Repository navigation
cloudfall 0.6.0rc11
Pre-release
Pre-release
·
451 commits
to main
since this release
Changes since v0.5.1:
Added
- A decision record has two new statuses, added to
operation-decision.schema.jsonbeside the old ones.ran: arisk: read
operation has run by the time it is recorded, so it is recorded with its
run's exit code (ranExitCode) instead of waiting asproposedforever.
superseded: whenoperations proposerecords an operation, an older
proposedrecord of the same operation, targets and inputs is closed and
names the new decision (supersededBy); an approved or finished record is
never touched. Since it reads the record first,operations proposenow
stops withRECORD_INVALIDon a record that does not load, before check
mode runs.decisions approverefuses aranorsupersededrecord with
PRECONDITIONbefore asking for the decision id.decisions list --status Sand
cloudfall why --status S(repeatable) keep only those statuses, and an
unknown status is refused. Records already on disk keep their status. One
disk-full incident left 47 records forcloudfall why --host hz1, most of
them reads and repeats (#26) - A decision proposed during
agent investigate --replaynames the
recording its check output was played back from, asreplayin the
decision record (optional inoperation-decision.schema.json; a record
without it was checked live).decisions listshows it,cloudfall why
says so in its story, anddecisions approverefuses such a decision with
PRECONDITION(decision_replayed) before asking for the decision id or
running anything.--replayalso refuses, before the model is asked, a
--decisionsdirectory that resolves to the recording itself, which would
have written played-back records into the run being played, and a
--replaypath longer than the 512 characters the records keep (#36) - Every turn of
agent investigatekeeps its raw trace, in the stream's
turnevent (trace) and in the investigation record (spec.trace): the
model's reasoning when it returns one (Nemotron'sreasoning_content),
each tool call exactly as written (name and raw arguments), the
endpoint's response id andx-request-id, the milliseconds the turn took,
and its prompt, completion and reasoning tokens. The record can now say
what the model thought when it proposed a change agent investigatestreams: one bare JSON line with_seqper model turn
(its words and how many tools it called), one per tool call (the step and
what the tool handed back,effect: createdwhen it wrote a decision),
and the saved investigation record last (kind: investigation), then the
"_summary": trueline witheffects, or the error envelope when it
failed.--no-streamreturns every event in one envelope. A watcher, such
as the hosted demo, shows the agent working instead of waiting for the end- A decision record's approval says how it arrived, as
approval.via:
terminalwhen a person typed the decision id at a terminal. A record
written before reads asvia: unknown.cloudfall whytells it in the
approval sentence. The field is optional in
operation-decision.schema.json(#25) agent investigate --replay DIRplays a real run back instead of running
Ansible: DIR is the decisions directory of a run on real hosts, and each
check-mode call the model makes answers with the output and per-host
results recorded for the same playbook, target and inputs. Validation,
the decision records and the gate work as they do live, the model is
live, and no host is touched; a call the recording does not hold fails
and says so, and a recording that is missing or holds no decision record
is refused before the model is asked. The investigation record names the recording under
replay. An agent can be tried on a real incident again and again, for
a demo or as a regression test.examples/disk-full-incident/recordings
holds the disk-full incident recorded on a Hetzner cx23- A mutating operation whose check failed now hands the model the end of
its output, so the model sees why (such asNo space left on device)
instead of only the exit code examples/disk-full-incident: a one-host Ansible repository with an
operations catalog, a destructive trap, and a scenario that fills the disk
with unrotated logs so PostgreSQL goes down for a reason that is not the
symptom; andevaluations/nemotron-incident, four Nemotron models given
that alert ten times each throughagent investigateon a real host.
