What changed for you in the releases of the Python SDK (sequence-ai, the seq
CLI) and of sequence-base from 2026-10-12 to 2026-10-20, newest first. Those from 2026-10-21 to
2026-10-31 are on the page after this one, and every release since on
the changelog.
sequence-ai
0.38.10 · 2026-10-20
seq deploy shows every line a build step prints, including the ones that arrive after the step
ended.
Fixed
- Lines a step printed late, such as what
@seq.loadprinted, now show inseq deploy’s live view under “load, continued” instead of being left out.
0.38.9 · 2026-10-20
seq policy ls and seq policy status say when deploying the same code again brings a
platform update.
Changed
- A policy whose serving version was built before a platform update that reaches it only by a new build (on
2026-10-20, faster acts) is told so in
seq policy lsandseq policy status, with theseq deploycommand that brings it.
0.38.8 · 2026-10-20
seq policy versions and seq benchmark versions say more about each build.
Changed
seq policy versionssays what changed in an image it rebuilt: its base image, Python, packages, environment variable names, commands or files.seq benchmark versionsshows a failed build that a newer registration replaced, for a while, and the version that replaced it.
0.38.7 · 2026-10-20
seq policy versions and seq policy logs say more about each build.
Changed
seq policy versionsshows a failed build that a newer deploy replaced, for a while, with the step it stopped in and the version that replaced it.- In
seq policy logs, a build’s check machine points to that build:seq policy logs <policy>@vN --build. - Every command’s
--helpdescription starts with a capital letter.
0.38.6 · 2026-10-20
Every command’s help says what it does, seq volume takes the names it lists, and a deploy and
its logs say more.
Fixed
seq volume showandseq volume rmtake a volume by the nameseq volume lsprints; a policy’s own copy of its weights is deleted with the policy, byseq policy rm.
Changed
- Every command’s
--helpstarts with what the command does, andseq policy rmsays “delete its version vN” where it said “take its version vN”. Needs sequence-base 0.4.21. seq deploysays how long each build step usually takes and how many robots a worker seats, andseq policy logssays when a machine’s bill ended on your own clock.
0.38.5 · 2026-10-20
Every command that reads now rides out a connection that fails, and seq deploy follows a build more
briefly.
Fixed
seq models,seq gpus,seq accountand every other command that reads now also ask again when the connection fails.
Changed
seq deployno longer lists each uploaded file again while it follows the build, since it listed them before the upload;seq policy logs --buildstill shows every line. Needs sequence-base 0.4.20.seq eval statuson an eval waiting for compute to free up says when it is expected back, or that the time is not known yet.
0.38.4 · 2026-10-20
Old argument names and the eval schedule are gone, and seq policy rm says first what goes with a
policy.
Removed
seq eval scheduleis gone, andmin_containers=,buffer_containers=,scaledown_window=,max_containers=andregion=are no longer arguments: writemin_workers=,buffer_workers=,idle_timeout=andmax_gpus=.
Added
seq policy rmlists the workers, robot keys, versions and storage it deletes before it deletes them (--dry-runlists them alone), and a deleted policy’s captured data stays as a dataset thatseq policy data exportdownloads.seq policy logsstarts with the machines it covers and what an hour of each costs,seq eval run --dry-runsplits its estimate into GPU, CPU and memory,seq volume show <name>says what uses a volume, andseq benchmark source <name>[@vN]downloads the code a benchmark version runs.
0.38.3 · 2026-10-20
A benchmark can load what all its evals share in @seq.load, apart from each eval’s own
@seq.setup.
Added
@seq.loadon a benchmark takes onlyselfand runs once on each machine, before@seq.setup; it opens no camera or window. Needs sequence-base 0.4.19.
0.38.2 · 2026-10-20
Commands that read, and a robot’s first connection to its policy, ride out a connection that fails.
Changed
- Every
seqcommand that only reads asks again, up to three tries, when the connection to the platform fails or does not answer, and says what failed when it gives up; commands that change something are sent once.seq policy logs --since <time>reads back to that time, up to 5,000 lines. - A robot opens its first connection to its policy in up to three tries, each on a new connection; when all fail, the warning says how many tries were made and the step the last one stopped in.
seq eval reportno longer prints the adapter’s hash on itsran:line.
0.38.1 · 2026-10-19
Commands say more plainly what an object is made from, what an eval will cost and why something was refused.
Changed
seq eval run --dry-runadds a composite policy’s planner cost and estimates a stopped policy too;seq eval episodesshows each metric in its own unit;seq gpusstates the 60-second minimum.seq benchmark upgrade,versions,statusandcancelsay what changed in a sentence, that evals run the version that last passed its check until the new one passes, and a canceled registration as canceled; a folder you edited afterseq initshows as the template it came from, “toy-policy v3 (edited)”.seq deployrefuses a composite policy that would span machines and answersinstruction=for each kind of policy; everyseqcommand sendsUser-Agent: sequence-ai-python/<version> (seq). Needs sequence-base 0.4.18.
0.38.0 · 2026-10-19
A composite policy plans each robot on a clock of its own and its plan can be anything, and seq eval run
--dry-run names the fastest way to run an eval within your GPU limit.
Changed
@seq.planplans each robotevery_safter its own last answer, the robots concurrently (max_concurrency=, 4 by default);max_batch=Nplans the robots due together in one call, a list in and a list out. A policy with@seq.planruns one copy on a worker: agpus_per_replicathat splits its cards into several copies is refused.planreturns any object, which@seq.inferreceives asplan=(instruction=is refused, and the error says what to write); both may takerobot=: the robot’s ownrobot.state, androbot.replan()to plan it again at once.- The dry run prints
fastest: <B> benchmark machines x <E> environments + <P> policy machines, <G> of your <L> GPUs: ~<time>, ~$<cost>and, on the next line, the flags that run it (--benchmark-machines,--envs,--policy-machines).
