Skip to content

OpenRig 0.7.16

Latest

Choose a tag to compare

@mvschwarz mvschwarz released this 15 Nov 09:55
· 71 commits to main since this release

[0.7.16]

OpenRig 0.7.16 is a patch for a person's first team and for recovering the same
work. A seat whose policy is the built-in standard or open keeps the team's
launch default, so choosing standard no longer brings more prompts than
choosing nothing. The rigs skill tells a person's own agent to bring a seat's
permission, trust and hook-review prompts to them as questions, then to run the
rig seat continue a seat shows after its startup prompt is answered. A failed
handover's running successor can be adopted, a seat's runtime home can be
recorded without a relaunch, an upgrade keeps OpenRig's packaged recovery
skills, and a send to a missing seat is kept and says so.
There are fixes across delivery, workflows and connected assistants, many
contributed by people using OpenRig.

Before you upgrade

  • Stop the daemon, back up the database, install 0.7.16, then start the
    daemon.
    Your agents keep running through a daemon restart; don't use
    rig down, which ends them.

    rig daemon stop
    cp "$(rig config get db.path)" "$(rig config get db.path).before-0.7.16"
    npm install -g @openrig/cli@0.7.16   # or your existing installation method
    rig daemon start
  • Two migrations run before the daemon listens. Migration 167 adds
    seat_handover_attempts (for rig seat handover --recover), and migration
    169 adds seat_runtime_home_records (for rig seat record-runtime-home). Going back to
    0.7.15 after them isn't established; keep the backup. (#3044, #3220)

  • The plugins: the openrig plugin carries the CLI's version, 0.7.16.
    openrig-core moves to 0.1.31, because its skills changed.

A person's first team

  • A seat whose effective policy is the built-in standard or open keeps
    the team launch default
    , as a seat with no policy does. Any recorded policy
    used to turn the team default off, so choosing standard brought more prompts
    than choosing nothing: ordinary rig reads, git reads and tests all asked.
    The member's policy decides over the rig's. An explicit seat permission
    choice, a Codex envelope or named Codex profile, a Claude seat's own rules,
    none, locked, custom, auto and yolo, and kernel, Pi and terminal seats
    keep their meaning. This keeps the team default; it doesn't apply the whole
    policy, and a seat that keeps it can still ask before the project's own
    commands (see Known limits). The Claude team default includes ask rules for lifecycle commands
    such as rig up and rig down. The Codex team default does not add
    equivalent per-command ask rules; native settings and explicit permission
    choices still apply. The launch plan now says whether a seat gets the team default's allowances,
    instead of a bare launch_posture=floor. OpenRig can't see rules applied in
    project or user native settings files; those seats keep the team default
    alongside them, and keeping it doesn't loosen those rules. (#3230)
  • The rigs skill tells a person's own agent to bring a seat's permission,
    trust or hook-review prompt to them as a question
    by default, instead of
    answering it itself. It tells the agent to show the prompt, capture the seat
    again before sending, and send the digit or letter the screen shows for their
    choice, never their words, and to ask when the screen shows no key. A yes to
    one prompt isn't a yes to the next, a "don't ask again" choice is picked only
    when they picked it, and an instruction they gave outright lets the agent
    decide that kind of prompt within what they said. The agent is told never to
    record its own answer as the person's. The getting-started guide names each authentication, trust,
    hook-review or permission prompt as the person's to answer before that seat
    gets work. (#3206)
  • After the person answers a prompt a seat stopped at while starting, the
    rigs skill tells their agent to check rig ps --nodes again and, when it
    shows Startup details (<seat>) ending in rig seat continue <seat>, to run
    that command, which delivers the seat's startup instructions in the same
    conversation. When the person's choice is Escape, it tells the agent to send
    the key itself (rig send <seat>@<rig> $'\e' --dangerously-interact --reason "…"),
    which is sent once, with no Enter added, and to repeat the prompt check while
    it waits for a seat's result. (#3271)
  • OpenRig's MCP instructions tell the assistant to list the team and ask which
    assistant a person means when more than one could be, instead of picking one
    from its own history. rig mcp serve --front-door says a request is picked up on
    its own only by a running team seat. (#3179)
  • A person's assistant on an external grant sees the team by the name the
    person knows it by, and its members by their roster label. The team's name
    comes from the rig's network roster or, when there's none, from the roster
    that declares the rig. A label over 200 characters, or with control
    characters, line or paragraph separators, bidirectional embeddings, overrides
    or isolates, certain invisible formatting characters, or angle brackets, is
    set aside with label_problem; the member and the roster still load. (#3221, #3249)
  • The rigs skill's setup path teaches an assistant that installed OpenRig for
    a person, and that they keep talking to, to keep its own return address, so
    the team's questions come back to it and it answers them with the person's
    exact words. (#3134)

