Repository navigation
Releases
Releases · irythill/orq
v0.10.1
Fixed
- A
-Tagent on herdr is no longer flaggedreason=no-hookwhile it is still working: herdr shows Claude idle while a subagent or a background task runs, so the no-hook signal now also needs the screen quiet (no busy marker on its status line,agents.<name>.tui_busy, built in for Claude) and the screen and worktree unchanged foragents.<name>.no_hook_graceseconds (default 90, was a fixed 60). The signal is tentative,status=unconfirmed reason=no-hook(not loud, not a failed round):orq waitkeeps waiting on such a prefix while the agent is open and returns the late hook's final status, and the/issueplaybook says how to handle it (#145). orq boardno longer shows a failed run asfail rc=0:orq agentrecords why a run failed when its exit code doesn't say (reason=empty,no-revieworfilterin.done, besidesno-verdict, and in the run's state) and TRACKS shows that reason, also for older runs from their.done(#133).orq boardMIXER shows quota of an account or agent no agent with a role (inagents.<name>.rolesor a[roles.*]section) uses as off instead of as a live meter (#135), and folds per-model rows (antigravity/<model>) into their agent's unless one is at 70% or more (#137).orq boardTRACKS keeps a space between the agent and model columns (antigravity gemini-…) and showsdefaultfor a finished run started without a model, as for a running one (#136).orq board --once(and a piped board) collects CI first for a workspace never collected, instead of printing MONITOR ascollecting...every time (#137).orq gckeeps a worktree whose branch has an open PR (KEEP — PR open (#N)) instead of reporting it asSTALE;STALEnow saysPR closed without merge (#N)orno PR, with the age of the last commit (#51).
Install: npm i -g @irythill/orq then orq setup.
v0.10.0
Added
orq ciand the board follow open PRs for issue worktrees after claim release, report the latest bot review, and notify the live or last owner of new head reviews and CI failures (#141).
Changed
orq ciand the board list only issues with an open PR: a claimed issue with no PR yet, or a merged/closed one, no longer shows asnone(#141).
Fixed
orq agentrefuses to start a review (review-<repo>, its rounds-<n>and redos like-2b) while another review of the same repo in the same issue folder is still alive, so a conductor can no longer stack concurrent reviewers on one issue (exit 3, nothing started) (#134).
Install: npm i -g @irythill/orq then orq setup.
v0.9.0
Added
- Manual waiting strategy for conductor harnesses that can neither be woken by a finished background task nor keep a foreground wait open (
agents.<agent>.conductor.capabilities.foreground_wait = false): the conductor reports the pending prefixes and says it will not continue by itself. The foreground loop now polls a running shell session handed back by the harness instead of starting a duplicateorq wait. New guide:docs/conductor-continuation.md(#119). - Conductor wait/resume strategy by harness capability: only a harness whose finished background task starts a new turn (Claude Code) ends its turn after dispatch; Codex, OpenCode, Copilot, Antigravity and unknown harnesses stay active in an
orq wait --any --timeoutloop and never re-dispatch on a timeout. The conductor instructions and the AGENTS.md block carry the selected strategy;orq instructions strategy [--conductor <agent>] [--json]reports it;agents.<agent>.conductor.capabilitiesoverrides it (#119). - Any configured agent CLI can conduct:
orq conductor [agent] [account|-] [model], with per-agent interactive launch/resume templates, model aliases and account environments. Setup saves the conductor choice, migrates the workspace map toORQ.mdwithout overwriting user instructions, and supplies the issue playbook for Codex and other conductors. Reserve and optional[conductor].headroomfollow the active conductor; legacy Claude launches remain supported (#67).
Fixed
- Worktree cleanup stops processes running inside the worktree before removing it, including processes whose working directory is the worktree root; orq and its own ancestors are preserved (#116).
- Release planning recognizes the original dev tree represented by a squashed release tag, so previously published features are not counted again in the next version or release notes (#118).
Install: npm i -g @irythill/orq then orq setup.
v0.8.0
Added
orq pickscores (agent, model) pairs: a level inmodels.<role>may list several candidate models (medium = ["a@medium", "b@medium"]; a string stays valid). Each pair is scored on its own stats row, with its own exploration and per-model quota check;--explainshows one line per pair. Fixed roles take the first model (#103).
Fixed
- Release check: the last version is the highest
v*tag, not the nearest one reachable fromdev(release tags sit on main's merge commits), so the dev → main PR no longer asks for the version already released.
