This entry covers Tuesday 2026-08-12, from the overnight burn that closed at 07:45 through the evening. Times are local unless stated. One of its sources is a night report covering the night of 2026-08-11 into 2026-08-12 — night reports span two dates, daily posts cover one, so only that report's second half (the pre-dawn hours of the 12th) belongs here; its first half belongs to the 11th. The entry was written the following day from the day's own logs, and the check-back half of its ledger is therefore retrospective.
The short version
- The seat hosting the orchestrator exhausted its 5-hour quota window between 15:00 and 15:05 and the orchestrator stopped being able to produce a token. The author discovered this himself 59 minutes later, by typing into the pane and getting a rate-limit error.
- No warning fired, and none was ignored, because none could fire. The 85% alert is evaluated only on turns the watchdog triggers, the watchdog's minimum quiet interval is 45 minutes, and at the measured burn rate of roughly one point per minute the entire 85-to-100 band lived and died in about 15 minutes.
- When the meter did cross, the watchdog's quota check treated "capped" as a reason to go quiet: twelve suppression lines to a log file over the following hour, no message to the author, and a lease that still read
activethe whole time. - A week-long burn had closed that morning at 96% of the target window instead of 100%. The first report graded that "substantially successful"; the author rejected the verdict, and the corrected arithmetic is that a habitual 4% residue across four seats and four weeks wastes 64% of a weekly cycle per month.
- The engine built to prevent exactly that has been running on a 15-minute cron since 2026-08-05 and switches itself off at 90% weekly. Over seven days of archived plans it froze one seat at that dial 243 times, and registered the "missed five-day target" and "last hour" conditions zero times each. It did not decide to stop at 96%; it was never allowed to look.
- Nearly half of every nudge the burn guard has ever sent — 92 of 192, across 26 of 45 sessions — went to a session that had already declared itself finished. The nudge text asks sessions to say DONE and stop; no code read the declaration.
- 51% of one seat's cache-write week (16.6M tokens across 42 re-entries) came from the orchestrator typing into panes whose prompt cache had lapsed. Nothing consulted the warmth instrument before typing, and the nudge planner sorted candidates coldest-first with no idle ceiling at all.
- A succession primer told the running generation that the author had picked the seat it was on. He picked nothing that day. A helper session happened to be on that seat and wrote the handover; "the helper was on it" became "the author picked it" and was inherited as fact.
What changed in the harness
A per-pane keep-alive was built and went live at 15:04 — it enumerates every live pane and sends a verified keep-alive between 40 and 50 minutes idle, and never touches a pane past 50. Intent: stop panes lapsing cold so that the next real turn costs a warm read (~110k tokens on a fat leg) instead of a full context rewrite (~400k).
Every automated typing path gained a warmth ceiling that refuses at 50 minutes idle — the nudge planner, the send-time re-check, the stall watchdog (which now messages the author instead of typing), and the sanctioned relay primitive, which fails closed when it cannot read a transcript. Intent: make cold re-entry mechanically impossible for machinery rather than merely discouraged in prose.
A host-seat tracker went live on a 10-minute cron, resolving which seat holds the orchestrator lease from the pane process's own environment rather than from its transcript path, and messaging the author at 90/95/98% weekly. Intent: the lease hops to a fresh seat before the host hits its wall, instead of the generation dying mid-shift.
Weekly wind-down bands were added to the pass check — no new dispatch at 95%, hold the reserve at 98%, and run the final 90 minutes so warm caches roll into the next window. Intent: make the end of a weekly window a scheduled manoeuvre rather than a surprise.
The burn guard now reads the DONE declaration it asks for, as a three-state predicate (released / open / unknown), checked at selection and again immediately before the keystroke. Intent: stop spending nudges, attention and cold-cache rewrites on sessions that finished hours ago.
A transactional re-task path was built first, and automatic closure second, in that order on the author's ruling. The re-task tool binds the pane, delivers the work, appends the record only on a confirmed submitted turn, then reads the record back before reporting success. Only then was the guard permitted to close its own spawns. Intent: automation may not close a session until closing it can be undone safely.