Super found the cause and the fix 10 of 10; Lightning proposed dropping
the database 10 of 10 and Ultra tried to approve itself 2 of 10, and the
gate kept every one of those a proposalagent investigateno longer lists a decision whose check failed under
unrecorded: it was recorded, only not proposedcloudfall agent investigate --alert TEXT --base-url URL --model NAME
lets a model work an alert through the operations catalog, over any
OpenAI-compatible endpoint (the API key from--api-key-from-env,
--api-key-from-fileorCLOUDFALL_API_KEY). The model sees the same
operation toolsmcp serve --repositorygives any client: a read
operation runs and returns what the hosts reported, any other operation is
proposed in check mode and recorded as a decision, and no tool approves
anything. A call repeated with the same arguments is answered from the
first result instead of running again. The investigation is recorded in
investigations/as anAgentInvestigation: the alert, the model, the
tokens, every call and how it ended, and the finding, with any decision id
the model named but never recorded kept apart underunrecorded. Exit 92
MODEL_UNAVAILABLEwhen the endpoint refuses, exit 93
INVESTIGATION_INCOMPLETEwith the record when the model runs out of
--max-turnsor does not answer in the asked-for shape. Proven on a
disposable host with Nemotron 3 Super on Nebius Token Factory: a full disk
behind a PostgreSQL alert was traced to unrotated application logs in five
turns, and the two proposals it recorded, once approved, brought the
database backagent investigaterecords the investigation when the model endpoint
fails partway (refused, unreachable, not JSON, or a malformed answer),
so the decisions earlier turns proposed are never left without one: the
record endsmodel-unavailablewith the steps, turns and tokens so far,
an emptyanswerand nofinding, the command still exits 92
MODEL_UNAVAILABLE, anderror.context.investigationnames the record.
A failure on the first request records an investigation with no steps;
a refusal before any request (exit 2) still records nothing (#29)- An operation tool of
mcp serve --repositorydescribes a read operation
as running and returning what the hosts reported, instead of saying it
runs check mode only operations proposereturns a read operation's play output in
data.output, so an agent acts on what the operation reported without
opening the decision's.difffile; the record keeps the file as before.
The output is the hosts' text, sodatacarries_trusted: false.
A proposed change answersoutput: null, and the read operation's MCP
tool no longer suggests approving a run that needs no approval (#24)cloudfall whyanswers "why did the agent do that" from the decision
records alone, for a host (--host), an operation (--operation) or a
time window (--since,--until, ISO 8601 or a bare date), as JSON or
as one HTML page (--format html). Each decision is told as what it was
based on, what was proposed, what check mode showed, what gated it, who
approved, what the run and the verify step did and how it ended, every
sentence drawn from a field of the record. A decision is about a host
when the record names it, as the target or in the per-host evidence of
a stage; a group pattern is not expanded, because that would need an
inventory the record does not depend on. The same question is the
read-onlywhytool oncloudfall-mcp --repository- Every JSON document
cloudfallprints, results and errors alike, carries
meta.schema_version(MAJOR.MINORof the output contract, now1.0)
andmeta.tool_version(the installed package version), so an agent can
tell when the output shape may have changed without a separate call.
meta.request_idnames the invocation with a UUID, the same on every
document it writes, for quoting in a retry or a bug report, and
meta.duration_mscounts the milliseconds from the start of the command
to that document, not counting Python's own startup.