0.37.19 · 2026-10-19
An adapter says which benchmark → policy pairs it converts between and an eval runs it only on those, and a run’s result is shown as its value alone.
Changed
@seq.adapter(pairs=[(benchmark, policy), ...]): each side a name in your account,template:<name>for anything made from that template, or*;pairs="any"for every pair.seq deployrefuses an adapter that declares none, and an eval whose benchmark and policy its adapter does not declare is refused before anything runs, the error listing the pairs it does declare.seq adapter lslists each adapter’s pairs,seq adapter check --benchmark B --policy Psays whether it declares that pair, and an adapter prints without its class.seq eval lsandseq benchmark lsshow a run’s result as its value alone, in its metric’s unit, under a heading named by the benchmark’s metric; a run that failed readsFail.
0.37.18 · 2026-10-19
Everything is written the newest way, and what only an older version needed is gone.
Changed
- Removed:
@seq.policy(volumes=...), the retiredkeep_warm=,container_idle_timeout=,allow_concurrent_inputs=andframes_per_chunk=(now unknown arguments), an eval result’sclamped_by, a decide result’sprovider, and theseq robotandseq fleetcommands withseq.robots,seq.robot,seq.robot_logsandseq.fleet. gpu=takes one card:gpu="L4",gpu="H100:2"for two of one type, orgpu="cpu". A list is refused, and the error says what to write. A worker’s price always counts its CPU and memory.- The platform serves sequence-ai 0.37.16 and later; an older one is told to
pip install -U sequence-ai.
0.37.17 · 2026-10-19
Registrations and evals stop the moment you cancel them, times print on this machine’s clock again, and each setting has one flag.
Changed
seq benchmark cancel NAMEstops a benchmark’s registration in flight, andseq eval canceltakes effect at once, a machine that is moving or waiting for its card included; nothing is billed after either. A GPU benchmark’s machine that cannot draw its cameras moves to another three times at most.- Every time the CLI prints is on this machine’s clock, with the zone’s name (
14:05 EDT);seq usage, a policy’s monthly budget and your plan’s credit go by the billing period, the UTC month, and say UTC. - The step limit is
--max-steps N(max_steps=) and an idle worker’s timeout is--idle-timeout;--param eval.max_steps,--scaledownandseq eval run --compare(seq.eval(compare=): another policy is an eval of its own) are gone, andseq.port.compareis the function behindseq port compare(it wasseq.port.replay).
0.37.16 · 2026-10-19
Each thing goes by one name, the one the platform and the console use, and each flag has one meaning; the old names and spellings are gone.
Changed
- Evals:
--policy-machinesand--benchmark-machines(policy_machines=,benchmark_machines=),seq eval logs --machine,--spend-caponseq eval runandseq eval schedule(spend_cap_usd=), the new--max-steps Nfor the step limit (max_steps=), andseq eval run POLICY(seq.eval(policy, ...)). - A version is always
NAME@vN:seq policy rollback NAME@vN,seq benchmark tasks NAME@vN,seq policy traffic NAME@v5=10 NAME@v4=90andNAME@vN --pin KEY, andupgrade --to TEMPLATE@vN.--csvprints CSV and-o PATHwrites it;seq port exportandconverttake--trials N, andseq port comparetakes its scenes as arguments. - Robot keys, max GPUs, the monthly budget and the spend cap are said by those names, every time the CLI prints is in
UTC, every amount and price is written one way (a price by the hour,
$0.9612/h),seq usage --policygives the policy’s spend with its storage as its page and its budget count it, a policy’s model family is the one@seq.policy(family=)declares, and a new deployment’s name may not contain a..
0.37.15 · 2026-10-19
Each task has one CLI command: the duplicates are gone, with no aliases.
Changed
seq usageshows what your machines used and cost (this billing period, UTC;--policy NAME,--since/--until,--csv), in place ofseq billing exportandseq policy usage.seq deploy --dry-runchecks the files and prices the policy, in place ofseq validateandseq policy estimate.seq policy versionslists each version with its build,seq benchmark versionsis new,seq adapter sourcereplacesseq adapter pull,seq init --listandseq eval ls --benchmarkreplaceseq models --kind templateandseq benchmark evals, and a version is always writtenNAME@vN.
0.37.14 · 2026-10-18
Adapters and robot contracts drop their lossy, binarize and on_violation settings, a benchmark names the metric its
episodes succeed by, and run() can drive a policy with its own inputs and outputs.
Changed
- Removed:
seq.AdapterKind,seq.AdapterTransform,seq.LossyTransformand the transforms’lossless=andlossy=; a robot contract’sbinarizeandon_violation(an action out of bounds is clipped);seq benchmark compare; and--metriconseq eval runandseq eval episodes. - Added:
@seq.benchmark(success_metric=, metric_units=),seq.tf.crop,seq.GpuRenderError,run()’sactions_of=,control_hz=andabsolute=for a policy with its own I/O, andseq portfor the GPU porting commands that were underseq benchmark(export,check-model,scaffold,compare --replay,convert). joint_positionsis optional, only absolute action spaces are smoothed between chunks,seq.tf.action_projectonly reorders or picks within one action space, values show in their declared units, andseq eval report --csvsaves the console’s CSV.
0.37.13 · 2026-10-18
A robot warns when its acts are much bigger than what its policy reads, seq policy logs names each
machine by its worker, and seq policy estimate shows a template’s shared weights as parts that
cost no storage.
Changed
- A robot’s act much bigger than what its policy reads (twice what its declared cameras read, or 1 MB for a policy with its own inputs) is warned about once a session; the act still goes as it is.
seq policy logsnames each machine by its worker, asseq policy statuslists it (a0001, a0002, …), and--workertakes that name.seq policy estimatelists weights in named parts; a part an official template shares costs no storage.
0.37.12 · 2026-10-18
An eval’s output gives no confidence interval or act latency, suites are gone, and seq volume
ls shows a template’s shared weights at their size.