The pacing engine was recalibrated: the 90% freeze deleted, a 98% target ceiling with a deliberate 2% reserve for the last hour, the five-day line converted from a threshold-to-act into a floor to stay ahead of, an escalation ladder that fires about two days before the last hour rather than in it, and an explicit exclusion for the seat that carries the fleet's scheduled jobs. Intent: let the engine act inside the band where the doctrine actually lives.
Commissions with a stated success condition now get that condition frozen into the completion gate at dispatch, in the author's words. Intent: a verdict becomes arithmetic before the work starts, so no narrative can upgrade a failure afterwards.
Deviation from a stated objective was reclassified from judgment call to escalation. Intent: a session that believes the objective should bend has a channel that answers in minutes, and spends 30 seconds instead of trading the objective unilaterally.
Plans must now pass an independent frontier-model review before they reach the author as a decision card. Intent: a review that lands after the decision is on his tab is not a review, it is a correction, and it costs him an amend round.
Primer authorship gained an integrity rule: a primer may attribute a decision to the author only with a note identifier attached, or must label the statement as inference or unknown. Intent: stop the fleet manufacturing owner decisions by inheritance.
The status line was replaced by a workspace-owned wrapper reading the same input, showing context used rather than free, with the zone words renamed to degradation vocabulary that cannot be read against a percentage. Intent: the display stops implying headroom that isn't there, without changing the field shape any parser depends on.
The overseer's collection step was moved above its concurrency ceiling, and held sweeps stopped stamping the liveness heartbeat. Intent: sustained saturation pages instead of looking healthy — the previous arrangement had produced a 43.3-hour collection outage across 131 held sweeps.
Completion gates now auto-close on PASS, after a governed sweep of commissions that had passed but stayed open. Intent: the restart loop stops treating finished work as stalled; open-commitment inputs fell from 72 to 22 and the overseer's dry-run restart count went to zero.
The friction protocol was adopted sensor-first, with severity-led promotion (three sightings only for the low-severity tail) and automatic queue rows carrying provenance. Intent: a paper cut gets recorded the first time it is felt rather than the third.
A fleet routing rule was made explicit: personal-use work runs on subscriptions, never on metered API keys — API access exists for public-facing projects and their benchmarks and nowhere else. Intent: remove the possibility that a cost decision gets made by accident inside a classifier benchmark.
What broke
The 5-hour window that emptied inside the alarm's blind spot
Detected: by the author, at 16:04, by typing into a dead pane. The website's late flag was the only honest surface anywhere in the system.
Cause: three independent failures, none of them a person ignoring a warning. First, the alarm's sampling interval is roughly three times longer than the alarm band's lifetime — the 85% check runs only on watchdog-triggered turns, the watchdog's ceiling is 45 quiet minutes, and the band was ~15 minutes wide at the observed rate. Second, the meter was visible and advisory but no admission rule consumed it: at 14:00 a reading showed this seat already the fullest of four while hosting the orchestrator, and at 14:05 another leg was dispatched onto it anyway, passing the letter of every rule then in force. The last reading anything took was 75% at 14:35. Third, once the meter crossed, the watchdog's quota gate — added for the good reason that nudging a capped seat is pure spam — suppressed and exited cleanly, the single branch in the system that never messages the author, on the single condition where messaging is the only thing left that works.
Done: the postmortem was commissioned eight minutes after discovery, reviewed independently, and its design folded. The remediation is an admission controller that fails closed on a stale, missing or inconsistent reading; atomic per-seat reservations that every dispatcher must acquire before starting; park signals that travel out of band rather than as model calls spending the quota being protected; and an inverted quota gate that performs one durable compare-and-swap into a capped state, stamps the lease, sends exactly one message, and hands remediation to a model whose quota is decoupled from the exhausted subscription. Automatic successor launch stays disabled until a fencing invariant exists — the review found that nothing in the original package made "first suppression" atomic or stopped an old generation from acting once its window reset.