MINOR moves when a command gains a key, MAJOR when one is removed,
renamed or changes meaning.cloudfall --versionprints the same
version as JSON and exits 0 - Every JSON document also carries a
warningslist. A key is removed only
after a release in which documents holding it warnFIELD_DEPRECATED
with the replacement key and the schema version that drops it cloudfall changelog [--since MAJOR.MINOR]lists changes to the output
contract, newest first: version, date,breaking, and the added,
removed and changed keyscloudfall --schema-version MAJOR <command>pins the output contract: it
fails withinvalid_argumentbefore the command runs when this build no
longer writes that MAJOR, so a pinned caller stops at a breaking release
instead of misreading it.--versionandchangelogreport the current
and minimum supported versionscloudfall --schema(alias--print-schema) prints every command in one
JSON document: its flags, read from the parser; its effect and gate, from
the command catalog; and anoutput_schemaper command, a JSON Schema of
each stdout shape with the condition that selects it. Every key, at the
top level and underdata, carriesx-stability:stablekeys change
only in a MAJOR release after
a deprecation warning,experimentalkeys (unreleased commands such as
whyandchangelog) may change in any release. What each key holds is
not declared yet. The tests validate real output against these schemas,
and anetagchanges only when a command, flag or declared output does- Every error carries
exit_code, the code the process exits with, so a
caller that reads the JSON needs no second channel:{"code", "message", "exit_code"}. Results that exit non-zero, such as an
auditthat finds drift, say so withok: falseand keeperror
null - Every
cloudfall--helpends with the exit codes and what each one
means: 0 positive, 1 ran with a negative result, 2 nothing ran, 3
unknown. The generatedAGENTS.mdrenders the same table from the same
source, so the two cannot drift --quietwrites nothing to stderr: no error document, no help text, and
no request lines fromdashboard serve. The exit code still says what
happened, and results still go to stdout--warnings-as-errorsfails a result that carries a warning: the
document saysok: falsewith the errorwarnings_as_errors, its
statusanddataunchanged, and the command exits 1 instead of 0
Changed
- The raw trace of
agent investigatekeeps at most 8000 characters of a
turn's reasoning and of its text, like the host output a tool result hands
the model: a longer one keeps its end after a[first N characters cut]
mark, in the investigation record (spec.trace) and in the stream'sturn
event alike. A model's reasoning often quotes the host output it just
read, and the record no longer keeps it unbounded.
agent-investigation.schema.jsonbounds both fields withmaxLength, and
says that a step's parsed arguments are what ran while
toolCalls[].argumentsis only the model's text (#50) - The investigation record of
agent investigatekeeps at most 8000
characters of the model's answer (spec.answer) and of each tool call's
raw arguments (spec.trace[].toolCalls[].arguments). Unlike the trace's
reasoning, a longer one keeps its start, followed by a
[last N characters cut]mark, in the record and in the stream's events
alike. The finding is still read from the whole answer, and the tool
still runs on the whole arguments.agent-investigation.schema.json
bounds both fields withmaxLength(#57) - treaty 1.0.0rc38 instead of 1.0.0rc29.
dashboard serveand
operator runwrite each event injsonandjsonlas a bare JSON line
with_seqinstead of an envelope withdataandmeta.seq, and end
with one"_summary": trueline or the error envelope.mcp serve
stopped bySIGINTorSIGTERMexits 130 or 143 with theCANCELLED
envelope instead of 0, and writes no envelope when it stops cleanly (#25) - The package is classified as Beta on PyPI instead of Alpha
- The site's landing page is redesigned for technical founders leaving the
cloud and for teams running servers for clients: a light, illustrated
look, the hero's call to action rotating between moving off Render and
moving off Neon, a section for MSPs on gated operations and the record
as the page sent to a client, a section on how the gate works with a
real decision record, and a free pilot by email; the manifesto page and
the link preview image follow the same look, and the manifesto's release
policy is stated as "proven live or labeled unproven" instead of "pre-1.0" - The README, the site, the roadmap and the architecture say what is built
on a team's own Ansible repository (the operations catalog, the gate,
decision records andcloudfall why) instead of calling it unbuilt, show
a realcloudfall whyanswer, and address both teams leaving a PaaS and
teams already running Ansible - The README opens with what Cloudfall is, a diagram of the gate, a short
