Download packages/daemon/assets/plugins/openrig-core/skills/openrig-user/SKILL.md from the publisher, or copy the path and open the repository. PlugMCP lists the skill and does not host the file.
Where does openrig-user live?
The publisher keeps it at https://github.com/mvschwarz/openrig.
Which agents can use openrig-user?
Claude, Cursor, and Codex can read a SKILL.md you add from the publisher's repository.
>-
name: openrig-user
description: >-
Use when a specific rig command, subcommand, or flag is already known and you need its exact
syntax, JSON shape, defaults, or error meaning. NOT for natural capability discovery, open-ended
how-do-I questions, or choosing which OpenRig move applies.
OpenRig User
This is an as-built guide to the shipped rig CLI.
Use current code and rig ... --help as ground truth if anything here ever conflicts with older planning docs.
Use the OpenRig builder guidance when changing OpenRig behavior, doctrine, or release posture.
Coordination trust boundary
Coordinate as openly as the trust boundary you actually control allows. Within a set of machines you
own that already trust each other — they can reach each other's daemon URLs and the daemon itself
requires no authentication — direct , , and should just work.
Don't invent bearer placeholders, allowlists, trust handshakes, or guest-reach restrictions the
environment doesn't require.
rig send
rig capture
rig queue
Do not add or defend a restriction without naming the protected asset, the actual adversary, the
blocked attack path, and the concrete consequence. Missing capability is not protection — repair the
route, registration, runtime posture, or product defect directly. Arbitrary boundaries create
constraint-induced circumvention: under goal pressure, agents build shadow transports, alternate
credentials, and weaker proof instead of the direct authorized path. Keep only explicitly stated
external or irreversible boundaries — public push / PR / publish, and destruction of canonical data.
Coordination primitives — when to use which
Three coordination surfaces, used together for forward-momentum work. Internalize this
on first read — it shapes every turn you take in a rig.
rig send <seat> "<text>" — intra-pod direct messaging / nudges
Use when you need to ask a quick question or give a teammate context that does not
carry handoff semantics. NOT for durable work. NOT for state that must survive
across turns. The message lands in the target's pane; there is no durable queue
record. The CLI prints Sent to ... (and Verified: yes/no with --verify); read
the receipt and move on.
Example:
rig send dev-reviewer@example-project "The import report is ready; the durable handoff names its evidence." --verify
Use for any substantive work that must not fall through chat — slice handoffs,
guard verdicts, QA results, full-tip reviews, multi-item batches. Survives agent
restarts. Tracked in the daemon SQLite schema. Surfaces in Project / queue views
in the destination seat's inbox. Tag with mission / slice / gate / checkpoint
so future-you (and any peer) can find it.
Body discipline: substantive bodies go through --body-file <path> (or - for stdin) — the
purpose-built, corruption-safe surface (it kills the backtick-shell-corruption class for multi-line
bodies). Do NOT inline a backtick-heavy or multi-line body via --body: rig queue create body
parsing breaks on unescaped backticks and rejects flag-like tokens.
Use when you have completed your turn on a qitem and the work moves to the next
owner. This is forward momentum. The ball passes to the destination seat;
chain-of-record (the prior qitem id) is preserved so the verdict trail is intact;
tags carry the selected work context forward. Gate tags describe checks actually
selected for that work; they do not require a fixed sequence of roles.
A turn ends by passing the ball, never by going idle holding the slice waiting
on a confirmation the selected process does not include. Follow the current
mission-slice-sop: proportional owner checks are the default; independent review
runs when selected, at the authored work boundary. Role names do not add per-commit
guard, QA, or orchestration gates. Do the authorized work, run its selected checks,
and return the outcome through durable custody.
Valid pauses are only:
A genuine blocker — file a blocked-state qitem against the blocking peer or
surface explicitly to orch.
A scope-or-architecture question that requires owner input and changes the
plan — surface to orch with the specific decision needed.
Implementing already-authorized work is neither of these. Proceed without
phantom-gating on an imagined "next prompt" or "operator confirmation" that the
process does not require.
Anti-patterns
Using rig send for durable work → use rig queue create instead. Sends do
not survive restarts and do not show up in queue/project views.
Idle-holding a slice for an imagined "next prompt" or "operator confirmation"
that the process does not require → pass the ball via rig queue handoff and
proceed to the next slice or stand by for the inbound verdict. See the §1b
doctrine above.
Inlining a multi-line / backtick-heavy body into rig queue create --body
→ use --body-file /tmp/<descriptive-name>.txt (or - for stdin), the
corruption-safe surface. The body parser does not tolerate raw backticks or
flag-like tokens inline.
Runtime-Gated Coordination Primitives
OpenRig v0.3.1 is published publicly as @openrig/[email protected] and GitHub Release
v0.3.1. It includes the bundled PL-004 Coordination Primitive System: Phase A
rig stream / rig queue, Phase B rig project / rig view, Phase C
rig watchdog, and Phase D rig workflow / workflow-keepalive.
These are shipped product surfaces in v0.3.x, but they require a compatible
v0.3.x daemon and matching SQLite schema at runtime — the installed package
version is not automatically the version of the daemon serving you. If a
coordination command behaves unexpectedly, confirm the running daemon with
rig whoami --json and daemon status before assuming a product bug.
Default posture:
Treat daemon rig queue, rig stream, rig project, rig view, rig watchdog, and
rig workflow as the product coordination surfaces when the active daemon is v0.2.0 or newer.
Use daemon-backed rig queue for durable routing. update / show / list
complement create / handoff for inspection and state changes; records in an
unrelated store are not evidence that this daemon owns the work.
If a daemon-backed coordination command fails, debug the command/runtime/schema edge directly;
do not assume the right workaround is to drop back to a config-layer primitive.
Do not perform daemon stop/start, production DB copy/mutation, release, publish, or other
consequence-boundary actions unless the operator/workstream has granted that specific gate.