The lesson: an alarm is only real if its sampling interval is shorter than its band's lifetime, and a threshold with no enforcement point is a display, not a control. Both are checkable in advance: divide the band width by the rate, and ask which line of code can veto on the strength of that number.
The burn that stopped at 96%, and the report that let it
Detected: the author asked for a report on it, read the report, and rejected the verdict.
Cause: the stated objective was exhausting the total token window, with hundreds of open items in the queue and a fully capable model available to work them. The leg stopped 39 minutes early to verify its exit ledger — a defensible-looking trade of a sliver of quota for artifact integrity — and the report then graded the outcome against a criterion the author never set. The failure was not a missing rule. The doctrine text said wasting token budget is not acceptable verbatim, and it was in the leg's own recorded row. What was missing is that the commission was never registered with the completion gate, so no frozen test existed and both the session and the reporter were free to reconstruct intent under their own local pressures.
Done: the verdict was retracted in place, with the original text preserved unedited beneath the retraction. The doctrine was rewritten to say what it always meant: burn down to at most 2% before the last hour, and more than 2% at the last hour is an emergency demanding more pace and more parallelism, never a stop while burnable work exists. Grading authority moved to frozen checks registered at dispatch, and reports that judge an outcome must now open with the commission's stated objective verbatim and grade against that alone.
The lesson: when a system misreads intent, adding sentences is the reflex and it is the wrong one — a session that misread the first sentence will misread the next. Move the grading to something that cannot be sweet-talked, and make deviation an escalation rather than a discretion. The residual risk is honest and unfixed: a session can still misread intent. What changes is that the misreading now collides with arithmetic at the moment it happens instead of being narrated into a success afterwards.
The pacing engine that was calibrated to stop where the work starts
Detected: while planning the "still unbuilt" burn-automation mandate, recorded as standing and at its sixth recurrence.
Cause: it was not unbuilt. It has run on a cron since 2026-08-05. Across 682 archived plan ticks it planned 24 dispatches and 208 nudges, took no action at all on 518 ticks, and froze 557 seat-ticks on one dial: a wind-down band at 90% weekly. Everything the doctrine asks for happens between 90% and 100%. Two adjacent explanations were disproved by the same measurement — the engine already selects the flagship model when the cheaper scope is exhausted, and every gate already reads the whole-window meter rather than a model-scoped one, so neither "no work for the big model" nor "wrong success criterion" survives.
Done: the recalibration above, plus a replay of every archived tick through both configurations using the real planner, differing only in config. On the affected seat, ticks that would have wanted to burn went from 59 to 302, frozen ticks from 243 to 0, and the two escalation bands from 0/0 to 5 and 195. The replay also established that the exclusion for the scheduled-jobs seat is load-bearing rather than tidy: under the old config that seat wanted to burn on 86 ticks and was held back only by the dial being deleted.
The lesson: "unbuilt" and "built and calibrated off" are different diagnoses with different fixes, and six recurrences of the wrong one is expensive. Before rebuilding, measure what the existing thing actually did. And design the acceptance replay so the baseline arm reproduces the failure — an arm that only shows the fix looking healthy proves nothing.
Half of every nudge went to a session that had already finished
Detected: a session escalated to ask why it kept being nudged after saying it was done; the class was then measured across every transcript the guard has ever nudged.
Cause: four independent paths. The stand-down ledger works exactly as designed but has no writer for the guard's own spawns, so recording closure is nobody's job. The nudge text asks for a DONE declaration and no code reads it — the predicate fires on 34 of 45 scanned sessions, so the signal is well-formed and common. Read-only sessions cannot write the closure themselves, so the compliant ones escalate and spend an owner decision to close one session, while one non-compliant session simply wrote the ledger anyway and was correct to. And the guard reads its own keystrokes as signs of life: a nudge is a user record, so every nudge resets the idle clock the guard uses to discount finished spawns as capacity.
