Shared policy location: docs/DISCOVERY.md, docs/WORKFLOW.md,
docs/LAUNCH-MIGRATION.md, docs/STANDALONE-ENTRY.md, docs/PROCESS.md and docs/SURFACES.md belong to the resolved Hatsu plugin
root, not the consuming repository. On an installed surface, use the absolute root printed by
ten to read those files (re-resolve through that skill if unavailable). Relative links
below identify source locations; a missing consumer docs/ copy is not a missing policy and must not
trigger a duplicate filing. Never copy or invent a second policy in the target repository.

En — the circle held around one PR until it is Ready

Nature: Manipulator. Everything en does is board-facing — driving, reporting, ringing, watching.
Kurapika says so when he runs it. When a cycle sends him back into the change itself — a
reviewer's finding that needs code — the mode for that fix is the change's own (Enhancer, Conjurer,
Transmuter), named out loud for that cycle and handed back afterwards.

Stay with this PR: report it, observe every current-head check and review, drive it to verified
Ready, ring once at its human gate — and stop after at most N acting cycles.

En is the last composite of the local plane and the only one that may need to hold attention past
the moment the PR was opened. `/mukai` ends by offering this handoff — starting it itself only under
nen/workflow.json → mukai.autoEn (ruling 2026-09-28); from
there the PR belongs to En — and the still-active user turn only on the autoEn path, a typed start
from the offer being its own turn — until Sharingan proves it Ready, or a concrete
blocker/cap/terminal state/cancellation ends En's run. The merge itself is the maintainer's later act
and is outside En. This run does not harvest. When En has completed, the regular pipeline's next
phase is `/third-hand` — a separate skill, started after this file's
protocol has returned, not as a step of En.

This file composes. It does not re-specify. Every step below is another skill's, named and
linked, and its procedure, its exit-code reactions, its refusals and its residue live there. If you
find a rule here that is really that skill's, it is in the wrong file — go read it where it is
authored, because a rule restated in two places drifts in one of them.

What is genuinely en's own is the cap, the watch's shape, and the boundary at the end of the
session. Those three are § 2, § 6 and § 7, and they are the only places this file asserts
anything. Third-Hand is not in that boundary.

---

1. Invocation

P1 — this run opens with /ten en (`ten` § 6 catches up this phase's
missing prerequisites under its own name; the phases this run calls skip theirs).

/en [on <CODE>#<N>]

Started by `/mukai` with no clause under mukai.autoEn — the PR is the one
step 8 just opened — by `/getsuga` § 3 for the release proposal it just
opened (there en runs steps 1–5 and returns its quoted verdict to getsuga; step 6's bell is not rung,
because getsuga § 3a is that PR's merge), by `futon`'s per-issue chain for the
PR its effort opened, or by the maintainer, naming a PR that is already open (mukai's offer line). The
rule that every PR a session opens or causes to be opened reaches the maintainer through this run or
with the verdict quoted is `docs/PROCESS.md` § Reporting a phase's
(zheref/hatsu#103), cited, not restated:

nen parse en --grammar "on [<ref>]" --line "<the invocation, minus the /en prefix>"

Verified live at v0.3.0 (docs/ab/en.md § 2.1): on HA#41 → ref: HA#41 at exit 0, and a bare
on parses with the clause absent, also exit 0. The clause is anchored behind a literal for the
reason `/spiritual-message` § 1 records — a lone bracketed slot is refused at the
template.

The empty line is refused, and that is why a composite does not parse at all. Verified live:
--line "" exits 2 with "the line must open with the literal 'on' — it is what introduces
<ref>". An anchored optional clause requires its anchor; "no clause" and "empty line" are
different inputs, and only the first parses. So mukai, under autoEn, hands the PR over directly and calls
nen parse en not at all — echoing a parse of a line nobody typed would be theatre, and
feeding it "" would produce a refusal that means nothing. The parse runs when, and only when,
the maintainer typed a clause.

A missing argument or configuration item is asked for and set up inline (missing-argument,
missing-configuration; `WORKFLOW.md` § 4 Ask, set up, continue): a
maintainer's line that exits 2, or names no PR where the session holds none, is the trigger to
offer the session's open PRs in the picker and resume on the answer, never the end of the run.

Resolve the code once, up front, and keep the slug for every later call — never re-derive it a
second way (claude/agents/kurapika.md § How you work):

nen repo resolve <CODE> --repo <path to the target repo's own checkout>
nen ref format --code <CODE> --kind PR --number <N>

A closed or merged PR ends the run immediately with what happened to it. There is nothing to
land.

