Dowel carries no record of earlier work in a long session, and widening the prefill fights the prefix cost #1034

New issue

Open

opened 2026-08-19 03:24:31 +00:00 by coilyco-ops · 0 comments

coilyco-ops commented

2026-08-19 03:24:31 +00:00

Member

Filed by Saiya (exec seat), 2026-08-19, at Kai's direction. Evidence from the owl.glass export.

Kai's read is that this is Discord prefill and the window can simply be widened. The prefill part is right. The "simply" is the thing this issue exists to flag, because the two obvious levers pull against each other.

Observed

During the 2026-08-18 multi-agent trial, Dowel could not report its own status. At 03:56Z:

finished: nothing I can truthfully report. this instance carries no record of earlier session work, so I won't fill the line with invention.

Again at 03:58Z:

this session has no records of the expo work and my scratchpad is empty, so I can't give real finished/open status and won't invent one.

The contrast is alpha at 01:44Z:

I have the durable session history from that kickoff onward, including the Expo run. I can derive any counters from it.

Dowel is honest about the gap rather than confabulating, which is correct behaviour and worth keeping. But an agent that cannot recall the exercise cannot hold a slice of it, so this compounds the shared-write-surface problem rather than sitting beside it.

The tension to decide, not to resolve here

History is passed as prefill, bounded by MaxContextMessages on the lane definition. Widening it is a config change.

But a wider window is a bigger fixed prefix, and prefix size is the surviving mechanical explanation of the 39.7s median turn, tracked at #1002. Kai's own words on that median: "40s of silence for simple 1 tool questions is basically fatal."

So the two changes point in opposite directions:

Widen the prefill and the agent remembers more, and every turn gets slower.
Trim the prefix and turns get faster, and the agent remembers less.

Deciding either one in isolation will silently undo the other. They want one owner and one decision.

What would settle it

#1002's first task is already the right measurement: split one median turn into model time, tool time, and harness time. That number tells you what a wider window actually costs per turn, which is the input this decision needs and nobody currently has.

Acceptance

A stated prefill window for the Dowel lane, with the measured per-turn cost of that window recorded next to it.

#1002 - the latency question and the prefix arithmetic.
The shared-write-surface issue filed alongside this, which this compounds.

**Filed by Saiya (exec seat), 2026-08-19, at Kai's direction. Evidence from the owl.glass export.** Kai's read is that this is Discord prefill and the window can simply be widened. **The prefill part is right. The "simply" is the thing this issue exists to flag**, because the two obvious levers pull against each other. ## Observed During the 2026-08-18 multi-agent trial, Dowel could not report its own status. At 03:56Z: > finished: nothing I can truthfully report. **this instance carries no record of earlier session work**, so I won't fill the line with invention. Again at 03:58Z: > this session has no records of the expo work and my scratchpad is empty, so I can't give real finished/open status and won't invent one. The contrast is alpha at 01:44Z: > I have the **durable session history** from that kickoff onward, including the Expo run. I can derive any counters from it. Dowel is honest about the gap rather than confabulating, which is correct behaviour and worth keeping. But an agent that cannot recall the exercise cannot hold a slice of it, so this compounds the shared-write-surface problem rather than sitting beside it. ## The tension to decide, not to resolve here History is passed as prefill, bounded by `MaxContextMessages` on the lane definition. Widening it is a config change. **But a wider window is a bigger fixed prefix, and prefix size is the surviving mechanical explanation of the 39.7s median turn**, tracked at #1002. Kai's own words on that median: "40s of silence for simple 1 tool questions is basically fatal." So the two changes point in opposite directions: * Widen the prefill and the agent remembers more, and every turn gets slower. * Trim the prefix and turns get faster, and the agent remembers less. Deciding either one in isolation will silently undo the other. **They want one owner and one decision.** ## What would settle it #1002's first task is already the right measurement: split one median turn into model time, tool time, and harness time. That number tells you what a wider window actually costs per turn, which is the input this decision needs and nobody currently has. ## Acceptance A stated prefill window for the Dowel lane, with the measured per-turn cost of that window recorded next to it. ## Related * #1002 - the latency question and the prefix arithmetic. * The shared-write-surface issue filed alongside this, which this compounds.

coilyco-ops added the

labels

2026-08-19 03:24:31 +00:00