Repository navigation
Releases
v0.72.0
Changed
- Records refuse steps their lifecycle does not have. A session's status, a session's archive, a waiting message, an agent, an idempotency record, a trigger's firing and a trigger's run each follow a written table of states. Every store checks its writers against the table: an append that would take an ended session along a step the table lacks, or a second answer written over a stored one, changes nothing and is refused. Tests drive each writer through every pair of states and every pair of racing writers, on every store that has one and on Postgres. The tables are rendered into Lean proofs and TLA+ models that the quality gate checks, together with models of the session lease and the trigger lease.
- The session status type reads text and names its empty value.
session.ParseStatusreads a status from text andsession.NoStatusis the empty one a list uses to narrow by none.
Fixed
- An idempotency answer already stored is no longer overwritten. A second
Finishof a record wrote its answer over the first, so two requests that held one key could leave the answer of the second. It is now refused and the first answer stays. - A slow request no longer overrides the retry that replaced it. A request that outlived its idempotency reservation could store its answer over, or remove, the reservation of the retry that took its place. Its answer is now refused and its removal leaves the retry's reservation alone.
Release evidence
- Tag:
v0.72.0 - Commit:
a68645cb271746d4298bc4ee07f022422a2f0615 - Deploy: https://github.com/latere-ai/topos/actions/runs/43804484006
- Origin: https://api.latere.ai/v1/agents
- Smoke:
/openapi.yamlanswered 200 with an OpenAPI description - Served version:
v0.72.0, equal to the tag
v0.71.0
Added
- The documentation has a page for each reader.
docs/README.md
indexes every page by who asks;docs/local.mdruns the server on one
machine with everything it needs;docs/install.mdcovers the image
and its signature, the database, keys, probes, the manifests and
upgrading;docs/api.mdcovers the concepts, the flows a client
follows and the errors to branch on;docs/authorizer.mdis the
contract an authorizer answers; anddocs/security.mdsays what Topos
sees and keeps and what an operator keeps safe.
Changed
-
Built with Go 1.27.4, which fixes GO-2026-6859.
-
A short deadline no longer leaves the next call without a service
token. A call to another service that gave up while its token was
being fetched, as the 300 ms Oversight check does after each token
expires, now leaves the fetch to finish and hold the token, so the
next call is answered at once. -
SECURITY.mdnames the test that holds each commitment. It lists
what Topos protects against a caller, a steered model, the code a
session runs and a runner, and what it does not defend. Report a
vulnerability through GitHub's private vulnerability reporting or by
email; the organization's policy sets the response times.
Fixed
make tokenprints the token alone. The build's output goes to
standard error, soTOKEN=$(make token)holds a token a server
accepts.- Stopping
make runstops the server. The target now execs
out/toposd, so the server ismake's own child whatever the shell
does with its last command, and theSIGTERMthatmakepasses its
child when it is stopped reaches the server. A shell that stayed the
server's parent took that signal itself and left the server running
on its ports. - A server with no sandbox service creates a session of an agent that
names no machine. With noTOPOS_CELLA_URLandTOPOS_HOST_SESSIONS
off, such a session was refused at create withmachine_unavailable,
although a session that calls no tool needs no machine. It is now
created on Cella's kind, as with a Cella, and its first call of a tool
that acts on the machine answersmachine_unavailableas the call's
result; the turn goes on. An agent that names more of ahostmachine
than its kind is still refused at create. - A list filter that names no owner narrows no list. An allow whose
filterhas noowners, such as{}, listed nothing foragent.list
and for atrigger.listthat names an owner, whilesession.listand
an unnamedtrigger.listread it as no narrowing. Every list now reads
it as the shared authorizer contract does: an emptyownersdoes not
narrow by owner. An authorizer that sent{}to hide every agent must
deny the list instead. - The repository's conformance runs hold every list to a decision.
They now passconformance.WithPageActions(), asdocs/authorizer.md
tells an authorizer's author to, so an endpoint that answers a list
with a page of its own and noallow, whichtoposdcannot read,
fails them. - The documentation says what the server does in four more places.
docs/configuration.mdsays a boolean is read as Go's
strconv.ParseBoolreads one, so1andTRUEaretrue, lists
OTEL_EXPORTER_OTLP_ENDPOINTand the OpenTelemetry variables beside
it, and links the section of spec 028 that designstoposd check.
docs/holding-triggers.mdsays an authorizer on the shared scaffold
that does not knowtrigger.holdmakes the hold routes answer
authorizer_unavailable, notforbidden. Theauthorizerpackage
says a fork whose send is refused is only logged, sincetoposdsets
no event sink.