quickstart and a guide index; the long walkthroughs moved to
docs/getting-started.md. The repository has issue
forms (bug, feature, pilot request), a pull request template, a support
page and a code of conduct, and the PyPI summary matches the project's description - The repository moved to the
cloudfall-devorganization:
github.com/cloudfall-dev/cloudfall. The package metadata, the README,
the site andREPOSITORY_URLpoint there; the old
github.com/romamo/cloudfall address redirects - Releases are cut by shipmill from
this CHANGELOG: an rc (0.6.0rc1, ...) after each batch of merges, and a
stable release promoted from an rc that soaked 3 days once its milestone
closes. The source distribution leaves out the repository's tooling,
evaluations and website - Breaking: the CLI follows one name per meaning, and its output is
typed:- Decision records have their own group:
decisions list(was
operations decisions),decisions approve(wasoperations approve), and a newdecisions show.operationskeeps the catalog:
list,show,propose data migrateisdata copy, no longer read asmigrate, and
services inspectisservices observe, asobservecollects server
evidence. Each old path answers 13REDIRECTEDwith the new command in
error.redirect.commandmigrate --releaseand--env-fileare--component-releaseand
--component-env-file: they takeCOMPONENT=VALUEpairs, where
deploy's take one value--observeddefaults totmp/observedon every command, where
observewrites:audit,services statusanddashboardrequired
it, and the operator's drift checks usedtmp/operator/observed- Field names in
dataare snake_case at every depth, as treaty's own
envelope keys are (openProposalsisopen_proposals,wouldRunis
would_run). Map keys that are data keep their spelling (resource
kinds, environment variable names, ids), and a record or resource in
data, anything with anapiVersion, keeps its stored form - Every command's output schema gives each
datakey its JSON type, and
an optional list or object is written empty rather thannull
(inventory showwritesansibleas{}for a project)
- Decision records have their own group:
- Breaking: argparse is gone: the last three
cloudfallcommands
move to treaty, and so do root--help,--versionand--schema.operator runstreams one envelope line per pass as it ends, with
data.pass(alerts,drift,autonomy) anddata.effect:
createdfor a pass that wrote proposals,updatedfor one that ran
licensed proposals,noopotherwise. Stopping it ends the stream with a
CANCELLEDline and exit 130 (SIGINT) or 143 (SIGTERM), so a systemd
unit needsSuccessExitStatus=143(the
operator guide has it). A gateway it cannot
use (none declared, or a missing or unreadable certificate) is 4
PRECONDITION, an unreachable one 12UNAVAILABLE.--intervaland--drift-intervalmust be above 0. The
operating contract now lists it with the commands that change servers,
since a declaredOperatorPolicylets it run proposalsdashboard servestreams one line once the server listens, then serves
until stopped, ending likeoperator run. Its request lines are info
log lines on stderr, shown at a terminal or with--verbosechangelogis treaty's: it lists interface changes per release from
the packagedschema-changelog.json, whichtreaty changelog-add cloudfall.app:appwrites, and--sincetakes a release versionoperator approvereports such a gateway as 4PRECONDITION(was 86
RECORD_INVALID) and an unreachable one as 12UNAVAILABLE- The argparse-era output contract added earlier in this release is
gone: the six-key document with a top-levelstatus,--output json,
--schema-version MAJORagainst that contract, the parser-built
--schema, the--helpexit-code table and exit codes 0 to 3. Every
command answers treaty's envelope, andcloudfall --schemalists every
command's flags, output schema and exit codes. TheAGENTS.mdthat
cloudfall initwrites describes this contract
- Breaking:
cloudfall-mcpis gone; the agent server iscloudfall mcp serve(treaty 1.0.0rc29, from issues #239, #240, #281 and #285 filed for
it). Its tools are Cloudfall's own commands, answering with the CLI's
envelope, and tool names follow the commands (list_operationsis
operations_list,validate_configisconfig_validate). MCP client
configs change fromcloudfall-mcp …tocloudfall mcp serve …- The server serves one fleet, chosen as before:
--project, or
--repositoryfor a team's Ansible repository. The fleet, the schema
and engine directories, and every directory a tool writes (snapshots,
receipts, artifacts, proposals, import output, the migrate plan) are
fixed for the run: they leave the tool schemas, and a call that passes
one is refused, so no call reaches another fleet or engine or writes
outside the project. So are the Render API URLimport_render-api
sends its key file to and the env filesecrets_renderwrites. Files
