without running a model.
Calling boundary: ADMIN.
--key-stdin is explicitly rejected through RPC. Supply credentials only through the authorized host terminal/keyring flow; never place real keys in these request examples.
RPC args template: ["--json","classify-config"]
All arguments, defaults and help
Inspect or configure classification without running a model
Usage: rsrs classify-config [OPTIONS]
Options:
--backend <BACKEND> [possible values: jev, ds]
--json Machine-readable output: JSON only, without human progress
--api-base <API_BASE>
--model <MODEL>
--set Persist the selected backend and optional endpoint/model
--key-stdin Read one API key from redirected stdin and store it in the OS keyring
-h, --help Print helpinject
Distribute the inject source: --targets lists 14 hosts; --id <target> injects; --remove --id <target> uninstalls; no args in a TTY opens the picker.
Calling boundary: ADMIN.
Use --targets, --id and --preview for noninteractive integration. --tui needs a terminal. Codex changes support preview revision checking through --expected. Files are read/written on the host, and a hook or injection may affect another application.
RPC args template: ["--json","inject","--targets"]
All arguments, defaults and help
Distribute the inject source: --targets lists 14 hosts; --id <target> injects; --remove --id <target> uninstalls; no args in a TTY opens the picker
Usage: rsrs inject [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--targets
--tui Force the interactive TUI picker (errors if there is no TTY)
--all Inject every detected target (skip the TUI; same as no-args when there is no TTY)
--id <ID> Target id (dsh/opencode/codex/claude/codebuddy/workbuddy/kylinbot/pi/zigcode/deepseek/qwen/doubao/generic)
--remove
--preview Preview a Codex inject or uninstall without writing (requires --id codex)
--expected <EXPECTED> Apply only if the config still matches this preview revision (requires --id codex)
-h, --help Print helpplugin
Plugin hooks (plugins.json): list shows config; test <EVENT> fires an event.
Calling boundary: ADMIN.
All arguments, defaults and help
Plugin hooks (plugins.json): list shows config; test <EVENT> fires an event
Usage: rsrs plugin [OPTIONS] <COMMAND>
Commands:
list List hooks configured in plugins.json (event -> command/timeout/failure policy)
test Fire one event to verify a plugin (--payload is business JSON; a test hook need not actually write the store)
help Print this message or the help of the given subcommand(s)
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpplugin list
Calling boundary: ADMIN.
RPC args template: ["--json","plugin","list"]
All arguments, defaults and help
List hooks configured in plugins.json (event -> command/timeout/failure policy)
Usage: rsrs plugin list [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpplugin test
Calling boundary: ADMIN.
RPC args template: ["--json","plugin","test","<EVENT>"]
All arguments, defaults and help
Fire one event to verify a plugin (--payload is business JSON; a test hook need not actually write the store)
Usage: rsrs plugin test [OPTIONS] <EVENT>
Arguments:
<EVENT> Event name: pre-remember | post-remember | post-recall | post-forget | post-sync
Options:
--json Machine-readable output: JSON only, without human progress
--payload <PAYLOAD> Business payload JSON (default {}); combined with event name/timestamp into the full event on plugin stdin [default: {}]
-h, --help Print helpmcp
MCP server over stdio JSON-RPC. The internal runtime serves HTTP/SSE at /mcp and /sse.
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","mcp"]
All arguments, defaults and help
MCP server over stdio JSON-RPC. The internal runtime serves HTTP/SSE at `/mcp` and `/sse`
Usage: rsrs mcp [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpweb
Open the hosted user dashboard without starting the local runtime.
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","web"]
All arguments, defaults and help
Open the hosted user dashboard without starting the local runtime
Usage: rsrs web [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--no-open Report https://dash.rsrs.rs without opening a browser
-h, --help Print helpModels and indexes
model
Model management: install, verify or uninstall BGE-M3.
Calling boundary: RPC.
install downloads/verifies artifacts without activating an index; activate validates and activates a rebuilt index. install-m3/uninstall-m3 are aliases. reset-cpu and install-engines explicitly require the host. engine changes configuration; probe performs actual inference.
