Skip to content

OpenRig 0.7.14

Choose a tag to compare

@mvschwarz mvschwarz released this 13 Nov 16:51
· 177 commits to main since this release

[0.7.14]

OpenRig 0.7.14 is a patch for steadier circulation on a busy team. A burst of node
reads now takes turns. In the change's own tests, a health check or a queue read
then waited for one step of it instead of the whole burst; a comparison under
load didn't attribute that gain to this change. Bounded queue listings over a
long retained history use indexes instead of scanning it. An assistant connected to
your team can reach a seat by the name its team shows, connected assistants
appear in rig ps and the TUI, and setup guidance tells the agent to fit a
first team to your logins. There are fixes across delivery, seats and the TUI,
many contributed by people using OpenRig. See Known limits for what this
release doesn't fix.

Before you upgrade

  • Stop the daemon, back up the database, install 0.7.14, 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-0.7.14"
    npm install -g @openrig/cli@0.7.14   # or your existing installation method
    rig daemon start
  • Three migrations run before the daemon listens. Migration 163 adds the
    network_installations table and four columns to network_bindings.
    Migrations 164 and 166 build indexes. The author's measurement of 166 took
    3.47 s on 28,976 rows, not a real installation's size, so a large database can
    take longer. The write overhead, and going back to 0.7.13 after these three
    migrations, aren't established; keep the backup. (#3007, #3008, #3090)

  • If the start is reported pending: the CLI waits 20 s, then may print
    "Daemon startup is pending after the readiness deadline". Readiness remains
    unconfirmed. It leaves the child running and the startup reservation
    retained, without recording the daemon. Inspect the address and log it
    prints. The child may later answer, and status can then report it running,
    but that does not finish this start's recording or release its reservation.
    Don't start it again until the steps below are done. The message's restart
    steps stop that child and clear the reservation: stop the named child PID and
    wait for it to exit, inspect daemon-start.lock, and only when both recorded
    launcher and child PIDs are absent, archive the lock and run
    rig daemon start. On a large database the new start can be reported pending
    again; reaching a recorded start this way hasn't been tested.
    rig daemon stop cannot stop that unpublished child.

  • The openrig plugin carries the CLI's version, 0.7.14; openrig-core stays
    0.1.29.

Under load

  • Node reads (rig ps --nodes, the TUI, several agents at once) take turns,
    one step at a time. In the change's own tests, health checks and queue reads
    were no longer held behind a whole burst. A comparison under load didn't
    attribute a gain to this change: health checks and queue reads were faster
    for the release as a whole, rig ps --nodes showed no change, and sends were
    inconclusive. (#3101)
  • Bounded queue listings over retained history use indexes for their ordering
    and for live and archived owner notifications. (#3090)
  • Codex seats re-checked in one readiness poll share process listings where
    each seat's observation order and the 2 s freshness allow; each keeps its own
    later observation and its own fallback. (#3061)
  • New tmux servers raise their soft open-file limit to 8,192 or the hard limit
    where they can, keeping an already higher limit. This makes descriptor
    exhaustion less likely as panes and viewers add up; servers already running
    keep their limits. (#3070)
  • A delegation parent gets one standing notice per child seat for unclaimed
    rows, not one per row, and held work gets its notice back when its typing
    hold clears. (#3049, #3082)

Your team and the people it works for

  • An assistant connected over MCP reaches a seat by its plain name, stored
    under the seat's actual session; an unknown name, or one several seats match,
    says so. (#3088)
  • Connected assistants appear, with their check-in, in rig ps, rig_team and
    the TUI's tables, and their person sees plain tool titles. (#3057, #3058, #3092)
  • rig doctor and setup say a login is needed when a seat you start runs on
    that harness. The operator is taught to check each seat's runtime against
    your logins before launch, name any missing sign-in and let you choose.
    (#3077)
  • When the readable person registry holds nobody, rig daemon start and
    rig gateway participants tell the setup agent to offer registration. The
    kernel operator is taught to include that offer in its greeting and register
    only on your yes. (#3041, #3071)
  • A team your agent hands work to knows how to ask you a question, and your
    answer wakes it. A click on an ask with several questions records that
    choice; the ask stays open until the last one. (#3059, #3045)
  • A pairing request's description of itself is quoted wherever it's shown. (#3067)
  • rig queue list takes exact --tag, --mission and --slice filters, and
    several tags must all match. (#3066)
  • A skill-only edition of the openrig plugin, for Claude's plugin directory,
    follows handed-on work and keeps its reply route. (#3056, #3075)