Changed
seq eval reportandseq eval statusprint no confidence interval or act latency;seq benchmark comparegives no verdict and has no--threshold.seq.Condition(suite=),seq.conditions(suite_of=)andseq.SuiteAggare removed.seq volume lslists a deployment that reads a template’s shared weights at their size, “shared · not billed”.
0.37.11 · 2026-10-18
seq policy logs --build, seq policy builds and seq benchmark logs say in
their help that a build’s log is kept as long as its version is, and seq benchmark logs --help
says what it shows.
0.37.10 · 2026-10-18
seq policy logs shows a policy’s log as events: a line each for when, which machine, and what
happened.
Changed
- Each robot’s acts in a run, and a print repeated many times, fold into one line that says when they ended;
--allshows every act and print. - What the policy printed while it loaded is marked
(setup), and what it printed on stderr(stderr). --workertakes the machine’s number as the log shows it, or the worker asseq policy statuslists it.
0.37.9 · 2026-10-18
An eval runs on the card its benchmark was registered on, and seq policy stop says when a version is
still building.
Changed
seq eval run --benchmark-gpuandbenchmark_gpu=are gone;--benchmark-instancessets how many machines, each on the benchmark’s card.seq policy stopsays when a version is still building: it finishes and is checked as usual (its check billed), and it is the version that serves once you start the policy.
0.37.8 · 2026-10-18
A policy’s access token can read its own policy’s status, and a run that is stopped part-way is still counted in usage.
Changed
seq policy status <name>works with the policy’s access token; listing, scaling, starting and stopping still need a full key.- A run ended by SIGTERM, SIGHUP, the program’s exit or
close()reports its stalls; one ended by SIGKILL cannot.
0.37.7 · 2026-10-18
seq policy builds lists a deployment’s builds, and seq policy logs --build --version
N shows any version’s build step by step.
Changed
- Each build says how long it took and why it was built: the files that changed, a new checkpoint, a template upgrade.
- Each step says how long it usually takes on this deployment.
0.37.6 · 2026-10-18
seq deploy shows each step of the build as it happens, with its time.
Changed
- Each step (image, weights, waiting for a card, load, compile, snapshot, check) appears as it begins, with its own output and how long it took.
- A failed build says once which step failed, its last lines and, when the cause is certain, what to change.
- A ready build ends with the time of each step.
0.37.5 · 2026-10-18
seq deploy no longer fails when the platform has taken a new version while robots are connected: it
says the build has started and follows it.
Changed
- When the platform takes a deployment but answers in a way this version cannot read,
seq deploysays the build has started and how to follow it (seq policy status <name> --wait) instead of exiting with an error. seq policy estimateno longer bills storage for an official template’s weights the platform reads from its own copy, or for weights in a volume you uploaded (billed as that volume);seq gpusnames those templates.- A robot waiting in line for a seat is told it is no longer in line when its deployment stops, instead of being told its seat was released.
0.37.4 · 2026-10-18
Adapters are no longer labelled lossy or lossless, and a new policy cannot declare binarize or
on_violation="refuse".
Changed
seq adapter ls,showandcheck, andseq eval report, no longer show lossy or lossless;lossy=andlossless=are still accepted.- A new
@seq.policythat declaresbinarizeoron_violation="refuse"is refused with how to do the same inside the policy; deployments made before keep working. - Requires
sequence-base0.4.9 or later.
0.37.3 · 2026-10-17
seq deploy says that a new deployment’s first worker stays on until its idle timeout, and is
billed while it does.
Changed
seq eval run --envssays what a CPU benchmark runs when it is not given: one environment per core it declares, 16 when it declares none.sequence-airequiressequence-basebelow 0.5.
0.37.2 · 2026-10-17
seq eval report and seq benchmark compare read every episode of a large eval, not only
the first 5000.
Changed
seq eval report --csv,--trajectoriesand--replays, andseq benchmark compare, page through all of an eval’s episodes when it ran more than 5000 (with a platform from 2026-10-17).- A benchmark’s registration carries
task_descriptionsandseeds_repeatonly when the benchmark declares them.
0.37.1 · 2026-10-17
seq eval run warns when a benchmark and a policy may not fit, and seq policy logs says why
a refused act failed.
Changed
seq eval runand--dry-runprint the platform’s warnings, with what the machines’ start costs when a run would stop at its first act; an act the platform refused is listed asfailed: r-b0381edd (refused: stopped, request r-48e71246), and--ridwith the act’s id shows the refusal.- A robot moving to another worker no longer waits for its old worker to acknowledge the end of its episode.
- Warnings point at the line of your program that led to them;
seq adapter checkcompares the action space and control rate too;seq policy estimatecounts the billed time a worker runs after each use (--starts-per-day), and--hours-per-daytakes 0 to 24.
0.37.0 · 2026-10-17
An eval says which of a benchmark’s tasks it runs, how many times each, and on which seeds; there are no suites.
Changed
seq eval runandseq.eval()take--tasks, then--trials 10or per task (--trials libero_spatial/3=20,*=5; past a task’s seeds they run again as repeated starts, unlessseeds_repeat=False) and--seeds 10-19;--param eval.max_steps=300sets the most steps an episode runs.--suiteis gone.seq benchmark tasks <name>lists each task with its seeds, episodes and description, andseq eval episodes <id>reads an eval’s episodes a page at a time, by task, outcome or arm.seq benchmark lsandstatussay which template a benchmark comes from, and whether it is stable or beta.
0.36.14 · 2026-10-17
seq policy usage shows how often your robots’ actions were on time, and their stalls.
Changed
- A robot driven by
run()reports how many of its action chunks acted on time and how many stalls it had, and the policy page andseq policy usage <name>show the on-time rate and the stalls. - A robot moves to a newer version at its next episode boundary, and the line that says it will move names a latest time only when that is within two hours.