All arguments, defaults and help
Model management: install, verify or uninstall BGE-M3
Usage: rsrs model [OPTIONS] <COMMAND>
Commands:
install Download or verify BGE-M3 without rebuilding the index [alias: install-m3]
activate Build a resumable index generation and activate it after complete validation
reset-cpu Force CPU and terminate the runtime, even when inference or RPC is stuck
engine Show or select inference engine. Failures do not switch engines; default CPU
install-engines Install compatible NPU execution providers through Windows ML
probe Run real BGE-M3 inference and report the selected backend
uninstall Delete the user-installed BGE-M3 files [alias: uninstall-m3]
help Print this message or the help of the given subcommand(s)
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpmodel install
Calling boundary: RPC.
RPC args template: ["--json","model","install"]
All arguments, defaults and help
Download or verify BGE-M3 without rebuilding the index
Usage: rsrs model install [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--mirror <MIRROR>
-h, --help Print helpmodel activate
Calling boundary: RPC.
RPC args template: ["--json","model","activate","<MODEL>"]
All arguments, defaults and help
Build a resumable index generation and activate it after complete validation
Usage: rsrs model activate [OPTIONS] <MODEL>
Arguments:
<MODEL> [possible values: m3]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpmodel reset-cpu
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","model","reset-cpu"]
All arguments, defaults and help
Force CPU and terminate the runtime, even when inference or RPC is stuck
Usage: rsrs model reset-cpu [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpmodel engine
Calling boundary: RPC.
RPC args template: ["--json","model","engine"]
All arguments, defaults and help
Show or select inference engine. Failures do not switch engines; default CPU
Usage: rsrs model engine [OPTIONS] [ENGINE]
Arguments:
[ENGINE] [possible values: npu, gpu, cpu]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpmodel install-engines
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","model","install-engines"]
All arguments, defaults and help
Install compatible NPU execution providers through Windows ML
Usage: rsrs model install-engines [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpmodel probe
Calling boundary: RPC.
RPC args template: ["--json","model","probe"]
All arguments, defaults and help
Run real BGE-M3 inference and report the selected backend
Usage: rsrs model probe [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--model <MODEL> [possible values: m3]
--text <TEXT> Text to compare against the CPU baseline, including long-input checks [default: 本地推理引擎验证：记忆检索]
-h, --help Print helpmodel uninstall
Calling boundary: RPC.
RPC args template: ["--json","model","uninstall"]
All arguments, defaults and help
Delete the user-installed BGE-M3 files
Usage: rsrs model uninstall [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpreembed
Recompute all embeddings (for embedder upgrades / dim changes; re-embed by semantics).
Calling boundary: RPC.
RPC args template: ["--json","reembed"]
All arguments, defaults and help
Recompute all embeddings (for embedder upgrades / dim changes; re-embed by semantics)
Usage: rsrs reembed [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpHost lifecycle
runtime
Host-owned foreground runtime and opt-in current-user sign-in startup.
Calling boundary: HOST.
runtime status and runtime autostart status inspect existing state. run/start/stop/restart and autostart enable/disable/remove explicitly reject worker execution. Autostart is opt-in and current-user only; --now and --stop add immediate lifecycle effects. No call installs a system service.
All arguments, defaults and help
Host-owned foreground runtime and opt-in current-user sign-in startup
Usage: rsrs runtime [OPTIONS] <COMMAND>
Commands:
run Serve in the foreground; service managers supervise this process
start Start the host through its configured lifecycle owner
stop Stop the owned host without automatic resurrection
restart Drain and restart the owned host through its configured lifecycle owner
status Inspect the existing host without creating a profile or starting it
autostart Manage opt-in current-user sign-in startup; never installs a system service
help Print this message or the help of the given subcommand(s)
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpruntime run
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","runtime","run"]
All arguments, defaults and help
Serve in the foreground; service managers supervise this process
Usage: rsrs runtime run [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpruntime start
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","runtime","start"]
All arguments, defaults and help
Start the host through its configured lifecycle owner
Usage: rsrs runtime start [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpruntime stop
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","runtime","stop"]
All arguments, defaults and help
Stop the owned host without automatic resurrection
Usage: rsrs runtime stop [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpruntime restart
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","runtime","restart"]
All arguments, defaults and help
Drain and restart the owned host through its configured lifecycle owner
Usage: rsrs runtime restart [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpruntime status
Calling boundary: ADMIN.