Security
- An authorizer over plain
httpon another host needs the
operator's consent.TOPOS_AUTHORIZER_URLwithhttp://on a host
other than loopback now stops the start naming
TOPOS_AUTHORIZER_INSECURE, asTOPOS_CELLA_URLdoes with
TOPOS_CELLA_INSECURE, because the authorizer's bearer and every
question with the caller's claims would travel in the clear. An
installation that uses such a URL must, before it upgrades, either move
the authorizer tohttps, or setTOPOS_AUTHORIZER_INSECURE=truefor
an authorizer reached over a network it trusts. A loopback URL, such as
http://127.0.0.1:9000/authorize, needs nothing. - The network policy no longer admits the Dash0 collector. The collector
is retired, so the egress rule to it is gone. Telemetry export stays off
until a deployment setsOTEL_EXPORTER_OTLP_ENDPOINT.
Release evidence
- Tag:
v0.71.0 - Commit:
ca6464f49d2ae5a8a34fd53c1e2dab66911e91ef - Deploy: https://github.com/latere-ai/topos/actions/runs/43750946967
- Origin: https://api.latere.ai/v1/agents
- Smoke:
/openapi.yamlanswered 200 with an OpenAPI description - Served version:
v0.71.0, equal to the tag
v0.70.0
Added
-
A release can deploy itself. The release workflow has a
deployjob
between the build and the publish stage, which appliesdeploy/prodwith
the image pinned by digest, waits for the rollout and runs
tools/smoke/release.shagainst the live origin. It runs only where the
repository variableTOPOS_RELEASE_DEPLOYis set, so a fork publishes
every artifact and deploys nothing.deploy/prodholds the maintainer's
production manifests anddeploy/prod-bootstrapthe identity the job
deploys with. Neither is part of an installation: the base and the
bootstrap are still what to install from. -
GET {base}/versionis public. WithTOPOS_BASE_PATHset, the
build'sversion,commitandbuild_timeare also answered under
the base path with no token, for example/v1/agents/version, so a
deployment behind an ingress that forwards only that prefix shows which
release it runs. The release smoke compares the served version with the
tag.
Fixed
-
A session id has one spelling. Ids are read only in the uppercase
form topos mints, so a lowercase spelling of an id no longer reaches the
same session directory. -
One-line text fields refuse hidden characters. A session title, app
name or metadata value with a bidirectional control character or invalid
UTF-8 is refused, and a long line the read tool cuts ends in an ellipsis. -
Retry-Afteris read and written safely. A very large value from a
model provider no longer turns into a negative wait, and topos writes the
header in whole seconds, rounded up. -
Request ids are checked. An
X-Request-Idthat is empty, longer than
128 bytes or not printable is replaced with a minted one before the
authorizer, the logs and the response see it. -
A list filtered by labels lists nothing. When the authorizer's filter
names labels, a list returns no objects, since no object carries a label
to match. -
Postgres migrations take a clean URL. A
pool_parameter on the
database URL no longer reaches the migration connection.
Release evidence
- Tag:
v0.70.0 - Commit:
c0edfb27c01e2076b314e95bc50cd40b45dac1a5 - Deploy: https://github.com/latere-ai/topos/actions/runs/42088852816
- Origin: https://api.latere.ai/v1/agents
- Smoke:
/openapi.yamlanswered 200 with an OpenAPI description - Served version:
v0.70.0, equal to the tag
v0.69.0
Added
- A call records what it reaches.
agent.tool_userecordsscope,
the reach the call was decided on, before the call waits for a
person: whether it can be undone (undo), the boundaries it crosses
(crosses), the class and kind of its target, whether it is a first
contact, andsource,servicewhen the decision service read the
call andtoposwhen the scope is Topos's own claims alone. A client
can show "cannot be undone" or "makes the app public" beside the
command. No member holds a path, a host, a branch or an app's name. A
bashcall without a decision service records no reach but its
source: show its reach as not known. See
docs/how-a-call-was-decided.md. - A confirmation can say how the person answered. A client can add
answer_ms,details_opened,suggestion_shownandsurfaceto a
user.tool_confirmation. The server records them, and refuses a value
of another type or outside its range asinvalid_request. Send them
only to a server whoseGET /v1/openapi.yamllistsanswer_ms; an
older server refuses them. - Repeated refusals hand the session back to the person. After 3
refusals in a row or 20 in a session, toposd records
session.denials_reached. On a turn a person began, every call that
changes something then waits for the person, withdecided_by: denials, until their next message. A turn a trigger began stops with
the new stop reasondenials, and the person's next turn asks the
same way, and so does a run of a rollout. Subagents share the counts.
