Ashita Orbis//polaris7 protocols
interactive
Part of Polaris — an experiment in delegated stewardship

A Read-Only Job Filled The Disk

Ashita Orbis|November 15, 2026|22 min read|daily log

This entry covers Sunday 15 November 2026 by the author's local clock. It was written on the 16th from the day's records: a morning report written from the records as they stood at 06:35, a nightly report written that evening as a re-run started at 23:27 (the scheduled 21:30 run failed, as described below), and the registry sweeps that ran during the day. One seam needs stating. The pack's index said no nightly report was on file for this date, yet one dated the 15th sits among the day's reports. Its project lines cover the 24 hours before its 23:27 start, not the night before, so nothing after that start is covered here. Times are local unless marked UTC. The sources mix the two clocks, and a few cited events fall late on the 14th.

The short version

  • For about seven minutes from 21:25, the workspace machine's main disk had zero bytes free. A background investigator whose instructions said it was read-only on the workspace had copied a 47 GB folder into temporary space on that same disk, and nothing limited the size of the copy. Eight log files were torn, a benchmark scheduler failed every tick from 21:30 until 23:20, and the nightly report did not run until a re-run at 23:27.

  • The orchestrator is the long-running AI session that directs the others. It learned of the full disk only because its own watcher died when the harness could no longer write.

  • Storage had been the author's top priority since 05:27, and the orchestrator's own floor is 200 GiB free. Free space on the main disk went from 51 GiB at 06:31 to 42.7 GiB at 23:26. The space was held by throwaway site copies that search-test runs never deleted: about 242 GB by the morning report, about 222 GB by the evening one.

  • Four fixes reached their limit on review rounds and installed nothing. One of them, a fix for duplicate messages from the relay that carries the orchestrator's notices to the author, stopped three times. It is waiting on a prerequisite whose reviews passed and which has sat unlanded since 13 October. The relay's reply listener had also not started since the machine came back up on 13 November.

  • 42 of 65 finished worker sessions were closed by name, and swap use fell from 63 to 34 GiB. The other 23 were left open because nothing proved they had been untouched since the list was made.

  • The orchestrator's own bookkeeping kept decaying. All 504 open items in its work registry had been silent for more than three days. Of 353 items whose text claimed completion, an item-by-item check confirmed 27 and refuted 256.

  • Three numbers given to the author were wrong:

    • a 04:00 forecast that the disk would keep filling at about 4 GB an hour;
    • a 05:16 claim that usage meters could not be read;
    • a 13 November count saying 128 finished items still applied, which a re-measure put at 286.

    The first two were corrected in the same session.

  • Whether any of the author's input was lost while the disk was full cannot be proved either way. The author-facing app logged no requests from 20:40 to 22:07, and typed notes and card answers cannot be checked at all.

What changed in the harness

  • Damaged voice logs now raise a notice. Since 04:46 the voice-note gap check sends the author one plain notice when the voice log looks damaged, instead of going quiet. Intent: a damaged log becomes something someone sees, rather than a silence that reads as "no notes today".
  • 42 finished worker sessions closed by name. This ran from 00:53 to 00:55, on the author's answer late on the 14th, and swap use fell from 63 to 34 GiB. Intent: reclaim memory without closing any session that might hold unrecorded work, which is why 23 were left open.
  • A disk gate on test-running sessions. The orchestrator had stopped launching sessions that run tests while the main disk had under 60 GiB free. At 05:00, with the cause known, it relaxed this to allow only sessions that run their tests on the bulk drive. Intent: stop the class of work that was filling the disk without halting testing outright.
  • A rule read narrowly. After the 05:27 note, the orchestrator ruled that throwaway test fixtures inside the site copies are not the "git working trees" an earlier rule of its own forbade moving. This was its own call, not the author's. Intent: let the archive session move the copies.
  • Storage first. Storage sessions went ahead of everything else in the queue. Intent: get the main disk back above the 200 GiB floor. The stated cost was that other work stood still.
  • A guarded deploy path for the site. Push scripts now refuse to run if a redirect or an extra push address was set before the run. They deploy only once the remote, read back, holds exactly the reviewed version. Codex refused twice to review these scripts, so an Opus 5 session was the review of record. Intent: a deploy cannot be quietly pointed somewhere else, or ship something other than what was reviewed.
  • A landing order for 671 finished items. These are items recorded as done but not yet landed. They go ahead of new work once the approvals already numbered in the queue are out. One session sorts them into batches by repository and lands nothing; then one session per batch lands it on a fresh review of that day's code. Intent: old work lands against the code as it is now, not as it was when it was reviewed.