RPC args template: ["--json","runtime","status"]
All arguments, defaults and help
Inspect the existing host without creating a profile or starting it
Usage: rsrs runtime status [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpruntime autostart
Calling boundary: HOST.
All arguments, defaults and help
Manage opt-in current-user sign-in startup; never installs a system service
Usage: rsrs runtime autostart [OPTIONS] <COMMAND>
Commands:
enable Register current-user sign-in startup; start now only with --now
disable Disable future sign-in startup; stop the existing host only with --stop
remove Remove only this installation's startup registration; preserve data and keys
status Read registration and runtime status without changing either
help Print this message or the help of the given subcommand(s)
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helpruntime autostart enable
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","runtime","autostart","enable"]
All arguments, defaults and help
Register current-user sign-in startup; start now only with --now
Usage: rsrs runtime autostart enable [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--now
-h, --help Print helpruntime autostart disable
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","runtime","autostart","disable"]
All arguments, defaults and help
Disable future sign-in startup; stop the existing host only with --stop
Usage: rsrs runtime autostart disable [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--stop
-h, --help Print helpruntime autostart remove
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","runtime","autostart","remove"]
All arguments, defaults and help
Remove only this installation's startup registration; preserve data and keys
Usage: rsrs runtime autostart remove [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--stop
-h, --help Print helpruntime autostart status
Calling boundary: ADMIN.
RPC args template: ["--json","runtime","autostart","status"]
All arguments, defaults and help
Read registration and runtime status without changing either
Usage: rsrs runtime autostart status [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
-h, --help Print helprestart
Restart the resident runtime; kill the verified process if graceful shutdown times out.
Calling boundary: HOST.
Host CLI argument template (not an RPC call): ["--json","restart"]
All arguments, defaults and help
Restart the resident runtime; kill the verified process if graceful shutdown times out
Usage: rsrs restart [OPTIONS]
Options:
--json Machine-readable output: JSON only, without human progress
--timeout <TIMEOUT> Seconds to wait for graceful shutdown before force termination [default: 10]
-h, --help Print helpResults and side effects
All RPC business calls retain the two response layers described in the integration guide. This table explains how to consume each result family; it does not pretend evolving business details are a fixed schema. Inspect the actual version, status, summary, errors, actions and details; do not fabricate success when optional fields are absent.
| Feature family | Result and handling |
|---|---|
status, doctor, audit, witness | Summary and diagnostic rows; warnings and failures are meaningful |
recall, list, show, chain, tree | Permitted record/relationship data in summary/items/details; titles mode deliberately omits bodies |
remember, update, retitle, attach, forget, restore | Action/identifier/revision or pending decision; verify a reported write rather than assuming from exit alone |
diary, diary-search, diary-fragment | Time-chain or fragment data with locators, source versions and budget metadata |
resort, split, classify, tree-deepen, tree-cure, tree-float | Plan or application report; confirm which execution option was requested |
sync, sync-conflicts, sync-history, sync-resolve, sync-restore | Counts, retained revisions, explicit conflict decisions and possibly pending synchronization |
snapshot | summary.action and snapshot details; create/list can report broken witness or partial snapshot problems |
model, reembed | Artifact/index state, selected engine or inference probe; scheduled preparation is not completion |
query-log, run-log, bench | Queries, explicit verdicts and evaluation metrics; feedback is not inferred from request success |
import, export, backup, extract, share-import | Paths, counts and item outcomes; paths point to the host, not file-transfer bodies |
grant, session, account, space | Action-specific identity/access details; grants and invites can contain credentials |
secret, keys-export, super-reset | Explicit sensitive output; do not treat these like ordinary diagnostic commands |
book-material, portrait-material, share | Plaintext material/prompt data; command names may differ from CLI spelling |
agent-config, config, classify-config, inject, plugin | Configuration, target/file actions or hook results; external side effects are possible |
Health and background fields
These fields are from the DEV reference and can be absent on older hosts. They are diagnostics, not a stable business-record schema. The backend supplies the inference object; consumers must accept additions, nulls and explicit errors without depending on undocumented Core internals.