First-user workspace setup
When booting into a rig on a host where the workspace is unset, gap-ridden, or
points at a stale layout, address that before substantive project work. The
shipped surface is small + bounded — reach for the canonical commands rather
than improvising.
rig workspace validate walks the workspace root and emits a structured
frontmatter-gap report against the v0 contract. Exit code is non-zero when
gaps exist (operators chain into hygiene fix loops). Default root is the
current directory; pass a positional path to validate elsewhere. --kind
scopes the contract to a specific workspace kind; omit for a kind-agnostic
structural check.
If rig workspace validate reports a non-zero gapCount OR the workspace
root is unset / unwritable, the workspace needs instantiation — see the next
section.
Instantiate the canonical workspace scaffold
Agent-actionable. The operation is additive and preserves existing files.
rig config init-workspace scaffolds the canonical workspace layout at the
configured workspace.root (default ~/.openrig/workspace):
missions/ — release missions + slices
exhaust/ — project-local coordination exhaust
SPEC.md — project intent
project.yaml — project catalog selections and mission root
workspace.yaml — project registration
.gitignore — local OpenRig state and exhaust exclusions
--root <path> targets a non-default root for this call; --dry-run reports
what would be created without writing. --force is deprecated compatibility
and still preserves existing files.
Redirect the workspace root
Operator-gated when persistent. Agent-actionable when one-shot via env-var.
For a single command:
OPENRIG_WORKSPACE_ROOT=<path> rig <command> ...
For a persistent host-level redirect, the operator changes the config file or
runs the setter:
rig config set workspace.root <path>
ConfigStore precedence: OPENRIG_WORKSPACE_ROOT env > config-file
workspace.root > built-in default ~/.openrig/workspace. The same
precedence governs OPENRIG_WORKSPACE_SPECS_ROOT → workspace.specs_root
(default <workspace_root>/specs).
Prefer the env-var form for one-shot redirects (transparent to operators);
reserve rig config set for changes the operator owns.
Build a workspace from scratch
Agent-actionable. Same surface as the canonical scaffold above; the
workspace.root cascade handles non-existent host paths.
The command additively creates any missing canonical entries and preserves
every existing one; only a complete six-entry scaffold is a no-op. Run
rig workspace validate /path/to/new/workspace --json after to confirm the
contract holds.
Create a workflow inside an existing workspace
Authoring is operator-or-agent; validation + instantiation are
agent-actionable.
Workflow spec files live at:
<workspace_root>/specs/workflows/<name>.yaml
<workspace_root> resolves via the ConfigStore precedence named above.
There is no rig workflow create verb in v0.3.x — the spec YAML is authored
directly. Template by hand from the documented schema, or copy a built-in
starter from <openrig install>/dist/builtins/workflow-specs/ and adapt.
Once written:
Both --root-objective <text> and --created-by <session> are REQUIRED
on instantiate — omitting either yields a Commander required-option
error before the daemon is contacted. --entry-owner <session> is an
optional override for the entry-step owner; default routing is per the
workflow spec.
validate returns a structured ok/error report; instantiate creates a
workflow instance + entry-step qitem. Inspect existing surface state with:
rig workflow specs --json # list registered specs (built-in + operator-authored)
rig workflow list --json # list active workflow instances
rig workflow show <instanceId> --json # inspect one instance
rig workflow project <instanceId> # ADVANCE an instance — projects the next-step packet
rig workflow continue <instanceId> # read-only inspector of an instance (does NOT advance it)
(Surface note — the current rig workflow command group registers 13 subcommands: validate, instantiate, project, list, specs, show, trace, continue, run, watch, route, resume, status. There is still no create verb — the spec YAML is authored on disk. project is the advancing verb (it projects the next-step packet); continue is a read-only inspector, NOT an advance — do not conflate them. The 13-verb set and the project-vs-continue semantics are verified against current product main d37a08ad (packages/cli/src/commands/workflow.ts, 13 registered .command(...) entries; the earlier "6-verb surface / continue-advances" claim here was stale). Verify individual subcommand flags with rig workflow --help.)
Permission policy — pick one at setup (onboarding)
OpenRig sets only a minimal usability floor on your harness permissions and otherwise stays out of the way — then it ships recommended policies you opt into. It never bakes a permission policy for you. On install / onboarding this is a required choice, presented as a top-level pick:
POLICY MODE — pick a built-in policy and have it applied:
Standard ⭐ (recommended) — routine development including push is allowed; PR creation, publication, merge/release, force-push and destructive actions ask.
Open — allow-by-default; everything except explicitly-destructive, which ask.
The built-in definitions ship as read-only policy spec files (Locked / Standard / Open); applying your pick is the job of the applying-a-permission-policy skill — it translates the chosen spec into your live harness config (Claude settings.json / Codex config.toml), interactively, showing the diff before it writes.
YOLO MODE — done with permissions, just want it to work: OpenRig boots every seat with the harness full-bypass launch flag. No config policy is applied (the bypass overrides it). This is a deterministic OpenRig setting, not a skill.
No choice = the floor — the minimal usability baseline (Claude acceptEdits / Codex workspace-only / Pi --no-approve), one consistent minimum, nothing more.
The floor and YOLO are launch flags OpenRig sets deterministically; the Locked / Standard / Open policies are config-file policies the skill applies (agent-driven, because harness config formats drift). A rig carries its chosen policy on its spec and boots with it — see Lifecycle → Bring a rig up. To (re)apply or change a policy, open applying-a-permission-policy.
v0.3.x Starter, Workspace, And Plugin Surfaces
OpenRig v0.3.0 adds rig agent-image, rig context-pack, rig workspace, and
rig config init-workspace. (0.5.0: the rig context-pack alias is retired — the store + compose library is the single rig context noun; see "Context packs and paced delivery (0.5.0)".) It also shifts fresh-user starter guidance toward
product-team for human-directed work and conveyor for workflow-oriented
work. Treat demo as legacy/test content unless a task specifically asks for
the old demo spec. (0.6.6: the built-in teams are starter and factory, plus the
specialist teams code-review, research and pm. product-team, conveyor and
demo are gone, and first-project is starter's old name. workshop is a rig
bundle.)
