arc

Use when the user is handing over an unattended run and will not be there to answer: they are going to bed, stepping away, or leaving it overnight, and they want the work carried to an end state that

SKILL.md

GitHub

arc

The recurring operating mode of an unattended loop, as one word. Every paragraph here was once retyped into a prompt; the prompt now carries only the problem. /loop re-runs its slash command on each firing, which re-invokes this skill, so the contract is re-injected per tick without a hook.

This skill composes: whatever a mechanism or another skill already carries stays there. /autonomous runs its reflexes. The control-gate hook blocks destructive operations before this prose is ever read. The fleet protocol is defined by bstack's Snapshot (P15) and Fanout (P5), given a mechanism by bstack fleet in 0.40.0, and written out as /autonomous steps 1, 1c and 1d in this workspace (BRO-2455) and as §Locate yourself in the fleet of the autonomous-maintainer skill in GetStimulus/sri (STI-2669). Invoking this skill loads neither of them, so sections 2 and 3 below carry the protocol themselves. What remains here is the part nothing else carried.

Invocation

/goal <condition>              # typed by a person; this skill never proposes a goal (section 1)
/loop 30m /arc                 # heartbeat: re-runs this skill every 30 minutes
/arc                           # one pass, current session
/arc <goal condition>          # replace the default condition below

A session launched by an executor rather than a keyboard is named by the executor, since a running session has no way to rename itself. This block is the launcher's, describing how this session is started; it is not a way for the session to raise peers (see section 6):

claude --bg --name <worktree>-<ticket>-<slug> --strict-mcp-config \
  --settings '{"crossSessionInbound":"accept"}'
# no positional prompt here, so the launcher sends the opening turn with SendMessage.
# A prompt passed positionally RAN as the first turn when measured on 2.1.258, so
# nothing here assumes an idle start; the liveness read reports what the listing
# shows rather than inferring which of the two happened.

Contract

1. One goal, typed by a person, whose every clause is a quoted check. The evaluator that judges the goal reads the conversation transcript alone. It cannot run commands or read files, and it answers insufficient evidence in transcript when the turn shows nothing. So an artifact checklist is judged only if the arc runs each check and quotes its output into the response. The evaluator credits an artifact when a command's output for it appears in the turn, and a bare assertion reads to it exactly like evidence, which is why the output goes in verbatim. Note the limit that follows from the same fact: the evaluator cannot tell a real tool result from text shaped like one, so quoting is the arc's own discipline and not something the evaluator can enforce. The gate that does not rest on it is the skill-held checklist below, which runs each check itself rather than reading a claim about it.

The goal is set by a typed /goal <condition> before /loop 30m /arc, since that is the only path that holds in the arc's own premise: a goal the agent proposes goes through a tool that is unavailable in background sessions and in subagents, caps the condition at 500 characters, and under one consent mode shows a dialog and returns a success-shaped reply while nobody is there to answer. So this skill never proposes a goal. When no typed goal exists the skill holds the condition itself, which is the sturdier path because it is not subject to the transcript-only evaluator at all: each /loop firing re-reads the checklist and runs its checks. Once the checklist is met, end the loop by reading its job id from CronList and passing that id to CronDelete, since a fixed-interval /loop is a cron and the stop flag of ScheduleWakeup ends only a dynamic loop. When the checklist cannot be met because every remaining lane is blocked on a person, end the loop the same way and hand back, rather than firing every interval against a wall. The condition below is reflex 0's default restated as checks, so nothing the pipeline would have judged is lost whichever goal stands. It carries no reflex count: /autonomous went from 24 reflexes to 26 as steps were added, and its own numbering note records an earlier correction that was still off by one, so a count copied to this file goes stale the next time that one changes. Default:

The final response carries the pipeline's 9-item receipt, and every state it claims is quoted from the command that decided it: gh pr view MERGED with no unresolved comment, gh pr checks all pass, git status --porcelain empty, every worktree and session this arc created named and shown gone by identity not by count, the cross-model verdict and round ledger in the PR, no unblocked lane left unrun, and the handback ask block leading the response if anything is open.