| Field | Meaning |
|---|---|
server, v, bin, pid | Runtime identity, protocol, binary version and process ID |
url, exe, data_dir | Bound URL, executable and selected library path |
store_ready, installation_root | DEV store readiness and installation-wide resource root |
managed_id, credential_state, snapshot_restore_id | DEV supervision, credential and restore diagnostics; optional/null values are meaningful |
recall_statistics.pending | Queued recall-statistics writes |
recall_statistics.enqueued, recall_statistics.persisted | Accepted and persisted statistics counts |
recall_statistics.rejected, recall_statistics.failed | Rejected and failed statistics counts |
retrieval_index.state, retrieval_index.scheduled | Index stage and whether another background pass is requested |
retrieval_index.worker_active, retrieval_index.running | Index worker lifetime and current activity |
retrieval_index.error, retrieval_index.model_setup | Index/preparation error and setup guidance, not a success fallback |
inference | Backend-provided global inference status, including host_recovery_required when present; not one command's progress |
rpc.available_parallelism, rpc.effective_parallelism | Detected CPU parallelism and configured worker capacity |
rpc.ordinary_running, rpc.lightweight_running, rpc.lightweight_capacity | Ordinary/lightweight occupancy and lightweight capacity |
rpc.queue_wait_timeout_secs, rpc.queue_cancelled | Queue admission deadline and cancelled count |
rpc.queued_or_cancelled_pending_removal, rpc.executing_until_reply | Queued/cancelled requests awaiting removal and executions retaining their slot until reply consumption |
rpc.recent_timing_micros, rpc.timing_scope | Recent queue/execution timing diagnostics and their measurement scope |
The runtime.result envelope.summary contains request_id, state, response and retention: use the original response when completed, wait when pending, and verify business state when unknown. model.prepare and index.prepare return the index-status fields above. model.control returns active and, for a matching task, id, phase, item, done, total, elapsed, idle and cancelled; a mismatched task reports stale. cli.progress phase data is separate from global inference; global waiting counts are not the current command's queue position.
Complete MCP adapter
This is the other protocol served by the same runtime. MCP uses JSON-RPC 2.0, unlike the custom /api/rpc envelope. The host runs rsrs mcp for stdio, one JSON object per line; client-only mode connects to the existing runtime. HTTP uses /mcp or /sse under the same actual-peer, Host and Origin checks.
This version implements initialize, tools/list, tools/call and ping, and reports protocol version 2025-03-26. Notifications do not produce ordinary responses. HTTP POST returns JSON by default; Accept: text/event-stream produces a message event, and GET produces an endpoint event using the actual bound endpoint. Do not assume durable sessions, server-pushed subscriptions, general cancellation, pagination or event replay. The HTTP adapter accepts batch frames; custom RPC does not use that batch format.
A successful tool invocation encodes the business envelope as a JSON string in result.content[0].text. result.isError indicates adapter/tool-call failure only; false does not exclude a business status of fail or pending. Check JSON-RPC error and isError, then parse text and inspect the business outcome. Results also include memory_context, identifying retrieved memory as historical reference rather than tool authorization.
These request templates target MCP endpoints only, never /api/rpc:
{
"jsonrpc": "2.0",
"id": 1,
"method": "initialize",
"params": {
"protocolVersion": "2025-03-26",
"capabilities": {},
"clientInfo": {
"name": "example-client",
"version": "1.0"
}
}
}{
"jsonrpc": "2.0",
"id": 2,
"method": "tools/list",
"params": {}
}{
"jsonrpc": "2.0",
"id": 3,
"method": "tools/call",
"params": {
"name": "memory_status",
"arguments": {}
}
}{
"jsonrpc": "2.0",
"id": 4,
"method": "ping"
}The catalog contains 18 tools; memory_forget and memory_restore are disabled by default. The host sets comma-separated names through RSRS_MCP_ALLOW and RSRS_MCP_DENY; DENY wins. ALLOW all or * still excludes those two sensitive tools, which require explicit names. Unknown names and empty items are errors. Legacy RSRS_MCP_ALLOW_TOOLS and RSRS_MCP_DENY_TOOLS are used only when the corresponding canonical variable is absent.