OpenRig v0.3.1 adds public package/source surfaces for Plugin Primitive v0,
Claude Auto-Compaction Policy, migration 040_workflow_specs_diagnostic,
Library Explorer finishing, Settings Destination Explorer, Dashboard/For You
vellum refresh, storytelling adapter, and action outcome + inline error UX.
rig plugin is read-only at v0:
rig plugin list
rig plugin show <id>
rig plugin used-by <id>
rig plugin validate <path>
There is no rig plugin install verb in v0.3.1. Plugin installation remains
explicit operator copy/symlink to $OPENRIG_HOME/plugins/<plugin-id>/.
The v0.3.1 package introduced opt-in Claude auto-compaction policy through
policies.claude_compaction.* ConfigStore keys. A package version alone says
nothing about a running daemon's configuration; inspect the selected instance
before relying on a policy or its default.
Compatibility checks:
rig down accepts a rig name or id. An ambiguous name matching more than one
active rig is refused with matching ids; use the intended id. It ends every
agent's session and any work in progress, so check rig ps --nodes --rig <rig>
and that the person asked; rig up <rig> --existing restores it.
For queue/view JSON or limit differences, compare the installed command's help,
the running daemon version and the actual response. A wrapper mismatch is not
by itself a daemon-health failure, and historical workarounds are not current
behavior guarantees.
After a startup timeout, inspect status and logs before retrying; a timeout
does not establish whether the underlying operation completed.
Recovery and Resilience (v0.3.4+)
v0.3.4's theme is Recovery + Resilience. The surfaces below compose into a
single boot-to-running-rig path that survives crashes, hand-resumed sessions,
profile-load drift, and partial workspace state without silently fudging
status.
rig start — recovery entrypoint
rig start is the top-level recovery sequencer. It does not invent recovery;
it composes existing primitives (daemon start + kernel verify + per-rig
restore) into one call.
rig start # interactive: daemon + kernel + pick-and-restore
rig start --last # headless: restore rigs that were last running; a rig stopped on purpose stays down
rig start --all # headless: restore all rigs with restore-usable snapshots
rig start --rigs <name> [<name>...] # headless: restore only the named rigs
rig start --json # JSON output for agents
rig start --last keeps down a rig that was stopped on purpose (rig down or rig seat stop).
Bring it back with rig start --rigs <rig>. A rig whose seat statuses can't be read is still
attempted, with a warning (seat_status_unreadable).
Framing: rig start is the RECOVERY entry point, not the getting-started
hero. The fresh-user boot hero remains rig up <team> (typically
rig up starter for a first change, or rig up factory). Reach for rig start after a host reboot, daemon
restart, or any "bring my rigs back" moment.
rig reconcile-session — no-launch adopt of a hand-resumed session
When an operator has externally resumed an agent session (e.g. attached a
shell, restarted a runtime by hand) and you want OpenRig to reconcile its
lifecycle state without re-launching or sending input, use:
This is a no-launch, no-input adopt. --rig/--node disambiguate when the
canonical session name does not uniquely resolve. --no-launch is accepted
for explicitness (it is the only mode this command has).
Five-term restore status vocabulary
The shipped restore vocabulary is intentionally honest. It surfaces in
rig up / rig restore / rig ps. Use the term that fits — do not collapse
to a generic "ok/failed":
resumed — seat resumed from its original session/snapshot and is live.
fresh-primed — seat opted into --fresh and was freshly started.
awaiting-decision — zero-session honest state. There is no resumable
session AND no --fresh opt-in was given; the seat is waiting for an
operator decision. Previously fudged as failed; that was wrong — nothing
is broken, the system is asking for input.
attention_required — seat is in a state needing operator attention; not
a transport failure. Clear via rig seat clear-attention once the
attention has been resolved.
failed — the send transport or launch genuinely failed.
This replaces the prior collapsed model (the v0.3.3 four-term vocabulary, in
which rebuilt was a term, is retired).
rig seat clear-attention — audited reconcile of stuck attention
When a seat is stuck in attention_required, do NOT hand-edit SQLite to
fake-clear the state. Use the evidence-gated, operator-attested, audited
reconcile:
--reason <text> can acknowledge startup-status and subset-restore attention.
It does not bypass full-restore continuity checks or an active pane-identity
mismatch/missing-pane check; those run first. Successful identity re-verification
can also clear coexisting startup or subset-restore attention. If neither earlier
path clears attention, omitting --reason requires activity or send evidence.
Startup and restore clears write their corresponding audit events; an identity-only
clear updates the current binding and identity verdict without a clear event.
Acknowledgment or responsiveness alone does not prove that the original
conversation resumed.
A 422 response names the uncleared class and its failed check. If the recorded
native token needs correction and you know the actual token, use
rig seat set-resume-token <session> --token-stdin --reason <explanation>, then
rerun rig seat clear-attention <session> to check the live evidence. The token
update records operator provenance; it does not itself prove continuity.
Stopping/relaunching a working seat is a separate disruptive operation, not a
required cleanup or proof of resumed lineage.
Periodic snapshots — crash-insurance floor
The daemon ships a periodic-snapshot scheduler. It runs independently of
teardown events and provides the crash-insurance floor that prior
event-only/teardown-only snapshots could not provide on hard crashes.
Config keys (SettingsStore):
snapshots.periodic.enabled — default true
snapshots.periodic.interval_seconds — default 300
snapshots.periodic.retention_keep — default 10
Newest-wins semantics: when both auto-periodic and auto-pre-down
snapshots exist for a rig, the freshest of the two is selected for restore.
A newer auto-periodic beats a stale auto-pre-down (the crash fix); a
genuinely-fresher auto-pre-down still wins on graceful cycles. Manual
snapshots are handled separately. See
packages/daemon/src/domain/snapshot-repository.ts for the ordering rule.
The last-snapshot floor surfaces in rig ps / status output so an operator
can see at a glance how recent the crash-insurance floor is.
Codex profile-v2 preflight
Profile-bearing launch/restore surfaces run a profile-load preflight. When
profile-load issues are detected, the failure is honest and actionable
(named error + remediation pointer) instead of a silent partial launch that
would later look like an attention_required seat with no explanation.
cmux launch readiness
cmux-backed launches no longer produce silent partial workspace state. When
parts of the workspace are missing, the launch surfaces partial state
honestly and the UI exposes a one-click open-missing affordance.