Install: npm i -g @irythill/orq then orq setup.
v0.7.0
Added
- Docs in sync with 0.7.0 (both READMEs, the
/issueskill, troubleshooting):orq pick's new filters, conductor session names, orq fetching/committing for agents that can't,orq <command>instead of the oldorq-*names; the Portuguese README's config example caught up. Dependabot PR titles are explicit Conventional Commits. - Release automation: every push to
devopens or updates thedev → mainPR with the version the commits call for; aversioncheck fails a PR intomainuntil package.json and both changelogs carry it; merging intomaintags, publishes to npm and creates the GitHub release. PR titles must be Conventional Commits (see CONTRIBUTING.md). - Agent capabilities
commitandnetwork([agents.<name>], default true). For a write run in a git repo, orq does
what the agent can't:commit = false→ orq commits the changes after the run (message from aCommit: ...line in
the answer, else the title) and reports it under## Commit (by orq);network = false→ orq runsgit fetch origin
before the run so the agent can rebase locally. The prompt says so, so the agent doesn't try. Built in: antigravity
commit = false, codexnetwork = false. - Agent capability
base_commit(default: true if the agent takes the prompt, false for a tool that ignores it). A tool that diffs against
origin/<base>itself can't review-b <round-n.sha>, soorq pick review --base <ref>skips it when the base is a
commit, andorq agent -b <commit>falls back to the branch base (--base-branch <branch>, elseorigin/HEAD), warns and
adds## Review scope (by orq)to the answer: the review covers the whole branch, not just the round delta (#76).
Changed
- Node 22 or newer is required (
engines >=22). Node 20 reached end of life in April 2026. CI tests Node 22, 24
and 26.
Fixed
orq statsnames every agent and model: from the job's argv, then the pick, then the log (review tools run as plain CLIs no longer show as?; a-minside a prompt is no longer read as the model; opencode's screen model comes from its config). Time on a screen run stops at the first turn (<prefix>.turns), so a screen left open no longer inflates it; killed runs and older screen runs are left out of the time median (-). opencode tokens are counted again on its newer event format (#59).orq pickskips an agent whose CLI doesn't resolve on PATH (e.g. installed under another Node version), with a reinstall hint, instead of failing after dispatch (#77).- Several conductors on one workspace get distinct session names,
conductor-<workspace>:<herdr label or ORQ_CONDUCTOR_ID>, so they can address each other;orq claim -lends each line with thatsession=(#79). orq agent -Twith Codex sometimes never signalled the end of a turn (#62): a.idleft on the prefix by an earlier
run (another agent or session) made--turn codexdrop every real turn as a side thread. A new launch now clears it
(-Rwrites the resumed session). On herdr, a turn whose hook still doesn't fire (screen seen busy, then idle for
60 s,.donenot rewritten) is signalled by the runner asstatus=fail reason=no-hook, with the screen's last lines
as the answer.policy.max_roundscounts only review rounds:orq agent -R --reason verify|unblock(defaultreview) neither
counts nor is refused; every follow-up goes to.rounds-<session>.logandorq roundsshows them per reason. The
round is counted at launch, so a dispatch refused later costs none./issuenames verify fixesverify-fix-<n>.md(#75).- A review (
orq agent -ron areview*prefix) that exits 0 without a verdict ("I'll start by exploring the repo...")
is nowstatus=fail reason=no-verdictinstead ofok. A verdict = a finding with a severity or an explicit approval
(APPROVED, LGTM, aprovado, no findings...).policy.review_verdict = falseturns it off. Findings written as bare
[blocker] file:linelines (no list marker) now count, fororq roundstoo (#78). - Fix-round reviews converge on large branches:
orq agent -b <commit>attaches only<commit>..HEAD(a review tool
gets the round's start as{base}, not the whole branch again); in/issuea finding outside the fix diff never
opens a round, it goes to a deferred list (security/personal-data leaks stop and ask the user);orq roundswarns
freshwhen two fix rounds bring new heavy findings — split the issue (#72).
Install: npm i -g @irythill/orq then orq setup.
v0.6.1
Fixed
orq setup --workspacewrites.orquestra-conductor.md(and the package ships its template), soorq conductor
starts in a freshly prepared workspace instead of failing with "no .orquestra-conductor.md here" (#63).orq pickblends each agent's prior with its measured delivery/economy/speed,(prior·k + observed·n)/(k + n)
with k =min_samples, instead of a flat prior belowmin_samplesruns: an under-sampled agent that costs far more
no longer wins on economy. Theexplorebonus fades with n and stops afterpolicy.weights.explore_streak(2)
consecutive picks;--explainshowsprior→observedper axis (#54).- Antigravity quota is tracked per model:
orq quotalistsantigravity/<model>lines, an exhausted bucket (no
remainingFraction, only aresetTime) reads as 100% used, andorq pickchecks the model it is about to choose,
falling back to a lighter configured model of the same agent instead of picking an exhausted one (#60). - An antigravity run that ends with
RESOURCE_EXHAUSTED/ "quota reached ... Resets in 5h3m" is recorded as an exhausted
bucket for the model it used (kept until its reset unless the API shows that model with quota left). The run is
status=failand its.out.mdsaysquota exhausted: <model> until <time>, so the conductor re-picks (#60).
