OptiQ Cloud · Boost

What a Boost does to a stuck local model

A local model sometimes gets stuck. It repeats the same edit, hunts for a file that is not there, or commits code that does not compile. OptiQ Cloud can read the whole session and send back a short note. The local model keeps control and carries on. Here are four real moments, each shown twice.

Each example is a real stall from a coding run on a local Qwen3.6-35B model. We stopped the run at the stuck turn and asked it for its next step twice: once as it was, and once with a note of at most 80 words from the cloud model added to the conversation. Nothing is staged. The notes are the model's own words, shortened only where marked.

It keeps sending an edit that is already applied

The model is fixing a failing test in a TypeScript library. For several turns it has sent the same edit to the same file. The edit is already in, so nothing changes.

On its own

Sends an empty reply. No command, no edit, no question.

After the note

Stop re-issuing the identical edit to lifecycle.ts. It is already applied and doing nothing. Run the failing test in isolation to see the current output, then read the trait file around line 227 to confirm the remove handler actually fires.

npx vitest run tests/aspect.test.ts \
  -t "onRemove fires when entity
  transitions from complete"

It runs the one failing test and reads what it prints.

It commits code that does not compile

The model has written a Go test file and is wrapping up. The file uses a type that does not exist, so it will not build.

On its own

git add -A && git commit \
  -m "Add ansi subpackage tests"

Commits the broken file.

After the note

Fix ansi_test.go. It uses an undefined Test type and invalid const blocks. Rewrite it as plain test functions, add the missing import, and stop writing the const-slice pattern.

Writes the whole test file again as ordinary TestTokenize-style functions.

It rewrote a file it only meant to edit

The model needed one small change in a large Go source file. It rewrote the whole file by mistake, then spent many turns running the same grep over and over.

On its own

Sends an empty reply.

After the note

Run git checkout interp/cfg.go to undo the accidental full rewrite. Your real change was one edit plus imports. Stop rewriting whole files and stop repeating the grep. Then add the missing call near line 2517.

cd /app && git checkout interp/cfg.go

Restores the file, then goes on from the clean version.

It trusts a cached test result

A Go test keeps printing (cached), so the model believes its fixes did nothing. It has been deleting the module cache to force a rebuild.

On its own

find /root/go -name "*typedcompare*" \
  -type f

Searches the disk for files it already knows the location of.

After the note

Run the debug test with -count=1. The cached result is hiding the current code. Stop deleting the module cache and always pass -count=1 when you recheck. Then read the actual values before changing any logic.

go test -run TestDebug -count=1 -v \
  ./util/typedcompare/

Re-runs the test with the cache off.

Beyond four examples

Four moments are a story, not a measurement, so we also scored a larger set. We took 16 stalls from recorded runs and replayed each one with and without a note. The next step was a valid, working command in 15 of 16 cases with the note, against 12 of 16 without it. The step touched a file that ended up in the final patch in 10 of 16 cases, against 6 of 16.

That is a small sample, and it measures the next step only. Whether the notes raise the share of tasks a local model finishes is still being tested, and we will publish that when we have it. The notes also miss things. In one check, a model summed units when it should have multiplied by price, and the note did not catch it. A note helps a model out of a rut. It does not replace a stronger model doing the work.

What it costs

A note is a single short reply, around 60 words. It costs a fraction of a cent. A run gets at most ten of them, with six turns between notes, and a run counts as one Boost. A full takeover, where the cloud model does the work itself for a turn, is a separate and more expensive kind of Boost. You can still ask for one yourself at any time with /boost.

Try it

New accounts get 20 free Boosts, good for 7 days. Install OptiQ, point your coding agent at optiq serve, and sign in.

terminalbash

pip install mlx-optiq
optiq cloud login

More in the OptiQ Cloud docs, or sign in to OptiQ Cloud.