(See also ## Token-Efficient Defaults (v0.4.0+) below for the compact-by-default read-command surface that lands in 0.4.0.)
Token-Efficient Defaults (v0.4.0+)
v0.4.0 flips the five most frequently invoked read-commands from firehose-by-default to compact-by-default, and rig queue list adopts the docker / kubectl read-command grammar. All defaults preserve breadth and capability — the firehose is one explicit flag away.
rig ps — scope-aware: bare rig ps = ALL rigs; --nodes = your rig only
rig ps # ALL active rigs on the host, one compact row each — RUN FIRST to know the world
rig ps --rig <name> # one named rig's summary
rig ps --nodes --rig <name> # per-node (seat) detail for a NAMED rig — the normal drill-in
rig ps --nodes # per-node detail — CURRENT rig ONLY (deliberately narrow; NOT the whole host)
rig ps --json # compact JSON (default = a bare array of ALL non-archived rigs)
rig ps --nodes -A # cross-rig node inventory (was v0.3.4 default)
rig ps --nodes --full # complete record (the v0.3.4 per-node default shape; resumeToken VALUE retained here for downstream consumers)
rig ps --nodes --session <sess> # narrow to one canonical session
rig ps --active # opt-in active-state filter (does NOT change the all-states default — ps surfaces topology/readiness, where stopped/recoverable/attention IS the actionable signal)
v0.4.0 breadth + projection changes:
Rig-level rig ps lists ALL active rigs (one row each — the cheap "know the world" view). The --nodes (per-seat) view defaults to your CURRENT rig only (from OPENRIG_SESSION_NAME's @<rig> suffix); --rig <name> picks another rig, -A widens --nodes to the whole host (expensive — prefer --fields/--limit).
Per-node TL;DR projection (compact) is the default; --full returns the raw byte-equivalent passthrough. Daemon-side recoveryGuidance relocated to a guidance-by-reference map (no longer duplicated per-node) — even --full benefits.
All-states stays default (different from rig queue list which defaults to active-only) — for ps, non-running states ARE often the actionable signal.
Resume-token security: --full JSON emits resumeTokenPresent (boolean) — the actual resumeToken value also remains in --full for downstream consumers that legitimately need it, but the compact default never carries it (an orch glance never accidentally leaks token material).
⚠ SCOPE-AWARENESS — the one that bites:rig ps --nodes (and --nodes --json) show ONLY your current rig's seats, by design — the narrow default protects your context window. Narrow output is not the whole world. Never conclude "my rig is the only rig on the host" from a --nodes read — run bare rig ps FIRST (cheap; it lists every rig), then rig ps --nodes --rig <name> for the one you need. (-A widens to the whole-host node view; choose it when that breadth is needed.)
rig whoami # compact: identity + peers names + edges + transcript path
rig whoami --json # compact JSON
rig whoami --json --all-peers # compact JSON with every peer, when output.compact bounds the list
rig whoami --full # complete payload (v0.3.4 default shape)
rig whoami --verbose # alias of --full
The first command every agent runs on boot AND every compaction-restore. The compact default keeps identity-recovery essentials (identity, peers names + sessionNames, peersNote, edges directional kind + to.sessionName, the transcript path). --full adds contextUsage, commands, runtimeContext and workspace. With output.compact on, compact JSON lists at most output.whoami_peers peers (default 8): edge peers first, then your pod, then roster order. peersNote and peersOmitted then say how many were left out, and --all-peers lists every one; edges stays complete. The compact-default is an ALLOWLIST projection — future payload fields default to --full and cannot silently re-bloat the every-boot path.
rig queue list — active-frontier + docker/kubectl grammar
rig queue list # active, compact, CURRENT-rig (docker-ps default)
rig queue list -a # + closed/done history within current breadth (docker -a)
rig queue list -A # cross-rig breadth (kubectl -A)
rig queue list --full # add body + chain-of-record + transition history
rig queue list -o json # compact JSON (token-safe, machine-parseable)
rig queue list --full -o json # full JSON
rig queue list --mine # just the caller's items
rig queue list --destination <s> # destined to <s>
rig queue list --source <s> # sourced by <s>
rig queue show <qitemId> # bounded single-item body preview
rig queue show <qitemId> --full # complete body and chain fields; a closed row adds closingTransition (the close's state, reason, target, actor, time and note)
Four orthogonal axes (scope × history × field-breadth × encoding), all composable. STOP using bare rig queue list as the cross-rig firehose. Default is now active + compact + current-rig. Cross-rig history with full bodies is opt-in via -A -a --full; request only the breadth and fields needed for the question.
rig restore-check # summary counts + not-ready seats (with reasons) only
rig restore-check --full # complete per-seat readiness across the fleet (v0.3.4 default)
rig restore-check --rig <name> # narrow
rig restore-check --as <session> # narrow to one seat
The summary retains not-ready seats and their reasons; --full adds ready-seat detail when needed. Scope the query before expanding its payload.
rig context — context-window usage viewer (0.4.x; REMOVED in 0.5.0)
rig context # compact summary (0.4.x only)
rig context --full # complete current payload
rig context --rig <name> # narrow to one rig
rig context --threshold 80 # filter to seats at/above 80%
Lower leverage than the others; keeps the read-command surface compact-by-default after the 0.4.0 upgrade. ⚠ 0.5.0: this usage viewer is removed entirely and the rig context name is reassigned to the context library (store + compose) — see "Context packs and paced delivery (0.5.0)" below. On a 0.5.0 host, bare rig context is the library, not this viewer.
Keep routine reads bounded
Choose scope, active/history breadth and fields before expanding a result. A
status question usually needs identifiers, owner, state and reason; open the full
body or artifact when it is relevant. Preserve full evidence on disk instead of
repeatedly loading unchanged output. Compact defaults reduce reading cost; they
do not remove the full-detail path or prove that nothing exists outside the scope.
