| name | next |
| description | One-step rpm orchestrator, or a bounded internal sequence when given a count or scope. Runs preflight maintenance, then starts the next obvious backlog action (or, in direct use, asks for clarification when nothing is clearly next). With no argument it runs one step; with `N`, `blocked`, `all`, or a group name it runs several steps itself, skipping the heavy preflight between steps, and is the recommended way to work several backlog items at once — cheaper than wrapping `/next` in an external `/loop`. How wide it fans out — one worker at a time or a concurrent batch — is the orchestrator's judgment call, sized to how independent the selected work is; it never waits for input mid-sequence. TRIGGER on terse forward-motion prompts — phrasings like "next", "next?", "next.", "next task", "what's next", "do next", "go next", "keep going", "continue" (when the prior turn was rpm work) all qualify and must route through this skill instead of being answered inline from the SessionStart preview. Use whenever the user wants the session to autonomously work the rpm backlog. |
| argument-hint | [N | blocked | all | <group> | status] |
| allowed-tools | Read Write Edit Glob Grep Bash Agent |
/next
Orchestrator that handles preflight maintenance, then tries to move the
backlog forward. With no argument a direct /next turn ends by
either starting the next obvious backlog action or asking the user to
clarify what should be next; a looped /next turn must not expect user
input, dispatches only unambiguous work, and otherwise idles until the
loop-exhausted guard stops it.
When given a count or scope (N, blocked, all, or a group
name), /next runs a bounded internal sequence instead: it works
several backlog items in one turn, skipping the heavy preflight between
steps (see Sequencing). How wide it runs — one worker at a time or a
fan-out batch — is the orchestrator's judgment call, sized to how
independent the selected work is (see Concurrency). It never waits for
user input mid-sequence. For an attended multi-step push,
reach for a sequence first — it is cheaper than an external loop (heavy
preflight only at the start and end, not every tick) and stays in one
turn. /loop /next still works and remains the right tool for genuinely
unattended or scheduled runs that span many turns; it benefits from the
same preflight cadence.
Project Amendments
At the start of every invocation, check whether
docs/rpm/skills/next.md exists in the consuming project. If it
does, read it and apply its contents as additional project-specific
instructions for this skill. Amendments may add preflight steps,
narrow dispatch criteria, add output requirements, or extend the
worker contract. They cannot remove or override plugin defaults —
on conflict, this SKILL.md wins.
Routing
Parse $ARGUMENTS into a first token and an optional second token:
-
empty → run one orchestrator step (the single-step flow below).
-
status → run the status formatter and stop:
!bash ${CLAUDE_SKILL_DIR}/scripts/status.sh
Render the output verbatim. Do not interpret, summarize, or add
commentary — the user wants the raw view.
-
integer N (≥ 1) → run a sequence of up to N cycles (see
Sequencing). N = 1 is identical to the single-step flow.
-
blocked → run a sequence until a task needs the user, or no
actionable task remains.
-
all → run a sequence until no actionable task remains, skipping
tasks that need the user and continuing with the rest.
-
anything else → treat the first token as a * Parent group
name in docs/rpm/future/tasks.org (match the heading text exactly,
then case-insensitively). An optional second token scopes the run
within that group: N, all, or blocked (default: drain the group
until it has no actionable task or one needs the user). If no parent
matches, print the usage block below and stop.
Reserved first tokens (status, blocked, all, and any integer) are
never interpreted as group names.
Usage block (print only when the argument is unrecognized):
/next — one orchestrator step
/next N — up to N steps; width per step is the orchestrator's call (one worker, or a fan-out batch); heavy preflight only at start + end
/next blocked — run until a task needs you, or nothing is actionable
/next all — run everything actionable, skipping tasks that need you
/next <group> [N] — work only that backlog group (drain it, or N steps)
/next status — in-flight subagents, recent decisions, idle streak, counters
(prefer the sequence forms above; `/loop /next` still wraps the one-step form for unattended, multi-turn runs)
Mode
There are three modes:
- Direct (single-step) — no argument, run interactively. May ask one
concise clarification question when no obvious next backlog action can
be started safely.