Read-only mode retains only read tools; paused mode retains only memory_status, and ALLOW cannot override service mode. A merge_ids merge also requires memory_forget permission. Use the actual tools/list rather than hard-coding the visible count. Direct RPC is not filtered by MCP allow/deny settings; the integration must enforce its own permissions.
Currently memory_query_log_mark forwards only ids/good/bad, not query-log-id/source. Feedback requiring explicit query identity should use cli.exec query-log mark; the tool does not expose every CLI argument. Below are every tool and its original inputSchema. Required fields are listed under required; permission and semantic checks described above still apply.
memory_status
Local store status (data dir, session, counts). Use to check that rsrs is ready.
inputSchema
{
"type": "object",
"properties": {}
}memory_extract
Import local coding-agent sessions (Codex, OpenCode, Claude Code, …) into rsrs memories. Batch write: imports up to limit newest sessions per source, seals each, then auto-syncs; pass confirm=true to write, or dry_run=true to preview
inputSchema
{
"type": "object",
"properties": {
"id": {
"type": "string",
"description": "One platform (codex/opencode/claude/…); omit for all detected"
},
"dry_run": {
"type": "boolean",
"description": "Scan and summarize without writing"
},
"limit": {
"type": "integer",
"minimum": 0,
"description": "Maximum sessions to import per source (newest first)"
},
"since": {
"type": "string",
"description": "today|yesterday|7d|YYYY-MM-DD"
},
"importance": {
"type": "string",
"enum": [
"important",
"trivial"
]
},
"parent": {
"type": "string",
"description": "Attach imported memories under this parent (id / tree / anchor title)"
},
"confirm": {
"type": "boolean",
"description": "Required to write (omit only with dry_run=true)"
}
}
}memory_remember
Store a memory (judge-then-store; --force writes immediately; merge_ids deletes the listed ids)
inputSchema
{
"type": "object",
"properties": {
"content": {
"type": "string",
"description": "Memory body; prefer cause, action, and effect sections"
},
"title": {
"type": "string"
},
"importance": {
"type": "string",
"enum": [
"important",
"trivial"
]
},
"parent": {
"type": "string",
"description": "Parent id or catalog title"
},
"force": {
"type": "boolean"
},
"merge_ids": {
"type": "string",
"description": "Comma-separated ids to merge into this content"
},
"supersedes": {
"type": "string",
"description": "Live important memory replaced by this conclusion"
},
"see_also": {
"type": "array",
"items": {
"type": "string"
},
"description": "Live important memories to link bidirectionally"
},
"type": {
"type": "string",
"description": "Memory kind; default context"
}
},
"required": [
"content"
]
}memory_recall
Semantic recall (local BGE + keyword fusion, no network)
inputSchema
{
"type": "object",
"properties": {
"query": {
"type": "string"
},
"limit": {
"type": "integer",
"description": "Default 20"
},
"related": {
"type": "boolean",
"default": true
}
},
"required": [
"query"
]
}memory_list
List recent memories
inputSchema
{
"type": "object",
"properties": {
"limit": {
"type": "integer",
"description": "Default 10"
}
}
}memory_show
Show one memory in full (including causal-chain context)
inputSchema
{
"type": "object",
"properties": {
"id": {
"type": "string"
}
},
"required": [
"id"
]
}memory_update
Update title/body/tags/kind/importance of an existing memory
inputSchema
{
"type": "object",
"properties": {
"id": {
"type": "string"
},
"title": {
"type": "string"
},
"content": {
"type": "string"
},
"tags": {
"type": "string"
},
"kind": {
"type": "string"
},
"importance": {
"type": "string",
"enum": [
"important",
"trivial"
]
}
},
"required": [
"id"
]
}memory_attach
Attach a child under a parent (explicit reparent)
inputSchema
{
"type": "object",
"properties": {
"id": {
"type": "string",
"description": "Child id"
},
"parent": {
"type": "string"