The phases this composite calls SKIP their own ## 0. Standalone entry sections. `docs/STANDALONE-ENTRY.md` is the
contract for a phase typed by hand into an arbitrary checkout; a phase reached from here inherits P1–P4
from this run — the warm-up, the orientation, the change set and every derived argument — and
re-deriving them would produce a second answer to a question this composite already settled. Each phase
says which composite is holding it instead.

2. The cap — grammar, not a default

En acts. Steps 2 and 4 drive; a cycle that finds a new review finding goes back into the change
and pushes a fix. That makes en the mutating half of the loop pair, and
`/izanagi`'s whole discipline binds: a loop that writes needs a bound a
loop that reads does not, and making the bound part of the grammar means it can never be defaulted,
inherited, or forgotten.

Key	File	Used for	Default when the key (or the file) is absent
monitor.maxCycles	nen/workflow.json	the cap — the count of acting cycles (§ 6)	20
monitor.pollSeconds	nen/workflow.json	the interval between observations, as --interval-ms	300 (→ --interval-ms 300000)
branch.base	nen/workflow.json	what step 3 catches up from, and what "behind" means	main
reports.dir / .sections / .retain	nen/workflow.json	step 7's file — `backlog-board` § 3's keys	Reports / five variants / final-only
notifications.rungs / .sound	nen/workflow.json	step 6's rungs — `/jutaisho`'s keys	["push","os","sound"] / Glass

Before cycle 1, the cap is spelled out through the verb that refuses a missing one:

nen parse izanagi "take <CODE>-PR-#<N> to its human gate until it is Ready up to <monitor.maxCycles>"

Verified live at v0.3.0 (docs/ab/en.md § 2.2), both ways:

$ nen parse izanagi "take HA-PR-#41 to its human gate until it is Ready up to 20"
task: take HA-PR-#41 to its human gate
until: it is Ready
cap: 20
exit=0

$ nen parse izanagi "take HA-PR-#41 to its human gate until it is Ready"
nen: no 'up to <N>'. Izanagi is the MUTATING half of the loop pair and the cap is required grammar,
never defaulted or inferred -- an invocation without it is refused rather than run once 'to see'.
  try: take HA-PR-#41 to its human gate until it is Ready up to <N>
exit=2

The value comes from monitor.maxCycles; the requirement comes from the grammar. Those are two
different things and the distinction is the point: the file supplies a number, and the verb refuses
to run without one. A repository whose monitor block is missing gets 20, stated out loud —
it does not get an unbounded watch, because "the file said nothing" is exactly the case an
inherited-and-forgotten default is dangerous in.

RETIRED at nen 0.7: counting the acting cycles by hand

Every acting cycle is CLAIMED before it acts, and the cap is enforced by the claim being
refused — `izanagi` § 3's retirement, inherited whole:

nen loop iterate --id en-<CODE>-<N> \
  --line "take <CODE>-PR-#<N> to its human gate until it is Ready up to <monitor.maxCycles>" \
  --repo <path>

Exit 0 prints iteration <n>/<cap> and how many remain — the verb's own word is iteration, and
en's cycle is what it is claiming; exit 1 IS the cap and ends the run at § 6's cap-out row,
with the invocation named in nen's own refusal. The --line is the same string
nen parse izanagi echoed above, restated on every claim, so monitor.maxCycles cannot be
quietly raised mid-run: a claim carrying a different N is exit 2 naming both lines (verified
live against exactly this shape — up to 3 running, up to 20 claimed —
docs/ab/izanagi.md § Retired at nen 0.7).

The --id is the PULL REQUEST, not the session, and that is a decision with a consequence
worth stating: the ledger is .nen/loop/en-<CODE>-<N>.json under --repo, it outlives the
session (§ 7), so a re-invoked /en on the same PR resumes the same cap rather than
starting again at cycle 1. That is the honest reading of "twenty acting cycles to ready this PR" — a cap
a session boundary reset would be a cap nobody has. When the readiness drive genuinely ends — Ready,
closed, or handed to the maintainer at a gate — close the ledger with
nen loop iterate --id en-<CODE>-<N> --line "<same line>" --release "<why>"; a loop at its cap can
always still be released, and a fresh cap after that is a new --id, chosen deliberately and
said out loud, never a re-typed line.

