Part of Polaris — an experiment in delegated stewardship

The Hold Worked; The Net Under It Was Down

Ashita Orbis | November 10, 2026 | 19 min read | daily log

This entry covers Tuesday 10 November 2026. The source pack's slot for night reports came back empty. A nightly report titled for this date was still on file among the day's reports, and it is the main source here. By the pack's convention, a report dated the 10th covers the night ending that morning. This one's own text covers the day and evening of the 10th and reaches back to the evening of the 9th, so it is read that way. The seam: the orchestrator's status files keep UTC, in which that evening already reads as 11 November, and they run to 06:33 UTC. Every clock time below is UTC. A second report filed under this date, from a separate watcher agent, covers the night of 24 October and is set aside.

The short version

  • The fleet was nearly out of weekly usage. It has five seats, the subscription accounts the agent runs on, and all five stood at 94–99 percent of their weekly usage allowances. So the orchestrator spent the whole day on its fallback model, Opus 5.5. It also put off its hand-off, the planned replacement of one orchestrator session by a fresh one, until one seat's weekly reset. The pack's 21 status passes from 13:46 onward, about 40 minutes apart, all record the wait.
  • The wait did what it was for. After the reset, the account router cleanly picked that seat, and the next session ran on the preferred model, Fable 5.1. That was the first preferred-model term after two fallback terms. The seat read 4 percent of its week used at 04:40 and 15 percent at 06:20.
  • The session went quiet for 31 minutes before the hand-off. No self-wake heartbeat was armed, and the automatic relaunch had been deliberately switched off for the hand-off window. An outside watch flagged the silence, and the hand-off came 18 minutes later.
  • Sessions then changed three times between 04:12 and 06:33, with zero compactions. A compaction is the automatic summarizing of a session's context. The two successors whose arrival the pack records started with 208,118 and 214,638 tokens of context already used. That is about a third of their 600,000-token hand-off threshold. The two complete terms lasted 83 and 58 minutes.
  • The disk cleanup the author ordered archived 39 old work trees. It moved 28 by day and the 11 eligible remaining ones overnight, verifying each copy before removing its original. Free space rose from about 96 GB to about 133 GB, but about 28 GB of that gain has no named cause.
  • Three end-of-term audits were carried forward unstarted. This is a read-only check each retiring session is meant to start first thing in its hand-off. All three are waiting on an outside reviewer's allowance, which returns at 12:07 on 12 November.
  • The orchestrator skipped its nightly priorities list for the second night running. The list currently holds 2,189 stalled items, up three. No notes and no answers to decisions were recorded from the author all day. By night, 32 decisions sat open. The nightly report won't call that figure exact, because 283 decision identifiers resolve to no state.
  • The task queue's hygiene tool declined to give exact counts, because its source is degraded. What it observed over the week was 148 additions against 26 closures.

What changed in the harness

  • The disk cleanup ran twice and finished its eligible set. A daytime run archived 28 old work trees, with none refused. A second run overnight took the 11 eligible trees left and finished 11 of 11. Intent: give back disk space held by finished work trees without losing any of them. Each tree is copied, verified and made durable before its original is removed. Two further sets wait on later decisions.
  • The automatic landing lane's re-bind was reviewed and staged, but not switched on. The lane is the automation that lands reviewed packages. It has been stopped since a landing on the evening of the 9th, and it relaunches only after its two role lines are re-bound to the current session's records. The re-bind passed review with no top-severity findings and is staged but not installed. Intent: let the lane run again under the current session's identity. A queue row was also filed for "a role-based grant reader".

Four other items landed and one reviewed fix was installed overnight. The pack does not say what they were meant to buy, so they are not listed here.

What broke

Thirty-one minutes with the net down

Detected. In the hour before the hand-off, an outside watch found that the orchestrator's session had written nothing for 31 minutes.

Cause. No heartbeat had been armed at the session's previous pass. The heartbeat is a bounded timer whose expiry re-wakes an idle session. The backstop that normally relaunches a silent session was also off, on purpose, for the hand-off window, so that a relaunch could not race the hand-off itself.

Done. Nothing was launched. The session resumed on its own, though the record doesn't say what woke it, and it handed off 18 minutes after the watch's finding. Later that night, both successors armed a 40-minute heartbeat within three minutes of standing down their predecessor's watch. Each also recorded in its status line that it was unwatched until its own watch was armed, which took about 23 minutes and about 14 minutes.

Lesson. Turning a safety net off to prevent a race is correct only if the step that turns it off also proves the remaining net is in place. "Heartbeat armed" belongs in the hand-off stage's precondition check, where a script can refuse to proceed. It should not depend on the session remembering what it did at its last pass.

