Effort, one generation at a time, and why we removed it
How muxcode once let Clef pick a seat's reasoning effort before every model call, what the live runs showed, and why muxcode 1.0.12 removed it.
At a glance
- Until muxcode 1.0.12, Clef could move a seat's reasoning effort before each model call, inside a band the host set.
- In a live comparison on 19 October 2026, 164 effort readings moved no level, and runs with Clef took 27% longer at the median.
- muxcode 1.0.12 removed the effort reading. Every seat now runs at the effort you set for it.
- Clef still reads the gate, the bypass guard and the rest, on any plan with your own Cloudflare token or through mux on Pro, Pro+ or a Pro trial.
Until muxcode 1.0.12, Clef, the classifier, could move a seat’s reasoning effort before every model call in a workflow, not once per task. In the two early live runs whose readings we kept, and in the comparison of 19 October 2026, it moved no level. So we removed it.
A seat is a role with its own model. The Architect plans, reviews its own plan and replans; until 1.0.12 it also judged. The worker writes the change. The retry takes over after a rejection or failed check. A roster is a workflow’s set of seats; muxcode ships a default. Plan first, then prove it walks through the seats in order.
A generation is one model call. Reasoning effort is how hard the model thinks: low, medium, high, xhigh or max.
pi, the open-source coding agent muxcode runs, does the work. Clef is Cloudflare’s decision model, then run as clef-flash on Workers AI and from muxcode 1.0.12 as the 27B @cf/cloudflare/clef: it answers typed questions with a probability, or a choice and its confidence. The host, muxcode’s own code around pi, sets the bar an answer must reach and applies the answer. The readings from September below were taken when the classifier was Jev, TypeSafe’s System One model.
The checkpoint
Before each generation, once any tool results were in, Clef read a bounded view of the seat’s work, never its private reasoning. The view held the seat’s prompt, new input and public replies, the last 6 tool calls, a count of new tool failures, and which context window the seat is on.
A bounded view. The prompt, including any contract (the Architect’s written terms for the change), counted up to 12,000 characters. Each of the last 6 tool results is cut to 4,000 characters from its start and end.
A context window is everything the model reads on one request. The host trimmed the oldest entries to fit 60,000 bytes and redacted secrets.
Two questions, one bar
Clef answered two separate questions: which level the next generation needed, and the lease. A lease was how many generations that level should hold: 1, 2, 5 or 10, counting the next.
Each answer counted only at a confidence of 0.85 or more, out of 1. Below that, the level stayed put and the lease was 1 generation.
During a lease the checkpoint took no reading. New input, a failed tool call, a model change or a level set elsewhere ended it early. So did a fresh context window, which pi opens when one fills, starting from a handoff, a short record of the task.
An example, not a recorded run. Each amber dot is a reading; its lease covers the generation it was read for and the ones after, with no reading. In both live runs every lease came back as 1.
Floors and ceilings
Clef moved a seat only inside a band the host set. A judge never dropped below its starting level, or high if that was unknown: nothing after the judge catches an accept it got wrong. A worker never dropped below its pin, the level the roster started it at. A low pin saves effort, and the checks, the judge and the retry catch an attempt that reasoned too little; that part still holds, with the pin now the seat’s only level.
A ceiling, “Up to” in Settings, capped the band and made the pin the floor on every step. The host offered Clef only the levels between, and never asked about a one-level band. muxcode 1.0.12 removed “Up to”; a roster saved with one loads its pinned levels.
When Clef was slow or silent
A missing reading left the level where it was. Each effort reading waited up to 2.5 seconds per attempt, with one retry. After two misses in a row, the host skipped that seat’s next 5 checkpoints, then asked again. Each pause that ended in another miss doubled, up to 40 checkpoints. An answer reset the pause to 5.
Two misses, then a pause that doubles. Each attempt waits up to 2.5 seconds, with one retry. After two misses in a row the seat rests 5 checkpoints, then 10, 20 and 40, and stays at 40 while Clef stays silent.
In QA on 2026-09-23, 17 of about 65 live Jev calls missed the 2.5-second wait, 12 of them effort readings. The next day we gave the once-per-run questions 8 seconds; effort readings kept 2.5, since each holds a model call back. In the comparison of 19 October 2026, the 164 effort readings added 235 seconds of recorded wait.
The idea, and the dials
We based the checkpoint on Ares, by Jingbo Yang and colleagues. Its lightweight router predicts the lowest reasoning level each agent step needs, from the history so far. The authors report up to 52.7% fewer reasoning tokens than fixed high effort, with minimal loss in task success.
In other tools, a person chooses. Amp’s dial, since 2026-07-09, offers low, medium, high and ultra. Claude Code’s opusplan plans on Opus and executes on Sonnet. Claude’s adaptive reasoning varies thinking per step, but within the effort level a person set.