Recovering the same work

  • An upgrade keeps OpenRig's packaged recovery skills. After rig down, an
    upgrade and rig up --existing, a seat's own unedited copies of OpenRig's
    packaged skills, such as claude-compaction-restore and
    seat-continuity-and-handover, come back at the current packaged contents.
    Before, a packaged skill whose bytes changed in the upgrade wasn't put back:
    the restore named it missing, and the seat came back without it. When the
    upgrade came before rig down, the copy used to stay at the earlier version;
    it's brought up to date now. An edited copy (its bytes or a file mode), a
    file OpenRig didn't write and a copy recorded from another source stay as
    they are, and the restore names each. A skill the seat's profile selects
    itself keeps its own restore policy. Other plugins' skills aren't changed.
    It's tested at the unit and source level; see Known limits. (#3279)
  • rig seat handover <seat> --recover adopts the running successor of a
    failed handover, instead of leaving the seat with its retiring occupant.
    Recovery is for Claude seats: it proves the successor's process and identity
    from the failed attempt's record, then records the conversation that process
    holds now as the seat's resume target. When that conversation can't be read,
    the conversation its launch named is recorded instead, with a warning that a
    /clear before the recovery would make it stale. Four results are new, all
    only on --recover: no_handover_to_recover, handover_attempt_superseded,
    recover_runtime_unsupported and successor_identity_unproven. Recovery
    from a real failed handover wasn't checked when this release was tested. (#3044)
  • rig seat runtime-home <seat> shows the runtime home (Claude Code's
    config directory or a Codex home) the seat's next launch uses.
    rig seat record-runtime-home <seat> --home <abs> --expect-occupant <gen> --expect-original <id> --expect-revision <n> --reason "<why>" records the
    one its existing conversation lives under, with no successor, relaunch or
    manual database edit. Recording sets the home used by the seat's next launch.
    A later explicit home change is retained; restoring a selected conversation
    shows and applies the home selected for that original. A stale expectation returns
    runtime_home_conflict and applies nothing. The result reports whether the
    home is present and whether the original conversation is found there, and
    reports the running process as unverified. (#3220)
  • After a host reboot or a tmux server restart, a rig seat set-runtime
    switch that was still before commit settles as abandoned once its deadline
    has passed, its staging session is gone and no observed pane shows the staged
    successor. Sends to the seat then go through again. Any sign of the successor
    keeps it pending, and the deadline is an estimate, not proof that nothing is
    in flight. It settles only at daemon start or on
    rig seat set-runtime <seat> --recover, and it wasn't tried across a real
    host reboot or tmux server restart. (#3109)
  • A manual rig compact whose preparation stops at a prompt returns to its
    bounded wait instead of disarming, and sends /compact once the seat is idle.
    Nothing is answered for the person, and a typed draft still stops the
    attempt. rig compact <seat> --state shows the prompt; an expiry while it
    still shows names preparation_needs_input. A manual compaction also no
    longer needs a context-usage reading, so rig compact <seat> --skip-map now
    compacts a seat at Claude Code's "Context limit reached" screen even with no
    usage reading (checked once, in Claude Code's default mode). (#3144)
  • When managed restore or preparation fails for a seat with no independent
    delegation parent, the recovery notice goes to the receiving agent or kernel
    operator, then to a registered person. Existing notices keep their ids and
    aren't rerouted. (#3235)
  • An attached or adopted Claude seat can capture its current conversation's
    resume token from process metadata when no seat-named status-line file
    exists, and Claude's conversation title no longer has to equal the seat name.
    A missing bound pane or an ambiguous session leaves the token unknown. Plain
    external-CLI attachment without a tmux pane stays unverified, and restore
    still relaunches an adopted seat under a derived session name. (#3242)
  • rig launch <rig> --seats <seat> no longer refuses because another seat in
    the rig is paused or being typed in; the other seats keep their pane, draft,
    guard mode and binding, and a paused target still refuses. rig launch also
    takes a seat's session name, including an adopted seat's raw tmux name or its
    <name>@<rig> form. A name matching several seats isn't launched. The
    single-seat form returns seat_ambiguous with candidates; under --seats it
    is listed in ambiguousSeats while other unambiguous selections proceed
    normally. If every selection is ambiguous, the reply is seat_ambiguous. (#3184)
  • For the default http://127.0.0.1:<port> target, the TUI's daemon-only
    start (from its crash cart or start-and-return page) passes only the port, so
    it honours the daemon's configured bind, or its default loopback-plus-Tailscale
    bind when none is configured, instead of loopback alone. It no longer promises
    a kernel that isn't listed. (#3181, #3227)