- Loop — this invocation is visibly part of
/loop /next, an
unattended dynamic loop, or repeated automatic invocation. Must not ask
for input. If there is no unambiguous action, log idle with the
reason and stop; after 3 consecutive idle ticks, emit loop-exhausted.
- Sequence —
/next was given a count or scope (N / blocked /
all / <group>). Runs a bounded internal loop (see Sequencing).
Like loop mode it never asks for input mid-sequence; it stops at the
argument's termination condition or the safety ceiling.
Sequencing
Triggered by a count or scope argument. A sequence runs cycles in the
foreground. Each cycle's width — a single worker or a fan-out batch — is
the orchestrator's call, sized to how independent the selected work is
(see Concurrency).
-
Full preflight once, at the start. Run the delegated preflight
(Preflight step 3) with phases=full — bypassing the cadence gate,
since a sequence always forces full at its start — act on its report,
and log a marker:
bash ${CLAUDE_SKILL_DIR}/scripts/log-decision.sh preflight-full "" "sequence start"
-
Cycle loop. Repeat until a termination condition (below) or the
safety ceiling:
- a. Cheap inline gate. Check the termination conditions and the
outstanding-user-blocker rule (Preflight step 1). Do not run the
drift scan, the contradiction check, or the preflight subagent
here — that overhead is exactly what a sequence exists to avoid.
- b. Task Selection. Recompute in-flight, walk the tree,
goal-aligned dispatch, apply the group filter if scoped, and run the
is-pre-completed.sh skip check.
- c. Dispatch. Dispatch using the Worker Contract, in the
foreground — await results, do not background. Log one
actionable-backlog per worker. Width is the orchestrator's call
(see Concurrency): one worker when the next candidates are
coupled, or a fan-out batch when they are independent — awaited
together before reviewing.
- d. Review inline. Evaluate each returned result, mark the backlog
entry DONE (or append changes-requested notes to the detail file),
then log
review-result and a paired backlog-result per worker.
Lean on the worker's terse report (it also persisted the full result
to the detail file) to keep orchestrator context lean.
-
Full preflight once, at the end. After the loop stops, run the
delegated preflight again with phases=full — it catches any drift
the workers introduced — and log a second preflight-full marker
("sequence end").
-
Emit the sequence summary (see Output Format).
Termination
N — stop after N completed cycles.
blocked — stop the moment a cycle would need user input (the top
candidate's only next move is a question), or when no actionable task
remains.
all — stop only when no actionable task remains. If a cycle's top
candidate needs user input, skip it (leave it as-is) and continue with
the next actionable task; collect skipped items for the summary.
<group> — restrict Task Selection to the matched * Parent;
terminate per the second token (N / all / blocked, default
drain-until-empty-or-blocked).
- Safety ceiling.
blocked, all, and unbounded group runs stop
after at most 25 cycles in one invocation. On hitting it, stop and
report ceiling reached — run again to continue. An explicit N above
25 is honored, still cycle-by-cycle; the ceiling only bounds the
open-ended forms.
Preflight cadence
Every single-step turn — a direct /next and a /loop /next tick —
decides its preflight depth by running
bash ${CLAUDE_SKILL_DIR}/scripts/preflight-due.sh first. The drift scan
and contradiction check are housekeeping (also run by /session-end and
/audit); re-paying them on a /next that follows one minutes ago, with
nothing changed, is the cost this cadence removes.
full → run the delegated preflight with phases=full and log a
preflight-full marker. This fires when there is no prior marker (so the
first /next of a session is always full), the last one is stale
(older than the freshness window), or a guidance input changed.
lite → skip the drift scan and contradiction check. Do the inline
blocker check, then run review-ready.sh:
- a worker result is pending → dispatch the preflight subagent with
phases=review-only to review just that result.
- no worker pending → there is nothing to delegate, so dispatch no
subagent at all. Go straight to Task Selection.
This applies to both direct and loop single-step turns; a Sequence
ignores the cadence and forces full at its start and end (see
Sequencing).
Orchestrator Flow
Do not treat preflight and tasking as mutually exclusive. Run preflight
first, in order, and continue to task selection when the preflight item
was resolved locally. Stop only when continuing would be unsafe, when a
direct-mode user question is required, or when loop mode has no
unambiguous next action.