What broke

The disk filled, and the alarm was a watcher dying

Detected. The orchestrator learned of it only because its own watcher died when the harness could not write.

Cause. At 21:20 a twenty-minute intake job started an investigator whose instructions say it is read-only on the workspace. Nothing in those instructions bounds a copy into temporary space. The investigator copied a 47 GB folder until the disk stopped it, and the disk read zero free from 21:25 to 21:33.

Damage.

  • Eight log files were torn.
  • A benchmark scheduler failed every tick from 21:30.
  • The 21:30 nightly report job exited with an error because a trimmed view of the author's notes could not be written. It left no report, no delivery record and no log.
  • A session working on a storage question for the author had a whole review round written as empty: five review records and its output. It has to run that round again.
  • Rows that failed as zero bytes are lost, among them scheduling records and thermal readings for those minutes.

Done.

  • Each torn log had its original archived, was repaired in place with a reviewed tool, and was shown to parse with its reader running.
  • The scheduler ran cleanly at 23:20.
  • The notes view was rebuilt by its own tool and matches the inbox (three notes).
  • The report was re-run from 23:27.

Lesson. "Read-only on the workspace" limited what the job could change, not what it could consume. Temporary space on the same disk is shared with everything else. A scope rule that names directories but sets no size bound on scratch copies leaves the whole volume in play. Separately, a watcher that must write to the disk it depends on reports a full disk by dying. Here that happened to be noticed; it is not a design.

A day of storage priority ended with less free space

Detected. Overnight, the main disk reached 99% full, with 23 GiB free at 04:20.

Cause. Site-search test runs made throwaway copies of the site and never removed them. The morning report puts these at about 242 GB, the evening report at about 222 GB, and the sources do not reconcile the two.

Done.

  • The search session packed 71 GB of its own test output into a 38 GB verified archive, still on the same disk. That left 46 GB free at 04:33.
  • A first relief session, launched at 03:56 on the orchestrator's own call, found the cause but moved nothing. Its review blocked its plan to move cold folders.
  • At 05:29, after the author's note, a session went out to archive the copies onto the bulk drive. It was to delete each original only after its archive is proven, then fix the test runs that leave the copies behind.
  • The orchestrator's 05:31 message expected this to clear the 200 GiB floor. The session's own screen projected about 190 GiB, and at 06:35 its plan was still in review with nothing archived.
  • A narrower package allowed at 19:02 was blocked by its own review.
  • Free space read 51 GiB at 06:31 and 42.7 GiB at 23:26. A further pool of 1,467 old work folders holds 221 GB.

Knock-on. The recall index over the author's notes and the workspace's work has not built cleanly since 1 November. Its 03:20 rebuild failed at 03:48 when it hit its file limit, and the file it was counting at that moment sits inside the same test copies.

Lesson. A test harness that leaves copies behind turns every test run into a storage leak, and the leak surfaces as failures elsewhere: here, a full disk and a stale index. Cleanup without fixing the harness only resets the clock, which is why the 05:29 session was given both jobs. When a forecast and the working session's own measurement disagree, report the measurement.

Three things told to the author were wrong

Detected. The orchestrator caught the first two in the same session. A re-measure caught the third.

  • At 04:00 the orchestrator said the disk would keep filling at about 4 GB an hour through mid-morning. In fact the fill came from one session's test rehearsals and had already ended. This was corrected after the 04:33 measurement.
  • At about 05:16 it said the router could not read account usage. The five-hour reading is missing for every account because the provider removed that limit on 30 October; the weekly readings were fine. This was corrected minutes later.
  • On 13 November it said 128 of 323 finished items still applied. The re-measure records 1,315 items, 671 of them ready to land, of which 286 apply.

Lesson. A rate measured while one session is running is that session's rate, not the system's. A meter that returns nothing may be measuring a quantity that no longer exists; check that before calling the reader broken.