TOPOS_DENIALS_CONSECUTIVEandTOPOS_DENIALS_TOTALset the limits,
0turns one off, and an authorizer'sdenialscan lower them for a
session.session.denials_reachedis a new event type, which a runner
of an earlier release refuses withschema_too_new: upgrade the
server and the runners together. - A subagent's refused calls warn its caller. When calls of a
subagent's run did not run because a rule or the person refused them,
the result of itsagentorsend_messagecall starts with a fixed
warning, and thetool.resultrecordsrefusedin itsmeta. - The decision service learns what each call reaches and how each
answer was given. Each call is sent with Topos's claims about its
reach (scope,target), the limits its machine enforces
(environment.holds:egress,protected_refs,checkpoints), the
visibility of the session's repositories (environment.repositories),
its subagent depth, and what anagentorsend_messagecall hands
over. Each answer carries the four fields of its confirmation,
waited_msfrom the log, andremembered, and never the note or the
pattern. - An authorizer can name the organization's hosts. An allow of
session.create,session.forkorsession.sendmay carry
boundary, the hosts the organization owns. The session keeps the
latest one, and aweb_fetchtells the decision service whether its
host is outside them. - A delivered repository records its visibility. The first machine's
session.machinenames each repository'svisibility,privateor
public, from one read with no credential at the git host when the
repository is delivered.
Changed
- In
learned, the decision service's new kinds of ask. An ask of
kindwider(the call reaches wider than any the person allowed)
waits for the person, assplitdoes. An ask of kindshort
(answers that agree, still in the ask band) orfew(approvals that
are not enough), and any kind Topos does not know, is decided by the
rules and the risk score. - A learned allow runs a call the risk score asks about only where the
person's approvals cover it.TOPOS_DECISIONS_LOOSENsets this for
learnedanddecide.covered, the default, lets anallowdecide
only when the prediction carriescover.covered: true, and leaves
every otherallowto the rules and the risk score, with the
suggestion recorded. A decision service that sends nocoverno
longer runs a call the risk score would ask about, so an installation
indecideorlearnedwith such a service sees more asks until the
service sendscover.neverlets the stricter of the allow and the
rules' verdict decide. - The decision service is not sent a subagent's task. An
agent,
send_messageor advisor call reaches the decision service without
its input.
Fixed
- A background command's files are gone once its log says it
exited. On a machine that runs commands in a stage sandbox, a long
command's script file was removed after the exit line was written, so
a reader that saw the line could still find the file. The file now goes
first, as it already did for a command run on the host itself.
v0.68.1
Changed
- The API document describes
decided_by,decided_by_ruleand a
create'srollout. Anagent.tool_use'sdecided_bylists the
layers that decide a call,decided_by_rulesays what rule within the
layer it names, and a create'srolloutstates its id's rule. A
generated client or reference now shows what each means. - The API document describes
decision_org,evidence_ratioand
ask_kind. Anagent.tool_use'sdecision_orgnames the
organization of the key that its decision was sent under. A
suggestion'sevidence_ratiosays how much the person's answers
settle it, and itsask_kindsays if an ask isunsureorsplit.
Fixed
- A
splitask does not mean that the answers disagree. The
documents, the code comments and the notes of v0.68.0 said that it
does. Asplitask has anevidence_ratioof 0.5 or more, and an
approval probability that is still between the profile's thresholds.
The person's answers can agree or disagree, and more answers that
agree can move the probability past a threshold. The behavior does not
change: inlearned, the service decides asplitask as before. See
docs/how-a-call-was-decided.md, which gives an example.
v0.68.0
Added
- A suggestion records why it asks. The
suggestionof an
agent.tool_usekeeps the decision service'sask_kindand
evidence_ratiowhen the service sends them: for anask,unsure
(too few answers) orsplit(answers that disagree), and how much the
person's answers settle the suggestion. A client can show why a call
asks without asking the service. A service that sends neither, and an
event written before this release, leave both out. See
docs/how-a-call-was-decided.md. - A session names the organization that owns its agent. A hosted
session of an organization's agent records the organization's ID as
org, fixed at its create and its fork. A session of a person's
agent has none. TOPOS_DECISIONS_ORG:initiatororagent.initiator, the
default, keeps the keys of v0.67.0. Withagent, the decision
service keys a call of an organization's agent by the organization.