This section describes a single step. A Sequence (see Sequencing)
runs the full preflight once at the start and once at the end; each
in-between cycle skips the preflight and goes straight to Task Selection.
Preflight
-
Outstanding user blocker — if the previous decision in
docs/rpm/~rpm-orchestrator-log.jsonl was blocked-on-user and the
most recent user message did not address the question:
- Direct mode: re-surface what we are waiting on, log
blocked-on-user if needed, and stop.
- Loop mode: do not ask again. Log
idle with rationale
blocked-on-user unresolved, then stop or loop-exhausted if the
idle streak reached 3.
-
Decide preflight depth (cadence gate). Unless this is a Sequence
start/end (which forces full), run
bash ${CLAUDE_SKILL_DIR}/scripts/preflight-due.sh and branch per
Preflight cadence:
lite + no pending worker (review-ready.sh finds nothing):
there is nothing to delegate. Skip step 3 entirely — do not dispatch
the subagent — and go straight to Task Selection. This is the
common fast path for a /next that follows a recent one.
lite + a pending worker: dispatch the subagent below with
phases=review-only.
full: dispatch the subagent below with phases=full and log a
preflight-full marker.
-
Delegated preflight — when the gate calls for it, dispatch the
rpm:preflight subagent (foreground) so the token-heavy work (raw
scan.sh output, contradiction classification, worker diffs) stays in
the agent's context, not this one. Resolve each value before sending —
pass real paths, not literals:
mode= direct or loop (this turn's mode)
phases= full (all three phases) or review-only (Phase 3 worker
review only; used on a lite tick that has a pending worker result,
per Preflight cadence), as the gate above decided
scan_script=${CLAUDE_PLUGIN_ROOT}/skills/session-end/scripts/scan.sh
contradiction_script=${CLAUDE_SKILL_DIR}/scripts/contradiction-check.sh
review_ready_script=${CLAUDE_SKILL_DIR}/scripts/review-ready.sh
log_script=${CLAUDE_SKILL_DIR}/scripts/log-decision.sh
memory_dir=$HOME/.claude/projects/$(printf '%s' "$PWD" | sed 's|/|-|g')/memory
instructions= the active directive files (CLAUDE.md,
AGENTS.md, MEMORY.md at project root if present, plus every
plugin/skills/*/SKILL.md)
today=$(date +%Y-%m-%d)
The agent applies obvious mechanical drift fixes, runs the
contradiction check, reviews the first ready worker result, and logs
each drift-fix, contradiction count, and review-result itself. It
returns ONLY a compact report (drift-fixes, drift-decisions,
contradictions, review, repo-safe, notes). Read that report;
do NOT re-run the scans, re-review the diff, or re-log what the agent
already logged.
Act on the report. If it reports repo-safe: no, an open question
the user must answer, or a drift-decisions item (e.g. an overridden
skill recommended for migration to the docs/rpm/skills/ amendment
path) needing a product decision:
- Direct mode: ask a concise question, log
blocked-on-user, stop.
- Loop mode: log
idle with the ambiguity, then stop or exhaust.
Otherwise carry the report's drift-fixes, contradictions (count +
top pairs), and review summary into your preflight: output line
and continue to task selection.
Task Selection
After preflight, recompute in-flight.
-
If in-flight >= 1, log idle with rationale
waiting on in-flight worker, output the waiting state, and stop.
-
Read docs/rpm/future/tasks.org and walk the whole tree, not
just the recurring/triage sub-task you most recently dispatched.
Each * parent should carry a Goal: body line stating the
measurable success metric for that tier; tasks are evaluated
against it (see "Goal-aligned dispatch" below). When the run is
scoped to a group (the <group> argument), restrict the walk to
tasks under the matched * Parent only — other parents are out of
scope for both selection and termination.
-
The clear next task is the topmost TODO or IN-PROGRESS entry
whose :BLOCKED_BY: deps all resolve to DONE or CANCELLED, as
long as it has an :ID: and a readable linked detail file. If
multiple tasks are actionable, the topmost one is the expected
next task; do not ask solely because more than one exists.