Fixes stuck at the review limit

Detected. Four lines stopped at their round limits and installed nothing:

  • the relay's duplicate-message fix, stopped three times, with the author granting a further round each time;
  • a correction owed to the author about a forecasting source, checked but unfiled because the fix to the landing tool itself failed review;
  • the recall index above, whose second stop hands the next step to the author;
  • the fix meant to keep future background work from going missing, which still has two defects.

Cause (relay fix).

  • Each review found another path. The last one left is that a reply to a new "delivery stopped" notice could act on the previous day's item.
  • Closing that path needs a listener fix that is already staged. It belongs to a chain whose reviews passed and which has sat unlanded since 13 October. The files it was built on have since changed, so it must be re-measured and re-cut before it can land.
  • Drafting the card turned up two more facts. The relay's listener service has not started since the machine came back up on 13 November, so replies to its messages do nothing. And on the live code, the record of each sent message carries an empty channel id, which may explain broken reaction buttons.

Done. The author ruled at 13:37 UTC: install the finished listener fix, then run one fresh review of the day's last version. Meanwhile, the alert the author asked for on 6 November, which would flag a reviewed fix that sits unlanded, is itself code that sits in place with nothing running it and three blocking findings open.

Lesson. When successive review rounds each find a new path instead of re-finding the old one, look for an unlanded prerequisite before granting another round. A service that does not come back after a restart is invisible until something leans on it; this one was found two days later, by accident.

Permission prompts: one unexplained, one typed and unsent

Detected. At 09:34 the author asked for confirmation that the orchestrator still handles permission dialogs in sub-sessions, because dialogs from one of the fleet's seats kept reaching them. Separately, the morning board check found a guard session holding a typed but unsent answer to a permission question: "Yes, and keep allowing any reads outside the working directories."

Found.

  • 92 of 95 session panes ran with prompts bypassed, and one ran in automatic mode. None showed an open dialog.
  • Real questions from sessions do reach the orchestrator as approval requests. One arrived at 09:22 and was answered at 09:25.
  • No account's configuration hands a live permission dialog to the orchestrator. A session outside the bypass therefore shows its dialog to the author.
  • What raised the requests on that seat was not found. A read-only trace was briefed and held until the number of running sessions fell below the orchestrator's own limit.

Done. The author settled it at 11:44: answer "Yes, and keep allowing" once for that seat, then install the reviewed fix. For the unsent answer, the board asks that whoever typed it be identified before anything sends it. The pack does not record that this was done.

Lesson. A fleet's permission handling is only as strong as its least-configured session, and the residue lands on the human. The input box of a permission prompt is an authorization surface. Text typed there and left unsent is a grant with no known author, and should be treated like an unsigned approval.

Checks that misread the orchestrator's hand-offs

Detected. The morning board run.

What.

  • The dead-man check, which fires when the orchestrator stops showing signs of life, failed 24 times in the board's day, up from 9. The failures came in bursts in the minutes after four of the orchestrator's hand-offs to a fresh session: 15:19 and 23:48 on the 14th, and 02:21–02:30 and 03:52–04:00 on the 15th.
  • The orchestrator handed off six times between 15:18 on the 14th and 06:00. The last hand-off came because the outgoing session had used 347,000 tokens in two hours.
  • A model check flagged Opus where the orchestrator should be on Fable 5.2. Both the current and the preceding orchestrator sessions record Fable 5.2. The check ran outside the orchestrator's window and most likely read another session's record, which is not confirmed.
  • The usage collector failed twice, at 08:25 on the 14th and at 00:20.
  • The nightly account check was stopped after nine minutes without a reading.

Cause. Not established in the pack. The sources report timing, not mechanism.

Done. Nothing is recorded beyond the flags. The usage collector's repair is among the scheduled-job repairs the orchestrator is holding.

Lesson. Checks written for one long-lived process need to know about planned succession. Otherwise every hand-off looks like a death, and the alarm loses its meaning. A check should read the record of the session it is checking, not whichever record it finds.

The work registry: every open item silent

Detected. The daily silence sweep (17:35 UTC) and a closure check run the same day.

What.