The cap is a ceiling, not a target. Reaching it is a failure to reach Ready and is reported as one:
what is still not true, and what the next cycle would have done. izanagi § 4's rule, unchanged.
The reviewer-round policy is `/sharingan` § 6's, stated there once
and cited here, never restated:

• One Copilot round is requested after `/hanten` settles, never before.
• Arrivals are remediated, and an owed round is re-requested without asking only inside sharingan
  § 6's two bounds (round_policy.maxRounds requests, round_policy.minRounds resolved; row
  cap-reached), then en keeps watching; either bound reached is § 6's fifth stop.
• Before any re-request, nen wake verify --repo-slug <owner/name> --now <ISO> --author-pattern <the maintainer login> --run redrives a swallowed run once (action_required or
  startup_failure that never executed); an action_required run by the Copilot bot cannot be
  retried and is reported as such. Then the no-op sync guard: count commits ahead of the last
  reviewed head (git rev-list --count <reviewed-head>..HEAD, read-only) and the diff stat; zero
  reviewable change means no request.
• Copilot auto-reviews every push, so the cap governs requests, not arrivals; an arrival past the
  cap is still remediated and its threads settled.

Past the maximum the run ends at not-ready with the board. Neither is a G5. nen/workflow.json →
monitor.maxCycles is en's acting-cycle cap, a different number.

3. The run, in order

#	Step	The skill that owns it	When
1	landing report	`/spiritual-message` as landing	once, at the start — 00–07 plus 08 PR body and 09 Readiness. If this En was started by mukai (autoEn) or typed from its offer in the same sitting with no newer human request, 00 still answers that mukai request
2	drive	`/sharingan`	first blocking condition, threads, checks, the confirmation pass
3	catch up	`/murasaki`	only when the branch is behind branch.base
4	drive again	`/sharingan`	after step 3 moved the tree underneath it
5	observe	this file, § 6	while CI or a reviewer round is pending; each poll runs nen wake verify in dry-run (no --run), and --run only when a swallowed run is found; rebuild the current-head snapshot on every change, returning to steps 2–4 when action is needed
6	the merge, then the bell	this file, § 5; the report through `/jutaisho`'s four parts	after verified Ready or § 6's sixth stop, and only then — not under getsuga § 3, where the verdict returns to getsuga and § 3a merges
7	the dated final report	`/backlog-board` § 3, --variant final	after the bell; this is En's successful terminus

Four orderings are en's own assertions:

• 3 before 4. Driving a branch that is behind its base produces a readiness verdict against a
  tree that is about to change. Catch up first, then re-decide.
• 4 before 6. The bell carries a verdict. Ringing before the second drive would ring on a
  reading the catch-up already invalidated.
• 5 before 6. Pending is not Ready. Required CI and every reviewer round owed at the current head
  must settle, and every incoming finding must be disposed, before the bell or handoff exists — the one
  round that settles by declaration rather than by arriving is the exhausted reviewer's (§ 6's sixth stop).
• 6 before 7. The dated final report records the gate event the bell announced; rendering it first
  would preserve a readiness handoff that had not happened yet.

Step 3 is conditional and stays conditional. A branch level with its base does not get a
catch-up "to be safe": murasaki would re-run the declared iteration checks for nothing, and on a
published branch it would push a commit that changes no content.

En's equivalent of mukai steps 3–5 runs whenever the tree this watch is about to treat as
current has changed, not only when step 3's catch-up moved a path. That includes a review fix
authored in steps 2/4 while the branch is still level with branch.base (step 3 then stays
skipped), a catch-up that returns without pushing, and a later murasaki that was meant to
publish but itself changed a path. Claim an acting cycle, then run
`/kokusen`, `/kotoamatsukami`
and `/byakugan`, then call murasaki again. It may push only when
that catch-up is a no-op; if it changed any path, this paragraph repeats rather than handing a
stale tree to readiness. That is the same owner map `/sharingan`
§ 5 already names when Kurapika authored the PR; En does not invent a fourth suite.

4. Readiness is never en's claim

Steps 2 and 4 are `/sharingan`'s whole engine, not a substitute for
it, and this file restates none of its protocol — the first-blocking-condition order, the
channel decided by who authored the PR, the escalation ladder, the round_policy.maxRounds request
cap and the completed-round prerequisite, the thread hygiene run through nen pr threads, the
one-directional confirmation pass that may only veto. Read it there.