Skip entries keyworded WATCH (deferred / observe-only) the
same way you skip closed DONE / CANCELLED — they are not
actionable. Entries marked BLOCKED without resolved deps are
likewise skipped. Before declaring a TODO or IN-PROGRESS entry
actionable, call
bash ${CLAUDE_SKILL_DIR}/scripts/is-pre-completed.sh <id>;
if it exits 0, the linked detail file already has a populated
## Worker Result — log drift-fix: pre-completed-todo: <id>
and continue to the next task instead of dispatching.
-
Goal-seeking — file the PROOF step, don't just grind gaps.
Before dispatching toward any * parent goal that is NOT MET,
reason backward from the goal — don't just pick the topmost
gap-task. Confirm both:
- the goal's definition-of-done is a DEMONSTRATION — a
test/artifact that PROVES it met (e.g. "N representative cases
pass end-to-end vs reference", "trace 38/38"), not a
gap-checklist or a vague description ("launch ready");
- a terminal VERIFY / SIGN-OFF task exists whose completion
DEMONSTRATES the goal met, with the gap-tasks as its
:BLOCKED_BY: deps.
If either is missing, filing it is the dispatch — file the
demonstrable DoD and/or the verify/sign-off task (decompose
backward: demonstrable DoD → verify/sign-off closer → the
gap-tasks it depends on) before grinding any gap-closer. This is
higher leverage than any gap-closer: a gap-closer only nudges a
NOT-MET goal; only the demonstration flips it to MET. Closing
every coded gap is necessary but not sufficient — without a
terminal proof the orchestrator gets "lost in the weeds,"
reporting progress while the actual capability is never verified.
Only once the chain terminates in a demonstrable proof, pick the
topmost actionable task by goal-aligned dispatch: score it
against its * parent's Goal: line. The task should either
(a) close a stated success-metric gap or (b) unblock a
critical-path item. Side-quest tasks that don't move the goal are
deprioritized in favor of goal-aligned ones, even when topmost in
file order. If NO actionable task in any parent moves a goal,
that's a gap — file a new research/triage/scoping task on the
gap before dispatching anything else, instead of idling.
-
If there is a clear next task, dispatch a worker using the contract
below and log actionable-backlog. Workers must persist their own
result; do not rely on a background notification as the only return
path.
-
If there is no actionable task, or the top task is missing an ID,
missing a readable detail file, marked watch/deferred, or otherwise
not scoped enough to start without guessing:
- Direct mode: ask a concise clarification question, log
blocked-on-user with that question as rationale, and stop.
- Loop mode: log
idle with the reason and stop. If this is the
third consecutive idle tick, emit loop-exhausted. Do not
collapse "visible queue=0" to "no work" before walking the
full tree against parent Goal: lines — the recurring/triage
queue is one node of one parent; many other goals may have
unaddressed critical-path items.
-
Saturation override. If the user has set a saturation
directive in memory (e.g. "1 tier3 + 1 tier2 at all times"),
the in-flight >= 1 rule from step 1 is replaced by per-slot
counts: dispatch as long as the directive's slots aren't all
full and the new dispatch wouldn't violate "never 2 daemon-using
simultaneously." When all slots are filled, log idle with
rationale saturated at directive ceiling — do not treat
this as exhaustion (it's intentional waiting, not "no work").
Output Format
Four lines, plus an optional fifth gaps: line when the goal-aligned
dispatch step (Task Selection step 4) found unaddressed critical-path
items:
action: <kind>
preflight: <none or comma-separated completed preflight actions>
in-flight: <N>
next: <hint or USER ATTENTION>
gaps: <none or comma-separated parent:metric pairs with no actionable task>
<kind>: the terminal outcome for this turn, one of
actionable-backlog, blocked-on-user, idle, or loop-exhausted.
<preflight>: none or short entries like
drift-fix: context.md broken-ref, review-result: task-id approved.
<N>: count of background subagents currently running (read from
the log; entries with kind: actionable-backlog and no
corresponding kind: backlog-result are still in-flight).