Composes with the progress command + scaffolding to update scope IDs and maturity vocabulary through rig scope. Agents update stage / verified / id through commands rather than hand-editing markdown and drifting. The --against MANDATORY rule on verified is the anti-stale keystone: bare timestamps are rejected because a bare timestamp is exactly what lets stale trackers lie while looking fresh. STOP hand-editing the stage / verified / id fields in scope frontmatter; use the new verbs. Existing missions / slices with id:null ghosts or missing PROGRESS.md are repaired idempotently via repair.
rig skill audit — skill cascade provenance
rig skill audit # human report of findings
rig skill audit --json # structured findings
rig skill audit --severity warn # stale + mirror-drift only
rig skill audit --rig <name> # narrow to embedded skill copies for one rig
Read-only audit of the skill cascade. Detects missing / stale / self-referential / invalid-date / mirror-drift across the canonical skills workspace → product mirror → hub cwd → installed plugin chain. Findings route back to the lifecycle for shaped propagation runs. False-green prevention: when audit evidence is unavailable, the CLI emits unable-to-audit with exit code 2 rather than reporting clean.
rig seat clear-attention — extended to derived projection staleness
clear-attention also reaches restore-derived attention even when
startupStatus=ready and sessionStatus=running. Full-restore attention requires
the restore reconciler's exact native-token and usable-pane checks; --reason
cannot replace them. Subset-restore attention can reach the attestation path
when no active pane-identity class takes precedence. That path records
operator_recovered with runtimeCwdVerified:false; it is an acknowledgment,
not proof of resumed lineage. Without an explicit stored continuity outcome,
inventory leaves continuity null (unverified in rig seat status) unless the
same restore attempt has a matching receipt and reconciliation with strict
native-token and usable-pane proof. Explicit stored outcomes remain historical
facts; acknowledgment or responsiveness alone cannot infer resumed.
Native Codex session id capture
Codex seats can now record the real native session id from the Codex
SessionStart hook instead of relying on scrape-shaped identity. Corroborate the id across the current hook payload, provider history and
managed record before relying on it. A release introducing native capture does
not prove that every existing seat uses it; retain any unavailable or conflicting
identity evidence explicitly.
Codex resume preserves approval posture
Resuming a Codex seat preserves the launching seat's approval/sandbox posture
and profile flags. Product-emitted resume commands carry the posture flags
instead of silently falling back to implicit-deny or an unrelated profile.
Do not "fix" a resumed Codex seat by relaunching it with broader approvals
unless the operator explicitly grants a bounded window. Verify the seat's
active posture first, and preserve it when composing recovery commands.
Use this command to set or restore a seat resume token. It replaces direct
SQLite edits, rejects unauthorized writes and bad/null token false-ready paths,
records redacted audit/provenance, and keeps token material out of command
arguments, stdout, logs, and normal status rows. Use stdin, not an inline flag,
when passing token material.
Core Loop
Most work in OpenRig reduces to this loop:
recover identity: rig whoami (compact default; add --full only when you need the heavy payload)
inspect inventory: rig ps --nodes (compact default; add --full only when you need the firehose)
An agent-managed app is a deployable OpenRig unit made of:
the software or service
one specialist agent dedicated to that software
Treat the specialist as the domain delegate for that app.
The current canonical example is:
rig: secrets-manager
pod: vault
member: specialist
logical ID: vault.specialist
session: vault-specialist@secrets-manager
Typical operator loop:
rig up secrets-manager --cwd /path/to/project
rig ps --nodes --json
rig send vault-specialist@secrets-manager "Check Vault health and report back." --verify
rig env status secrets-manager
rig env logs secrets-manager
Cross-rig communication is valid when the target session resolves uniquely.
Example:
rig send vault-specialist@secrets-manager "Read secret/data/dogfood and report the value." --verify
Use the specialist instead of teaching every peer the same app-specific toolchain.
For Vault, ask vault.specialist to do secrets-domain work rather than improvising curl or Vault CLI usage in unrelated agents.
Identity and Recovery
Start here after launch, compaction, or confusion:
If the daemon is unreachable but identity can still be inferred, --json may return a partial result instead of crashing.
WhoamiResult (v0.3.3+) carries a required peersNote field with three pointers
the agent can use to navigate the rest of the rig from a cold start. The
human-formatted CLI output preserves the literal Peers: line prefix verbatim
(parser/test compatibility) and surfaces the clarifier in-band beneath it; the
JSON form exposes peersNote directly for programmatic consumers.
Inventory and Monitoring
rig ps # ALL active rigs on the host, one compact row each (run FIRST to know the world)
rig ps --nodes # compact node inventory (current rig)
rig ps -A # all-rigs breadth (was the pre-0.4.0 default)
rig ps --nodes --full # complete per-node record (the firehose — opt-in)
rig ps --nodes --json # compact JSON node inventory (add --full for the full record)
v0.4.0 flipped these to compact-by-default — see the rig ps compact-defaults section above; STOP using bare rig ps --nodes --json as a fleet-wide firehose (scope and detail are separate choices). The compact rig ps --nodes node inventory (add --full only when you need the complete record, -A for cross-rig breadth) carries, per node:
session name
runtime
session/startup status
restore outcome (compact: resumeTokenPresent boolean; the token VALUE is in --full)
attach/resume commands
latest error
Other health surfaces:
rig status
rig daemon status
rig config
rig preflight
rig doctor
rig env status <rig>
rig env logs <rig>
rig env down <rig>
Bounded agent self-scout
Use the typed health projection before reading raw coordination history. The
default query is the current seat; widen deliberately when the evidence points
beyond it:
rig health --json
rig health --rig <rig-id> --json
rig health --instance --json
rig health explain <finding-id> --json
Follow the returned stable finding ID and suggestedInspection. Use explain
when the summary matters: it returns the same canonical record with its bounded
window, freshness, literal detector rule, evidence references, and next
inspection. Human output projects those same fields; it does not calculate a
second score.
An empty result means only that no records matched the bounded query. It is
not a healthy assertion. Stale, unavailable, contradictory, and
indeterminate evidence stays explicit. Never read raw SQLite for a self-scout,
and never turn a finding into an acknowledgement, notification, queue row, or
remediation automatically: rig health is strictly read-only.