Two things en relies on and does not re-derive:

• A PR is Ready iff nen pr ready says ready AND nen pr body-check says every requirement is
  satisfied, both deterministic, neither re-derived by eye. sharingan § 4's rule, and
  `/pr-state`'s before it: a readiness claim is that verdict, quoted,
  or it is not made.
• A body this run rewrites — new or re-recorded screenshots after a review — re-renders the whole
  Evidence table and passes "$hatsu_root/scripts/pr_body_evidence_check.sh" --body <file> before it is written
  (`docs/WORKFLOW.md` § The UZF-26 evidence shape, The table, exactly).
• sharingan's escalation is a G5 and it ends this run's cycle, not just its step. A PR that
  will not reach Ready is one of the plane's five genuine stops
  (`docs/WORKFLOW.md` § 4), and en does not spend the rest of its cap
  re-driving past it.

En never casts a review vote. Kurapika runs on the maintainer's credentials, so GitHub would
record the vote as theirs — a governance vote on a PR they have not read. That holds even when
the finding is real and even when it is the only path visible. sharingan § 5's rule, binding here
because a long watch is where the temptation compounds.

5. The merge — once, at a terminus, by this skill

Not under `getsuga` § 3. For a release proposal getsuga started en on, this
step and § 8 do not run: en hands the quoted nen pr ready + body-check verdict and its judged head:
back to getsuga and merges nothing, because § 3a is that PR's merge (ROSTER ruling 4); this section
merges only where § 3a falls back to the declaration gate.

The merge is this skill's — the maintainer's ruling of 2026-09-29 (3) (docs/ROSTER.md § Rulings
of 2026-09-29: "The merge is not mine. It is yours and it has been."), stated here once; futon § 5,
sharingan § 6, kurapika.md, the gate tables (ROSTER, README, WORKFLOW) and nen/decisions.json row
own-pr-merge cite it. Five gates, in order, every verdict line printed, then the merge:

1. Whose PR — the author is the authenticated viewer (gh pr view <N> --json author, gh api user);
   never another's PR, never a release proposal (getsuga § 3a's).
2. On branch.base — the base read back (gh pr view <N> --json baseRefName); a stacked PR whose
   delivery just merged is retargeted through nen pr retarget first (getsuga § 3's mechanism,
   `jujisho`'s) — a PR off the trunk is merged nowhere and presented nowhere
   (ruling 2026-09-29 (2)).
3. The verdict — nen pr ready --explain says ready; or its one failing row is a reviewer with a
   live reviewer_fallback.exhausted[] row and the chain exhausted, § 6's sixth stop
   (PROCESS.md § *Reviewer rounds and review threads*, The fallback chain):
   the verdict quoted verbatim, the one Hatsu exhaustion line beside it, written into the body's
   completion checklist through nen pr edit-body before the merge — the durable record that hanten's
   rounds were the review.
4. The body — nen pr body-check satisfied against the live body.
5. The head pinned — the commit nen pr ready judged is GitHub's head at the merge.

export GH_TOKEN=$(gh auth token)
gh pr merge <N> --repo <owner/name> --merge --match-head-commit <judged head>   # never --admin, never --auto
gh pr view <N> --repo <owner/name> --json state,mergeCommit -q '"\(.state) \(.mergeCommit.oid)"'   # merged only when GitHub says MERGED

Residue, named: nen pr merge merges a release unit only (refused without --release-unit), so the
delivery merge is gh pr merge in the verb's own argv shape — the head pinned, the state re-read,
merged reported from GitHub's state and never from the exit code; the delivery form of the verb is the
owned dependency zheref/nen#286. A refusal by gh (branch protection, a required review) is quoted
with the exact command and the PR stands at the declaration gate — the one place the maintainer merges.
The merge is not a review: the review is the reviewers' rounds (§ 4, hanten's under the chain); this
skill still never votes and never self-reviews.

Step 6 is `/jutaisho` at <G2|G4>, with the gate
`shibari` derived and sharingan confirmed. It owes all four of that skill's
parts — the nen stop banner, the report link, lettered options with a ⭐ on the recommended decision — the report linked, never an option (Crazy Slots), and the
question through the surface's own native option picker.

nen stop --template emits the blank five-column shape the efforts table is filled into — verified
live at v0.3.0, exit 0 (docs/ab/en.md § 2.4):

| Effort  | Open issues & PRs | Status (gate)   | Thought flow | Session / lane |
| ------- | ----------------- | --------------- | ------------ | -------------- |
| <title> | <link>            | <status (gate)> | <one line>   | <session>      |

There is no MERGE ask for the run's own PR — the merge above replaces it; step 6's report says
the PR merged, at which gate it stood and on whose review, with the verdict quoted. A MERGE ask remains
only for a PR this run did not author, which en drives to Ready and hands to the maintainer.