What the shipped roster did
On the roster shipped before 1.0.12, Clef read two seats and never took either below where it started. The Architect, Claude Opus 5.5, was fixed at high, a one-level band, so Clef never read it. The worker, also Opus 5.5, ran medium up to xhigh. The retry, Claude Fable 5.1, ran xhigh up to max. Since 1.0.12 the worker runs at medium and the retry at xhigh, with no band.
Clef could only add effort. Each band started at the seat’s pin, its floor, and stopped at its “Up to”. The Architect’s band was one level, so Clef was never asked about it.
Here Clef could only add effort above the pin, the reverse of Ares, which cuts it. Only the medium pin we set by hand saved anything against fixed high, and it still does.
In a QA run in mux’s cloud sandbox on 2026-09-23, on a test roster, Jev gave 8 level answers at 0.23 to 0.79 confidence. None cleared 0.85, so no level moved, and every lease came back as 1 generation.
The first live run with the shipped Architect and worker, filed 2026-09-24, had 4 worker readings from Jev. Each chose medium, its current level, at 0.92 to 0.98 confidence, in 208 to 252 ms. Every lease was 1 generation, at 0.36 to 0.61 confidence. Neither run reached the retry, so its band has no live reading yet.
Two small runs, and only the ends of each range kept. The worker readings cleared the bar but chose the level the worker was already on, so nothing moved. The runs used different rosters, and neither reached the retry.
The comparison of 19 October 2026 measured it: no level moved in 164 readings, and runs with Clef used 1.3% more model tokens. Its timing overlapped another benchmark on the same Mac, so the 27% longer median is an observation, not a clean cost.
What you can change
You set each seat’s model and effort in Settings → Agents, and the seat keeps them for the whole attempt. There is one Worker row, with no “Up to” and no list of workers to pick from. Clef still reads a workflow run elsewhere: intake, the contract screen, a worker that drifts or stalls, and the gate. How Clef works lists every reading.
Whose key pays
muxcode decides who answers each Clef call, checking in order. First, your own Cloudflare API token and account ID, set in Settings → Classifier: the call goes to Workers AI directly, on any plan, and spends none of your plan’s tokens. Next, a spent token window turns Clef off until it resets. Next, from muxcode 1.0.5, Clef runs through mux on Pro, Pro+ or a Pro trial; in a team workspace, your own plan counts. Only a debug build, which muxcode’s developers run, falls back to a token built into it; the download never does.
Without any of these, muxcode makes no Clef call, and the code’s own checks decide.
The first yes decides. Your own token comes before any window or plan, so a spent window never stops it.
Token windows
Through mux, Clef counts the input tokens it reads, per person, across all your workspaces: 280,000 a UTC day and 100,000 per five-hour window on Pro and the trial, ten times that on Pro+. Hobby includes none. The day resets at midnight UTC; the five-hour windows are fixed, and run apart from the day. A spent window pauses Clef only: a trial keeps running.
Two limits, both in force. Each window has its own scale. Pro+ gets ten times Pro’s tokens in both.
The working where each number comes from
Each number is set in muxcode 0.12.7’s code, published 2026-09-27 and read on 2026-09-28, and read again in muxcode 1.0.5 on 2026-10-06, unless noted.
- 10 of 12, 164 and 10 readings, 1.3% and 27%, 235 seconds: the live comparison of 19 October 2026 in muxcode’s benchmark post, on build
c086886f, with every seat on GPT-6 Astra; recorded before the removal and not rerun since. The removal is muxcode 1.0.12 (CHANGELOG/1.070.md). - 17 of about 65 calls: timeouts in the 2026-09-23 QA report; we didn’t keep its log.
- The test roster: GLM 5.3 planned and judged, and DeepSeek V4 Flash worked.
- The retry’s band: moved up one level on 2026-09-24, after the live run.
- 52.7%: the Ares paper’s best case, submitted 2026-03-09.
- Leases, the checkpoint view, the misses and pauses:
LEASES,TASK_CHARS,RECENT_CALLS,CALL_CHARS,STATE_BYTES,QUIET_READINGS,QUIET_RESTandQUIET_REST_MOSTineffort-runtime.mjs; the 2.5-second wait inworkflow/intake.rs. Read again at muxcode commit 08c9032 (2026-10-06) on 2026-10-07. - Whose key pays:
key_routeinworkflow/intake.rs, muxcode commit 08c9032 (2026-10-06), read on 2026-10-07. The stand-in model answers only low-stakes questions (set_stakes), and the effort question is not one. - Token windows:
PLAN_METERSin mux’spackages/types/src/index.ts, and the Clef guide on the docs site, read on 2026-10-07. A live trial gets Pro’s windows.