Install: npm i -g @irythill/orq then orq setup.
v0.6.0
Added
orq conductor [account|-] [model] [-r <session>]opens the conductor (Claude) in any terminal and OS: finds the workspace
(.orquestra-conductor.mdin the folder, elseCONDUCTOR_WS), mapssonnet|opus|haiku, resolves the account from the
claude-accountsfile, registers the conductor (inside herdr) and runsclaudewith the instructions (--append-system-prompt-file).herdr-conductoris now a thin
wrapper around it (still herdr-only, and renames the pane).
Changed
- The
/issueskill and the conductor templates use theorq <cmd>names (orq agent,orq verify,orq claim,
orq deps,orq baseline,orq pick,orq conductor) and no longer assume herdr; the completion signal is
[HERDR-AGENT DONE]in herdr andorq waitelsewhere. The permission allowlist gains theorq ...patterns and keeps
theherdr-*ones for old installs. - README: daily use is
orq conductorin any terminal. orq boardfolds finished and dead agents older thanui.board_hide_hours(default 6) into a count line;--json
keeps all of them.orq gcalso prunes the stale agent state that no longer shows on the board.orq boardshows an agent with no CPU forui.board_idle_minminutes (default 30) asidle.
Fixed
- Notices no longer go to a closed pane after a handoff:
orq handoffredirects are followed only to a pane herdr still
knows; when the target is gone the notice stays on the last live pane and the stale redirect is deleted (herdr
unreachable: the redirect is followed as before).orq claim --registerand a new handoff also clear redirects that
would point away from, or loop back to, a live conductor. - Finished
-Tagents no longer pile up: with the screen open and no follow-up for[machine] tui_idle_close_min
minutes (default 30,0= never) after the turn ended, the agent is stopped, cleaned up and its herdr tab closed. A
laterorq agent -T -Rreopens the session (tui_resume). On herdr, a turn typed straight into the screen counts as
activity; on tmux only the end of a turn does. Not applied withclose_tab_after = 0. A-Tagent killed from outside
now closes its herdr tab too (a tmux window closes with its runner). - New
examples/agy-impl.json: a default implementer allowlist for agy (git, npm, node, ... plus globs such as
command(/usr/bin/*)andcommand(*/node_modules/.bin/*)); point thewritetemplate at it with--settings <file>
(seeexamples/irythill/config.toml). The denied-commands answer below points to it. - Antigravity (
agy) runs where some commands were denied by its allowlist are no longer lost: the answer gets a
"Denied commands" section (each command and message, plus how to retry or extend~/.config/orq/agy-impl.json), and
a write run whose worktree has changes isokwith the diff stat attached (the session id is kept, so-Rworks).
Without changes (or read only) it staysfail, now with the denied list as the answer instead of an empty one. - A run that ends with rc=128+signal (SIGKILL, memory cap, ...) is
killed, neverok, even with anok_exitlist or a
leftover answer; the.donefile and the notice to the conductor saystatus=killed. - The Antigravity
.out.mdof a denied run names the denied command (taken from the lastCommandLine). orq agentrefuses a prefix that still has a live run (exit 3, nothing started); a dead run's stale results are
cleared. A-T -Rfollow-up typed into that run's own open screen still works.orq gcpath containment resolves a missing prefix through its real ancestor and works on macOS (/var→
/private/var) and Windows, where it ignores case.
Install: npm i -g @irythill/orq then orq setup.
v0.5.0
Fixed
- Notices to the conductor (verify/agent done, CI red) are now delivered by
orqitself, detached, and no longer fail
silently: the oldherdr-notifydepended onjq, gave up after 3 tries in a minute and printed its errors to a tab or
/dev/null. Each notice is logged (<state>/notices.jsonl;orq noticeshows it) with every attempt and its outcome,
retried with backoff (5s, 15s, 45s, 2min, 5min), waits out a dialog and sends to a working conductor (its input queues
the line). orq agentwithout-a(or$AGENT) now refuses instead of starting Codex: anorq pickwith no candidate gave an empty
$(orq pick ...), which started Codex past blocks and quota.orq boardshows each run's own model when a prefix is reused (the state file now records it), and?instead of a
growing duration for a run whose runner is gone.