The send-guard (v0.4.0) — the default is SAFE. A default rig send is guarded: it will NOT submit into an interactive prompt / permission block on the target pane. Flags:
--verify — delivery evidence.
--force — a back-compat no-op on the send DECISION: it never bypasses the interactive-prompt/permission guard and never changes whether a message is delivered (a mid-task/busy pane already sends-with-advisory by default). (It does NOT "bypass activity-risk checks" — that earlier teaching is retired.) It is not fully inert, though — it is still parsed solely to be rejected in combination with --wait-for-idle: rig send … --force --wait-for-idle <n> prints --wait-for-idle cannot be combined with --force, exits 1, and sends nothing. So do not read "no-op" as "--force --wait-for-idle is harmless"; that pairing errors. (Verified against current product main d37a08ad: the guard-bypass no-op is declared at send.ts and confirmed by runtime capture — a plain --force send delivers through the ordinary path; the --wait-for-idle rejection is enforced at send.ts, routes/transport.ts, and session-transport.ts, and confirmed by runtime capture — exit 1, nothing sent.)
--wait-for-idle <seconds> — wait until the target is explicitly idle before sending. Cannot be combined with --force (that pairing is rejected: exit 1, nothing sent).
--raw — send exact text/keystrokes without the From/To messaging envelope (still guarded against interactive prompts).
--dangerously-interact --reason "<why>" — the ONLY override of the prompt/permission guard: deliberately drive an interactive prompt/permission block (implies --raw, requires --reason, audit-logged).
--host <id> — send on a remote host declared in ~/.openrig/hosts.yaml (ssh hosts shell out; http hosts go CLI-direct to the remote daemon).
--from <session> — deprecated and ignored; it does not select the sender. Sender identity comes from the current seat and transport provenance; cross-host envelopes use the durable local origin, not the supplied flag or a relay identity.
--context <ref>(0.5.0) — attach a composed context pack/piece by ref (see "Context packs and paced delivery"). Small piece → send --context; a real pack → rig walk. The noun rig context composes the ref; the verb delivers it.
Durable work goes to the QUEUE, not send.rig send is an ephemeral message to a pane — it can be missed, and its delivery status is pane-render, not receipt. If you are assigning work, or the message is important enough that losing it would be a real bummer, use rig queue (below): it's durable, owned, tracked, and survives compaction and restart. Reach for send for a quick conversational nudge; reach for the queue for anything that must not get lost. Do not default to send for work — that's the most common mistake.
As of v0.3.3, content beginning with -- or - is safe:
rig send <session> "content starting with -- or - is now safe" delivers
literally. The daemon's send_text path carries an explicit --
end-of-options sentinel so tmux no longer parses dash-prefixed content
as its own flags. The CLI surface itself is unchanged. For multi-line
or large bodies handed off as durable work, use
rig queue create --body-file <path> (- for stdin) — that's the
queue-side surface, not rig send.
--verify delivery outcomes (v0.3.3+):
delivered — text + Enter both succeeded and capture re-confirmed the body landed.
rendered-unconfirmed — text + Enter both succeeded but capture could not re-confirm the body (TUI redraw race or scroll). The message landed; the post-send re-check could not prove it. Treat as landed-but-unconfirmable, NOT failure.
failed — the send transport itself failed.
The legacy Verified: yes/no line is preserved verbatim (parser/test
compatibility). A new Delivery: <outcome> line carries the named outcome
above.
Observed operator nuance for --verify:
Sent to ... + Verified: yes (Delivery: delivered) = strong positive delivery evidence.
Sent to ... + Verified: no + Delivery: rendered-unconfirmed = the message landed; capture could not re-prove it. Don't blind-retry — check reply / rig capture / transcript before sending again.
Sent to ... + Verified: no + Delivery: failed = send-transport failure.
no Sent to ... line or a hard error = send failure.
When you get Verified: no, do not immediately retry blindly. First check one of:
a direct reply from the target
rig capture <session>
transcript evidence
queue/outbox state if the message asked for a durable handoff
Use rig broadcast sparingly — prefer rig send fan-out. Without --rig or --pod, broadcast targets
every running session across ALL rigs (plus attached external_cli nodes) — the fastest way to cause a
broadcast storm. Reserve rig broadcast for small rigs or a genuine all-hands emergency. For the normal
"message several seats at once" case, use rig send, which scopes the fan-out and keeps the messaging
envelope, the delivery/interactive-prompt guards, and per-recipient results:
rig send --to dev-impl@my-rig,dev-qa@my-rig "message to specific seats" # named seats (comma-list or repeat --to)
rig send --pod dev "message to one pod" # scoped fan-out
rig send --rig my-rig "message to one rig" # scoped fan-out
history — retrieve with composable filters (sender, since, after, topic)
wait — block until new matching messages arrive (polls history, times out honestly)
clear — delete all messages for the rig (destructive, rig-scoped)
See something, say something
OpenRig has an observation stream — the fleet's zero-friction institutional memory, mined for
real product improvements. When you notice anything worth externalizing, say something and keep
working:
a bug, a rough edge, or something that needs fixing
a feature idea or an improvement
something that worked really well — a technique, tool, or pattern worth spreading
an observation, positive or negative feedback, or something genuinely cool, productive, or funny
rig stream emit --source <your-session> --body "what you noticed"
That's the whole reflex. Don't decide where it goes or who it's for — the intake router triages
(destination/type/urgency/tags are optional hints — --hint-type review|handoff|idea,
--hint-urgency routine|urgent|critical, --hint-tags — never required). One command, then carry
on; the value is the habit, not the polish. Don't overdo it, either: stream real signal, not
narration — a good observation beats ten noisy ones. It's a passing thought you externalize, not a chore.
topic — set a topic marker
watch — SSE or tmux-based live stream
Roundtable protocol:
Inspect old room: rig chatroom history my-rig --limit 5
Save if needed: rig chatroom history my-rig --json > /tmp/old-room.json
Clear if needed: rig chatroom clear my-rig
Set topic: rig chatroom topic my-rig "ROUND START"
Product-native switching of agent auth profiles from the CLI. The runtime is a flag (--runtime <codex>), not a command noun — never rig codex-auth.
rig auth status --runtime codex # presence / mode / parseability / login-state (never prints token contents)
rig auth list --runtime codex # saved profiles
rig auth save <profile> --runtime codex # snapshot the auth FILE (mode-guarded), never echoes contents
rig auth switch <profile> --runtime codex
rig auth validate <profile> --runtime codex
rig auth seats … --runtime codex # seat -> profile registry (metadata only; NOT proof of a live account)
Hard secret boundary: no token value is ever printed, logged, queued, streamed, or committed; status/validate report presence/mode/login-state only; seat labels are metadata, not live-account proof. MVP is --runtime codex; other runtimes use the same surface with a different --runtime, never a parallel command.