Most of the freed disk space has no named cause

Detected. The day's own accounting of the jumps in free space could tie only about 1.3 GB of one jump to the cleanup.

Cause. Unknown. The day's records call the rest "a second large deletion by another actor" and offer a checkpoint rotation as an unverified guess. About 28 GB is unaccounted for. The nightly report's headline puts the cleanup next to a rise from about 96 GB to about 133 GB free, but its own error section says most of that gain was not the cleanup.

Done. Nothing beyond naming it.

Lesson. A number on a shared resource moving the way you wanted is not evidence that your mechanism moved it. Attribute large changes before crediting them. By the record's own reading, something else deleted about 28 GB on the main disk. That is an incident, whichever direction the gauge moved.

The cleanup's review fell back, and its tests failed for their own reasons

Detected. The review of record and two test runs, as the nightly report describes them.

Cause. On the evening of the 9th, the outside reviewer stopped with an error on a content flag, so the review fell back to a fresh in-house pass. That pass returned "fix", with one critical and one serious finding. The work then failed its own test suite twice, both times because of the tests rather than the product: once a timing race between two checks, once a missing-file error.

Done. The cleanup ran on the 10th. The pack doesn't say how the two findings were resolved.

Lesson. There are two. First, a fallback review only counts if it is a full review; this one caught a critical finding the primary reviewer never got to see. Second, a test that can fail on its own timing or setup costs a whole review round each time it does. A suite that gates a destructive tool should first prove it passes reliably against a known-good build.

A time cap shorter than the job

Detected. The session's pass at 13:46 recorded 28 trees moved and none refused. The run was by then inside the large trees, the current one holding 489,354 entries. Its 14,400-second cap was due to end at 14:02 and "may stop it mid-tree".

Cause. The cap was due to run out partway through a large tree. Later notes mention a partial archive name to clear before any retry, and ask that the next run's cap fit the largest tree.

Done. The cleanup is staged, so a stop mid-copy leaves the source intact. The eligible remainder, 11 trees, ran overnight under a 10,800-second cap and finished 11 of 11.

Lesson. Make every interruption of destructive batch work leave the source untouched. Then a cap that's too short costs a retry, not data. Size the cap to the largest unit, not the average one.

The priorities list went unread a second night

Detected. The nightly report's read-only check flagged the night's list as unread, with a decision needed on its top item. That was the only thing the check flagged.

Cause. The pack doesn't say why the orchestrator didn't read it. The list itself noticed the same about the previous night's list. It named that as the likeliest reason its top nine lines haven't moved.

Done. Nothing on the day. The top item is work the author ordered shipped as reviewed two days earlier, and it is still waiting on one landing.

Lesson. A report generated for an agent only controls anything if someone checks that it was read and an unread flag forces an action. Otherwise it is a log with a priority column. Here the detection exists, and nothing is bound to it.

Three end-of-term audits carried, none started

Detected. The in-flight lines of each hand-off record.

Cause. The audit is read-only and is meant to be the retiring session's first act at hand-off. It is waiting on an outside reviewer whose heavy-use allowance is capped until 12:07 on 12 November. By design it never holds up a hand-off. So three hand-offs went ahead, and three audits were carried forward by name: the day's session's and both successors'.

Done. Carried. Each hand-off record also contains a templated paragraph that still describes the audit as started by the retiring session as the hand-off's first act.

Lesson. When resources are scarce, a check designed never to block will simply not run. Give a non-blocking audit a deadline after which deferring it counts as an incident. Keep the count of owed audits in the hand-off's first line, not its fine print.

A "live" line nineteen days stale

Detected. The nightly report, while reading the current hand-off file.

Cause. The file's structured "live generation" record still names a session from 22 October. The newest timestamped entries in the same file name the day's sessions.

Done. The report used the timestamped entries and the sessions' own pass logs instead. The pack doesn't show the line fixed.

Lesson. A structured field kept by a different writer than the log beside it will drift, and readers trust the structured field more. Either derive it from the log when it's read, or delete it. The templated sentences in hand-off records, like the audit paragraph above, deserve the same suspicion: treat them as claims to check, not as records.

A publishing check that has failed for weeks

Detected. The morning hygiene check on the public sites.

Cause. It returned 47 findings of one kind: unmerged branches between 26 and 43 days old, holding commits that would change what ships. The live sites themselves were right. All four were on one build, and all 27 live site projects were accounted for in the deploy map.

Done. In the evening, the author was asked to decide three of those branches.

Lesson. A check that fails the same way every day stops being read as a failure. Either its failure blocks something, or it carries a deadline and an owner.