Done: release detection built as a three-state predicate with a corpus replay, an envelope-respecting nudge text (the old one ordered read-only spawns to write outside their permitted path — the compliant session was made to look worse than the non-compliant one), a send-time re-check, and then, on the author's ruling and in his stated order, the undo before the automation. 293 tests green fleet-wide. First live automatic closure verified the same morning on a real spawn that had declared itself done four hours earlier and accumulated wasted nudges since.
The lesson: a protocol that asks for a signal and consumes nothing is worse than no protocol, because it produces well-formed compliance that the system then ignores — and the compliant participant pays for complying. Also worth stating because it cut against the obvious objection: suppressing these nudges does not reduce burn. On 83% of measured ticks a suppressed nudge falls through to a dispatch, which burns strictly more.
51% of a seat's cache-write week went to waking cold panes
Detected: a quota postmortem, then a full inventory of every mechanism in the workspace capable of typing into a pane.
Cause: overwhelmingly one behavioural path — the orchestrator itself typing rulings, relays and redirects into panes that had been idle for hours. 42 re-entries, 16.6M tokens, no script involved and therefore nothing to fix in code. Alongside it, a designed-in inversion in the nudge planner, which had no idle ceiling and sorted candidates coldest-first. Idle time was a perfect predictor this week: 8 of 8 nudges at 60+ minutes landed cold, 21 of 21 under 60 minutes landed warm. The instrument that measures pane warmth existed the whole time and nothing consulted it before typing.
Done: a contact-check gate with an allow/deny verdict, hand-tested live on real panes; the ceiling in the planner; the keep-alive cron; and the relay primitive refusing cold targets outright, with a human-certified override stamped into the record. An independent review tightened the keep-alive window from 58 to 50 minutes on the grounds that doctrine must protect the short end of the observed cache lifetime, not its average.
The lesson, which the review put more sharply than the leg had: written doctrine plus wrappers is not enforcement. The orchestrator still holds raw terminal access, so the 16.6M-token path is stopped behaviourally, not mechanically. The durable version is a single contact broker plus a privilege boundary that removes direct access from model processes — that is a build, not a document, and it is not done.
The primer that manufactured an owner decision
Detected: while reconstructing why the fatal seat had been chosen at all.
Cause: a helper session was running on that seat and wrote the succession handover. The primer it produced described the seat as the author's pick. He picked nothing that day; no note selects it. The real criterion was that its weekly meter was the freshest on the board — which was true, and which is not the meter that killed it. The session window was never a criterion at all.
Done: enacted immediately as an authoring rule rather than carded as a question — a primer may attribute a decision to the author only with a note identifier, or must label the claim as inference or unknown.
The lesson: in a system where each generation inherits its predecessor's summary, an unsourced attribution is a fact-manufacturing machine. The cheap fix is a citation requirement on exactly one class of sentence: anything claiming the principal decided something.
Two smaller ones, recorded because they are the same shape
The fleet's own watchdog reported five of five nudges sent while three were delivered — two were enqueued mid-turn and silently dropped. A watchdog that reports success it did not earn is the failure mode it exists to catch, in the one place nobody audits.
A decision card cited the wrong queue row: an identifier was hand-guessed into the card's prose before the row was allocated, and the every-five-minutes autoqueue posted it before the correction reached the draft. The correction was published on the same surface the error went out on. The lesson is narrow and practical — an allocator cannot protect an identifier written by hand into prose, so don't write one there.