Context packs and paced delivery (0.5.0)
Compose context once, hand it to a seat cleanly. A library primitive plus a set of delivery flags. The rule that keeps the grammar coherent — internalize this one: the noun stores and composes; the verbs deliver.rig context never sends anything; delivery is only ever rig send / rig broadcast / rig walk / rig queue.
Version note: this describes the library surface introduced in 0.5.0. Check the installed command and serving daemon before relying on it. In 0.4.x, bare rig context was a context-window usage viewer (above); in 0.5.0 that viewer is removed and the rig context name belongs to the library here.
rig context — the store + compose library (never delivers)
Manage and compose context (any text/markdown) into reusable packs. Every piece and pack has a stable, path-like ref — you address context the way you address files (skills/claude-compaction-restore, reference/rig-spec.md).
rig context list # what's in the library
rig context show <ref> # read a piece or pack
rig context add <source-dir> # install an existing pack directory into the store
rig context preview <ref> # assemble + show a pack WITHOUT delivering it
rig context sync # re-walk discovery roots, refresh the library index
rig context rm <ref>
rig context compose --out packs/<ref> --from <fileA> <fileB> ... # ordered pieces -> a durable pack
Sensible default store location; works unconfigured, can be pointed elsewhere later (another folder now; a machine or URL later).
compose (v1) is honest ordered concatenation of named files into a durable pack with a ref — "here's a file, read a file."
No delivery verb lives on the noun. To get a pack to a seat, hand its ref to a delivery verb below.
rig walk — paced delivery of a sequence
rig walk <seat> --through <ref | file ...> --pace 10s
Walk a seat through a pack: each piece is sent into the pane, spaced by --pace, so the agent processes between sends (the human paste → wait → paste rhythm). Its own top-level verb, push-direction — the walker leads and does not wait for replies; the spacing does the work. Reach for walk on onboarding, repriming, or a fleet update — anything absorbed in order rather than all at once.
The delivery grammar — send a ref, walk a pack, or attach it to a qitem
When
Verb
One thing, now
rig send <seat> --context <ref>
One thing, everyone
rig broadcast --rig <rig> --context <ref>
A sequence, absorbed
rig walk <seat> --through <ref> --pace 10s
Context riding a durable handoff
rig queue create … --body-context <ref>
Rule of thumb: small piece → send --context; real pack → walk. An oversized send --context warns "this is walk-sized" instead of blasting the pane.
--body-context snapshot rule: a qitem built from a ref stores the resolved content in its body plus the ref for provenance — the handoff carries what was actually sent, and a later library edit never silently rewrites a past handoff's history.
The orchestrator habit — assign work with its context attached:
Replace <brief-file> and <proof-file> with your existing local files. The curated context rides the durable handoff, survives compaction, and is auditable.
Skills tier vs context tier: skills are the HOT tier (ambient, finite, always-visible front-matter); context packs are the COLD tier (unbounded, fetched on instruction — "read rig context get onboarding-width"). Don't overrun the skill layer by using skills as context packs — that's what this primitive is for.
Lifecycle
Bring a rig up
rig up <source>
rig up <source> --plan
rig up <source> --yes
rig up <source> --cwd /path/to/project
rig up <source> --existing
rig up <source> --fresh <seat...>
rig up <source> --json
<source> can be:
a rig spec path
a .rigbundle path
a bare name
Bare names are special:
if they match a library spec, rig up launches from the spec library
if they do not match a library spec, rig up treats the name as an existing-rig restore/power-on target
if both exist, rig up fails loudly on ambiguity
Resume-original-by-default (v0.3.4+):
For an existing rig, rig up <name> resumes each seat from its original session/snapshot by default (operation A). Seats that successfully resume report resumed.
--fresh <seat...> is the per-seat opt-in for deliberate fresh-prime (operation B). Named seats are reported as fresh-primed.
--existing forces existing-rig restore semantics on a bare name, bypassing library-spec resolution. Useful when a rig name collides with a library spec name.
Example: rig up --existing my-rig --fresh dev-impl — resume everything in my-rig except dev-impl, which is freshly primed.
Seats with no resumable session land in awaiting-decision (zero-session honest state, NOT failed); see the five-term restore vocabulary in "Recovery and Resilience" below.
--plan (v0.3.4+):
rig up <source> --plan produces a read-only restore plan preview. It surfaces per-seat resume/fresh-prime intent and any awaiting-decision seats without mutating state. Honest async timeout: a stuck plan reports the timeout rather than hanging silently.
Current behavior notes:
--target <root> is only for .rigbundle / package installation. It does not change agent cwd.
rig up --cwd is shipped. rig up --cwd <path> sends a per-run cwd override for all members in that launch.
local:agent_ref values resolve relative to the rig spec directory, not your shell cwd.
if you copy a built-in spec elsewhere, keep its agents/ tree beside the YAML or rewrite those refs to path:/absolute/path
rig specs add <directory> installs a full spec tree when the directory contains rig.yaml or agent.yaml.
Permission policy: a rig carries a permission policy and boots with it (never changed on the fly). Default if none set = the minimum floor (Claude acceptEdits / Codex workspace-only / Pi --no-approve); otherwise a chosen built-in (Locked / Standard / Open) or deliberately none. When you spec or bring up a rig, decide its policy — apply it via applying-a-permission-policy (see also the onboarding menu, "Permission policy — pick one at setup").
rig archive marks a stopped rig as archived (sets archivedAt) without
discarding it. The rig is preserved for later restoration via rig unarchive,
which clears archivedAt and returns the rig to the active set.