}
},
"required": [
"id",
"parent"
]
}memory_tree
Causal tree: forest / subtree / outline
inputSchema
{
"type": "object",
"properties": {
"from": {
"type": "string"
},
"outline": {
"type": "boolean"
},
"depth": {
"type": "integer"
}
}
}memory_history
Entry change audit (create/update/delete/restore)
inputSchema
{
"type": "object",
"properties": {
"id": {
"type": "string"
},
"limit": {
"type": "integer"
}
}
}memory_diary_search
Explicit cross-date diary fragments, bounded output; scores are relevance, not credibility
inputSchema
{
"type": "object",
"properties": {
"query": {
"type": "string",
"minLength": 1,
"maxLength": 500
},
"limit": {
"type": "integer",
"minimum": 1,
"maximum": 20
},
"max_chars": {
"type": "integer",
"minimum": 512,
"maximum": 12000
},
"mode": {
"type": "string",
"enum": [
"lexical",
"hybrid"
]
}
},
"required": [
"query"
],
"additionalProperties": false
}memory_diary_fragment
Bounded read of a diary fragment; rejects changed or missing locators
inputSchema
{
"type": "object",
"properties": {
"source_id": {
"type": "string"
},
"fragment_id": {
"type": "string"
},
"neighbors": {
"type": "integer",
"minimum": 0,
"maximum": 1
},
"max_chars": {
"type": "integer",
"minimum": 512,
"maximum": 12000
}
},
"required": [
"source_id",
"fragment_id"
],
"additionalProperties": false
}memory_diary
Diary time chain (trivial zone included)
inputSchema
{
"type": "object",
"properties": {
"limit": {
"type": "integer"
},
"date": {
"type": "string"
},
"from": {
"type": "string"
},
"to": {
"type": "string"
},
"contains": {
"type": "string"
}
}
}memory_chain
Causal chain: ancestors + this entry + descendants
inputSchema
{
"type": "object",
"properties": {
"id": {
"type": "string"
},
"depth": {
"type": "integer",
"description": "Default 3"
}
},
"required": [
"id"
]
}memory_query_log_mark
Self-grade recall candidates: --good (chosen) or --bad (rejected)
inputSchema
{
"type": "object",
"properties": {
"ids": {
"type": "string",
"description": "Comma-separated ids; 8-char prefixes ok"
},
"good": {
"type": "boolean"
},
"bad": {
"type": "boolean"
}
},
"required": [
"ids"
]
}memory_forget
Tombstone a memory (propagates with sync)
inputSchema
{
"type": "object",
"properties": {
"id": {
"type": "string"
}
},
"required": [
"id"
]
}memory_restore
Restore a tombstoned memory by full id
inputSchema
{
"type": "object",
"properties": {
"id": {
"type": "string"
}
},
"required": [
"id"
]
}memory_taxonomy
List the 23 built-in catalog roots
inputSchema
{
"type": "object",
"properties": {}
}Hidden and reserved controls
A complete feature catalog does not promise compatibility for internal controls. The following options are not all shown in ordinary help but matter to integration boundaries. Classification controls are advanced parameters in this reference version and can change; do not depend on unpublished tuning interfaces. Compile-time native-fault methods and the private Core ABI are not third-party RPC features.
| Control | Purpose and boundary |
|---|---|
--progress-id | Integration ID used by cli.progress; independent of the RPC request ID |
--model-task-id | Matching model operation ID for model.control |
classify --samples | Requested classification samples; default 5 |
classify --batch | Requested tree classification batch size; default 12 |
classify --tree-depth | Maximum selected tree depth; default 2 |
classify --min-kids | Minimum child count for causal segments; default 3 |
classify --segments | Maximum causal segments; default 0 means all |
classify --rounds | Maximum auto-reparent rounds; default 5 |
--client-only, --direct, --runtime-internal | Host CLI pre-parser flags, not cli.exec arguments; --direct never authorizes a second writer |
--managed, --installation-root, --supervised-child | Reserved runtime-run bootstrap options; leave them to the host manager |
lifecycle | Reserved RPC request field for host coordination, not an authorization credential |