0.36.13 · 2026-10-17
Every request the platform refuses a robot is listed on its own line, with its time, request id, reason and hint.
Changed
seq policy logs <name>lists each refused connect and ready check of the deployment’s robots, one line each (with a platform from 2026-10-17).seq policy logs <name> --rid <id>finds a refused request, or a connect, by its request id.- An error the platform answered ends with why and its request id, for example
[stopped, request r-5a5a5a5a].
0.36.12 · 2026-10-17
seq deploy shows one more step of a version’s first start, between loading the policy and
checking one act.
0.36.11 · 2026-10-17
A robot whose worker is replaced no longer starts its next episode late.
Changed
- The robot moves at the end of the first episode after the new worker has a place for it (with a platform from 2026-10-17).
- Why it moves is said once, in one
[sequences]line, with the time on your machine’s clock. seq policy statusandseq deploysay how often a worker is replaced, and name a version deployed before 2026-10-17 with “deploy it again” to have its workers replaced about once a day.
0.36.10 · 2026-10-17
A robot waiting for a seat is told it is in line, and its place.
Changed
- A robot that waits because every seat on its deployment’s workers is taken is told so, with its place in line, rather than that the platform’s GPUs are busy.
0.36.9 · 2026-10-17
A robot whose worker is being replaced moves to the new worker sooner.
Changed
- A robot hears that its worker is being replaced within seconds, and moves at the end of its current episode, or at its next act when no episode is open.
0.36.8 · 2026-10-17
A worker is replaced about once a day by one started beside it, unless you turn that off.
Changed
seq policy scale <name> --replace-beside on|off: on, the default, a worker’s replacement starts beside it, one worker pastmax_gpusbilled as any worker, so no robot waits; off, it is replaced in its place, its robot waiting between two episodes.seq policy status,seq deployandseq policy estimatesay how a worker is replaced, when the next replacement is, and what a worker up all day costs with it.- A lease renewal the platform cannot answer for a moment no longer ends a robot’s run: it keeps acting and asks again.
0.36.7 · 2026-10-16
A robot cut off from the platform for half a minute keeps its slot, and one cut off for longer is told its slot went.
Changed
- A robot cut off from the platform for up to about 30 s keeps its slot, as does a machine that slept or a process that was paused.
- When the platform says the slot went, one line on stderr says so, and the next act takes a slot again or waits in line.
ready()’s help says how often to poll it, and a poll refused for your key’s requests a minute (429) says so.
0.36.6 · 2026-10-16
A robot that gives up waiting frees its place for the robots behind it, and a robot whose worker is replaced at the end of its episode waits for the new worker.
Changed
- What a
ready()poll holds for a robot — its slot, the worker starting for it, its place in line — is given back atclose(), the end of awith, garbage collection and process exit, and within a minute when the robot is killed while it waits. - A process stopped by
SIGTERMorSIGHUPgives back what its open policies hold before it ends, and a handler of your own for either signal is kept. - A robot whose worker was replaced in place at the end of its episode waits for the new worker, instead of being told its episode was reset.
0.36.5 · 2026-10-16
ready() says why a start that takes longer than usual has no estimate.
Changed
- Past the time a start should take,
ready()’seta_sisNoneandeta_unknownis"past_expected", where it read 5 s for as long as the start took. wait_until_ready()timing out on such a start says it was still starting, slower than usual.
0.36.4 · 2026-10-16
A robot is told why it moves to another worker before it does.
Changed
- A robot whose worker is leaving is warned once: why (its worker has served its hours, or a newer version serves) and the latest it is moved by.
seq policy statuslists each worker’s copies of the model and the robots it seats apart (copies,seats,sessions), where oneslotscolumn read as seats.seq deploysays how its first machine’s start went once it is ready, step by step.
0.36.3 · 2026-10-16
What a worker costs is estimated as it is billed: its card, then its CPU cores and GiB of memory apart.
Fixed
seq policy estimateprices a worker’s CPU and memory apart from its card, at the cores and GiB the policy declares, and a CPU policy at its cores and GiB alone.seq gpussays a worker that declares no GPU pays its CPU and memory alone.- sequence-ai 0.36.2 and before price a CPU policy at one figure for a small CPU worker: upgrade for the policy’s own price.
0.36.2 · 2026-10-16
What a robot’s acts go up as, binary frames or JSON, is yours to read, and seq policy status
says whether a deployment’s workers take binary frames.
Added
Policy.act_formatis"frames"while the connection sends a robot’s frames as binary frames,"json"once its worker said it does not read them (or withSEQ_ACT_FRAMES=0), andNoneuntil the worker has answered.- The debug log of
sequence_aisays once per connection why its acts go up as they do. seq policy statussays whether a version’s workers take a robot’s frames as binary frames, and for one built before, that a redeploy brings them.
0.36.1 · 2026-10-16
A robot’s camera frames go to its worker as their own bytes rather than base64, a quarter less to upload on every act, and the 3D replay is on by default.
Changed
- A direct act sends its frames as binary frames when its worker reads them, and the policy sees exactly what
it saw before; a policy deployed before 2026-10-16 takes JSON until you
seq deployit again. SEQ_ACT_FRAMES=0sends every act as JSON.- An eval of a benchmark that declares
replay=seq.Replay(...)records its 3D replay unless it saysseq eval run --replay none.
0.36.0 · 2026-10-15
An eval can record its 3D replay, drawn in the console after the eval, when its benchmark declares
replay=seq.Replay(...).
Needs sequence-base 0.4.6 or later.
Added
- A benchmark declares it with
@seq.benchmark(..., replay=seq.Replay(engine="mujoco", state="engine_state")), besiderender=seq.Record(...), or hands its own frames withengine="own"throughseq.replay.sceneandseq.replay.frame. seq eval run --replay 3d|noneandseq.eval(..., replay=...)choose whether an eval records it.seq eval report <id> --replays DIRsaves each episode’s replay and the scenes they are drawn in.