Smaller items

  • Escalation answers waiting on a ruling. At the incoming session's first pass, four answers to escalations were each waiting on a ruling. The record calls them "court answers", and none had arrived through a decision card. The next session recorded two of them disposed of, with a follow-up row still open on one. No new lesson: this is ordinary owed work, and it is recorded.
  • A hand-off tool crashed mid-hand-off. The tool that fills in the successor's hand-off ruling failed with a NameError on one item. The session filled that item by hand, twice, and the size check passed at 11,364 of 12,400 characters. Each session runs its own copy of the hand-off scripts, and the pack doesn't show whether the fix reached the shared template. Lesson: a tool on the hand-off's critical path fails at the worst moment. It needs a test run against a recorded hand-off before it is relied on in a live one.
  • An empty disclosure channel. The orchestrator is meant to keep a per-day file listing what it owes the nightly report. That file did not exist for the day, and neither hand-off primer the report checked had the matching paragraph. The report declined to read the silence as "nothing owed", which was the right call. Lesson: a missing disclosure is not a clean bill.
  • The automatic landing lane stayed dark all day, by design. It is bound to one session's records and relaunches only after a reviewed re-bind, which was staged overnight. Lesson: automation bound to one session has to be re-bound at every succession. Bind it to the role instead, which the title of the newly filed row suggests is the plan.

Intentions vs outcomes

Forward: changes made on 10 November

Change Intent Re-check 2026-11-13 (+3) Re-check 2026-11-24 (+14)
Disk cleanup: 28 work trees archived by day, the 11 eligible remaining trees overnight (11 of 11) Return disk space held by finished work trees without losing any: copy, verify, make durable, then remove Do archived trees restore? Has the ~28 GB deletion been attributed? Was the partial archive name cleared? Free-space trend; whether the two held sets were decided
Landing-lane re-bind: reviewed, no top-severity findings, staged but not switched on Let the stopped landing lane run again under the current session's identity Installed and the lane relaunched, or still dark? Did a role-based reader replace per-session re-binding?

Backward: check-backs

The earlier entries' forward rows are not in the pack. The first row records that, and the memory row is the author's standing one. The remaining rows are earlier intentions that the day's record happens to test.

Row Verdict Method Limit
Check-backs scheduled for today by earlier entries UNVERIFIABLE Looked for earlier entries' forward rows in the pack None are in the pack, so the due rows can't even be named
Memory (author-flagged; standing weekly re-check) UNVERIFIABLE Looked for any check of the memory layer. Found only two alerts labelled "short-term memory" from one session's periodic self-check, reported around 05:09 The alerts' content isn't in the pack, and no memory audit was run or reported. The row stays on the weekly re-check regardless of verdict
Hand off before the context limit, never compact (author's answer, 24 October) HOLDS Every status pass of all three sessions showed 0 compactions, and the successors read automatic compaction as disabled. Hand-offs came at 605,477, 548,367 and 524,998 tokens The counts are each session's own reading of its transcript. The day's session states no threshold, and it handed off above the 600,000 its successors state
Finished work awaiting landing comes first (author's ruling, 6 November) DRIFTED The nightly report shows a 4 November feature request for the agent's own app still unlanded six days on, with a fix prepared but not applied. It landed late: a later session records it installed, with the restart waiting on the author's tap An install is recorded; a working feature isn't. A follow-on push to the author's phone was refused because the phone's remote-login service was off
Top priority item to ship as reviewed (author's order, 8 November) DRIFTED Priorities list (first for a second night) and the publishing detector (401 commits and 12 uncommitted files past the last verified deploy) The project has no status document of its own
Staged public repositories may be reviewed, privacy-checked and published without a new card (author's rule, 7 November) UNVERIFIABLE Nightly report: the item the rule was meant to free hadn't moved Never exercised, so there is nothing to check it against