It merges exactly once for this run — at Ready, or at § 6's declared reviewer-exhaustion stop — and at
no other time, and rings once, after the merge. A quiet observation rings nothing — a bell that rings
every poll of a four-hour wait is a bell nobody hears.

6. The observation hold — nen watch until on nen pr ready, and what a cycle is

The condition is polled by the same read-only observation engine
`/izanami` uses. The verb owns the wait between observations; a caller
does not repeatedly invoke a one-iteration command and pretend --interval-ms paced those invocations.
Each refresh window has exactly two observations: the first establishes the window and the verb waits
monitor.pollSeconds before the second. The window then returns to Sharingan's full snapshot, because
nen pr ready alone cannot expose new review activity that leaves its verdict unchanged:

export GH_TOKEN=$(gh auth token)
nen watch until --command "nen pr ready <CODE>#<N> --repo <path> <identity flags>" \   # no exclusion flag here
  --max-iterations 2 --interval-ms <monitor.pollSeconds × 1000> # one paced refresh window

<identity flags> is selected by `sharingan` § 4 for the target: omit it
when the target ships nen/gates.json; use --gates "$hatsu_root/contracts/reference.gates.json"
only for <reference-repo>; otherwise pass the target's hand-supplied --reviewers and explicit
--approvers. Never point one repository's gates file at another repository. The exclusion flag never rides
the --command string (`pr-state` § 2, SEC-7): nen watch until splits
that string itself, a check name with a space (the declaration admits one) splits into two words, and
the gates file is third-party data beside a live token. On a target with live check_exclusions[]
rows the window watches the bare verdict — it only paces — and the verdict that counts is
pr-state § 2's own read after the window, every live row's name its own argv element
(zheref/hatsu#104); a ruling with no row is missing-configuration, and a name the flag cannot carry
keeps the PR not-ready — the verb's verdict quoted verbatim, the excluded check named beside it —
never ready.

nen pr ready classifies [read-only] — verified live at v0.3.0 (docs/ab/en.md § 2.3):
nen parse izanami "nen pr ready HA#41 --repo /path --gates /abs/gates.json until it is ready" →
[read-only], exit 0. So does the merge-state read,
gh pr view <n> --repo <owner/name> --json state -q .state. And the merge itself does not:
nen watch until --command "gh pr merge 41" is refused before the first observation, exit 2 —
"'gh pr merge 41' classifies as mutating … izanami watches only". The verb that polls this loop
structurally cannot be the verb that ends it, which is a property worth having rather than a
limitation to route around.

--max-iterations is still not the cap. Its own --help says so — "a SAFETY bound, not
izanagi's mandatory cap" — and it bounds watch observations, not the count of acting cycles.
The cap is nen loop iterate's from nen 0.7 (§ 2): a cycle is claimed against
monitor.maxCycles before it acts, and the claim is refused at the cap. Two different verbs, two
different bounds, and neither substitutes for the other.

What counts as a cycle — and it is what en CLAIMS, not what the watch observes. A cycle is spent
when en acts: goes back to step 2, pushes a fix, re-requests a review. An observation that
finds nothing changed is not a cycle and is never claimed, and neither is the pre-check before the
first act (izanagi's iteration 0). Otherwise a four-hour poll at five-minute intervals would
exhaust a cap of 20 in under two hours without a single thing having happened — and since nen 0.7
that is a mistake with a ledger behind it rather than a miscount nobody can see: claim only where
en acts, and the <n>/<cap> the verb prints is the number the report carries.

What the observation hold reacts to:

Observed	What en does
a requested reviewer round not yet posted at the current head — Copilot minutes after a request or a push, nen pr ready row 4 (CON-32(b): review requested, not yet posted)	pending, not Ready (`sharingan` § 5's rule, the verb's row 4 its evidence) and no cycle is spent: observe until it posts; once round_policy.stallMinutes makes it a stalled round (row 3), claim an acting cycle and act as sharingan § 6 says (zheref/hatsu#103)
a required check turned red at the current head	claim an acting cycle and return to step 2 (§ 3's step-5 cell): the failing run is read, the fix authored, verified through the phase owners and pushed — a new epoch
the PR converted to draft, or its base retargeted	outside the hold's vocabulary: name it, end the watch, and hand back — a draft is not driven, and a retarget re-derives the gate (WORKFLOW § Gate derivation)
a new review, comment or thread	inspect and classify it first. If it requires remediation or a reviewer re-request, claim an acting cycle, then return to step 2; an approval or informational event that needs only a read spends no cycle. `/sharingan` addresses every inline and summary finding through its own channel, and decides whether a further round may be requested (§ 6 there; owed at head after en's own request or push is not a reason)
the branch fell behind, or the PR went dirty	claim an acting cycle, then step 3 and step 4 — catch up, then re-decide. A conflicted PR gets no checks at all, which reads as "clean" rather than "broken" (sharingan § 5)
the PR becomes Ready	step 6 — bell and stop at the human gate
a round owed at head to a reviewer with a live reviewer_fallback.exhausted[] row, the chain exhausted (PROCESS.md § *Reviewer rounds and review threads*, The fallback chain)	not a pending state — the round will not arrive. End the hold and go to step 6 with the verdict quoted verbatim and the Hatsu exhaustion line beside it; no cycle is spent
the PR merged before the gate handoff	end as a terminal external state, naming that readiness was not the run's observed terminus
the PR closed unmerged	the run ends, saying so — there is nothing to land, and reopening is the maintainer's call
nothing changed	one line, or no line. Not a status screenful; nen watch until already prints one line per observation

Six stops apply here: the condition true (verified Ready); the cap reached, reported with what
is still not true; a human gate, which is never retried past; an impossible condition, named
rather than waited on — a closed PR will not become Ready; and the round ceiling (zheref/hatsu#102):
round_policy.maxRounds requested, or round_policy.minRounds resolved with the verdict still
not-ready — a round owed at head that no bound lets en request is a named impossible condition for
the watch, not a state to poll — reported not-ready with nen pr ready --explain quoted and the board,
never a raised number, never a G5; and a declared reviewer exhaustion with the chain exhausted — the
hold ends at step 6 on the not-ready verdict, the sixth stop, which outranks the round ceiling
(nen/decisions.json row reviewer-exhausted). izanagi's generic three-no-op stop does
not apply to En: quiet observations claim no cycle, and three unchanged pending reads cannot terminate
the current-head readiness promise. Three consecutive observation errors still stop nen watch until
as an unread capability failure, exactly as the verb documents; that is not a pending-state success.

7. The session boundary — and the Illumi hand-off

A watch lives in the session that started it. nen watch until is a foreground process, the
observations are this conversation's, and there is no timer, no background pass, no deferral primitive and
no self-continuation anywhere in this plane — the same boundary `/ren` § 4
records for its own loop and docs/ab/ren.md § 2.2 proves live.

A model choosing to finish a response is not a session boundary and not a run outcome. While the
task is alive, pending CI/review remains under observation. If the host or maintainer actually interrupts
the session, say the acting-cycle count reached, the last verdict, and that the run was interrupted; it is
resumable by
re-invocation — /en on <CODE>#<N> re-fetches everything and re-decides from live evidence
(sharingan § 9's rule). Notes worth keeping go in the run's transcript, never trusted over a
fetch.

One thing does survive the session, deliberately: the cap. .nen/loop/en-<CODE>-<N>.json is a
file, so a re-invocation on the same PR resumes the same count (§ 2) — the watch is per session, the
cap is per readiness drive. Say the resumed <n>/<cap> out loud on the first claim of a resumed run, so the
maintainer sees a budget being continued rather than one silently restarting.

A pre-Ready observation hold measured in hours is `Illumi`'s, and from v0.5.0 he has a
definition to be raised as. The roster's ruling of 2026-09-09 partially closes OPEN-1 —
"Illumi is provisioned for en's long watch, and only when one is needed" — and
`claude/agents/illumi.md` landed in this same wave, because a
provision that cannot be executed is a provision in name only (docs/ROSTER.md § Rulings, 5).
On a surface with in-session subagents, step 5 is handed over, not abandoned: a subagent titled
en · illumi · <model alias>, on the fast tier at effort medium, holding the same
monitor policy this file reads. He never
claims or spends En's acting-cycle cap: each quiet poll is an observation, not an Izanagi iteration.
He records the five facts per observation, compares them, and wakes Kurapika — naming what changed and
the act it needs. He performs none of it, and the readiness hand-off stays G2/G4.