Delivery, the queue and status

  • A send or row to a missing destination is kept and says so. A queue row
    whose rig or seat is absent keeps its exact address, with a destination
    advisory (destination_status, or unmatched_destination_seat for a missing
    seat in a known rig): missing, archived, hosted or unknown. A direct
    send that finds no receiver keeps the message in the sender's outbox; the CLI
    prints "Recorded, not delivered", --verify exits nonzero, and rig walk
    stops there. No forwarding, retry or reassignment is added. (#3142)
  • An otherwise-admitted state-changing park or close stores its supplied
    non-blank --summary or --evidence-ref, instead of refusing the whole
    update with summary_evidence_not_persistable; note-append and re-estimate
    rules remain unchanged. A blank value isn't stored, and the
    receipt names any requested field the stored row doesn't carry. (#3136)
  • A draft typed while a conditional Enter's own checks wait is no longer
    submitted with OpenRig's text: the guard refuses with draft_input_changed,
    sends no key and leaves the input as it is. The input read and the later key
    write are still separate steps; see Known limits. (#3176)
  • A send to a busy Codex seat (seen with 0.167.1 and 0.167.2) whose input is
    empty is delivered once, instead of being pasted and left in the input,
    neither delivered nor held. OpenRig now reads Codex's "tab to queue" hint row,
    and a person's draft under that hint is still held. That's shown by the
    change's own tests; it wasn't checked with the real programs when this
    release was tested. (#3232)
  • A notice from human@host no longer hints at replying there: the hint says
    that address names no person. (#3132)
  • rig ps ends each summary line with the daemon it answered from
    (· daemon <address>); JSON output is unchanged. (#3107)
  • A watchdog records the request's session header as its registering session
    when there is one. That's attribution, not an authenticated identity. (#3194)
  • rig edge add's cycle refusal no longer says restore refuses a cycle, and an
    open resumed ceremony episode keeps its name when what counts as coordination
    changes. (#3163, #3192)
  • The queue's waiting reads reuse their prepared statements per connection.
    Results don't change, and this isn't claimed to fix a daemon freeze. (#3214)