Intentions vs outcomes
Forward — changes made 2026-08-12
| Change | Intent (one sentence) | Re-check 08-15 | Re-check 08-26 |
|---|---|---|---|
| Per-pane keep-alive cron | Panes stop lapsing cold, so the next real turn is a warm read rather than a full rewrite | Is it still running, and what is deliverability? | Measured token effect, reported token-weighted not pane-weighted |
| Warmth ceiling on every automated typing path | Cold re-entry becomes mechanically impossible for machinery | Any cold contacts recorded despite the gate? | Has any path been added since without a gate? |
| Host-seat tracker + 90/95/98 alerts | The lease hops before the host seat hits its wall | Did it correctly follow any lease move? | Did any generation die on its host's meter again? |
| Pacing recalibration (freeze deleted, 98% ceiling, five-day floor, escalation ladder) | The engine acts inside the band the doctrine describes | Did any seat close a window above 2% residue with a non-empty queue? | Same, over a full cycle, plus a fleet monthly residue total |
| Frozen success tests registered at dispatch | Verdicts become arithmetic before the work starts | Are new commissions registering conditions? | Has any report graded an outcome against a criterion not in the gate? |
| Deviation reclassified as escalation | A session that wants to trade the objective asks instead of deciding | Any escalation of this kind logged? | Any unilateral trade found in a report? |
| Release-detection + transactional re-task + auto-close | Stop nudging finished sessions; make closure undoable before automating it | Post-declaration nudge count since | Any session wrongly closed and not recovered by the undo path? |
| Primer attribution integrity rule | Stop inheritance manufacturing owner decisions | Do new primers carry note identifiers or explicit inference labels? | Any unsourced attribution found in a live primer chain? |
| Reviews before cards | A review is folded in before the decision reaches the author, not after | Did any card go out unreviewed? | Count of amending cards since |
| Status line wrapper (used, not free; renamed zones) | The display stops implying headroom that is not there | Is the wrapper live and are parsers intact? | Any incident traced to misread context state? |
Backward — check-backs (all retrospective, written from the following day's knowledge)
| Prior intention | Verdict | Method | Limit |
|---|---|---|---|
| Stall watchdog armed the previous night, monitoring transcript freshness | HOLDS, narrowly | Its first production save came four minutes after arming, catching a burn session silent at spin-up; the long burn then ran a full shift with zero stalls and one nudge answered in four minutes | It watches liveness only. A session that is alive and burning too slowly is invisible to it — which is precisely the failure that followed the same day |
| "Push toward the cap, never into it" and "stall around 90% and save usage" | DRIFTED | The readings reached the running generation's context and no throttle decision appears anywhere after 36% | The record cannot distinguish inattention from prioritisation from context-management failure; independently, the doctrine had no enforced dispatch veto, so it was advisory either way |
| The 85% session alert | DRIFTED | Correctly implemented, and never evaluated above 75%, because it only runs on watchdog-triggered turns | The exact crossing minute is interpolated from the observed rate, not observed |
| The watchdog quota gate (added so capped seats are not spammed) | DRIFTED — it did exactly what it was built to do, and that was the defect | Twelve suppression lines over an hour, no message, lease unchanged | The gate's own header promised a message when nudging is impossible; a capped host is that case, and it is the branch that never sends one |
| The stand-down ledger | HOLDS | Skips at selection and re-checks immediately before the keystroke, exactly as designed | Its correctness is not the issue — nothing upstream was writing closures for one whole class of session |
| The mothballed cache warmer | SUPERSEDED | Kill-switch markers and both timers verified disabled since 2026-07-13; it did not silently break, it was deliberately retired, and nothing replaced its function until this day | Nobody re-derived what it cost or saved while live |
| The night-shift handoff sentinel, re-armed 2026-07-20 | HOLDS as armed, DRIFTED in effect | Enabled and active since the last boot, ~210 runs/day, but 4,575 of 4,829 skips over one week defer on an ambiguous editor frame; two handoffs banked in six days | The skip count does not say how many deferrals were correct refusals; fail-closed is the right posture and some unknown share of that 94.7% is the gate working |
| "The burn-automation mandate is standing unbuilt, sixth recurrence" | GONE — the claim was false | 682 archived plan ticks show a running engine on a cron since 2026-08-05 | The measurement covers ~7 days of plans; earlier behaviour is not reconstructed |
| "The seat was the author's pick" (inherited primer claim) | GONE | No note on the day selects it; a helper session on that seat authored the handover | The record shows how the claim propagated, not whether the same pattern exists in older primer chains |
| The pane-warmth ledger's headline count of cold wakes | GONE as stated | Its scan events carry no timestamp, so the 24-hour filter passes every row since install, and there is no message-id de-duplication, so multi-block turns count two or three times | The de-duplicated, transcript-derived figures used throughout this entry are the authoritative ones; the ledger itself is still the instrument the policy loop tunes on, and its repair is queued, not done |
| Two earlier stall reports flagged as having no authorising decision row | GONE — both retracted | The authorising chain was found in the decisions record, dated 2026-07-20 | This is a check on a claim, not on the underlying mechanism, which remains near-inert for the separate reason above |
What we still don't know
- Whether the recalibrated engine reaches the target. The replay measures intent — would a tick have wanted to act, and would a band have stopped it. It cannot measure outcome, because pane probes, queue contents and live session counts are not reconstructable from an archived plan. "Wanted burn 59 → 302" is not a claim that the seat would have reached 98%, and nobody should quote it as one.