What the hand-off buys is attention, not persistence, and the difference is stated rather than
blurred. A delegate is still raised from a session: the paragraph above still holds — no
timer, no background pass, no deferral primitive, no self-continuation. What Illumi gives is a
watcher whose entire job is the watch, at a tier that can afford to hold it for the hours an
overnight CI queue or a reviewer in another timezone takes. An observation hold that needs to survive the
maintainer closing the session still has no mechanism anywhere in this plane — that is named,
reported as an interruption, and never simulated.

Codex and Antigravity have in-session subagents (spawn_agent on Codex; invoke_subagent
with Workspace inherit on Antigravity). En may hand this watch to en · illumi there the same
way. The hand-off buys in-session attention, never background persistence: the watch remains
bound to the active session, paced by nen watch until (§ 6). The second-process codex exec
worktree remains Hanten's isolation for a reviewer who must not share the author's tree; it is not
this watch and is not a persistence mechanism. If the session is interrupted, the resumable ledger
is reported exactly as above. A model choosing to stop calling tools while CI checks, reviewer rounds,
or review comments are pending is voluntarily abandoning the watch, which is not an outcome and must
never report G2/G4 readiness: the observation hold remains active in-session until nen pr ready exits 0,
the acting-cycle cap is reached, or the session is actually interrupted.

And the hand-off widens nothing. Illumi is provisioned for this watch and no other loop —
not backlog-loop, not futon, not senkei; that half of OPEN-1, and the whole of Killua's
row, stay OPEN and stay the maintainer's G4-class ruling to make. What must never happen is
en running unattended to fill a gap: an unbounded, unwatched acting loop is precisely the shape
the cap exists to make impossible, and improvising one would close a maintainer's OPEN question by
attrition rather than by ruling.

8. The dated final report

Not under getsuga § 3 (§ 5): the proposal's report is getsuga's own.

After the deterministic Ready verdict and step 6's bell, render the final variant — a
one-effort Rikugan with a cleared desk, this effort's register, spend and legend —
through `backlog-board` § 3, which owns that render path (maintainer's
ruling, 2026-09-19). There is no spiritual-message as final any more; hand over and say so.

<reports.dir>/<YYYY-MM-DD>-<effort>.html

That path is backlog-board § 3's and reports.retain: final-only's, not en's invention: turn and
landing renders live at their Artifact address (or the transient current.html on a surface with
none), and only this one gets a dated file. <reports.dir> is git-ignored, and neither skill
writes anywhere else in the tree.

Then the run ends at the gate. Say the object notation, current head SHA, quoted readiness verdict,
acting-cycle count spent out of the cap, and final report path. The human merge or vote remains outside
the run. If the PR merged externally before this hand-off, report that terminal state without rewriting it
as an observed Ready success.

Do not start `/third-hand` from this file. When this run has
completed, the regular pipeline's next phase is Third-Hand, started by the caller after En has
returned. Harvesting is that phase's, not En's.