Workflows

  • When a queue handoff closed a workflow's frontier packet,
    rig workflow route to the same target adopts the queue's single handoff
    successor, writing nothing to the queue, and rig workflow project records
    the agent's exit on the closed packet without changing its queue history. A
    route that can't adopt returns 409. A write-free replay needs the same
    original --packet; a canceled packet isn't treated as terminal; a lone
    handed-off graph packet projected as waiting still reports the graph as
    active; and route doesn't mark a parent workflow recovered. (#3252, #3255)
  • rig workflow list filters by mission, project, rig, workflow, slice or tag;
    repeated tags combine with AND. (#3170)

Connected assistants, Codex and tools

  • MCP tools whose CLI verb takes --host accept a host for a registered http
    host: rig_ps, rig_send, rig_capture, and the queue create, show, list,
    update and handoff tools. Only an owner connection can name a remote host; a
    granted connection is refused, and an ssh host gets a teaching error. (#3010)
  • rig mcp token rotate <name> gives a file-backed MCP grant a fresh token in
    place. Clients must switch: the old token stops matching at the next
    authentication, and there's no grace window. (#3200)
  • Per-seat Codex filesystem policies: codex_permissions has an opt-in
    filesystem form with native read, write and deny rules for that seat's own
    invocation, carried through managed launch, resume, restore and rebuild. A
    concrete conflict in the seat's effective Codex configuration stops the launch
    and names a remedy; a configuration that can't be read keeps the authored
    invocation and reports "policy not verified". The configuration read and the
    start aren't atomic. Codex enforces the rules. A seat's resume command in
    rig ps and node detail carries the policy's flags, but not yet its model
    and effort. When this release was tested, enforcement wasn't checked with a
    real Codex seat, and neither were a manual resume with that command nor two
    seats keeping their policies apart. (#3108, #3239)
  • The TUI's Projects view can show missions and slices across every registered
    host, each with a status (ok, stale, unreachable, auth-failed,
    unsupported or unsupported-transport). A failed read shows that host's
    last-known rows marked stale; that cache is in memory, so after a daemon
    restart a failed host shows no rows. (#3191)
  • The openrig plugin's README gives GitHub Copilot CLI's install and seat
    steps; Copilot installs the plugin from the repository's main branch, unlike
    the other hosts' release installs. (#3224)
  • rig ps --nodes --json --fields activityState returns each seat's activity
    object, and daemon.seat_poll_interval_ms reports that a change applies after
    a daemon restart. (#3240, #3241)
  • rig context get reference/product-journey-sdlc.md resolves. (#3208)
  • The rigs skill says how to follow a task handed on to another host
    (rig queue show <id> --host <host>), and that rig queue list shows 100
    rows by default. (#3213)
  • scripts/portability-report.mjs --strict exits 1 when the report finds
    matches; without it nothing changes. (#3202, #3226)
  • Docs and tests: #3197, #3201 and #3203.

Behaviour that changed

  • standard and open seats keep the team default. A seat whose
    effective policy is the built-in standard or open now launches with the
    team default's allowances, as a seat with no policy does. Before, any
    recorded policy turned the team default off. (#3230)
  • A state-changing park or close carrying --summary or --evidence-ref is
    applied
    instead of refused. (#3136)
  • rig ps summary lines end with · daemon <address>. (#3107)