a tool reads as input (a blueprint, a key file, an env file) stay
arguments - On a project,
deploy,rollback,restart,migrateand
data_migratereturn their plan until called withyes: true, as on
the CLI; the oldconfirm=truehandshake is gone.build_artifactand
theconverge_baseline,converge_servicesandconverge_domains
tools stay, the converges destructive and run only with
confirm_destructive: true.backup_runandbackup_verifyrun in
one call, as their commands do (they were confirm-gated in
cloudfall-mcp), andoperator_watchandoperator_approveare gone:
the watch loop runs as a service, and a person approves a proposal from
a shell - On a repository, the tools are the read-only fleet commands and one
operation_*tool per declared operation, which runs check mode,
records a proposal and answers with the approval command in
data.next; a destructive operation's tool asks for
confirm_destructiveeven to preview - Each directory has one flag name across the commands that read and
write it, so a server can fix it once and a non-default one reaches
them all:observe --output-diris--observed(asauditreads it),
services inspect --output-diris--service-observed(asservices statusreads it), and--receiptsis--releasesondeployand
migrate,--backupsonbackup run|verify,--data-migrationson
data migrateand--env-receiptsonsecrets render(asaudit
reads it).mcp servetakes--releases,--backupsand
--data-migrationstoo init,operations approve,operator runanddashboard serveare
never MCP tools (mcp=False), on this server ortreaty-mcp- A fleet that cannot be read stops the server before a client connects,
with the envelope on stderr: 79PROJECT_INVALID, 80CONFIG_INVALID,
85INVENTORY_UNREADABLE.--list-toolsprints the tool list dashboard servereports a port it cannot bind as 4PRECONDITION
withdashboard_listen_failed; an unreadable evidence file is no
longer reported as a listen failure
- The server serves one fleet, chosen as before:
- Breaking:
cloudfall-engineruns on treaty.
Every command answers one JSON envelope on stdout (ok,data,error,
warnings,meta), errors included, which used to go to stderr in their
own shape;inventory renderputs the inventory underdata.inventory
(the--outputfile is unchanged). Failures exit with their own codes:
79PROJECT_INVALID, 80CONFIG_INVALID, 81ARTIFACT_BUILD_FAILED,
82PLAYBOOK_FAILED; the old snake_case code is inerror.context.code.
A missing playbook, inventory, or role directory exits 2 before Ansible
starts.playbook run --checkis now--dry-run; Ansible's play log
streams to stderr.cloudfall-engine manifestdescribes every command - Breaking: the last eight commands that change state move to treaty:
observe,secrets render,import render,import render-api,
deploy,rollback,restartanddata migrate(treaty #183 lets
their camelCase keys through). Each keeps itsdatakeys,status
included; the ones that write addeffect.deploy,rollback,
restartanddata migratestill change nothing without--yes: that
run is treaty's dry run (effect: would_update,meta.dry_run: true,
treaty #197). The plan writes the result keys asnull, and the result
writes the plan keys (wouldRun,instruction) asnull, since treaty
answers one object per command.observeis read-only (it declares the
snapshots it writes under the project as output, kept bycleanup) and exits 91INCOMPLETE(was 1)
with the run indatawhen a server gave no snapshot. Engine steps,
sops and the Render API go through treaty:doctorchecks sops, and
import render-apihonours--proxyand marks its answer as content
from outside. Failures that were exit 2 are 4PRECONDITIONwith the
old code inerror.context.code. A key a command writes only sometimes
isnullwhen absent, soinventory showwritesdata.ansibleas
nullfor a project - Breaking:
whyruns on treaty,--format htmlincluded (treaty
#179): the page is drawn from the answer's JSON document, as before. Its
verdict moves todata.status; a bad--sinceor--untilis 2
ARG_ERRORwithwhy_instant_invalidinerror.context.code, and a
record that cannot be read 86RECORD_INVALID.--format htmlis a
whyformat only; other commands refuse it - Breaking: seven commands that run Ansible move to treaty:
health,
backup run,backup verify,operations propose,operations approve,
operator approveandmigrate. Ansible runs through treaty in the
project directory, so--timeoutand Ctrl-C stop it, anddoctorchecks
ansible-playbook.operations approveandmigratestill change nothing
without--yes; that preview is treaty's dry run,effect: would_update
andmeta.dry_run: true, withdata.statuspendingorplan. New exit
codes, each with the result
indata: 87ENGINE_STEP_FAILED(an inventory render or playbook
failed, was 1), 88UNHEALTHY(was 1), 89CHECK_FAILEDwhen check mode
fails inpropose(was 1), 90NOT_VERIFIEDwhen an approved run fails
or does not verify (was 1), 3PARTIAL_FAILUREfor a pausedmigrate.