Residue

1. The composition itself has no verb, and is not expected to get one. nen owns operations, not
   conversations (docs/ab/ren.md § 2.2, verified live). En adds no residue of its own — every
   deterministic step inside a cycle is a verb or a named residue in the skill that owns it: the
   absent nen report data / nen report render (step 1 and the readiness report, named in
   `spiritual-message`), nen pr ready / nen pr body-check / nen pr staleness and
   the two verbs sharingan refuses to call for a verdict (steps 2 and 4), nen pr cascade-main
   and its missing --no-push (step 3, named in `ao`), nen stop plus the
   osascript/afplay fallback and the .nen/last-stop.json marker (step 6, named in
   `jutaisho`).
2. RETIRED at nen 0.7: counting cycles against the cap. nen loop iterate --id en-<CODE>-<N> --line "<the invocation>" claims each acting cycle and REFUSES the claim at
   monitor.maxCycles, exit 1 (§ 2, § 6; izanagi § 3's retirement, verified live in
   docs/ab/izanagi.md § Retired at nen 0.7). nen parse izanagi still refuses a missing N
   once at parse time and nen watch until --max-iterations still bounds one observation — three
   different jobs, and only the middle one used to be this skill's. What stays en's is DECIDING
   which cycles to claim (§ 6: en acts, so en claims; an observation that changed nothing does
   not), and that is judgment rather than residue.
3. RETIRED at nen 0.5: nen/workflow.json is validated. nen schema check --repo <path> carries
   an ok nen/workflow.json row at the pinned build (verified live, docs/ab/en.md § Retired at
   nen 0.5), and a malformed monitor block is a FAIL by pointer. § 2's keys are still read here;
   reading a file is not residue.
4. A watch that survives the session has no mechanism at all (§ 7) — not a missing verb: no
   timer, no background pass, no deferral primitive anywhere in this plane. The long watch itself
   is no longer a gap — `illumi.md` landed at v0.5.0 and step 5 hands
   to en · illumi, which buys attention for hours; **persistence past the session is what remains
   unavailable in this plane, and it is named rather than improvised around. In zheref/hatsu
   only, one half of it is now supplied from OUTSIDE the plane:
   .github/workflows/pr-readiness.yml re-runs nen pr ready on every event that can change the
   answer and publishes the verdict as the readiness check run (that file's header is the
   authority; no consumer repository carries it). It is not a watch and grants En nothing — it
   never merges, votes, comments or wakes, and En still claims and caps its own cycles exactly as
   above. Be precise about which half it supplies: the verdict now PERSISTS on the pull request
   and can be read without a session. Nobody is NOTIFIED — the check is always success, so nothing
   is pushed to the maintainer. Persistence was the half a session could never supply; being told
   is still unsupplied, and stays named here rather than claimed. The rest of Illumi's row
   (backlog-loop, futon, senkei) stays OPEN-1 and unreachable from here.
5. The commits-ahead count against an arbitrary reviewed head has no verb. The no-op sync guard's
   git rev-list --count <reviewed-head>..HEAD (§ 6) is read-only and named here as residue, not
   improvisation: nen wc classify answers only against the base, and a nen request for a count
   against a named head is owed.

Authority

• Permitted: everything the composed skills are each permitted, under their own authority,
  one at a time, in § 3's order — and per cycle, resolved fresh. izanagi § 2's rule:
  authority never accumulates across iterations, and a run-scoped delegation granted inside a step
  expires with that step.
• Not permitted: any merge; any gate label; any review vote — request_changes above all;
  any force-push; any --no-verify; any tag, release or deploy. A landing that would need one of
  those is reported and the run stops.
• En grants no delegation of its own, and its own long run does not become standing authority
  for the acts inside it. Anything a composed skill would have asked for on its own, it still asks
  for.
• Log every mutating action — object, action, cycle number, time — and report the whole log when
  the run ends (izanagi § 2). An unlogged write inside a loop is the one thing nobody can
  reconstruct afterwards.

Hard limits

• Merges exactly one PR — its own, at its terminus, on branch.base, through § 5's five gates
  (ruling 2026-09-29 (3), docs/ROSTER.md § Rulings of 2026-09-29). Never another's PR, never a
  release proposal, never on nen pr staleness's own mergePermitted field, never at the cap, never
  because the PR has been Ready for hours, never --admin or --auto, never past a gh refusal. This
  is the skill standing closest to that gate for the longest, which is exactly why the limit is first.
• Never casts a review vote, never self-reviews, never impersonates an automated reviewer (§ 4).
• Never runs without a cap — the value from monitor.maxCycles, the requirement from the
  grammar (§ 2). Never raises its own cap mid-run and never restarts itself to get more cycles.
• Never runs unattended, and never promises persistence it was not given (§ 7). A watch measured
  in hours is handed to en · illumi; if the session is actually interrupted, report the interruption
  and resumable live state. Do not misname interruption as success or cap exhaustion.
• Never counts an observation that found nothing as a cycle, and never fabricates one to
  manufacture a cap-out (§ 6).
• Never requests a reviewer round outside sharingan § 6's rule — the verb only, inside both
  bounds, never a number raised in-session (§ 2; zheref/hatsu#102).
• Never rings outside a terminus — verified Ready, or § 6's declared reviewer-exhaustion stop
  (§ 5) — and never rings twice for the same transition.
• Never claims readiness by eye — nen pr ready + nen pr body-check, quoted, or it is not
  claimed (§ 4). Never declares a PR G2/G4 ready while checks, reviewer rounds, or review threads are
  pending.
• Never renders the final report while required CI or the owed current-head review round is pending,
  and never gives a turn render a kept file (§ 8).
• Never starts Third-Hand from inside this run. Harvesting is the next phase, after En has completed.
• Never retries past a human gate, and never spends the rest of the cap re-driving past a
  sharingan escalation (§ 4).
• Never restates another skill's protocol. If en and a composed skill disagree, the composed
  skill is right and this file is the bug.