<hint>: short prose for what will likely come next. Use
USER ATTENTION needed for: <reason> only in direct mode or when
reporting a loop-exhausted terminal state.
After the four-line block, include follow-on output for the terminal
outcome: dispatch confirmation, clarification question, or loop idle
reason. For the delegated preflight, relay the agent's compact report
verbatim-ish (its drift-fixes, contradictions, and review lines) —
that summary IS the preflight view; do not re-derive it by re-running
the scans or re-reading the diff.
Sequence output
For a sequence, do not print the four-line block per cycle. Emit one
short line per cycle as it completes, then a final summary:
sequence: <arg> (cycles: <done>/<limit>, done: <X>, changes-requested: <C>, skipped-needs-user: <S>)
preflight: full@start, full@end
1. <task-id> → <approved | changes-requested | skipped>
2. <task-id> → ...
stopped: <count reached | nothing actionable | task needs you | ceiling reached — run again to continue>
next: <hint>
The two preflight-full runs are the only preflight entries — do not
report a preflight per cycle.
Cheaper-path tip (once per run)
Interior sequencing is cheaper than driving /next from an external
/loop /next: a sequence pays the heavy preflight only at its start and
end, while a loop re-checks it every tick. To steer users toward the
cheaper path without nagging, emit a one-time tip.
On a Loop mode tick, append one extra line to the output —
tip: /next all runs this in one cheaper turn (skips per-tick preflight)
— if and only if both hold:
- there are ≥ 2 actionable backlog tasks this turn (an interior
sequence would batch more than one — a single-task backlog gains
nothing from sequencing), and
- no
nudge marker appears in
docs/rpm/~rpm-orchestrator-log.jsonl after the most recent
loop-exhausted entry (or anywhere in the log, if there is none).
When you emit it, log the marker so it does not repeat this run:
bash ${CLAUDE_SKILL_DIR}/scripts/log-decision.sh nudge "" "suggested /next all"
This fires at most once per loop run. It never blocks, never gates
dispatch, and never repeats — under-suggesting is preferred to nagging
(soft suggestion only, per rpm's recommend-not-force stance). Direct and
Sequence modes never emit it.
Logging
Use the helper script — never hand-format JSONL:
bash ${CLAUDE_SKILL_DIR}/scripts/log-decision.sh <kind> [target] [rationale] [agent-id] [status]
The rpm:preflight agent writes the delegated-preflight entries
(drift-fix, contradiction counts as drift-fix, and review-result)
itself — do not duplicate them. The orchestrator logs the step-1
blocker outcome, the terminal outcome, and actionable-backlog.
Log with the appropriate kind:
drift-fix — pass the drift category as <target> (e.g.
context.md broken-ref) and the fix summary as <rationale>.
Emitted by the preflight agent.
review-result — after reviewing a pending worker result, pass the
same task ID as <target>, the reviewer action as <rationale>,
the worker ID as <agent-id>, and a status of
approved | changes-requested. This clears the review queue for
that worker result. Emitted by the preflight agent.
actionable-backlog — pass the task ID as <target>, the one-line
rationale, and the dispatched subagent's ID as <agent-id>. Match
volta's idiom: the agent ID is what the Agent tool returned (look
for agentId: in the dispatch result).
blocked-on-user — pass the open question text as <rationale>.
Use this for previously unanswered direct-mode questions and for
direct-mode task-selection ambiguity.
idle — empty <target>, short <rationale>.
loop-exhausted — empty <target>, fixed <rationale> like
3 idle ticks.
preflight-full — empty <target>, short <rationale> (e.g.
sequence start, sequence end, loop tick). Marks that a full
preflight ran. preflight-due.sh reads the most recent one to decide
whether a later loop tick can go lite. It is ignored by the
decisions view and the idle-streak count (status.sh and the Idle
Terminal filter only track terminal decisions), so it never resets the
loop-exhausted guard.
nudge — empty <target>, short <rationale> like
suggested /next all. Marks that the one-time Loop mode cheaper-path
tip fired (see Cheaper-path tip). Like preflight-full, it is
ignored by the decisions view and the idle-streak count, so it never
resets the loop-exhausted guard.
backlog-result — when a <task-notification> arrives for a prior
actionable-backlog dispatch, log it: pass the same <target> and
<agent-id> plus a <status> of
needs-review | plan-written | blocked | no-op. This pairing is what
powers the in-flight: <N> count. If the worker already wrote this
line itself, do not duplicate it.