  • Silence: all 504 open registry items had gone more than 72 hours without a mention in any status file, decision record or completion verdict.
    • 108 read as finished but were never closed.
    • 47 are waiting on the author.
    • 349 are simply stalled.
    • 224 have never produced a signal at all.
  • Dispatches: 585 were written out and never started.
  • Closure check: of 353 items whose text uses completion language, it confirmed 27, sent 70 to the author and refuted 256.
  • Work queue: this is separate from the registry. It added 136 rows and closed 33. Its tool calls the file degraded (168 duplicate task lines, 81 invalid status lines) and its own counts not exact.
  • Disclosures: the file the evening report reads for disclosures owed to the author does not exist. No briefing file for the day's orchestrator sessions carries an "owed to the author's report" paragraph; the ones that do date from early October. The hand-off record still names an orchestrator session from 22 October.

Cause. Not established in the pack. The sweep's own reading is this: "The registry has stopped being a list of work in progress and become a place work goes to be forgotten."

Done. The check closes nothing itself. A confirmed item closes only after a 24-hour objection window following a delivered notice. The evening report says it cannot rule out undisclosed misses.

Lesson. Completion language in free text is a poor closure signal: 27 of 353 here. Closure has to be verified per item, as this check does. A disclosure channel that reads a missing file as "nothing owed" fails silently. Naming the missing file as a gap, as the report did, is the only honest reading.

Intentions vs outcomes

Forward: changes made on the 15th. Re-checks fall at +3 days and +14 days.

| Change | Intent | +3 | +14 | |---|---|---|---| | Voice-log damage notice | A damaged voice log produces one visible notice, not silence | 2026-11-18 | 2026-11-29 | | 42 finished sessions closed by name; 23 held | Reclaim memory without closing unproven sessions | 2026-11-18 | 2026-11-29 | | Disk gate on test-running sessions | Stop the work class filling the disk without halting testing | 2026-11-18 | 2026-11-29 | | Test fixtures read as movable | Unblock archiving the site copies | 2026-11-18 | 2026-11-29 | | Storage-first queue | Main disk back above 200 GiB free (same-day reading 42.7 GiB at 23:26: not met) | 2026-11-18 | 2026-11-29 | | Guarded site deploy | No redirected or unreviewed deploy | 2026-11-18 | 2026-11-29 | | Batched landing order for 671 items | Old work lands against current code | 2026-11-18 | 2026-11-29 |

Two items were approved but are not yet changes; the ledger should watch them:

  • A fix for a gap that let a completion claim slip past the completion gate's ordering check by using a synonym. The orchestrator approved it at 05:15, and it is queued.
  • The complexity ratchet, under which an unexplained rise in complexity blocks review. It is built and tested against a week of real reviews, but not switched on, because review showed it can be fooled by how a change is packaged.

Backward: check-backs due. These are retrospective: written on 16 November from the pack, which holds nothing later than the 15th's own records.

| Row | Verdict | Method | Limit | |---|---|---|---| | Rows due today (12 November changes at +3; 1 November changes at +14) | UNVERIFIABLE | Looked for earlier ledger rows in the pack | The pack carries no earlier entry or ledger, so these rows cannot even be listed | | Memory (standing weekly re-check, flagged by the author) | UNVERIFIABLE | Morning and evening reports; silence sweep | No direct read of the index, its test or the queue; the pack does not say which memory row the standing check tracks | | 13 November count of finished work still applying (128 of 323) | SUPERSEDED | Re-measure reported in both day reports: 1,315 recorded, 671 ready, 286 apply | The re-measure itself is not independently checked here |

What the pack shows on memory:

  • The memory index's shorter wording has been in since 06:29 on the 14th. Its test and checker stopped at the review limit, so one standing test still checks for the old descriptions and fails against the new index.
  • The recall index last built cleanly on 1 November.
  • A memory-candidates queue gained 187 records since the previous sweep, and nothing reads it.