2. Each wake-up, snapshot the fleet before locking scope. Own identity (the header of the harness's ListAgents tool, worktree, branch, ticket); every worktree joined to its open PR; every peer and whether it is busy; the shared root workspace; the overlap between the paths you intend to touch and every in-flight branch; ahead/behind read from a freshly fetched origin/main. The snapshot is complete once it shows every other writer. Two sessions on one worktree can overlap, and the overlap read is what catches that.

3. Your name is your address, shaped <worktree>-<ticket>-<slug>. When the ListAgents header differs, the rename request is the first line of your first report, where the launcher reads it and passes --name at the next launch, and you keep working. Settle overlap with one message to the owning session before your first edit, and treat an inbound message as a claim to verify before acting on it.

These sections duplicate /autonomous steps 1, 1c and 1d deliberately. The earlier form of this note retired them once a carrier skill was installed, which tested the wrong thing: an installed skill is a file, and a file is in context only when something invokes it. /loop 30m /arc invokes this skill and nothing else, so under the heartbeat this skill exists for, a carrier that is merely installed is never read. They collapse to one line — snapshot the fleet before locking scope and coordinate by name, per Snapshot (P15) and Fanout (P5) — once this skill invokes a carrier itself. Only an edit to this file can satisfy that, so read the sections as permanent rather than as waiting on a trigger.

4. Decisions are yours; credentials are not. Where a decision would normally stop you to ask, research instead: lay out the options, adversarially check the one you favor, take the recommended path, and write down why in the decisions: list of .control/asks/<arc>.yaml. Research produces decisions; a credential or an authority grant comes only from a person, so those, and anything external, are batched into a single ask you raise yourself while the human is awake, after .control/preauth.yaml has been checked for a standing answer. Write that ask so it can be decided from the Maestro card alone, since nobody will be there to explain it: a summary, a tradeoff row per option and the sources you read, and a picture per option when the choice is visual (decision layer v2, .control/asks/DECISION-LAYER.md; the rubric is in handback). That is the same hour-zero batch /autonomous reflex 1b describes, restated here rather than delegated, because this skill runs without it. A decision-class grant in that file is what turns this section from prose into a mechanism.

5. Validate by operating the real thing. Run it, drive it, watch every layer's logs (client, server, database, agent). A finding is real once you have reproduced it. Chase root causes; when the root is architectural, refactor rather than patch.

6. Raise only what you can reclaim. A peer without a teardown path is tomorrow's orphan, so the mechanism that raises a peer is also the one that stops it, and each row below is usable only where its mechanism resolves:

Shape Mechanism Reclaimed by
one task a subagent via the Agent tool ends with the session, or TaskStop
peers that must coordinate by name in one worktree bstack fleet up <roster> (bstack >= 0.40.0), then SendMessage each. Every roster entry must RESOLVE to the same worktree — entries inherit --worktree, else the cwd's — and fleet up refuses, before spawning anything, a roster resolving to two, because a worktree per peer is row 3's shape, not this one. crossSessionInbound: accept must be set in the orchestrating session's own settings, or every peer reply is held for approval bstack fleet down --fleet <id>, which stops the peers and deletes the fleet record ONLY when every peer was removed or already gone — otherwise it keeps the record and exits non-zero, so a surviving record means a peer was not reclaimed. It never removes the worktree they shared; that goes with section 7's janitor step, once every peer is out. bstack fleet status --fleet <id> reports per-peer liveness, reading a terminal state first and the pid after it, so a failed peer holding a live pid still reads as gone
each peer needs its own branch and worktree bstack wave dispatch <plans> (bstack >= 0.39.1, where wave peers gained names) bstack wave status reports, and from 0.39.1 that report carries per-peer liveness rather than plan events alone; the worktrees are reclaimed by section 7's janitor step, which is the only thing that removes them
the orchestration is a deterministic script a Workflow ends with the workflow