The script writes to docs/rpm/~rpm-orchestrator-log.jsonl
(gitignored, ephemeral). Failures print a stderr warning and exit 0 —
never block the orchestrator.
Worker Contract
When dispatching actionable-backlog, include this contract in the
worker prompt. Workers must write a durable result to the orchestrator
log before finishing; chat notifications are only supporting context.
The orchestrator computes today's date at dispatch time
($(date +%Y-%m-%d)) and substitutes it into the today: field
below. Workers must use that exact value for any dated artifact —
they cannot infer the current date reliably.
You are an rpm backlog worker.
Task:
- target id: <task-id>
- task heading: <task heading from tasks.org>
- detail file: docs/rpm/future/<detail-file>.md
- orchestrator log: docs/rpm/~rpm-orchestrator-log.jsonl
- worker id: <agent-id if known, otherwise worker-unknown>
- today: <YYYY-MM-DD>
Rules:
0. Use the literal `today:` value above for every dated artifact you
create (filenames, frontmatter dates, log entries). Subagents do
not have reliable clock access; the orchestrator passes the date
in via this field. Do not infer it.
1. Read docs/rpm/future/tasks.org and the detail file before writing.
2. Do real work when the task is scoped enough to execute safely.
Keep changes limited to the task. If the task is ambiguous or too
broad, do not guess; write a plan instead.
3. Append a `## Worker Result` section to the detail file when you
changed files or otherwise completed the task. Include:
- Summary
- Files changed
- Verification run
- Remaining risks or follow-ups
4. Append `## Plan` if you only planned the work. Append `## Blocked`
if you could not proceed, with the missing information.
5. Before finishing, log your result:
bash ${CLAUDE_SKILL_DIR}/scripts/log-decision.sh backlog-result \
"<task-id>" "<one-line result>" "<agent-id>" "<status>"
`<status>` for `log-decision.sh backlog-result`:
- `needs-review` — you did work and want the orchestrator to review it
- `plan-written` — you only appended `## Plan`
- `blocked` — you appended `## Blocked`
- `no-op` — the task is already fully handled
If your project also uses an orch-job tracker (e.g. Volta's
`tools/scripts/orch-job.sh`), close it with the equivalent status.
Status enums may differ between the two helpers — check the script's
usage line. Common mapping: `needs-review` → orch-job `success`;
`plan-written` → orch-job `success` (with summary noting the plan);
`blocked` → orch-job `blocked`; `no-op` → orch-job `no-op`.
6. In your final response, state the detail file changed and the status
you logged. The file + JSONL log are the source of truth.
Idle Terminal
Read docs/rpm/~rpm-orchestrator-log.jsonl only if it exists. Filter
to kind in {blocked-on-user, drift-fix, actionable-backlog,
review-result, idle, loop-exhausted} and ignore backlog-result.
If the file does not exist, the idle streak is 0. If the last three
filtered entries are idle, this turn becomes loop-exhausted instead
of idle.
Saturation exception: if the current idle rationale starts with
saturated (i.e. workers are intentionally running at the saturation
directive's ceiling, see Task Selection step 7), the 3-idle threshold
does NOT apply — saturation idles are not "no work," they're "work in
flight," and exhausting the loop here is wrong. Log normal idle and
keep the cron armed.
Idle is for loop-mode ambiguity or waiting on in-flight work, not for a
direct-mode "no obvious next task" state. In direct mode, ask the user
when task selection is unclear. In loop mode, never ask; log idle and
let the 3-idle threshold stop runaway autonomous loops. The user can
resume by running /next directly; the next invocation reads the log
fresh and picks up wherever priority leads.
Killed workers. A worker that registers an orch-job but never
calls complete (process killed mid-run, runtime early-terminated,
quota exhausted) leaves a phantom running entry. The orchestrator
must close the orch-job (orch-job.sh complete <id> <success|no-op| failed> "killed: <reason>") and log a paired backlog-result so the
dashboard reflects reality and the in-flight count is accurate. Do
not rely on the worker's own logging — it didn't reach that step.