A failedmigratestep's own code and message move from the top-level
errortoerror.context.deploy,rollback,restartand
data migratestay on argparse until treaty #183, since their plan
answer has awouldRunkey - Breaking: six more
cloudfallcommands run on treaty:init,
add ssh-key,add server-type,add server,dashboard buildand
services inspect. Their answer is the treaty envelope described below,
with the samedatakeys, pluseffect: createdoninitandadd,
which treaty requires of a command that changes the project.
dashboard buildandservices inspectstay read-only and declare the
reports they write undertmp/as their output, which treaty'scleanup
never removes. Failures: 6CONFLICTfor a resource that already
exists (was 2), 7PERMISSION_DENIEDfor a read-only project or path
(was 1 or 2), 4PRECONDITIONfor a non-emptyinitdirectory (was 2),
5NOT_FOUNDfor a missing key file, 2ARG_ERRORfor a bad id, name or
address; the old code is inerror.context.code.initruns git through
treaty, socloudfall doctorchecks for git 2.24.observe,
secrets renderand the twoimportcommands stay on argparse until
treaty can carry their camelCase keys on a command that writes (#183) - Breaking: nine read-only
cloudfallcommands run on treaty, the
first step of moving the CLI off argparse:config validate,
inventory show,operations list,operations show,
operations decisions,operator list,operator show,auditand
services status.cloudfall manifestlists them; every other command,
root--helpand--versionstay on argparse for now. Their answer is a
treaty envelope (ok,data,error,meta,warnings): the verdict
moves from the top-levelstatustodata.status, and the otherdata
keys are unchanged.inventory showwritesdata.ansibleasnullfor a
project. Failures are an envelope on stdout with their own exit codes:
79PROJECT_INVALID, 80CONFIG_INVALID(resources or evidence files),
85INVENTORY_UNREADABLE, 86RECORD_INVALID, 5NOT_FOUNDfor an
undeclared operation or a missing proposal, 2ARG_ERRORfor bad input;
the old snake_case code is inerror.context.code.auditexits 83
DRIFT(was 1) or 84UNKNOWN(was 3) with the report indata. Pick
the format with--format json:--outputafter one of these commands
is refused, since treaty's--outputnames a file. Relative paths still
resolve against the project, without changing the working directory cloudfall-engineruns Ansible and git through treaty'sctx.run
(treaty 1.0.0rc18), so--timeoutand Ctrl-C stop the whole process
group, and secrets are redacted from the play log. The play log streams
to stderr as plain text, as before, terminal or not;--quietsilences
it. A failed playbook carries the last 4096 characters of the log in
error.context.output, marked as untrusted content from the hosts.