Known limits

  • Restore's overall result: a rig can still read fully_restored while a
    seat restore didn't start (one claimed or added after its roster or snapshot
    was recorded), or a seat kept in its external CLI attachment, goes unnamed.
    A seat restore left as a bare shell after "Harness not started" still counts
    as running in rig ps. restore-check can also say "Restore tries …" for a
    self-attached member that restore leaves out; that member isn't restored.
  • Draft-aware delivery (#3076): typing after the guard's final check and before
    the paste can still leave the notice beside the person's draft. In the tested
    append-mode schedule Enter was refused and nothing was submitted or deleted;
    this is not a guarantee for every interleaving. Before a conditional Enter,
    text typed after OpenRig's input read can still be submitted with OpenRig's
    text; #3176 moved OpenRig's own checks before that read. Neither #3097's nor
    #3176's narrowing is checked with the real programs on this release.
  • Draft-aware delivery with a wide Claude Code footer: with Claude Code's
    full-width auto-update-failure footer on screen (seen at 200x50), a seat in
    the opt-in draft-aware typing mode whose input is empty can hold an ordinary
    send instead of delivering it. The message is kept, not lost, and
    rig seat held-messages shows it.
  • Seats at the edit floor can stop for approval, OpenRig's own commands
    included.
    A Claude seat running at the harness's edit floor can ask for
    approval before shell commands, including OpenRig's own read-only rig
    commands such as rig whoami and rig queue show. A seat with the built-in
    standard or open policy now keeps the team default; other recorded
    choices keep their meaning, and the launch plan says whether a seat gets the
    team default's allowances. A seat that keeps the team default can still ask
    before the project's own commands, such as running its tests or git add
    and git commit. Each prompt waits until someone approves it, in the seat or
    through your own agent, so a team can stall while nobody answers.
  • A question about a seat's prompt can wait with the kernel operator. If
    you aren't registered as the team's person, OpenRig's question about a seat's
    prompt doesn't reach you; it waits with the kernel operator. The kernel
    operator's offer to register you mentions pairing, but seat questions need a
    registered person too. Registering wasn't tested as a fix for this. This is
    OpenRig's own question, separate from your agent reading a seat's screen and
    bringing its prompt to you.
  • Approvals your agent recorded: In one Claude setup, your agent recorded
    approvals it made itself as "Person approved". Those entries alone do not
    establish your approval; check the conversation to see who actually decided.
  • A new Claude agent's startup text can stay unsent in its input box while
    the seat reads ready.
    Press Enter in its terminal.
  • Managed compaction with Claude Code 2.1.322 (other versions weren't run;
    0.7.15 behaves the same):
    • If someone has typed text in a Claude seat's input box when managed
      compaction runs, OpenRig's /compact can be sent as part of that text
      instead of as a command. The seat isn't compacted, but
      rig compact --state reports the compaction, restore and audit as sent.
      Make sure the input box is empty before compacting, and check that the
      seat actually compacted.
    • With the draft-aware typing guard turned on, Claude Code 2.1.322 can leave
      managed compaction's long preparation prompt unsent (it shows as
      [Pasted text #1]), and the attempt stops as preparation_incomplete.
  • Packaged skills across an upgrade (#3279): a Claude seat that loads the
    vendored openrig-core plugin reads these skills from the plugin, so it keeps
    no seat copies to follow. A resumed Codex conversation isn't sent its skill
    catalog again until a fresh start; the files on disk are current. When an
    upgrade changes only one file's mode and another file's bytes, and
    rig down had already removed the first file, the copy is kept and named
    rather than followed. If saving a skill's new record fails, files whose
    source changed are named missing, as before, and a later restore tries
    again.
  • Linked Codex config files: when a Codex seat's config.toml or
    hooks.json is a symbolic link, OpenRig records its hook trust under the
    link's target, as in 0.7.14. A change to key it by the file's own name isn't
    in this release, and such a seat may be asked to review hooks again. With a
    linked config.toml, OpenRig can also miss a setting stored for it that
    turns OpenRig's hooks off, and set those hooks up anyway when it moves them
    into hooks.json. (#3078)
  • Daemon memory: 0.7.14's comparison under load established neither equal
    memory nor a regression against 0.7.13. This release makes no memory claim.
  • From 0.7.15: its known limits
    still apply, except where this release changes them: #3232 repairs the
    input read behind the busy Codex seat case, #3230 changes which seats launch at the edit floor,
    and #3176 narrows the conditional Enter window described above. 0.7.15's
    other limits weren't re-checked for this release unless named here.

Credits

Thanks to @DeryFerd, @billbenetti, @EZotoff, @Kayhusk, @dajiaohuang,
@cavalcanticampos and @w4rc0n for merged contributions. Thanks also for the
reports behind these fixes: @billbenetti (#2985, #2728), @KM199
(#3103) and @Kayhusk (#3193).