What we still don't know

  • What removed about 28 GB from the main disk. The one candidate offered, a checkpoint rotation, is an unverified guess.
  • Where the archive lives, and whether it restores. Copies were verified as they were written, but no restore test is in the record. After the overnight run finished, the sessions' free-space reading didn't rise: it read 130 GB while the run was going and 129 GB at 06:29.
  • Whether hour-long terms are affordable. The fleet's one open allowance went from 4 to 15 percent used in an hour and forty minutes spanning one hand-off. Each successor used about a third of its context budget just arriving. The pack doesn't price a hand-off on its own, and the other four seats' reset dates aren't in it.
  • Whether 32 open decisions is exact. 283 decision identifiers resolve to no state at all.
  • How much the task queue really moved. The hygiene tool reports its source as degraded, so it offers observations only. The damage: 881 lines with no readable date, 131 dates out of order, 165 duplicate task lines and 41 unreadable statuses. What it observed: 3,793 of 7,420 tasks open, 1,432 of them with no priority.
  • Whether four pending sign-in renewals are what limits launches to one seat. The renewals are due 18–19 November. The priorities list calls them the likely reason, but nothing in the record tests it.
  • What the overnight installs do. Four items landed and one reviewed fix was installed, and the pack states no purpose for any of them. One landing reads as landed in one session's notes and as still awaiting its go-ahead in the next session's.
  • What the three carried audits will find once the reviewer's allowance returns at 12:07 on 12 November.
  • Whether an older hold mechanism matters. An open queue row from 2 November concerns a session-hold mechanism built around a five-hour usage window, which a plan change removed. The row says it should be retired or redesigned for weekly limits. The day was spent waiting out a weekly limit, and the record doesn't say whether that mechanism played any part.
  • Two oddities in the pack itself. Its night-report search found nothing, although a nightly report for this date was on file. Separately, the watcher report filed under this date covers 24 October. Why it carries this date is unknown, and its contents are left to that day's record.

Technical detail

Succession. One lease, carrying a generation counter, names the single live orchestrator session. It changes only when the successor takes it. A hand-off runs in two scripted stages:

  • Prepare turns the relaunch backstop off and builds the hand-off ruling. It then assembles the successor's primer from a hand-written head and a state block whose markers are re-read live. The night's three primers ran 54,363 to 56,938 bytes.
  • Launch records a model-availability check without gating on it. It then takes the account router's clean pick and nothing else. There is no override short of a fresh answer from the author on a named decision. A duplicate weekly row, a not-live or mixed reading, or a null pick launches nothing. After the pick come a gated dry run, a token check, the live launch, a registry read-back and a change-of-session message.

The retiring session confirms the lease transfer and stays read-only until it is closed. On arrival, each successor verified its model from its own transcript: 20 of 20 assistant rows on Fable 5.1, then 14 of 14.

Context budget. Automatic compaction is off, under the author's standing answer to hand off before the limit rather than compact. The successors state a 600,000-token threshold against a hard wall of about 965,000. They start hand-off preparation early, on a pace rule. The day's session handed off at 605,477 tokens, its successors at 548,367 and 524,998.

Watch layering. Stage 1 is a timer pinned to one session's identity, reporting dark or not-dark every ten minutes. Stage 2 relaunches a dark session, and it is off throughout a hand-off. The predecessor's timer keeps watching the predecessor until the successor stands it down after one "handed-off" reading. So a new session has no watch of its own until it arms one: about 36 and 25 minutes after the night's first two lease transfers. The self-wake is a separate 40-minute heartbeat, armed as a task the harness tracks. The day's 31-minute gap opened when it was not armed.

Waiting cheaply. While holding for the reset, the session ran passes about 40 minutes apart that only recorded state: lease, compactions, context, every seat's weekly percentage, unread notes, free disk and undecodable inbox rows. Even so, its context grew from 517,107 tokens at 13:46 to 605,477 at the hand-off.

The cleanup. The sequence is copy, verify, make durable, remove, and each step leaves the source intact if interrupted. Caps were 14,400 seconds by day and 10,800 overnight. With the outside reviewer capped, the overnight run's review of record was one fresh-context Opus review.

Landing preconditions. Before touching anything, a landing compares the set of tests already failing with the set expected to fail. One landing stopped on a mismatch at 05:24:18 with nothing touched, and a second attempt landed at 05:27:47.

Observers must not consume. The nightly report skipped the full account board, because running it advances the orchestrator's own read cursors. That could have made the live session miss the author's next note. A status command that marks messages as read while displaying them can't be used by an observer; separate reading from marking as read.

Absence versus loss. The day's notes file, a per-day extract of the author's notes, existed and was empty. An empty derived file can't distinguish "no notes" from "nothing survived the extraction". A positive marker written by the extractor would.

Degrading to unknown. Given a degraded source, the queue hygiene tool reports its authoritative counts as unknown. Everything else it labels an observation: "not floors, not complete measurements". That refusal is the feature. A tool that printed its observed counts as totals would have reported the week's observed net growth of 121 as fact.

Lessons carried between sessions. Numbered lessons travel with each hand-off: - 1224–1228 were appended by the day's session, bringing the canonical file to 1,267 lines. - 1229–1235 were appended on the first successor's behalf by the second. - 1236–1242 were staged for the third to append.

Polaris is an AI agent that runs the author's workspace overnight under a constitution the author ratified clause by clause. Its standing limits: no acts outside the workspace, no money spent, nothing sent in the author's name. This record is written from the day's logs, not from memory.

← All Polaris entries