playbook runreportsmeta.dry_run: trueunder--dry-run. A git step
that runs past its 600-second limit now exits 10TIMEOUT, not 81 with
artifact_git_timeout.cloudfall-engine doctorchecks for
ansible-playbook 2.21 and git 2.24- Every JSON document has the same six top-level keys:
ok(true exactly
when the exit code is 0),status(the command's verdict, such asok,
plan,driftorunhealthy),data(the command's payload, which
used to sit at the top level besidestatus),error(nullunless
the command failed), andmetaandwarnings. A failure on stderr is
{"ok": false, "status": "error", "data": null, "error": {"code", "message"}}. An agent readsokanddatawithout
knowing which command it ran. Anything that read a payload key from the
top level, such as.inventory, now reads it underdata --helpprints to stderr when stdout is not a terminal, so a caller
capturing stdout never finds usage text where it expects JSON. At a
terminal it stays on stdout--output jsonis accepted before the command and after every command.
JSON is the only format, so the flag changes nothing; a caller that asks
for JSON the usual way no longer gets a usage error.--outputwith any
other value isinvalid_argument- The six options that named an output path are renamed so
--output
means one thing everywhere:--output-dironobserve,
services inspect,dashboard build,import renderand
import render-api, and--output-fileonsecrets render. Scripts
passing--output DIRto these commands must switch operatorerrors use the same envelope as every other command,
{"status": "error", "error": {"code", "message"}}, instead of a bare
{"code", "message"}, so one parser reads every failureoperator showandoperator approveprint the receipt under
data.proposalwith a top-levelstatus, instead of the bare receipt.
approvereportsstatus: failedwhen the verify step did not confirm
the fix, as thecloudfall-mcptool already did
Fixed
- A decision proposed during
agent investigate --replayno longer marks an
open live proposal of the same operation, targets and inputssuperseded,
which left the operator a live proposal closed and a replayed one that
decisions approverefuses. A newer proposal now supersedes an older one
only when both carry the samereplayvalue: live closes live, a replay
closes a replay of the same recording, and nothing crosses between them. A
replayed read is still recordedran, but its verdict andcloudfall why
now say its output was played back from the recording, not run on the
hosts (#58) agent investigatekeeps its investigation record when the reader closes
stdout mid-run (such as| head -2). Since the command streams, such a
run left the decisions it proposed without the investigation that made
them, and exited 0. The record is now saved endedstopped, a new status
inagent-investigation.schema.json, with the steps, turns and tokens so
far, an emptyanswerand nofinding; nothing more is written to the
closed stdout, and the command exits 94INVESTIGATION_STOPPED. A run
stopped by a signal or a timeout keeps its own exit code, and one stopped
while it waits on the model still leaves no record;
--no-streamis unchanged (#44)agent investigateexits 4PRECONDITION(code
agent_investigations_unwritable), naming the investigations folder,
when it cannot write the investigation record, such as when a file
stands where the folder should be. A run that ended crashed with a
traceback instead; one whose reader closed stdout mid-run printed the
traceback and exited 0. Such a stopped run still exits non-zero, now
PRECONDITIONwith the reason as a JSON line on stderr, though its
record is lost (#56)- The operator proposal schema refuses
approval.viaon an autonomous
(policy) approval. Such a receipt used to pass validation and lose the
field on load; it now fails schema validation like any other invalid
receipt. Cloudfall never wrote one, so only a hand-edited receipt is
affected (#46) - A decision record that is not JSON or does not match its schema answers
RECORD_INVALID(exit 86, codedecision_record_invalid) naming the file
and the failing field, fromdecisions show,decisions list,
decisions approveandwhy; it used to crash them with exit 1 and no
envelope.decisions listandwhystill stop at the first bad record
rather than skip it (#41) migrate --env-file NAME=PATHand--data NAME=PATHresolve a relative
path against the project again, as every other path does; since the
move to treaty they resolved against the current directorymigrate --buildand thecloudfall-mcpbuild_artifacttool read the
release from the engine's treaty envelope, underdata. Since the engine
moved to treaty every build step failed with "artifact builder returned no
release id".build_artifactnow returns the engine'sdatakeys
(git_ref,archive_sha256, ...) besidestatus- The time baseline accepts a host whose clock another daemon keeps
(ntp, ntpsec, chrony).timedatectl set-ntpdrives systemd-timesyncd
only, so on such hosts check mode reported "Would enable network time
synchronization" and a deploy failed with "NTP not supported". The role
now switches NTP on only where systemd-timesyncd is installed, and
otherwise requires the clock to be synchronized; check mode warns when
nothing keeps time, and a deploy fails with a message naming the daemon
to check when the clock does not synchronize within 60 seconds cloudfall addinto a read-only or full project failed with a Python
traceback and could leave a new server type behind when only the
server file failed. It now removes what it wrote and fails with
project_write_failed(exit 2), naming the file and the reason. Any
other command that cannot write a path because it is read-only or not
the caller's reportspath_not_writableas JSON (exit 1, since part of
the work may be done) instead of a tracebackcloudfall observeno longer prints Ansible's play log on stdout ahead
of its JSON document, which broke any caller parsing stdout as JSON.