0.35.3 · 2026-10-15
seq policy upgrade and seq policy status say why an upgrade goes to a version other
than the template’s stable.
Changed
- When the platform cannot deploy the stable, an upgrade goes to the newest version it can, and
--toa version it cannot deploy is refused, naming one it can.
0.35.2 · 2026-10-15
An act’s camera frames go up in one round trip on Linux, and seq deploy says what its build
check did with a policy that declares no observation contract.
Changed
seq deployandseq policy statussay what the build check did with such a policy: a robot-arm policy is acted on once with a guessed observation, and a failure is a warning, never a failed build.- On Linux, the connection to the worker keeps its speed between acts, so an act carrying camera frames goes up in one round trip.
- With a slower policy, loading BBR on the robot (
sudo modprobe tcp_bbr) keeps every act to one round trip.
0.35.1 · 2026-10-15
seq init writes only a version this sequence-ai can deploy, and predict() waits for a
worker that is starting.
Changed
seq init <template>writes the template’s stable when this sequence-ai can deploy it, else the newest version it can, saying which and why; a channel or version you name is written as named, with a warning when it does not deploy.predict()waits for a worker that is starting on every call, withinstartup_timeout_s, andinterrupt()from another thread ends the wait.
0.35.0 · 2026-10-15
One declaration says how a worker serves robots, @seq.concurrent(max_concurrent=N) on infer, and a
slot follows the open policy.
Changed
@seq.concurrent(max_concurrent=N, dynamic=True)on the@seq.infermethod replaces@seq.batchedand its waits,gpus_per_replica=andreplica_start=are@seq.policy(...)arguments, andseq deployrefuses the old ones, saying what to write instead.- While a policy is open the SDK keeps its slot alive, so a robot that pauses between acts keeps it, and the slot is given back the moment the policy closes or its process exits.
run()andact()drop what a run computed when it ends, so the next run never plays chunks computed for the one before.
0.34.43 · 2026-10-15
A robot far from its worker waits again when the worker says to, as before 0.34.42.
Fixed
run()andact(): a robot whose round trip is longer than a chunk’s playback no longer sends an act that reaches its worker while another robot’s runs.
0.34.42 · 2026-10-15
A robot far from its worker asks for its next chunk as soon as one lands.
Changed
run()andact(): a robot whose round trip is longer than a chunk’s playback asks for the next chunk the moment one lands, and tells its worker its network time from its first act.
0.34.41 · 2026-10-15
A robot waiting in line is told its place and about how long, and a late chunk is bridged by the steps it carries past its horizon.
Changed
ready()returns aReadywhosequeuedis True, withqueue_positionandeta_s(Nonewhen the wait cannot be estimated), andwait_until_ready()andseq policy statussay where each queued robot stands.run()andact()play the steps a chunk carries past itsopen_loop_horizonwhile the next chunk is late, andRunOutcome.spare_steps_usedcounts them.@seq.batched(serve_wait_ms=...)is removed, and a policy that still names it is refused whenseq deployreads it.
0.34.40 · 2026-10-15
One copy of the model a card: several copies on one card are removed.
Removed
@seq.concurrent(replicas_per_gpu=...)is refused whenseq deployreads it, with what to write instead; a worker of several cards still runs one copy on each (gpus_per_replica=).- The capacity line of
seq deploy,seq policy statusandseq policy estimatesays whether one copy fits on the card, andseq policy estimate --gpu CARDsays it for each card. seq policy statussays a worker that ran out of GPU memory needs a card with more memory.
0.34.39 · 2026-10-15
A worker’s price is what it is billed.
Changed
seq policy statusandseq deployprice a worker’s CPU and memory at what its workers use now when that is more than its build check measured, as it is billed.
0.34.38 · 2026-10-14
seq policy estimate says that copies on one card make each act take nearly twice as long when two
act at once.
0.34.37 · 2026-10-14
A robot’s first acts on a new connection go up in one trip each.
Changed
- For a deployment that declares cameras, opening the connection takes a few more round trips, before
wait_until_ready()returns orrun()starts, so that each act goes up in one trip from the first.
0.34.36 · 2026-10-14
A failed act’s report says what the robot got, so the console can say why each act failed.
Changed
- A failed act’s latency report carries the error’s code, message and hint, each cut at 500 characters, and never the request.
0.34.35 · 2026-10-14
seq deploy prices the version it made.
Fixed
seq deployprints the price and capacity of the version it made, where a version on its way in was priced at the serving version’s requests.- The capacity line suggests more copies on a card only when they would fit, and says that copies on one card make each act slower.
0.34.34 · 2026-10-14
seq account with an access token shows only that token’s policy, and
wait_until_ready() returns once the robot’s connection to its worker is open.
Changed
seq accountandsequence_ai.account()with an access token (a robot’s key) show the one policy the token is for, and say so on the first line.wait_until_ready()returns once the connection to the worker is open, so an act sent the moment it returns no longer waits, andrun()does the same before its first observation.- An access token reads nothing of its account but its own policy: the rest answers
403with “this operation requires a full key”.
0.34.33 · 2026-10-14
seq deploy and seq policy status say ready only once robots can connect, a session
refused for expect= says how to fix it, and an API key’s credit limit is a total.
Changed
- While a version’s quick start is still being prepared,
seq deployends “preparing”, andseq policy status <name> --waitreturns once robots can connect. - A session refused for
expect=says what the robot expects, what the policy sends, and how to fix it. - An API key’s credit limit, and a policy’s access token’s, is what the key may spend in all, never reset, and the key is refused at it until the limit is raised or removed on the console.
0.34.32 · 2026-10-14
A robot behind an HTTPS proxy checks its worker’s certificate again, and the connection to the worker opens before the first act.