Archive vs delete:
rig down --delete — permanent removal; not recoverable.
rig archive — recoverable; the rig is hidden from the default active view but its record + snapshots are preserved.
Visibility in rig ps:
rig ps — active rigs only (default).
rig ps --include-archived — includes archived rigs, marked with *.
SSE events rig.archived / rig.unarchived drive Project / dashboard updates;
consumers that depend on the rig list should subscribe rather than poll.
Environment services
rig env status <rig>
rig env logs <rig> [service]
rig env down <rig>
Use these for service-backed rigs and agent-managed apps.
For secrets-manager, these are the fastest CLI surfaces for:
confirming whether Vault is healthy
reading Vault container logs
stopping the Vault env without tearing down the specialist session first
Release management without killing live claimed sessions
Use rig release for adopted/claimed-session rigs when you want OpenRig to stop managing the rig but leave the tmux sessions alive.
This is the safe recovery/reset surface for the "sessions still exist, management is broken or stale" case.
If the rig contains OpenRig-launched nodes, rig release refuses loudly instead of pretending the mixed rig is safe to detach.
Spec + bindings is the proven recovery pair for adopted rigs.
Spec gives OpenRig the intended topology. Bindings tells OpenRig which discovered live session belongs in each logical node.
Proven adopted-rig recovery workflow
This workflow is proven for the case where the external tmux sessions are still alive:
the new logical IDs appear in rig ps --nodes --rig <rigId>
rig export includes the new pod
Mixed-origin rigs are allowed
One rig can contain both:
adopted nodes bound from already-running sessions
OpenRig-launched nodes created later with rig expand / rig launch
Current safety rule:
rig release is for claimed/adopted-only rigs
if a rig contains launched nodes, rig release fails with contains_launched_nodes
Manager-assisted recovery
The proven operator pattern is:
keep one OpenRig manager session outside the rig it manages
address the target by rig name, not cached rig ID
find the target rig with bare rig ps (lists all rigs), then resolve its owner from rig ps --nodes --rig <target> (a bare --nodes read is your current rig only, not the target's)
send the manager the spec path, bindings path, and verification steps with rig send
This lets ordinary agents ask the manager for OpenRig help instead of every agent needing to be an OpenRig expert.
rig grow is the simplest way to add seats: no YAML, the default agent spec, and one
runtime and working directory for the named seats (--new-pod creates a pod for them).
Check rig grow --help on your installed version. Use a fragment with rig expand (a pod)
or rig add (one member) when a seat needs an explicit model, a permission policy, a
different agent spec or role profile, per-seat runtime or cwd, or startup files.
Node-granular managed partial restore (v0.3.4+):
rig launch <rig> <nodeRef> relaunches a single seat by logical id or node id through orchestration. <rig> is the
rig's name or ID; a name shared by more than one active rig stops with the candidate IDs.
rig launch <rig> --seats <a,b,c> relaunches a comma-separated subset of seats.
--hold-reason <text> records a reason for holding non-target seats during the partial launch.
This is a SUPPORTED managed path. The prior pod_aware_launch_unsupported dead-end is retired; pod-aware narrow launch now goes through this surface rather than ad-hoc rebuilds.
rig add is the top-level verb for the add_member converge op. It adds a
single member to an existing pod from a YAML/JSON member fragment file. The
daemon resolves the named pod, validates the member, runs preflight, and
launches the member in place.
<rig> is the rig id or its exact name (the rigName that rig whoami --json
reports). A name shared by more than one rig is refused before anything is
added, with the matching ids listed (rig_ambiguous under --json). rig add
always targets the local daemon: while a remote host is selected it refuses to
resolve a name (remote_host_selected); an exact local id still works.
HTTP outcomes:
201 — member added; per-node launch state included in the response.
400 — validation_failed or preflight_failed (the fragment or its launch posture is rejected before any state change).
409 — member_conflict (a member with that identity already exists in the pod).
Use rig add when you want additive growth inside a pod without re-running
the full rig expand pod-fragment path or rebuilding the rig.
When the CLI behaves strangely, use the smallest truthful check first:
rig whoami --json
rig daemon status
rig ps --nodes --json
Specific operator rules:
Sent to ... + Verified: no is ambiguous delivery, not automatic failure. Check reply, rig capture, transcript evidence, or queue/outbox state before retrying.
partial rig whoami --json can happen when identity is still inferable but the daemon-backed path is degraded.
the unified-exec-process warning is a host/tooling-layer signal, not automatic proof that the OpenRig topology is unhealthy.
If you hit the unified-exec warning, inspect for stale one-shot helpers before touching live seats:
orphaned one-shot wrappers like tmux send-keys ...
Do not mass-kill:
tmux attach ...
codex ...
claude ...
JSON and Error Posture
Design assumptions that hold in the shipped CLI:
many operator commands support --json
error messages are intended to say what happened, why it matters, and what to do next
daemon-backed commands fail loudly when the daemon is stopped or unhealthy
restore failure is not something you should silently reinterpret as success
After-Compaction Recovery Checklist
rig whoami --json
rig transcript <your-session> --tail 100
rig ps — lists ALL rigs on the host (know the world FIRST); then rig ps --nodes --rig <your-rig> for your seats. ⚠ rig ps --nodes --json alone is your CURRENT rig only — do NOT mistake it for the whole host (a freshly-compacted agent has no other context to catch the lie).
rig chatroom history <rig> --limit 50
Commands That Do Not Exist
Do not assume these exist unless the shipped help starts listing them:
rig claim
rig blame
rig replay
How do I install openrig-user?
Download packages/daemon/assets/plugins/openrig-core/skills/openrig-user/SKILL.md from the publisher, or copy the path and open the repository. PlugMCP lists the skill and does not host the file.
Where does openrig-user live?
The publisher keeps it at https://github.com/mvschwarz/openrig.
Which agents can use openrig-user?
Claude, Cursor, and Codex can read a SKILL.md you add from the publisher's repository.