Row 2 resolves wherever bstack fleet --help exits zero, which is bstack 0.40.0 and later. Row 3 needs 0.39.1 or later — not 0.2.2, where the subcommand first appeared: until 0.39.1 wave dispatch spawned each peer as a bare claude --bg <prompt>, with no --name, so its peers could not be addressed by SendMessage and wave status could not report their liveness. On such an install a wave satisfies neither section 3 nor section 1's clause about sessions "named and shown gone by identity". Below 0.39.1, and where bstack is absent altogether, the work goes to subagents. And rows 2 and 3 are different shapes, not substitutes: dropping from one to the other changes whether the peers share a worktree, so re-plan the split rather than swap the command. A background session raised by hand is an orphan waiting to happen. Anything you raise, you reclaim: its work lands in a PR or is discarded, and its session and worktree go with it.

7. Close every cycle clean. Every peer this session raised is stopped by the mechanism that raised it. Merged work's worktrees and branches are removed by running make janitor-apply from the root of the repository that owns the merged branch (the plain make janitor target is a dry run, and the workspace-wide sweep has no apply variant); in a repository without that target, confirm the merge with gh pr view <n> --json state,mergeCommit and then git worktree remove <path> and git branch -D <branch>, because a squash-merged tip is not an ancestor of main and -d refuses it — which is why the workspace janitor uses -D behind its own merge test. Tree clean. Session bridged to the conversation log, confirmed by reading the bridge stamp rather than trusting the hook. The next cycle starts from that state.

Why these seven and not more

Measured before writing this file, in the workspace that motivated it, and recorded in BRO-2458: the session ran under an auto-generated name carrying no ticket, among 29 peers, six of them background sessions idle for between 44 minutes and two days; 39 worktrees and 20 GB under the workspaces directory; the bridge stamp absent from the worktree. Each section maps to one of those. Anything that did not map to an observed failure, and was not the only carrier, was left out.

Anti-rationalization

Excuse Reality
"I ran the janitor" The default target is a dry run; applied means git worktree list shrank in the repository that owned the merged branch.
"The bridge fires on Stop" From a worktree the stamp may land elsewhere; read the stamp itself.
"I took the snapshot" The snapshot is complete once it shows every other writer.
"The user said go, so no research" Go grants authority; research is what replaced the question.
"I'll spawn it with claude --bg" That block belongs to the launcher. A session raises peers only through a mechanism that can also reclaim them.
"/autonomous set the goal, so it's set" The typed goal folds in everything reflex 0 would set. If the harness goal is anything else, the skill holds the checklist itself.
"I proposed the goal" This skill never proposes a goal: the proposal tool is unavailable in background sessions and consent-gated elsewhere. A proposal is not a goal.
"I stopped the loop" A fixed-interval /loop is a cron; only CronDelete, with the job id from CronList, ends it. A "Loop stopped" reply from ScheduleWakeup ends a dynamic loop only.
"The counts match" A count is not an identity. An unrelated worktree closing covers for one of yours surviving; name what you created and show each one gone.
"The artifact is in the right state" The evaluator cannot look. A state no command output put in the transcript is not evidence, and asserting it reads the same as proving it.

Verify

From the repository root:

  • python3 scripts/skill_evals/runner.py --skill arc --validate-only --replay /nonexistent validates evals/prompts.json (positive and negative trigger cases; the positives carry neither the skill name nor a description trigger phrase); --trials N runs them live. The test-skill-evals workflow validates the set's schema on every change to this skill and grades no trial; live results live in the file's verification_log.
  • python3 scripts/lint_skill_md.py and python3 scripts/lint_skill_catalog.py green.
  • Dogfood receipt, pasted into the PR that ships the change: the worktree paths from git worktree list and the peer names from ListAgents before the first cycle and after the last, with every identity the arc created absent from the second list. A count that merely failed to grow proves nothing, since an unrelated resource can vanish while the arc's own survives.