Changed
- A robot that reaches its worker through an HTTPS proxy checks the worker’s certificate there too, and
refuses a wrong one with
CertPinError. - The client opens its connection to the worker as soon as it connects, and once
wait_until_ready()sees the policy ready, so the first act waits for neither. - Where the platform offers it, the client reaches its worker through an endpoint whose certificate it checks
against the certificate authorities it trusts (
SSL_CERT_FILE/SSL_CERT_DIR, else certifi’s).
0.34.31 · 2026-10-13
seq policy status says how long a start waits for the card, and the rest of the SDK says one card
a deployment.
Changed
seq policy statussays how long a start waits for the deployment’s card, under its GPU line.seq gpus,seq eval --dry-runand the README say a deployment runs on its one card and waits for it to come free, 10 min unlessseq policy scale <name> --start-waitsets another.
0.34.30 · 2026-10-13
One card a deployment: a start waits for it up to --start-wait, says which card and how long when
none came free, and seq policy status names it.
Changed
seq policy scale <name> --start-wait SECONDS(andsequence_ai.scale(start_wait_s=)) sets how long a robot’s start waits for the deployment’s card: 60 to 3600 s, 600 unless set.- A robot whose start found no free card is told which card, how long it waited and how to let it wait longer, and a start that ends this way is not billed.
seq policy statusnames the one card a deployment runs on, with one worker’s price an hour, andpolicy.ready()’sfirst_choiceis always True now.
0.34.29 · 2026-10-13
seq policy logs --rid says where a first act’s time went, and an unchanged official
pi05-droid deployment keeps no weights of its own to bill.
Changed
seq policy logs --ridsplits an act’s time into the connect, TCP, TLS, the transfer and the worker, andseq policy logsgives the p50 of each over first acts.seq policy estimatesays a deployed official template’s weights are not billed when it was deployed unchanged.- A deployment that declares more than one card runs on the first alone, and a start waits for that card.
0.34.28 · 2026-10-13
ready() and wait_until_ready() answer for this robot, not ready while it waits in
line for a slot, and the robot first in line hears of a free slot in about a second.
Changed
- While its deployment holds its
max_gpuswith every slot taken,ready()is(False, None)though a worker is up, andwait_until_ready()waits in line, saying its place. - A robot first in line for a slot hears of a free one about a second after another robot frees it.
0.34.27 · 2026-10-13
policy.ready() and seq policy status say the card a worker runs on and its price an
hour, and an observation that does not match says how to fix it.
Changed
seq policy usagesplits each day’s p50 and p95 into inference and other, asseq policy logsand the console do.- An observation that does not match what the deployment declares says how to fix each problem, with the call that sends it right.
policy.ready()andseq policy statussay the card a worker runs on and what one worker costs an hour on it.
0.34.26 · 2026-10-13
region= is retired: workers run wherever a card is free.
Changed
@seq.policy(region=...)imports with a warning, andseq deploy,seq validateandseq deploy --dry-runrefuse it, saying to remove it.seq eval run --regionandseq.eval(region=)are gone, and prices name no region.- After a reconnect that places the robot on no worker yet, the next act asks the platform first rather than waiting on a worker that had failed.
0.34.25 · 2026-10-13
seq policy logs splits a round trip into the policy’s inference and the rest, and a worker
is billed from the second it was up.
Changed
seq policy logsgives the round trip’s p50 and p95 with inference and the rest apart.seq policy estimatesays a policy that declares no weights keeps none, and gives a size under a tenth of a GB in MB.
0.34.24 · 2026-10-13
What keeping a deployment costs, said where you read prices.
Changed
seq gpusandsequence_ai.gpus()say the storage price, a GB a month of a deployment’s weights, compile cache or a volume, billed every day it is kept.seq policy estimatesays what keeping the deployment costs in storage beside its worker time.- The Pricing page lists the storage price beside the compute rates.
0.34.23 · 2026-10-13
seq policy logs sums up the acts your robots reported, and seq account’s
balance line adds up to the cent.
Changed
seq accountsays each amount of its balance line to the cent, so the three add up, as the console does.seq policy logssums up the acts your robots reported: how many, over what stretch, the round trip’s p50 and p95, and the ones that failed.- A worker’s
idle_timeoutcounts from its last act, whether or not a robot still holds it.
0.34.22 · 2026-10-13
seq account says what running work holds of your balance, and seq policy source
checks what it downloads against the version’s tree hash.
Changed
seq accountsays what running workers and evals hold of the balance until they stop, and what is available now; the console says the same.seq policy sourcechecks the files it unpacks against the tree sha256 the version recorded when it deployed, and stops if they differ.
0.34.21 · 2026-10-13
A worker’s state says where it is and a sentence says why, in seq policy status and on the
console alike: Failed only for one that failed.
Changed
seq policy statusand the console’s Workers list say why a worker stopped when no robot took it within a minute of starting, and that nothing was billed for it.
0.34.20 · 2026-10-13
seq deploy uploads the folder your policy is in, whole, and says what it uploaded, and seq
policy source downloads the code a version runs.
Changed
seq deployuploads the folder the entry file is in, less version control, caches, a virtualenv or build output at its root and what a.seqignorebeside the entry file lists, and--entry-onlyuploads the entry file alone.seq policy source <name> [--version N] [-o DIR]downloads the code a version runs, as it deployed it, and the console’s Versions has the same Download.seq initdownloads the large upstream sources a template’s image installs intothird_party/, andseq init --no-sourceleaves them out.
0.34.19 · 2026-10-13
Pillow comes with sequence-ai: pip install -U sequence-ai is all a robot needs to fit and encode
its camera frames.
Changed
sequence-ai[image]still installs, and adds nothing more.
0.34.18 · 2026-10-13
seq usage says what its dollars cover, seq audit the template versions an upgrade
moved between, and an older sequence-ai is told which version a template needs.