Concurrency — the orchestrator decides how wide to fan out
The orchestrator owns width. How many workers to run at once is its
judgment call, made from the shape of the selected work — not a fixed
ceiling. A no-argument /next typically dispatches one actionable-backlog
worker in the background and reviews it next turn; a sequence (N /
blocked / all / <group>) works foreground. Either may fan out to as
many concurrent workers as the batch genuinely supports, or stay at one.
What narrows effective width. Fan-out pays off only to the extent the
batch lacks couplings. Weigh these when sizing a batch — each one that
applies pulls width back toward one:
- Overlapping edits — workers touching the same files/directories risk
concurrent-edit clobbers and non-FF push conflicts.
- Shared serialized resource — tasks competing for one render queue,
build lock, dev server, port, or fixture serialize anyway.
- Cross-cutting verification — a corpus-wide check that one worker's
change can invalidate for another means fixes must land and verify one
at a time to stay honest.
- Needs user input — an ambiguous task can't be dispatched unattended
regardless of width.
Genuinely independent work (disjoint edits, no shared resource, per-task
verification) can fan out wide; tightly coupled work stays narrow or
single-threaded. Judge per batch.
For a no-argument turn that stays single-worker: if <N> >= 1 after
preflight and you are not fanning out, do not dispatch another worker.
Output action: idle with next: waiting on in-flight worker.
Rationale (2026-05-04 audit) — why width tracks coupling: on a
game-fix corpus, saturating to 4 produced an 8.4% dispatch hit rate (107
dispatches -> 9 shipped fixes) because (a) workers competed for the daemon
render queue, (b) concurrent edits created non-FF push conflicts, (c) each
fix verified against one game shipped before contradictions in other games
surfaced. That corpus was maximally coupled on all three axes above — so
its right width was one. A batch without those couplings has no such tax;
size it to the work. The lesson is width follows coupling, not width is
capped.
What this skill does NOT do
- Does not pick up tactical user requests —
/next is for unattended
autonomous work, not the conversational thread.
- Does not modify
tasks.org ordering — that's /backlog review's
job. /next only reads.
- Does not run audits, releases, or session-end. Those are explicit
user-invoked operations.
- Does not retry failed dispatches. A failed subagent logs a
backlog-result with status: blocked; the next /next tick picks
the next unblocked item.
Example Sessions
Looped run with mechanical preflight and a clear next task:
$ /loop /next
action: actionable-backlog
preflight: drift-fix: context.md broken-ref
in-flight: 0
next: review worker result when it returns
Fixed: docs/rpm/context.md broken-ref to legacy plugin path.
Dispatched general-purpose subagent <id> on `sync-codex-scripts`.
Will append a result to 2026-04-30-sync-codex-scripts.md when it
returns.
Looped run with no unambiguous task:
action: idle
preflight: none
in-flight: 0
next: no unambiguous backlog task; loop mode will not ask for input
No unblocked TODO/IN-PROGRESS backlog entry has a readable detail file.
Direct run with no unambiguous task:
action: blocked-on-user
preflight: review-result: sync-codex-scripts approved
in-flight: 0
next: USER ATTENTION needed for: choose next backlog item
Reviewed worker <id> for `sync-codex-scripts`; approved.
Which backlog item should rpm start next?
Sequence run (/next 3):
sequence: 3 (cycles: 3/3, done: 3, changes-requested: 0, skipped-needs-user: 0)
preflight: full@start, full@end
1. sync-codex-scripts → approved
2. fix-broken-ref → approved
3. update-readme → approved
stopped: count reached
next: backlog top is now `dutch-bakeoff` (needs a scope decision)
Sequence run that stops early (/next blocked):
sequence: blocked (cycles: 2, done: 2, changes-requested: 0, skipped-needs-user: 0)
preflight: full@start, full@end
1. fix-broken-ref → approved
2. sync-codex → approved
stopped: task needs you
next: USER ATTENTION needed for: `dutch-bakeoff` scope (v1/v2/native?)