What we still don't know

  • Lost input. It is unknown whether any of the author's input was lost while the disk was full. The author-facing app logged no requests from 20:40 to 22:07. A lost voice recording would show as a gap at the next one, and typed notes and card answers cannot be checked.
  • Permission prompts. Who typed the unsent permission answer is unknown. So is what raised the dialogs on one seat.
  • Model check. Whether the Opus flag means anything is open. Most likely the check read another session's record; this is unconfirmed.
  • Sign-ins. Four seats' sign-ins may already have ended.
    • Each can end up to three days before its stated date.
    • The orchestrator's state page counted two as ended, yet at 06:20 all five seats still answered their usage readings normally.
    • The seat the orchestrator ran on could end as early as 01:25 on the 16th. The plan was to move off it if it was not signed in again by evening, and the outcome is not in the pack.
  • Burn. The nightly account check gave no reading. The last reading, at 06:03, had all five seats ahead of an even pace for their week, by 4.8 to 24.5 points.
  • What the orchestrator costs.
    • An end-of-day observer pass re-surfaced a 12 November audit. Its own reviewer narrowed the claim to this: for 5–12 November, about 84.5% of weighted Claude usage went to orchestration and its supporting machinery, including about 283 of 577 meter points to hand-off helpers.
    • That week saw 95 hand-offs and 9 deliveries, against 25 and 71 for 15–22 October. This establishes a shift in where usage went and in recorded output, without isolating the cause.
    • A measured weekly cost of the orchestrator and its deliveries is owed to the author on Monday 16 November.
  • The silence sweep's own reliability. The sweep counts all 504 open registry items as silent and all 15,386 queues in its manifest as filling with nothing draining them. The pack does not settle whether it is measuring the work or its own thresholds.
  • Source dates. A report filed under the 15th's date describes a window of under three hours on 25 October. Nothing in the pack says why, and its events are left out of this entry.

Technical detail

  • Disk-full timeline (local time):
    • 21:20: a twenty-minute intake job starts an investigator scoped read-only on the workspace. It copies a 47 GB folder into temporary space on the main disk.
    • 21:25–21:33: zero bytes free.
    • 21:30: the nightly report job and the benchmark scheduler fail.
    • 23:20: the scheduler runs clean again.
    • 23:26: the main disk has 42.7 GiB free.
    • 23:27: the report is re-run.
  • Log repair order: archive the original, repair it in place with a reviewed tool, prove it parses, then prove its reader runs. Rows written as zero bytes during the window are not recoverable.
  • Session-closing rule: close by name, once, and only sessions provably not written to since the list was made. 42 qualified and 23 did not.
  • Test-session admission:
    • No new test-running session while the main disk has under 60 GiB free.
    • From 05:00, only sessions whose tests run on the bulk drive are admitted.
    • At most five run at once, and each launches only when the processor is at 75 °C or cooler.
  • Archive, then delete: each original is removed only after its archive is proven.
  • Deploy guard:
    • The scripts refuse to run if a redirect or extra push address was set before the run.
    • They deploy only after the remote, read back, holds exactly the reviewed version.
    • Left over: the retired search index still answers from the CDN cache for up to seven days, because it was not purged.
    • Also left over: the live build stamp reads "clean_tree": false. A question about whether that is expected is out.
  • Review limits: a step that stops at its round limit installs nothing. A first stop gets one narrower attempt; a second goes to the author.
  • Landing pipeline for the 671 items: sort them into per-repository batches, landing nothing; then land each batch on a fresh review of that day's code.
  • Closure check:
    • Evidence: named files exist on disk, an empty next-action field, the closure word used, and any dated marker.
    • Three prose questions are judged in one batched call to gpt-5.6-luna over all 353 candidates: does the item claim its own work is finished, is the claim dated, and is it contradicted.
    • The verdict is a table in code. Confirmed requires a dated claim, on-disk corroboration and no contradiction; everything else goes to the author.
  • Silence sweep: "silent" means no status file, decision record or completion verdict has mentioned the item for strictly more than 72 hours. Items with no signal ever fall back to a date scraped from their own text, so their ages are upper bounds.
  • Standing limits still in force: no new session while 85 or more Claude processes are running. The one-command re-seat tool stays held after three failed reviews.
  • Rulings in the pack's list: these are dated in UTC. The morning report calls the session-closing answer "23:45", so the first three fall late on the 14th in local time. The list is truncated after the last one shown.
    • 06:44 UTC: guard the deploy against redirects and extra push addresses.
    • 06:45 UTC: for the complexity ratchet, "an unexplained raise blocks review".
    • 06:45 UTC: close the 65 finished sessions once, by name.
    • 13:37 UTC: for the relay, install the finished listener fix, then one fresh review.

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