2026-08-18: The Prewarm That Ate Its Own Tail

Pencil sketch on a dark ground: one long ribbon of tractor-feed printout, sprocket holes down both edges and rows of illegible marks across its panels, curves into a closed loop and ends in a serpent's head whose open jaws bite down on its own tail — a small teal halo glowing at exactly the point where the record swallows itself

For eighteen days the code seat answered every real question with a refusal, and nobody heard it. Qui-Gon — OpenClaw’s coding brain, routed through proxyd’s council-code alias into cathedral-Devstral on :3301 — returned HTTP 413 to roughly fifteen hundred turns a day, from 2026-08-01 to 2026-08-18. Its /health endpoint returned 200 the entire time. Every canary passed. The canaries send 544-token prompts, and only large prompts were failing, so the seat looked immaculate from every angle anyone had thought to look.

Two medians, one log file

The outage was not found. It surfaced sideways, while someone was measuring whether a quantized-KV cache deploy would pay off on that seat — the same stopwatch discipline that had parked the native-MTP port the day before.

Two views of the same log disagreed, badly. Filtering for request completed lines gave a median prompt of 322 tokens: small, healthy, boring. Grepping the raw prompt_tokens= field gave about 49,684. The gap between those two numbers was 21,389 lines that had never completed anything, because they were rejections. A filter is a claim about which lines mean what you think they mean. That one had quietly excluded every line that mattered.

The remedy that became the disease

devstral-prewarm-sentinel.sh was not a blunder. It existed to solve a genuinely nasty problem: Qui-Gon’s roughly 19,000-token system prompt cold-prefills for three to four minutes, and OpenClaw’s idle watchdog gives up at a hard 120 seconds. The arithmetic is merciless. The first turn after any restart could never succeed.

So the sentinel fires a sacrificial turn on a timer — one turn whose failure is expected, discarded, and useful only for the warm cache it leaves behind. Its own header calls that turn throwaway.

Its arguments said otherwise. The script passed --session-id devstral-prewarm, and a named session in OpenClaw is persistent and resumed, not thrown away. Every fire therefore appended about 1.3 KB of [assistant turn failed before producing content] to a transcript that the next fire would faithfully re-send as its prompt.

The ratchet

From there the shape draws itself. A dead seat serves nothing, so it sits idle. Idle for twenty-five minutes is precisely the sentinel’s trigger, so it fires. The fire fails, appends its failure, and the prompt grows. A larger prompt fails harder and keeps the seat idle, which guarantees the next trigger.

compactionSummary went from 2,141 bytes to 16,907 in under four hours. The session file reached 28 MB; its trajectory, 10 MB.

What 52,887 tokens actually were

Nobody had to guess at the composition. It was measurable, so it was measured.

Component	Tokens
Accumulated failure transcript	34,300
System prompt	12,826
Tool schemas	5,761
Total	52,887

Two thirds of the request — 34,300 tokens, 64.9 percent — was the seat’s own failure transcript.

That arithmetic killed two comfortable theories on the spot. Skills were not accumulating: progressive disclosure keeps SKILL.md bodies out of the prompt entirely, which is the whole point of it. Tool schemas had not bloated either — they shrank. Across twenty-four days the fixed part of the prompt got 145 tokens smaller. Every byte of growth came from one place, and that place was the seat quoting its own failures back to itself.

The curve is a straight line into a wall. On 07-25 the prompt measured 20,069 tokens. On 08-01 it hit 33,377 and crossed the 32,768 cap. By 08-11 it stood at 52,970, and there it plateaued, because a request that is already refused has no room left to grow.

Version one said the outage was six days old. The log had simply never rotated, and six days was the age of the oldest line anyone had bothered to read; the true onset was 08-01, eighteen days. The second correction was the dangerous one, because it arrived dressed as evidence. “Maximum completed prompt: 4,011 tokens” looked like proof the seat had never served a large turn in its life. It was a filtering artifact — request completed is emitted only by the non-streaming path, so every streaming turn was invisible to it. Measured properly, the seat had been serving real work at 19,000 to 23,000 tokens, around a hundred turns a day, until 08-04, when it collapsed. The rejections had begun three days before that, on 08-01, the day the prompt first crossed the cap. Same verdict, sounder evidence, and a precise date of death in place of a vague one.

The fix, and the seatbelt already in the drawer

The repair costs nothing and takes no memory: purge the session files before each fire. The prewarm then sends exactly the 18,600-token system prompt it exists to warm, and nothing else. Verified live on the running seat: the prompt fell from 52,887 tokens to 18,588, and the 1,775 rejections logged earlier that day became zero after. No GiB spent, no restart, no cap raised.

The second half of the fix had already been invented and never fitted. council-code and council-devstral carried no max_prompt_tokens at all, while council-heartbeat has carried one since its own oversized-caller incident — with a comment in the config describing this exact bug class, in advance, in writing. A seatbelt bolted into one seat and left out of the others. max_prompt_tokens: 32000 now sits on the code seats too.

One more rule went in that should never have needed writing down. A 413 is deterministic: the same prompt is too large forever, and retrying it is not resilience, it is noise. proxyd was retrying about seventeen times per real turn, amplifying a refusal into a load. Commit cd0ef69 in sanctum-config carries all of it.

Why the cap was the wrong lever

The tempting fix is always the cap, and the arithmetic is what refuses it. Lifting the ceiling to 53k costs an extra 6.25 GiB of KV pool on a Mini with 3.10 GiB free. Even in a world where that memory existed, what it buys is the privilege of prefilling 34,300 tokens of failure text on every single turn — paying full price, at full speed, to read the seat its own obituary.

Alive is not the same as serving

Every probe pointed at that seat asked one question: is it up? /health answered honestly. The process was running, the weights were loaded, the socket accepted. Not one probe asked whether it was serving anybody — the haus has walked into this exact room before, the day Force Flow’s health endpoint said OK for twenty-five hours over a dead enforcement loop.

A rejection-rate gate would have fired on 08-01, seventeen days before a human noticed. A prompt-growth tripwire would have fired around 07-28, before the first rejection existed to be counted. That sentinel is being built, and it is the only part of this episode that is new work rather than restored work.

Because the fix does not make Qui-Gon healthy. It makes it July again. The cold start underneath is exactly as real as it ever was: nineteen thousand tokens still take three to four minutes to prefill, the watchdog still quits at 120 seconds, and the haus still needs a sacrificial turn to bridge the two. What changed is that its failures finally have somewhere to go. A remedy designed to fail needs a place to put the failing — otherwise it finds one on its own, and the place it finds is the prompt.