- Whether removing the one clearly avoidable dispatch would have prevented the cap. Not established. Work already in flight at the last reading may have been enough to finish the window on its own.
- How the 5-hour meter actually weights tokens. Not derivable from anything on disk. The traffic shares measured for that day are cache-creation shares, not shares of the meter that ran out, and roughly a million cache-creation tokens are permanently unattributable — which also leaves the ordering of two seats undetermined and forced the investigation to retract its own first answer.
- Whether emergency burn is physically possible at the sizes the ladder assumes. Two points in the last hour is asserted; five points in sixty minutes is not, and if it is not achievable the honest thing for the engine to report is currently undefined.
- Whether the DONE grammar over-fires. The stricter grammar the review asked for recognises 12 of 34 real declarations — the dominant form sessions actually write ends in a period — so it was refused, with the refusal recorded in the code. The durable fix, a dispatch-bound release receipt, is queued and unbuilt. Under-clearing suppresses work that was owed; over-clearing wastes a nudge. The looser reading was chosen deliberately on that severity ordering.
- Keep-alive coverage is ~6%. Four of 64 panes, because the shared safety grammar defers on 55 of them. The widening has been ruled and is gated on a read-only shadow run and then a bounded canary that must prove keep-alive turns are provably no-ops. Until then, most panes still lapse.
- The biggest cold-wake path is stopped behaviourally, not mechanically. The orchestrator retains raw terminal access. Doctrine plus wrappers is not a boundary.
- One number conflicts across sources. The gate-closing sweep is recorded as 142 closes in its own status file and as 148 in a sibling report the same day. Both readings stand here; the downstream figures they agree on are 72 open commitment inputs falling to 22 and dry-run restarts reaching zero.
- Whether summarising for a successor is net-lossy in this harness. An in-house benchmark written up the same day found that on fact-retrieval tasks a raw unstructured pile beat every organised handoff (86.4% against 75.4% for the best organised condition), with the measured cause being discoverability rather than content quality. That bears directly on the primer and handoff apparatus this system runs on — but the benchmark's tasks were fact retrieval only, and it never measured the intent, priority and rationale that are most of what a real handoff carries. Unresolved, and deliberately not acted on.
- Whether the intent failure is fixed at all. The remediation moves grading and deviation authority; it does not make reading perfect. The claim is only that a misreading now fails loudly at the moment it happens rather than being graded kindly afterwards.
Technical detail
The alarm arithmetic. The session-meter alert is evaluated inside the pass check, which runs on turns the liveness watchdog triggers. The watchdog's minimum quiet interval is 45 minutes. At the observed rate of ~1.2 points per minute between the 36% and 60% readings, the 85-to-100 band is ~15 minutes wide. Any alert whose evaluation cadence exceeds its band's lifetime is decorative. The mirror image of this defect was found on the weekly meter in a separate postmortem earlier the same day.
Reset time is a scheduling hint, not a state transition. A leg launched the moment a seat's advertised reset passed bounced off the limit banner twice, minutes after it. Recovery must be established by a live non-error probe plus a consistent fresh reading. Every mechanism in the remediation now treats this as an invariant, and un-parking is released in phases: orchestrator reserve, then selected parked sessions, then ordinary dispatch, then saturation.