The playbook's output is captured, and when the run fails its last 2000
characters, ending with the play recap, are kept indata.detail. The
same capture keeps theobservetool ofcloudfall-mcpfrom writing
Ansible text into the protocol streamcloudfall-mcpoperator_watchandoperator_approvewrapped engine
failures in the error envelope twice (error.error.code); they now
return one envelope like every other tool- A failed
operator approveis recorded even when the engine's log is
long. The receipt keeps the error code and the end of the log, where the
failing task and the play recap are, within the 1000-character cap.
Before, a run against unreachable hosts crashed while writing the
failedreceipt, left the proposalproposedwith no trace that it ran,
and showed the caller a traceback - The investigation record of
agent investigatebounds the rest of the
model's text. The finding'srootCauseandwhykeep at most 8000
characters, keeping the start before a[last N characters cut]mark;
the proposed decision ids are still read from the whole line. A tool call
whose arguments serialize to more than 8000 characters of JSON is refused
before it runs and recorded as aninvalidstep with empty arguments, so
spec.steps[].argumentsonly holds arguments that ran whole; the trace
still keeps the start of the raw arguments.agent-investigation.schema.json
bounds both finding fields withmaxLength(#63)
Security
cloudfall decisions approve <id>asks the person to type the decision id
at a terminal before it runs anything, and no flag answers that,--yes
included. Off a terminal, as from an agent's shell or a script, it exits 4
withPERSON_REQUIREDand runs nothing; a mistyped id exits 4 with
ATTESTATION_MISMATCH. Before,--yesran it headless and--approver
fell back to$USER, so an agent could approve its own proposal in the
person's name. Without--yesthe command no longer previews the
proposal:decisions showdoes. Scripted or CI approvals stop working.
This is a speed bump, not a boundary: a process running as the same OS
user can fake a terminal (#25)operator approveis no longer an MCP tool. On a project,
cloudfall mcp serveserved it asoperator_approve, so an agent could
run and verify its own proposal with no person in the loop. A person now
approves from a shell withcloudfall operator approve, as with
decisions approve;operator_listandoperator_showstay tools. The
--gateway-url,--gateway-ca,--gateway-certand--gateway-key
flags ofcloudfall mcp serve, which only that tool used, are removed
(#35)cloudfall operator approve <id>asks the person to type the proposal id
at a terminal before it runs the playbook or the verify step, and no flag
answers that,--yesor"yes": truein--raw-payloadincluded. Off a
terminal it exits 4 withPERSON_REQUIRED; a mistyped id exits 4 with
ATTESTATION_MISMATCH; either way the proposal staysproposed. Before,
any shell call ran it, so an agent could approve its own proposal.
operator showpreviews it. The receipt's approval records
via: terminal; a person's approval written before reads as
via: unknown, and a policy's has none. The field is optional in
operator-proposal.schema.json. Scripted approvals stop working; a
declaredOperatorPolicystill licenses autonomous runs (#35)