Added
- Failures are loud, independent of the conductor reading its pane: a herdr notification with sound, a
✗title on the
conductor's pane (display only, clears after 30 min or when acknowledged) and a desktop notification (notify-send;
[ui] desktop_notify = true|false, default true, Linux only). Successes stay quiet (plain herdr notification). orq board: a red line on top lists failures from the last hour whose notice was not delivered, or that the
conductor has not acted on. Acknowledged = the conductor ranorq wait(which printed the result),orq verifyor
orq agentfor that issue after the notice.orq block <name> --quota: blocks until the quota window resets, read from the collector, instead of a fixed period.- Watchdog layer 2:
orq rounds <ID> --diagnose(only on warn/stop) has a cheap, read only agent of another vendor than the
implementer diagnose why the rounds don't converge, from the metrics,fix-<n>.mdand the review outputs, and recommend
continue / split / change-implementer / ask-user. Saved to.orquestra/<ID>/diagnosis-<rounds>.md;orq roundsshows its
RECOMMENDATIONline. New pick rolewatchdog(own pool or[roles.watchdog], else theopinionpool, elsereview). orq agent -R -Fpastpolicy.max_roundsnow refuses (exit 4) without a diagnosis for the current round count that is newer
than the latest fix/review file.-F -Fbypasses the check (the user's own call).
Install: npm i -g @irythill/orq then orq setup.
v0.4.0
Fixed
orq roundsnow counts review findings written with a leading bold severity (**Bloqueante**,**Sugestão —file:7–8:**, PT/EN),
not only[blocker]tags, and reads markdown-link paths, en-dash line ranges and per-round.worktrees/<dir>/prefixes so
the same file matches across rounds.orq quotano longer reports the antigravity login as expired after an hour idle: the access token only lives 1h
and agy renews it on its next run, so with a refresh token in the keyring the note is "token idle (agy refreshes on
next run)" andorq pickkeeps the account. "login expired" stays for a missing refresh token or an API 401.
Changed
install.sh --uninstallnow also runssystemctl --user disable --nowon the orq timers.orq picklimits: 95% of the week (policy.quota_week) and, on the 5h window, 90/85/80% for simple/medium/heavy
tasks (policy.quota_5h), for every agent. An agent is never picked to review or give an opinion on its own code,
whenever another declared agent exists for the role.policy.quota_limitcaps the 5h pick limit and sets the!inorq quota --line.
Migration:quota_limitused to also cap the weekly window; it now only caps 5h, so configs that lowered it to
keep accounts away from the weekly limit should setpolicy.quota_weekexplicitly instead.orq quota's human reset times ("resets 14:30") moved from right after each window to the end of the line, so
<name> 5h=NN% week=NN%always stays intact and easy to grep.
Added
orq rounds <ID> [-w <workspace>] [--json]: watchdog metrics per issue with no LLM, from the files in
.orquestra/<ID>/(fix-<n>.md,review-*.out.md): fix rounds vspolicy.max_rounds, time total and per round,
reviewer findings that come back across rounds (same file + word overlap) and an alert level (ok / warn / stop).
orq boardshows a marker such asr2/2 ⚠next to each conductor's issues.- Scope guideline in the
/issueskill and the workspaceCLAUDE.mdtemplate: split big issues into sub-issues or
sequential batches (one PR each) before dispatching, and stop and split when reviews hitpolicy.max_rounds. - Pacing in
orq pick: an account whose usage, at the pace so far in a window (used ÷ elapsed share, needs the reset
time and 10% of the window elapsed), projects past that window's pick limit at the reset losespolicy.pacing_penalty
(default 0.3, 0 = off) of score. Soft: the hard limits are unchanged; fixed roles are not scored. headroom = Nin an[agents.<name>]section: children are dropped from that agent's accounts once usage reaches
the limits minus N points (5h and week), leaving that share for conductors; account choice respects it.orq pickskips an agent whose quota reading says the login expired, with the fix command as the reason
(e.g.antigravity: login expired: run `agy models`);orq quotakeeps showing it.orq ci [-w <workspace>]... [--json]: CI/CD signal per claimed issue with a PR — queued, running, passed, failed
or none, fromgh pr view's checks on the branch the issue's worktree carries, collected live (it's the user
asking right now). Cached in the state dir like quota;orq boardreads that same cache for a new CI column
and never blocks on the network. A run turning passed or failed pops a herdr notification, and a failure also
sends the owning conductor one line (notify(), or a stdout line for a non-herdr conductor), once per head
commit — a later commit that is still red notifies again.- Quota lines carry each window's reset time (epoch seconds) when the source gives one (Claude, Codex, Antigravity,
MiniMax, Copilot's monthly reset), asreset5h=/resetweek=fields that old cached lines (and anything without
them) still parse fine.orq quotashows them as a human time ("resets 14:30", or weekday + time past today).