Fixes

  • An ordinary send is refused as needing input while an open Claude Code or
    Codex /model picker is positively identified; nothing is typed. Close the
    picker before trying to deliver the message. This is shown by the change's
    own tests; it hasn't yet been checked with the real Claude Code and Codex
    programs on this release. (#3083)
  • On a resume of a Codex seat's own conversation in a rig with pods, including
    rig up --existing, OpenRig avoids adding missing guidance blocks to the
    repository's tracked AGENTS.md when Codex is 0.164.0 or later, no guidance
    file or developer instructions are configured, and the configuration and
    block sources can be read. (#3039)
  • With recovery.auto_drive_provider_prompts on, Claude Code's gated-grants
    trust check is never answered for you, and the plain trust prompt is answered
    "Yes". (#3073)
  • rig terminal open --provider herdr reuses a live Herdr workspace, and Rex
    tiles show the seat's agent and its activity. (#3068, #2992)
  • rig broadcast prints a held recipient as retained, not sent. (#3054)
  • Account erasure also removes OpenRig's own wakes and notices for that
    account's seats. (#3089)
  • A snapshot without sessions returns its invalid_snapshot_data blocker
    instead of throwing a TypeError; the restore still stops there. A retried
    no-launch import recovers the original rig by its operation ID. (#3069, #3042)
  • A rebuilt briefing limits optional superseded recap paths to the newest
    three by default when one folder covers that history; the current recap and
    all artifact records remain. Other layouts keep the full list, and
    output.rebuild_recaps set to 0 lists all paths. (#3055)
  • rig compact --state names the input check that stopped a managed
    compaction, and the TUI's graph shows a seat the daemon reads as unknown as
    unknown. (#3053, #3046)
  • Setting a context-pressure threshold explains when an environment override
    inverts the pair, and rig workflow validate warns that harness pins aren't
    checked against seats. (#2567, #3102)
  • The config browser's entries say each setting's type, choices, range, format
    and write route. (#3050)

Experimental

  • rig network is experimental, unsupported and at your own risk. Its connector
    isn't inert: the daemon starts it, sweeps leftover credential folders and
    schedules local rounds even with no binding. (#3007, #3084)
  • A grant service can hold back only new work under its fair-use limit, and an
    app can read what needs a person, and the answers to it, by link. (#3048, #3008)

Behaviour that changed

  • Node reads: a nodes response is no longer one snapshot across its steps,
    and in a burst each nodes response comes later: about 43 ms against 21 ms
    median in a 24-request burst. One very large rig is still one step. (#3101)
  • An assistant naming a seat now hears three errors: destination_ambiguous
    (several seats match; the candidates are listed), destination_not_found (no
    seat it can reach has that name) and destination_unchecked (the team
    couldn't be read; try again). (#3088)

Known limits

  • Queue listings: the new indexes aren't a cure for every stall. Listings of
    active work showed no gain, one case was slower, and writes weren't measured. (#3090)
  • Daemon memory: in a before-and-after comparison under load, the daemon's
    resident anonymous memory and its memory when idle afterwards had medians
    about 23-25 MiB higher on 0.7.14 than on 0.7.13, but the two versions' runs
    overlapped. Neither equal memory nor a regression is established.
  • Draft-aware delivery (#3076): an automatic OpenRig notice can still be left beside
    a draft started during delivery. In the reproduced paste-window cases the person's draft
    stayed intact and no Enter was sent, but OpenRig's text stayed in the input.
    Those cases don't cover a later window. On a conditional send, one that names
    the conversation it expects, an isolated test with a scripted terminal showed
    characters typed while OpenRig ran its last checks before Enter submitted
    together with OpenRig's text, and the send reported success. The same test
    reproduces it on 0.7.13. It hasn't been tested with a real client or model.
  • The TUI's tables can drop a connected assistant's check-in line when its
    rig's seat inventory can't be read; rig roster still prints it. (#3105 isn't in this release)
  • Repository guidance on restore: rigs without pods, old or unknown Codex
    versions, configured guidance and unreadable configuration keep the existing
    file path. Fresh, fork and handover launches are outside this resume fix.
    Existing managed blocks may still refresh, and the seat-skills line still
    belongs to the file writer; check git status. A block whose source can't be
    read is reported missing.
  • 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 0.7.13. 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)
  • From 0.7.13: its known limits
    still apply, except where this release fixes them. The runtime fixes cover
    adding missing guidance on the eligible Codex resumes described above
    (#3039) and the unquoted pairing description (#3067). The fix for
    recognized model pickers (#3083) is shown by the change's own tests and
    hasn't yet been checked with the real programs on this release.
    This release also changes the operator-first registration (#3041,
    #3071) and pre-launch login guidance (#3077). In a few test
    runs with real agents, the operator offered registration once in its
    greeting, on Claude Code and on Codex (one run each); a no registered nobody,
    and a yes registered the default person. The person saw the offer in 3 of 4
    runs; in the fourth, the setup agent didn't pass the operator's offer on.
    The login guidance worked in both runs tested, once from the setup agent and
    once from the operator. 0.7.13's other limits weren't re-checked for this
    release unless named here.

Credits

Thanks to @Kayhusk, @billbenetti, @cavalcanticampos, @Aummadour, @Tankard,
@w4rc0n and @Totopo27 for merged contributions. Thanks also for the reports
behind these fixes: @Kayhusk (#3035), @Tankard (#2583),
@dajiaohuang (#2529), @billbenetti (#2728), @emilio-gagliardi (#3087), @korallis
(#161) and @w4rc0n (#3017).