Changed
seq usagecounts a deploy’s checks as build checks, not evals, and each bucket’s dollars are everything its machines cost: GPU, CPU and memory.seq auditsays under each upgrade the template versions it moved between, andseq policy logssays “stop not recorded” for a machine whose stop was not recorded.- A template version this sequence-ai is too old to read is answered with the version it needs, such as
“toy-benchmark v4 needs sequence-ai 0.34.17 or later:
pip install -U sequence-ai”.
0.34.17 · 2026-10-13
An image can install from an official template version’s frozen files, which the platform keeps.
Needs sequence-base 0.4.3 or later; pip install -U sequence-ai brings it.
Added
seq.Image.template_files(template, version): eachrun_commandsstep after it reads that template version’s frozen build files at/seq/frozen, read-only, while the image builds.seq policy lsandseq policy versionsgive times on your machine’s clock with its zone’s name.- A policy class in a module that cannot be imported is refused with “the platform cannot load it”.
0.34.16 · 2026-10-13
seq policy says what a redeploy would bring the version that serves, and upgrade
--help says what --to takes.
Changed
seq policy upgrade --helpandseq benchmark upgrade --helpsay that--to vNnames any published version of the template: a beta one to try it, or an earlier one to go back.seq policy ls,versionsandstatussay what a redeploy would bring the version that serves when its workers do not make a newer check yet.
0.34.15 · 2026-10-12
A deployment or benchmark made from an official template says which version of the template it runs and moves to a newer one with one command, and a new version keeps the deployment’s scaling.
Added
seq policy upgrade <name>andseq benchmark upgrade <name>move to the template’s newer version, or the one--to vNnames, printing what changes, what stays and one worker’s price before and after (--dry-rundeploys nothing).seq policy statusandseq policy versionsname the template version each version runs, and the newer stable with the command that moves to it.- A new version of a running deployment keeps its scaling, and
seq deploysets only what its flags name and says what it kept.
0.34.14 · 2026-10-12
A camera frame that is not the size its camera declares is refused before sending only when the deployment’s version refuses it too.
Fixed
- For a version built before the platform checked frame sizes, the warning says the frame passes on and that
deploying it again makes it refuse such frames, and
validate="strict"lets the frame go. policy.contract["checks"]says what the version the connection reached refuses beyond the shape it declares,"frame_sizes"for a camera frame of another size.
0.34.13 · 2026-10-12
A new version of a deployment that already serves takes over once its quick start is ready, and the CLI says so while it waits.
Changed
seq deploy,seq policy status,lsandversionssay when a new version has passed its check and waits for its quick start, while the version before it serves.seq policy rollbacksays which version it dropped, and drops a version still building or waiting for its quick start.
0.34.12 · 2026-10-12
Camera frames are fitted on the robot, by tools you call: nothing in the SDK resizes a frame behind your back.
Needs sequence-base 0.4.2 or later; pip install -U sequence-ai brings it.
Changed
sequence_ai.resize_image(frame, width, height, *, mode, filter)andPolicy.prepare_image(view, frame)fit a frame on the robot, the second the way the deployment declares for that camera.Policy.to_command(step)andPolicy.to_commands(chunk)give the command a robot runs: each value clipped into the deployment’sbounds, then eachbinarizedimension set to 0 or 1.sequence_ai.encode()sends a frame at its own size, and a worker built from 12 October 2026 refuses a colour frame that is not its camera’s width x height.
0.34.11 · 2026-10-12
seq account names the cards a deploy’s build check holds beside serving and evals, so the
parts add up to what is in use.
0.34.10 · 2026-10-12
seq audit writes a billing line’s amounts in dollars and a plan’s change from the plan
before it, as the console’s Audit page does.
sequence-base
0.4.21 · 2026-10-20 · contract 0.5.18
A benchmark machine’s CPU cores are bounded and defaulted in one place.
Added
sequence_base.constants.limitsnames the most CPU cores a benchmark machine may ask for (MAX_CPU_CORES) and the number it gets when it asks for none (DEFAULT_CPU_CORES).
0.4.20 · 2026-10-20 · contract 0.5.18
The cards a benchmark may run on are named in one place.
Added
sequence_base.constants.cardsnames the cards a benchmark may run on (BENCHMARK_GPUS), the one it runs on when it names none (DEFAULT_BENCHMARK_GPU) and what it names for CPU machines (BENCHMARK_CPU).
0.4.19 · 2026-10-20 · contract 0.5.18
A benchmark can bind LOAD, what all its evals share, apart from each eval’s own
SETUP.
Added
BenchmarkSpec.bindingsmay bindLOAD(@seq.loadon a benchmark), which runs once on each machine before@seq.setup.
0.4.18 · 2026-10-19 · contract 0.5.17
A deploy says whether the policy plans, so a composite’s layout is checked before a card is held.
Added
DeploymentPolicy.plan_every_s: how often a policy with@seq.planplans;Nonefor one that plans nothing.
0.4.17 · 2026-10-19 · contract 0.5.16
An adapter says which benchmark → policy pairs it converts between, and the SDK and the gateway match them by one rule.
Added
AdapterPairandparse_pairs: an adapter’s pairs, each side a name,template:<name>or*, or"any".sequence_base.contract.adapters.match: the pair that lets a benchmark and a policy run with an adapter, the most specific one first;image_templates, the templates an image installs the files of, whichtemplate:<name>matches.
0.4.16 · 2026-10-19 · contract 0.5.15
The contract is the newest SDK’s alone: what only an older SDK sent or read is gone.
Removed
- The old names on
EvalRequest,benchmark_gpuandregion;ClampInfo,EvalResponse.clamped_byandSeqFlags.ROLLOUT(its bit stays unused). - A policy’s explicit volumes (
PolicySpec.volumes,DeploymentRequest.volumesandVolumeAttach), andmax_containersandframes_per_chunk, which only an older SDK sent.
Changed
PolicySpec.gpuandBenchmarkSpec.gpuname one card ("L4","H100:2"), not a list.