orq pick --explainand its exit-3 message say when an agent out on quota comes back. orq handoff: when a conductor's Claude account reachesconductor.handoff_at(default 85% of the 5h window, or the weekly limit), opens a new
herdr space on an account with room, moves its issues and children's notices there and tells the new
conductor to resume. Targets spread across accounts instead of always filling the lowest-usage one: it counts alive
conductors per account first and picks the one with the fewest (then the lowest usage), skipping any at
conductor.max_per_account(default 2). A move is skipped when every window overconductor.handoff_at(5h and/or
weekly) resets withinconductor.reset_grace_min(default 15) minutes, unless the quota reading is unreadable. It saves a
trimmed recent transcript from the old pane as untrusted context
for the new conductor; a failed read does not stop the move. It registers idle conductors
doing PR/merge work.setup/orq-handoff.timerruns it every 5 minutes.orq agent -Renforcespolicy.max_rounds: it counts follow-ups per session (whatever the prefix is called) and
refuses the next with exit 4;-Foverrides with the user's OK./issuefixes thefix-<n>/review-<repo>-<n>
names and limits visual validation (screenshots) to one cheap attempt.orq claimrecords each conductor's name (its herdr workspace label, elseORQ_CONDUCTOR_ID), shows it in-l,
andorq claim --from <name> [-f]finds (and takes over) a conductor's issues by any part of that name./issue
uses it when the user names another conductor, and resumes from.orquestra/<ID>/instead of starting over.
orq claim -l/--fromandorq boardnow show the herdr workspace's current label while the pane is alive
(it may have been renamed since the claim), falling back to the name stored at claim time.orq pick review|opinionexits 2 with a clear message when more than one agent is eligible for the role (declared, not excluded or blocked, with a model for it; before quota checks) and
neither--impl-agentnor--impl-vendoris given (--no-record/--explaindo not change this): nothing to
filter the reviewer down by otherwise.agents.<name>.no_review: a regex that, when it matches the first line of a review tool's answer (or all of a short one) (e.g. revtool's "Nada para
revisar"), ends thatorq agentrunstatus=failinstead ofok— "nothing was reviewed" is not a verdict, so
orq pick's fallback and the conductor treat it as not run.install.sh:orq-gc.service/orq-handoff.servicenow findnodeunder mise and nvm. TheirPATHalways
includes mise's shims directory (stable across versions); whennodeinstead resolves to a version-specific
directory (mise'sactivatePATH, or nvm, which has no static shims dir) the installer writes a drop-in with
that real directory. A plain, system-package install is untouched either way.
Fixed
- Multi-account
orq pick: prefers an account whose pace does not project past the limit before falling back to lowest
usage (whenpolicy.pacing_penalty > 0), and keeps expired accounts with their fix command in the quota rejection note
when the remaining accounts are over the limit. - Switching an agent's account no longer shows the old one's quota or block. The "last reading" quota cache is now
tied to the login (Antigravity), andorq blockrecords which login it was made for: when the agent or Claude
account is logged into another one, the block is dropped on its own (claude, codex, antigravity, copilot, opencode).
Install: npm i -g @irythill/orq then orq setup.
v0.3.0
Added
orq agent -T(andherdr-agent -T): the agent's own screen in TypeScript, in a herdr tab or a tmux window
(it used to be the bashherdr-agent-tui, herdr only). Claude, Codex and OpenCode; any agent viatui/
tui_resume/turn_hookin[agents.<name>]. Every end of turn runsorq agent --turn, which writes
.out.md/.id/.doneand notifies;orq waitworks as with headless runs.-Rtypes the follow-up into the open
screen (a bracketed paste in tmux, so multi-line prompts arrive whole) or, once it is closed, reopens the session.
machine.tui = trueorHERDR_AGENT_TUI=1make it the default where it can run.
Fixed
- OpenCode 2.x: MCP servers were not turned off in child runs (v2 silently drops an MCP entry without
typeand
command); the end-of-turn plugin uses the v2 plugin API. - A screen closed before its first turn ends writes
status=fail, soorq waitno longer waits forever; closing the
tab or the tmux server cleans up the throwaway review worktree.
Removed
setup/herdr-agent-tui(the bash -T); a stale~/.local/bin/herdr-agent-tuilink from an earlier install can go.
Install: npm i -g @irythill/orq then orq setup.