The organization's cost profile then applies, and a person's answers
there stay apart from their own context. Every other call keeps the
initiator's subject. Set it after the clients and the authorizer of
the decision service read organizations.- A call records the organization it was decided under.
agent.tool_userecordsdecision_org, the organization of the key
that its decision was sent to the decision service under. A client
that reads the service's records of a call reads the organization
from it, and the initiator's subject from an event without it.
Changed
- A decision names the layer that decided it. Topos sends each
decision to the decision service with the source
topos/<mode>/<decided_by>, such astopos/progressive/learned,
topos/progressive/scoreortopos/confirm/always_allow, in place of
topos/confirmortopos/progressive. The service can tell a call its
suggestion decided from one a rule or the risk score decided, and the
source matches the one a replay of the session's log names. A decision
without a layer keepstopos/<mode>. A consumer that matches the
whole source string, or reports reliability by one source, must accept
the longer form. Reliability over the verdicts applied is the same.
Fixed
- The documents say what each
TOPOS_DECISIONS_MODEdoes when the
decision service fails. Onlydecidemakes the call wait for a
person, recorded asfallback. Inlearnedthe rules and the risk
score decide with no random review, and inshadowthe rules decide.
topos rundoes not read the mode and always decides asdecide. The
behavior did not change.
v0.67.0
Added
TOPOS_CELLA_INSECURE, the explicit consent forTOPOS_CELLA_URLto
behttp://on a host other than loopback, for a control plane
reached over a network you trust; bumpslatere.ai/x/cellato
v0.22.6, whose client refuses such a URL without it.
v0.66.0
Added
TOPOS_DECISIONS_MODE:shadow,learnedordecide. Inlearned,
the decision service decides aprogressivecall only where it has
evidence about the person: itsallow, itsblock, and anaskon
answers that disagree. Where it has too few answers, or does not
answer, the rules and the risk score decide as without a service, and
automatic verdicts are still picked for review at the service's audit
rate.decideis the behavior of v0.63 to v0.65 without shadow.- A person's answer to a network approval goes to the decision service
as an answer to the call whose connection was refused, after the
call's decision, the same as a confirmation of the call. Each approval
is one answer. An allow that adds the host to the session's network
answers this call only. A fetch outside the network already sent its
confirmation as the fetch's answer.
Deprecated
TOPOS_DECISIONS_SHADOW=trueis read asTOPOS_DECISIONS_MODE=shadow
for this release, and refused besideTOPOS_DECISIONS_MODE.
v0.65.0
Added
agent.tool_userecords which layer decided the call,decided_by:
network,always_confirm,tool,plan,read_only,
always_allow,remembered,sandbox,score,confirm,memory,
question,learnedorfallback; and the rule within it,
decided_by_rule: the pattern that matched, the tool that held the
call, the score's band, or the decision service's model (spec 012). A
client can tell a rule from a learned decision without reading the
reason.
Changed
- A rollout's run ends when its machine is unavailable. In a session
created withrollout, a tool call whose machine cannot be opened with
machine_unavailable, such as a sandbox Cella refuses because its owner
runs as many sandboxes as their limit allows, is answered and recorded
as before; once the step's other calls are answered, the session ends
withstop_reasonfailedanddetailmachine_unavailable, and the
model is not asked again. No person attends a run to resume it, and its
caller counts it failed for the machine whatever the agent does next,
so the caller can run it again sooner and nothing is spent on an agent
working around the missing machine. A subagent of a run that meets it
ends the run the same way. A session of no rollout keeps the failure as
the call's result, and its agent goes on.
Fixed
- A session's end deletes the sandbox a failed machine open left. When
Cella created a session's sandbox and the runner then could not reach
it, as when Cella's environment briefly drops its connections, the
session opened no machine, and its end left the sandbox running until
Cella stopped it at the environment's idle stop. Meanwhile it held a
place in its owner's limit of running sandboxes, so the owner's next
session was refusedmachine_unavailablewithquota_exceeded. The
end now deletes that sandbox and the Secrets applied for it, as it does
an open machine's.
v0.64.1
Fixed
- The release publishes its image. The v0.64.0 release built and signed
the image but stopped at its attestation, which could not find the
registry login, so v0.64.0 has no published image; use this release.