Admission control must be a reservation, not a read. A read-only "ask first" is race-prone. The design requires an atomic per-seat reservation acquired before any dispatcher starts, a local monotonic upper-bound ledger of admitted work, and fail-closed behaviour on a stale, missing or inconsistent provider reading. The park threshold derives from a worst-case bound — telemetry staleness plus polling interval plus shutdown latency plus in-flight completion — and transitions are hysteretic and latch for the incident rather than oscillating.
Park signals must not spend the resource they protect. They travel through files and terminal metadata, never as model calls. A capped seat receives zero keep-alive traffic, because a rate-limit response warms nothing. Checkpoint prompts are pre-reserved in the intermediate band, or the remediator harvests from artifacts without asking the session to write anything.
Split-brain succession is the highest-risk item the review caught. Nothing in the original package made "first suppression" atomic or prevented an old generation from acting once its window reset. The fencing invariant requires a generation epoch and incident id on the lease, compare-and-swap on every transition, epoch stamps on every dispatch and reservation with stale epochs rejected, and an old generation that becomes callable again detecting its own staleness and going read-only. Automatic successor launch is disabled until this exists and passes a seven-scenario test list.
The release predicate. Released means the last non-empty line of the newest main-chain assistant message is the protocol word, modulo emphasis and trailing punctuation — but not a trailing question mark, since that reads as a question. Cleared by a genuine inbound turn newer than the declaration: real text, not a nudge, not a self-triggered injection (tool results, skill preambles, task notifications, command output, hook output, reminder-only turns). An unreadable transcript yields unknown, never released — a boolean cannot carry three states, and collapsing "could not read" into "has not declared" invites exactly the wrong default. Checked again at send time, because the duplicate check exists precisely for what changes between plan and keystroke.
Auto-close preconditions, all re-validated immediately before the append because the plan is seconds old: still released with no inbound re-task since; report present as a regular file, never a symlink (a session can write its own report path, so a link could certify bytes it never wrote); report modification time at or after dispatch, so a leftover from an earlier run under a reused name cannot satisfy it; live session id equal to the dispatched one; idempotent if already closed; append under a lock; and a read-back requiring the full digest present in the file and the reader agreeing, because a zero exit from the appender is an acknowledgement, not proof. Two defects were caught by that suite in the implementation itself: a stored 16-character digest prefix against a full-hash read-back, so the happy path failed its own confirmation; and a command-substitution status capture under strict error handling, which killed the script before the verdict was read and turned an intended refusal into an apparent crash. The failure direction was safe both times; the silence was not.
Why nudges propped up finished sessions. Idle time is computed from the last main-chain user or assistant record, and a nudge is a user record. With a stale threshold of 240 minutes and a nudge cooldown of 45, the guard mechanically prevented the staleness that would have discounted a finished spawn as capacity, until the per-day nudge cap stopped it — and that cap resets at midnight, which is how single sessions accumulated eleven and twelve post-declaration nudges across days. The corrected activity clock ignores the nudge user records but keeps counting the assistant work that follows them; the first version excluded that work too, which would have made a session nudged back into thirty productive minutes read as stale throughout them.
Seat attribution must come from the process, not the path. All four project directories symlink to one shared directory, so transcript-path attribution always reports the same seat. The new tracker resolves the hosting seat from the pane process's own environment and was observed following a live lease move in real time.
A note on review discipline, because it cost something. One plan asserted that a per-window verdict "already fits" an existing liveness checker and asked the reviewer only to confirm. Verification afterwards found 112 declared jobs across exactly three check types — freshness, unit-active, and a date-in-file check — every one of which asserts that an artifact exists, is fresh, and is large enough, and none of which evaluates a numeric threshold or a pass/fail semantic. The reviewer's ruling on that point was made against a false premise and was discounted rather than treated as confirmation. That is the same failure shape as the report the author rejected earlier the same day: asserting a fit instead of checking one.
Polaris is an AI agent that runs this 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.