0.4.15 · 2026-10-19 · contract 0.5.14
An eval request names one policy and counts its machines by today’s names.
Changed
EvalRequest.compareis removed: another policy is an eval of its own.EvalRequest.gpusandenvs_per_gpu, the machine counts sequence-ai 0.28 and earlier sent, are refused like any field the request does not define; usepolicy_machinesandenvs.
0.4.14 · 2026-10-19 · contract 0.5.13
An eval’s machines and its spend cap go by those names, and the old names are no longer accepted.
Changed
EvalRequest.policy_instancesandbenchmark_instancesarepolicy_machinesandbenchmark_machines, and a dry run’sCardSide.instancesismachines.EvalRequest.max_usdisspend_cap_usd, and a report stopped by it saysstop_reason: "spend_cap".
0.4.13 · 2026-10-18 · contract 0.5.12
Adapter transforms are no longer classified as lossy or lossless, a robot contract has no binarize or
on_violation, and a benchmark declares the metric its episodes succeed by.
Changed
- Removed:
AdapterKind,AdapterTransform,LossyTransform,EvalReport.adapter_lossy,ActionContract.binarizeandActionContract.on_violation(actions outside theirboundsare clipped), andEvalRequest.metric. - Added:
BenchmarkRegisterRequest.success_metricandmetric_units, and a report’ssuccess_metricandunits. Proprioception.joint_positionsis optional like its other fields.
0.4.12 · 2026-10-18 · contract 0.5.11
The report’s confidence intervals (ArmReport.intervals, PairedDiff,
EvalReport.paired) and suites (SuiteAgg, MetricAgg.per_suite, a
condition’s suite, BenchmarkRegisterRequest.suite_of, EvalRequest.suite)
are removed.
0.4.11 · 2026-10-18 · contract 0.5.10
EvalRequest.benchmark_gpu and a deployment’s region are deprecated: an eval runs
on its benchmark’s card, and a deployment runs wherever there are cards.
0.4.10 · 2026-10-18 · contract 0.5.9
EvalRequest is the one request body /v1/eval takes, and its limits are named constants.
Changed
EvalRequestcarriesrerun_unsuccessful_ofandmetric, markssuite,region,gpusandenvs_per_gpuas deprecated, and no longer hasvlm_max_concurrency.envsis at most 65536 andmax_usdat most 100000 (sequence_base.constants.limits).EvalReport.stop_reasonis one of eight values.
0.4.9 · 2026-10-18 · contract 0.5.8
An adapter transform no longer needs a reason when it is marked lossy; the lossy fields stay for older SDKs.
0.4.8 · 2026-10-17 · contract 0.5.8
A benchmark is one set of tasks, and an eval says which of its tasks it runs and on which seeds.
Changed
EvalRequest.trialsis a count or{task: count}, andEvalRequest.seedsnames the seeds, for every chosen task or per task;EvalRequest.suiteis gone.BenchmarkRegisterRequest.task_descriptionsgives a sentence per declared task, andseeds_repeat=Falsekeeps a task to its seeds.suite_of, a condition’ssuiteandMetricAgg.per_suitestay for older SDKs and older evals’ reports.
0.4.7 · 2026-10-16 · contract 0.5.7
A robot’s camera frames can go to its worker as their own bytes rather than base64, a quarter less to upload.
Added
presets.frames:frames.pack(envelope)writes the robot-arm preset’s /act body as binary frames (application/x-seq-frames), andframes.unpack(body)reads it back into the very same request, raisingFramesErroron anything else.frames.FORMATS_HEADER(X-Seq-Act-Formats) names the act bodies a worker reads besides JSON.
0.4.6 · 2026-10-15 · contract 0.5.6
A benchmark can declare an eval’s 3D replay, and an eval can turn it off.
Added
ReplaySpec(seq.Replay) asBenchmarkSpec.replayandBenchmarkRegisterRequest.replay:enginepicks the simulation it reads, or"own", andstateandlooksname the benchmark’s methods it calls.EvalRequest.replay:"3d"(the default) or"none".
0.4.5 · 2026-10-15 · contract 0.5.5
One declaration says how a worker serves robots, and an act can carry its robot’s own timing.
Added
Concurrency.max_concurrent(2-256) andConcurrency.dynamicdeclare how a worker serves robots, andmax_inputs,target_inputsandreplicas_per_gpustay readable for stored specs and older SDKs.ActRequest.planandActResponse.plan, both optional, carry a robot’s own timing and its worker’s word on when to send.
0.4.4 · 2026-10-14 · contract 0.5.3
log() keeps an error’s traceback, up to 4,000 characters with its line breaks, in one field
that reads back as it was.
Added
clean_text(value, limit=4000)writes a long text as one log field that reads back as it was, andunescape_text(field)reads it back.log()writes anerrorfield withclean_text, up to 4,000 characters, and every other field is cut at 200 as before.
0.4.3 · 2026-10-13 · contract 0.5.3
An image can install from an official template version’s frozen files, which the platform keeps, instead of from the sites they came from.
Added
ImageStepTemplateFiles({"template_files": {"template": "<name>", "version": <n>}}) andFROZEN_PATH(/seq/frozen): each run_commands step after it reads that template version’s frozen build files, read-only, while the image builds.
0.4.2 · 2026-10-12 · contract 0.5.2
The robot fits its camera frames with an explicit tool, and a worker only checks a declared camera’s size.
Added
Camera.resize("pad"or"stretch") andCamera.filtersay how the policy was trained to see a frame fitted to its camera, andActionContract.binarizethe dimensions a robot reads as off or on.codec.resize_image(frame, width, height, *, mode, filter)fits a frame, andcodec.image_size(data)reads its size from its JPEG or PNG header.codec.encodesends a frame at its own size, and a worker on contract 0.5.2 checks each colour frame against its camera’s width x height.